ASSESSING THE INTERACTIONS BETWEEN RED SNOW ALGAE AND BLACK
CARBON FROM FOREST FIRE SOOT USING A FIELD EXPERIMENT AND
REMOTE SENSING METHODS IN WESTERN CANADA
by
Charlie Bourque
B.Sc., University of Northern British Columbia, 2021

THESIS SUBMITTED IN PARTIAL FULFILLMENT OF
THE REQUIREMENT FOR THE DEGREE OF
MASTER OF SCIENCE
IN
NATURAL RESOURCES AND ENVIRONMENTAL STUDIES

UNIVERSITY OF NORTHERN BRITISH COLUMBIA
June 2025
© Charlie Bourque, 2025

Abstract
My thesis investigates the relationship between snow algae and black carbon from forest fire
aerosols. In Chapter One, I discuss the motivation and objectives of this thesis, followed by a
background on the importance of glaciers in Western Canada, processes of glacier surface
darkening, and a review of the ecology of snow algae and forest fire aerosols. In Chapter Two, I
describe and analyze the results from a field experiment conducted on Place Glacier, in which I
fertilized snow with wood ash. I observed no difference in snowmelt from the treatment, and due
to a loss of data from early snowmelt on the last field visit, the wood ash impact on red snow
algae growth was inconclusive. At the beginning of the field experiment, I measured snow
reflectance, which I then converted to instantaneous radiative forcing. The average instantaneous
radiative forcing for the snow algae at Place Glacier on June 3 was 192 ± 12 W m-2 with a
maximum of 274 ± 16 W m-2. This radiative forcing translated to a melt potential of up to 27.8
mm w.e. d-1. In Chapter Three, I use remote sensing data from the Harmonized Landsat Sentinel
(HLS) project and reanalysis products from MERRA-2 and ERA-5 Land to model the
occurrence of red snow algae on glaciers in British Columbia and Alberta for the summer
seasons of 2015 to 2023. I modelled the red snow algae using the random forest regressor and
XGBoost algorithms, and used the permutation feature importance and SHAP to evaluate the
variables’ influence on the model output. The model explained 60% of the variance in the data
and the three most important variables were longitude, black carbon and temperature. These
results suggest that black carbon may promote red snow algae growth. Finally, in Chapter Four I
conclude with the major findings, limitations and recommendations for future research.

ii

Table Of Contents

Abstract .......................................................................................................................................... ii
Table Of Contents ........................................................................................................................ iii
List of Tables................................................................................................................................ vii
List of Figures ............................................................................................................................. viii
Acknowledgement ...................................................................................................................... xiii
1

Chapter 1: Introduction ....................................................................................................... 1
1.1

Motivations and Objectives ............................................................................................ 1

1.2

Importance of glaciers..................................................................................................... 2

1.3

Glacier Darkening ........................................................................................................... 3

1.4

Life on Snow ................................................................................................................... 5

1.5

Remote Detection of Snow Algae ................................................................................. 10

1.6 Forest Fire & Black Carbon ................................................................................................ 12
1.7 Research Gap ...................................................................................................................... 14
2

Chapter 2: Quantifying the interaction between wood ash and snow algae on Place

Glacier, British Columbia........................................................................................................... 15
2.1

Introduction ................................................................................................................... 15

2.2. Methods.............................................................................................................................. 18
2.2.1 Study Site ..................................................................................................................... 18
iii

2.1.1

Snow algae and black carbon experiment ............................................................. 19

2.2.3 Cell abundance ............................................................................................................. 23
2.2.4 Empirical measurement of snow reflectance ............................................................... 24
2.2.5 Radiative Forcing and Red Snow Algae Melt Estimate............................................... 24
2.2.6 Total metals, carbon and nitrogen content ................................................................... 26
2.2.7 Statistical Test .............................................................................................................. 27
2.3. Results & Discussion ......................................................................................................... 29
2.3.1 Wood ash effect on snow algae abundance and melt ................................................... 29
2.3.3 Wood ash Nutrients ...................................................................................................... 34
2.3.4 Reflectance and Radiative Forcing .............................................................................. 36
2.4. Conclusion ......................................................................................................................... 40
3

Chapter 3: Satellite and airborne remote sensing of snow algae in relation to black

carbon deposition from forest fires in Western Canada.......................................................... 42
3.1

Introduction ................................................................................................................... 42

3.2

Study area...................................................................................................................... 45

3.3

Data ............................................................................................................................... 47

3.3.1

Harmonized Landsat Sentinel ............................................................................... 47

3.3.2

Aerosol Optical Depth .......................................................................................... 48

3.3.3

ERA5-Land – Climatic parameters ....................................................................... 49

3.4

Data Acquisition and Methods ...................................................................................... 49
iv

3.4.1

Image acquisition .................................................................................................. 50

3.4.2

Mask and threshold ............................................................................................... 51

3.4.3

Snow Algae Index ................................................................................................. 52

3.4.4

Data aggregation ................................................................................................... 52

3.4.5

Trend analysis ....................................................................................................... 53

3.4.6

Data Preprocessing................................................................................................ 53

3.4.7

Models................................................................................................................... 54

3.4.8

Evaluating Variables Importance .......................................................................... 59

3.5

Results ........................................................................................................................... 60

3.5.1

Correlation ............................................................................................................ 60

3.5.2

The relation between Algae Ratio and the Environmental, Climatic and

Geographic Variables ............................................................................................................ 63
3.5.3

Main Model Performances Evaluation ................................................................. 69

3.5.4

Annual Models Performance and Feature Importance ......................................... 73

3.5.5

Lag-year model performance and feature importance .......................................... 76

3.6

Discussion ..................................................................................................................... 79

3.6.1

Variables Influence from the Main Models .......................................................... 79

3.6.2

Annual models and black carbon lag year models................................................ 85

3.6.3

Models Performance and Associate Limitations................................................... 87

3.7

Conclusions ................................................................................................................... 90
v

4

Chapter 4: Thesis Conclusions........................................................................................... 92

5

References ............................................................................................................................ 96

6

Appendices ..........................................................................................................................113
6.1

Appendix A: Spring surface air temperature at Place Glacier .....................................113

6.2

Appendix B: Elemental metal concentration of the wood ash, control and treated

samples. ....................................................................................................................................114
6.3

Appendix C: HLS tile names per Mountain Regions ..................................................115

6.4

Appendix D: Parameters used for Random Forest and the XGBoost models for the

Annual and lag year model ......................................................................................................118
6.5

Appendix E: Correlation matrix significance values ...................................................119

6.6

Appendix F: Annual model permutation feature importance ..................................... 121

6.7

Appendix G: Annual model SHAP Value ................................................................... 130

................................................................................................................................................. 131

vi

List of Tables
Table 2-1: Description of measurements conducted per dates on the experiment plots. Snowmelt
is not included, as each plot on each field visit was measured. No*: Data loss due to human error
during processing of the sample. .................................................................................................. 21
Table 2-2: Shapiro-Wilk Test results for the snowmelt and cell concentration data and sample
size (n). .......................................................................................................................................... 28
Table 3-1: Glaciated area, and elevation of the nine regions ....................................................... 47
Table 3-2: Parameters description and best value used for the Random Forest Regressor models
....................................................................................................................................................... 56
Table 3-3: Parameters description and best value used for the XGBoost models ....................... 58

Appendix Tables
Table C 1: HLS tile names per Mountain Regions .....................................................................115

Table B 1: Elemental metal concentration for the wood ash, control and treated samples. Rows
in bold represent elements with significantly higher concentrations in the treated samples. ......114

Table D 1: Random forest parameters used for the annual models and the lag model. The ‘F’ and
“T’ stands for True and False. Table 3.2 describes the parameters. .............................................118
Table D 2: XGBoost parameters used for the annual models and the lag model. Table 3.3
describes the parameters. .............................................................................................................118

vii

List of Figures
Figure 1-1: Hypothesized Chlainomonas sp. life cycle in Bagley Lake Habitat (Matsumoto et
al., 2024). Phases denoted with a letter on a white background were observed in the field. Phases
not observed in the field are denoted with a “?” on a dark background. The dotted lines represent
the transition process that remains unknown. ................................................................................. 7
Figure 2-1 A. Map showing the location of Place Glacier in the southern Coast Mountains of
British Columbia, Canada. Blue polygons are glaciers from the Randolph Glacier Inventory V. 7
(RGI Consortium, 2023). B. Hyperspectral mosaic RGB of Place Glacier on 22 August 2022
produced by the Airborne Coastal Observatory. The glacier boundary is outlined in red. ........... 18
Figure 2-2: Map of the experiment plots and their location on Place Glacier, BC, Canada.
Location for P8-2 and P8-1 missing. ............................................................................................ 20
Figure 2-3: Mean cells per mL for the 4 plot settings from 4 June, 11 June and 2 July 2023 on
Place Glacier, British Columbia, Canada...................................................................................... 31
Figure 2-4: Snowmelt and temperature trends on Place Glacier, British Columbia, Canada. A.
Daily temperature, average measure melt and positive degree day melt model (PDM) from Shea
et al. (2009) B. Melt per plot setting from 4 June and 11 June 2023. C. Melt per plot setting from
11 June and 2 July 2023. ............................................................................................................... 33
Figure 2-5: Spectral reflectance measured on 3 and 4 June coloured according to the different
plot settings A. Uncolonized and untreated [Clean snow], B. Colonized and untreated [Red
snow], C. Uncolonized and treated [ Clean snow with wood ash], and D. colonized and treated
[Red snow with wood ash]. The green lines point out to the chlorophyll absorption feature (680
nm), and the red rectangles point out to the carotenoid absorption feature (400-600 nm) ........... 37

viii

Figure 2-6: Instantaneous radiate,ve forcing or wood ash, clean snow and red snow as measured
on June 4 and July 11 from the field spectrometer reflectance data. ............................................ 39
Figure 3-1: The glaciated region of the Canadian Cordillera below 60ºN is divided into nine
regions. The small glaciers of the Insular Mountains on Vancouver Island are omitted in this
study. Details about the regions are found in Table 3.1. ............................................................... 46
Figure 3-2: Flow chart of the workflow used to acquire the algae ratio from the HLS image
collection. The section mentioned in the workflow corresponds to the next three sections. ........ 50
Figure 3-3: Pearson correlation coefficient between black carbon flux and red snow algae ratio
for the nine subregions A. Monthly mean correlation and B. Monthly max correlation .............. 61
Figure 3-4: Correlation matrices of the A. Monthly mean and B. Monthly max predictors and
target (algae ratio) values. ............................................................................................................. 62
Figure 3-5: Beeswarm plots for the SHAP variable importance for the algae predictors with the
Random Forest and XGBoost models........................................................................................... 66
Figure 3-6: Permutation feature importance scores for the Random Forest and XGBoost for the
A. Monthly mean dataset and B. Monthly max dataset ................................................................ 68
Figure 3-7: Boxplot representing the A. Coefficient of determination (R2) B. Root Mean
Squared Error (RMSE), and C. Mean Absolute Error (MAE) for each of the models using the
mean and max dataset. The scores were calculated for 50 training and testing dataset iterations.
....................................................................................................................................................... 70
Figure 3-8: Plots of actual vs predicted algae ratio value for A. Random Forest Monthly Mean,
B. Random Forest Monthly Max, C. XGBoost Monthly Mean and D. XGBoost Monthly Max.
The dash line is a 1:1 line and the red lines are linear regression model. ..................................... 72

ix

Figure 3-9: Random Forest (teal) and XGBoost (brown) R2 Score when applied to the model
run on a singular year. All models were run on the mean dataset except for 2016, which was run
on the max dataset. ........................................................................................................................ 73
Figure 3-10: Feature importance for the model run annually. A. Random Forest Permutation
Values, B. XGBoost Permutaiton values, C. Random Forest Shap values and D. XGBoost Shap
values. ........................................................................................................................................... 75
Figure 3-11: Permutation feature importance for the lag-year model for A. the random forest and
B. the XGBoost. ............................................................................................................................ 77
Figure 3-12: Beeswarm plots of SHAP values performed on the lag-year dataset for the A.
random forest and B. XGBoost model.......................................................................................... 78
Figure 3-13: Burned area from 2015 to 2023 from MODIS Burned Area Monthly Global 500m.
....................................................................................................................................................... 81

Appendix Figures
Figure A.1: Surface air temperature at Place Glacier in April and May 2023 from HOBO
temperature sensor, elevation 1924 m. a.s.l. The red arrow indicates the approximate date when
the air temperature sensor gets uncovered. ..................................................................................113

Figure E 1: Mean variables correlation’s significance for the correlation matrix. The green
indicates a correlation with ρ< 0.05, thus significant. .................................................................119
Figure E 2: Max variables correlation’s significance for the correlation matrix. The green
indicates a correlation with ρ< 0.05, thus significant. ................................................................ 120

x

Figure G 1: SHAP Values for the random forest on the annual model 2015 ............................. 130
Figure G 2: SHAP Values for the XGBOOST on the annual model 2015 ................................ 130
Figure G 3: SHAP Values for the random forest on the annual model 2016 ............................. 131
Figure G 4: SHAP Values for the XGBOOST on the annual model 2016 ................................ 131
Figure G 5: SHAP Values for the random forest on the annual model 2017 ............................. 132
Figure G 6: SHAP Values for the XGBOOST on the annual model 2017 ................................ 132
Figure G 7: SHAP Values for the random forest on the annual model 2018 ............................. 133
Figure G 8: SHAP Values for the XGBOOST on the annual model 2018 ................................ 133
Figure G 9: SHAP Values for the random forest on the annual model 2019 ............................. 134
Figure G 10: SHAP Values for the XGBOOST on the annual model 2019 .............................. 134
Figure G 11: SHAP Values for the random forest on the annual model 2020 ........................... 135
Figure G 12: SHAP Values for the XGBOOST on the annual model 2020 .............................. 135
Figure G 13: SHAP Values for the random forest on the annual model 2021 ........................... 136
Figure G 14: SHAP Values for the random forest on the annual model 2021 ........................... 136
Figure G 15: SHAP Values for the random forest on the annual model 2022 ........................... 137
Figure G 16: SHAP Values for the XGBOOST on the annual model 2022 .............................. 137
Figure G 17: SHAP Values for the random forest on the annual model 2015 ........................... 138
Figure G 18: SHAP Values for the random forest on the annual model 2015 ........................... 138

Figure F 1:Permutation feature importance for the random forest on the annual model 2015.. 121
Figure F.2: Permutation feature importance for the XGBoost on the annual model 2015 ........ 121
Figure F.3 : Permutation feature importance for the random forest on the annual model 2016 122
Figure F.4: Permutation feature importance for the XGBoost on the annual model 2016 ........ 122
Figure F.5: Permutation feature importance for the random forest on the annual model 2017 . 123
xi

Figure F.6: Permutation feature importance for the XGBoost on the annual model 2017 ........ 123
Figure F.7: Permutation feature importance for the random forest on the annual model 2018 . 124
Figure F.8 : Permutation feature importance for the XGBoost on the annual model 2018 ....... 124
Figure F.9: Permutation feature importance for the random forest on the annual model 2019 . 125
Figure F.10: Permutation feature importance for the XGBoost on the annual model 2019 ...... 125
Figure F.11: Permutation feature importance for the random forest on the annual model 2020 126
Figure F.12: Permutation feature importance for the XGBoost on the annual model 2020 ...... 126
Figure F.13: Permutation feature importance for the random forest on the annual model 2021 127
Figure F.14: Permutation feature importance for the XGBoost on the annual model 2021 ...... 127
Figure F.15: Permutation feature importance for the random forest on the annual model 2022 128
Figure F.16: Permutation feature importance for the XGBoost on the annual model 2022 ...... 128
Figure F.17: Permutation feature importance for the random forest on the annual model 2023 129
Figure F.18: Permutation feature importance for the XGBoost on the annual model 2023 ...... 129

xii

Acknowledgement
This works was supported by the Natural Sciences and Engineering Research Council of Canada
(NSERC), the BC Graduate Scholarship program and the Tula Foundation. I would like to thank
my supervisor, Dr. Brian Menounos, for your mentorship, for pushing my coding skillsets and
for your support in learning how to use remote sensing data for research purposes. Thank you to
my supervisory committee member, Dr. Lisa Wood, for providing guidance in the development
of my field research, access to your laboratory equipment and challenging the biological side of
my research. Thank you, Dr. Joseph Shea, for questioning my methods and keeping my science
reproducible. Another special thank you to the Airborne Coastal Observatory (ACO) for landing
your field spectrometer and the hyperspectral dataset as well as Rob White, Nick Viner, Steve
Beffort and Dereck Heatfield from the Hakai Geospatial team for your assistance with geospatial
data and your help with fieldwork. A huge thank you to Maria Kroeger, without whom I could
not have successfully completed the field component of my research. Thank you for keeping me
organized, taking care of all the logistics for me in Prince George's, driving winding roads, and
carrying the green monster over 1500+ metres of elevation. Another special thank you to Chris
Donahue for answering my one-million questions about radiative forcing and hyperspectral data,
your support was indispensable to the radiative and albedo section of Chapter 2. Thank you,
Caleb Andrew Mathias, for keeping me accountable and on track, and to Adam Hawkins, for
your support in the field. Another special thank you to my partner, Doug Massel, for believing in
me and putting up with the chaos of being a graduate student. And lastly, thank you to my
parents for your eternal support even when I moved across the country to follow my mountain
dreams.

xiii

1

Chapter 1: Introduction
This research sought to improve our understanding of processes of glacier surface darkening

by investigating the relationship between black carbon from wildfire soot and blooms of red
algae using a field experiment and remote sensing methodologies. This thesis focuses on fire
emissions, which are commonly referred to by a range of terms (smoke, ash, soot, etc) and are
typically made up of a complex mix of constituents, including black carbon, organic carbon,
char, trace metals, and mineral dust, among others. In this thesis, an assumption is being made
that black carbon is the dominant aerosol, and the most important fire emission aerosol for
impacting snow cover, and therefore, for simplicity, is being used as the primary term for
aerosols originating from fire
1.1 Motivations and Objectives
The formation of snow algae blooms and black carbon deposition from forest fire activity
during the snowmelt season represent two processes that diminish glacier albedo. Yet, the
interaction between snow algae and forest fire soot remains uncertain. My research objective was
to examine the interaction between black carbon from forest fires and snow algae on glaciers in
western Canada. In the following chapters, I:
A) Describe the field experiment during which I quantify the snow algae response to the
application of wood ash on the mid-elevation bench of Place Glacier, British Columbia,
Canada for the 2023 melt season (Chapter two);
B) Use remote sensing imaging from the Harmonized Landsat Sentinel datasets and
reanalysis data from MERRA-2 and ERA5-Land to investigate the interaction between

snow algae and black carbon and produce a regional analysis of glaciers in BC and
Alberta using machine learning (Chapter 3), and;
C) Synthesize and discuss the research limitations and recommendations for future work
(Chapter 4).

1.2 Importance of glaciers
Snow and ice account for about 70% of the global storage of freshwater (Stephens et al.,
2020). Loss of snow and ice cover will thus have major ramifications on accessibility to water
worldwide. In western Canada, glacier mass loss has great implications for society and the
ecosystem, as glacier runoff provides an integral source of freshwater to regional inhabitants
(Bonsal et al., 2020a), freshwater ecosystems (Moore et al., 2009a; Schoen et al., 2017), and
supply water for agriculture and hydropower purposes (Hock et al., 2019a). Glaciers deliver
cold, ample flows to many streams during late summer when conditions are typically hot and dry
(Bonsal et al., 2020b; Jost et al., 2012). Glacial contributions to streamflow, even for drainages
with glacier cover as low as 1 or 2%, are crucial during this period (Huss, 2011; Jost et al., 2012)
as they provide needed inputs of cold water. Yet, due to anthropogenic climate change, glacier
extents are rapidly declining (Hugonnet et al., 2021). As glaciers shrink and lose mass, runoff
contributions to streamflow are reduced (e.g. Huss, 2011; Moore et al., 2020; Stahl & Moore,
2006). For example, Anderson & Radić (2020) models predict unprecedented low streamflow for
many communities established along glacier-fed rivers in Alberta as glaciers disappear.

2

1.3 Glacier Darkening
Since the turn of the twenty-first century, rates of glacier mass loss accelerated (Hugonnet
et al., 2021). While increased global temperatures are largely responsible for accelerated mass
loss, there are other factors that may explain regional variations (Cook et al., 2020). Numerical
models based on increased global mean temperature project Western Canadian and US glaciers
to lose 60 to 100% of their mass by 2100, supporting that glaciers are highly sensitive to
temperature changes and that every degree of warming will have significant impacts on glacier
mass loss (Clarke et al., 2015; Rounce et al., 2023). Recent studies also suggest that changes in
glaciers’ albedo – the index measuring the proportion of reflected light at the Earth’s surface
ranging for 0 (full absorbance) to 1 (full reflectance)– is an important factor influencing surface
melt in the ablation season, along with air temperature (Box et al., 2012; Dumont et al., 2012;
Skiles et al., 2018). Thus, improving our understanding of the physical processes that drive
glacier darkening could provide insight into factors that could explain observed accelerated mass
loss in western Canada.
The energy balance at a glacier’s surface controls melt (Hock, 2005). Darker surfaces
absorb more solar radiation, increasing glacier melt (Box et al., 2012). Changes in the density
and snow grain size of snow and ice during melt season (Warren, 1982) and the deposition of
light-absorbing particles (LAPs), such as mineral dust, black carbon, and microbial growth,
darken glacial surfaces. The excess energy absorbed from solar irradiance due to surface
darkening from LAPs is characterized as radiative forcing and is measured in W m-2. LAPs lower
the albedo in the visible wavelength, the part of the light spectrum where snow is the most
reflective (Flanner et al., 2007; Benning et al., 2014; Biagio Di Mauro, 2020; Aubry‐Wake et al.,
2022). During the melt season, LAPs can accumulate on the snow surface if not entrained by the
3

melt runoff, increasing the darkening process and further decreasing the albedo. Similarly, the
bare ice surface on glaciers accumulates LAPs over the year, contributing to the ongoing
darkening process (Di Mauro, 2020; Flanner et al., 2007). Dust deposition on glaciers is mainly
due to the aeolian transport from arid and semi-arid landscapes (Clow et al., 2016), land-use
changes and from local lithology. Black carbon, the optically absorbing portion of soot, is
produced by incomplete combustion of fossil and biofuels and is emitted from anthropogenic
sources and forest fires (Kang et al., 2020). Finally, biological darkening is mainly caused by
blooms of snow algae (Chloromonas, Chlainomonas and Chlamydomonas sp.) and ice algae
(Mesotaenium and Ancylonema sp.) (Williamson et al., 2019).
On average, alpine glaciers albedo is 0.80 in the spring prior to the onset of melt
(Marshall & Miller, 2020), and typical summer snow albedo is about 0.5 due to changes in snow
grain sizes (Wiscombe & Warren, 1980). Blooms of red snow algae alone can further reduce the
albedo to 0.38, increasing snowmelt by 40% (Cook et al., 2017). In Western North America,
snow algae blooms are estimated to occur on 5% of the total glaciated area (Engstrom &
Quarmby, 2024). Similarly, on the western margin of the Greenland Ice sheet, studies found
biological darkening from ice algae to be the main albedo-reducing agent of the conspicuous
dark area known as the "dark zone" (Ryan et al., 2018; Cook et al., 2020; McCutcheon et al.,
2021). Meanwhile, Williamson & Menounos (2021) show that albedo for Western North
American glaciers is declining due to rising temperatures and propose that this decline might also
be linked to an increase in black carbon deposition from forest fires, as suggested by the change
in aerosol optical depth in MODIS global aerosol product.

4

1.4 Life on Snow
Glaciers and perennial snow patches worldwide support microbiological life, such as algae,
bacteria, and fungi (Hoham & Remias, 2020). During melt season, blooms of snow algae colour
the snow surface red, green and sometimes orange. Blooms of algae have been observed on the
Greenland Ice Sheet ( Uetake et al., 2010; Williamson et al., 2018; Cook et al., 2020) and
glaciers in Antarctica (Fujii et al., 2010; Huovinen et al., 2018; Khan et al., 2021), Europe
(Remias et al., 2009; Di Mauro et al., 2020), Asia (Yoshimura et al., 2000; Takeuchi et al., 2006;
Tanaka et al., 2016; Nakashima et al., 2021), Chile (Takeuchi & Kohshima, 2004) and North
America (Painter et al., 2001; Takeuchi et al., 2006; Ganey et al., 2017; Hamilton & Havig,
2017; Engstrom et al., 2020).
Organisms that subsist on the cryosphere are restricted to periods when liquid water is
available (Laybourn-Parry et al., 2012). Thus, algal blooms typically appear following surface
melt and will remain on the surface for the duration of the ablation season (Ganey et al., 2017).
Liquid water limits snow algae growth; therefore, algae bloom typically appears near the
transient snow line (Jones et al., 2001; Takeuchi, 2001; Takeuchi et al., 2013). Some studies
speculate that snow algal cells, during the incipient of snow melt, undergo a motile, haploid,
vegetative stage that allows them to 'swim' up to the surface (Laybourn-Parry et al., 2012; Davey
et al., 2019; Matsumoto et al., 2024). Using a collection of cell types from Chlainomonas
species, Matsumoto et al. (2024) hypothesized a life cycle for snow algae (Fig. 1.1). The life
cycle includes large quadriflagellate swarmers (A) in the early season that develop into large
non-motile red cells (F) that will remain on the snow surface during the melt season and have the
potential for asexual cell division through a protoplast exclusion process (K). The large red
immobile cells then transition into spores (I) when in wet snowpacks and will settle in the
5

snowpack or other surface, typically at the end of the melt season. At this stage, sporangia (L, M)
will develop from the spores. Pill cells (O) are then released from the sporangia and will develop
as flagellated pill cells (O+) to start the life cycle again.

6

Figure 1-1: Hypothesized Chlainomonas sp. life cycle in Bagley Lake Habitat (Matsumoto et al., 2024).
Phases denoted with a letter on a white background were observed in the field. Phases not observed in the
field are denoted with a “?” on a dark background. The dotted lines represent the transition process that
remains unknown.

7

Snow and ice surfaces present challenging conditions for life as they are subject to high
levels of irradiance, low nutrients, and freezing temperatures (Laybourn-Parry et al., 2012).
Organisms that live on snow and ice surfaces require many adaptations to cope with this harsh
environment. Multiple snow algae species from the genera Chlamydomonas and Chloromonas
produce secondary carotenoid pigment (mainly astaxanthin) that protects them against strong
solar radiation, freezing temperatures and as a response to low nitrogen levels (Remias et al.,
2005; Laybourn-Parry et al., 2012). The algal cells accumulate cytoplasmic secondary
carotenoids around the chloroplast, which attenuates the light travelling through their internal
photosynthetic apparatus and protects the cell against damage from overheating and
photodamage (e.g. Bidigare et al., 1993; Leya et al. 2009). Healy & Khan (2023) analyzed the
pigment composition of snow algae collected over an early and late season survey. They
observed an increase in photoprotective pigment, astaxanthin, to the detriment of photosynthetic
pigment such as chlorophyll α in snow algae cells in the late survey. The photoprotective
pigment absorbs part of the solar radiation and dissipates it into radiant energy as heat. The
added heat melts the surrounding snow and ice crystals and increases the local liquid water
content, promoting algae growth (Dial et al., 2018; Williamson et al., 2020). Secondary
carotenoid production also offers protection against freezing temperatures and is linked to
nitrogen starvation (Bidigare et al., 1993; Remias et al., 2012). When deprived of nitrogen, the
algae redirect their energy to produce nitrogen-free metabolites, such as fatty acids and
secondary carotenoids. The increase of production in secondary carotenoids often parallels a
decrease in cellular chlorophyll and primary carotenoids, leading to the colour change from
green to red (Müller et al., 1998). The presence of unsaturated fatty acids, which can be esterified
with astaxanthin, a hydroxyl member of secondary carotenoids (Leya et al., 2009), lowers the

8

water content in the alga cell and reduces damage from crystalline ice formation, providing
protection against freezing temperatures (Bidigare et al., 1993; Dial et al., 2018).
The production of secondary carotenoids results in red pigmentation of the snow algal
blooms, which darkens the snow and ice surface. Snow algae blooms have the potential to
significantly reduce glacier albedo and, thus, increase melt. A study looking at snow algae
blooms in all Western North America glaciated areas found that glaciers with blooms covering
up to two-thirds of their surface have an estimated increase of snow meltwater of 3 cm across the
entire glacier surface (Engstrom & Quarmby, 2024). On average, glacier and snow algae
accounted for 9 to 13% of surface melt on the south-western edge of the Greenland Ice Sheet
(Cook et al., 2020), 17% of snow melt on the Harding Icefield in Alaska (Ganey et al., 2017),
and 13% of surface albedo reduction in the Arctic (Lutz et al., 2016).
A few studies measured the radiative forcing of red snow algal blooms on glaciers. They
found that red snow algae's maximum instantaneous radiative forcing was 175 W m-2 in Alaska
(Ganey et al., 2017), 185 W m-2 in Antarctica (Khan et al., 2021), 295 W m-2 in the Purcell
Mountains, British Columbia (Engstrom et al., 2022), and 235.56 W m-2 in the North Cascade,
US (Healy & Khan, 2023). The cell density was higher in the Purcells than in Alaska, with a
maximum algal cross-sectional cell area of 1779 cm2 mL-1 compared to ~360 cm2 mL-1 (Ganey
et al., 2017; Engstrom et al., 2022), suggesting that high cell density at the snow surface has
greater impacts on the albedo than low cell density. In addition, higher radiative forcing from
snow algae in the mid-latitudes suggests that algal blooms' influence on snowmelt is greater in
the mid-latitudes compared to the high latitudes (Healy & Khan, 2023).
The parameters controlling snow algal spatial distribution and bloom intensity are poorly
understood, although the presence of nutrients has been linked to increased algae growth.
9

Previous studies found that nitrogen limitation, among other factors, often triggers the synthesis
of secondary carotenoids (Muller et al. 1998; Leya et al. 2009). On the Harding Glacier, Alaska,
snow algae blooms are more predominant on the continental side of the Icefield. Researchers
speculate that inputs of nutrients from terrestrial aerosols are higher on the continental than the
coastal side, promoting increased growth of algae blooms on the continental side (Takeuchi et
al., 2006). Likewise, experiments on red snow show that algae growth increased following the
addition of nitrogen, phosphorous and potassium (NPK) fertilizer in Alaska (Ganey et al., 2017).
On the Greenland Ice sheet, phosphorus-rich mineral dust from the local lithology promotes ice
algae growth (McCutcheon et al., 2021). Another study suggests that algae growth increased in
the presence of Fe-rich minerals (Phillips-Lander et al., 2020). Finally, photosynthetic activity on
glaciers from algal blooms was linked to a decrease in nutrient concentration in the meltwater,
especially nitrogen (Holland et al., 2019; Jones et al., 2001). Those studies suggest that the
presence of nutrients in the snow and ice is likely to be linked with the spatial distribution of
snow and ice algae.
1.5 Remote Detection of Snow Algae
Previous studies used remote sensing to map the spatial and temporal distribution of red
snow algal blooms. The first detection of red snow algae from remote sensing used airborne
hyperspectral images to map the algae distribution from a snowfield in the Sierra Nevada,
California (Painter et al., 2001). They observed a linear relationship between the spectral
reflectance depth of the chlorophyll absorption feature at 680 nm and algae abundance collected
on the field. Painter et al. (2001) created a model using the integral of the scaled chlorophyll
absorption feature and cell concentration from field observations and applied the model to the
hyperspectral image to map snow algae. Several studies used satellite optical remote sensing
10

such as SPOT, Landsat and Sentinel to map the extent and duration of red snow algae blooms
(e.g. Engstrom et al., 2022; Ganey et al., 2017; Takeuchi, 2013). Due to the coarse spectral
resolution of the sensors, the 680 nm chlorophyll absorption feature cannot be detected by
satellite. Takeuchi et al. (2006) used satellite images from SPOT-3 and targeted the broad
carotenoids absorption feature at 400-600 nm to estimate the spatial distribution and abundance
of red snow algae on the Harding Icefield in Alaska. They found a correlation between snow
algae abundance and the reflectance ratio of the red (610-680 nm) and green (500-590 nm) SPOT
satellite bands and used this ratio to estimate the distribution and abundance of algae blooms on
the icefield. Similarly, Ganey et al. (2017) utilized a normalized-difference spectral index
between the red and green Landsat-8 bands to target the carotenoid absorption feature to estimate
algae abundance over the Harding Icefield. The advantage of normalization that ratio is that it
considers varying light intensity between the different images. Others successfully used the redgreen normalized difference (RGND) index to map the extent of snow algae blooms from
satellite in Antarctica (Gray et al., 2020) and in British Columbia (Engstrom et al., 2022). Healy
& Khan (2023) used a multispectral camera mounted on an uncrewed aerial vehicle (UAV) to
map snow algae near Mt. Baker in Washington, US. They used a principal components analysis
and an optimized RGND to map the abundance of snow algae under differing bloom states. They
found that the RGND yields better distinction between snow and snow algae during early season,
likely due to clean early snowpack being more spectrally distinct from the algae patches than
aged snow filled with other impurities. A few studies pointed out that mineral dust and dirt have
a similar spectral signature to snow algae when using a ratio between the red and green bands,
which can potentially lead to an overestimate of algae abundance if other impurities are present
in the snow (Di Mauro et al., 2015; Di Mauro et al., 2024; Huovinen et al., 2018). A new spectral

11

index between the green and blue bands was developed to distinguish mineral dust from snow
algae when both are present (Di Mauro et al., 2024) and could be used to mask mineral
impurities when using the RGND index.
1.6 Forest Fire & Black Carbon
Prolonged summer dry seasons and reduced moisture in vegetation due to climate change
are leading to more frequent forest fires with increased duration, size and severity (Hanes et al.,
2018; Westerling et al., 2006). Forest fires in 2017, 2018, 2021 and 2023 in British Columbia,
exceeded the 10-year average in hectares of forest and land burned. Notably, over 2,800,000
hectares were burned in 2023 (Government of British Columbia, 2024). Forest fires release large
amounts of aerosols into the atmosphere, including black carbon (Kondo et al., 2011). Through
wet and dry processes, black carbon gets deposited on snow and ice surfaces, accelerating melt
due to increased radiative forcing and reduced surface albedo (Flanner et al., 2007; Skiles et al.,
2018). Additionally, the excess warmth released by the black carbon particles in snow increases
snow grain size, further reducing the snow albedo (Hadley & Kirchstetter, 2012a). Williamson
and Menounos (2021) observed that declining albedo from mountain glaciers in the last two
decades in western North America correlated with increasing air temperatures and deposition of
aerosols from forest fires. Similarly, in the North Cascade Mountains, Washington, USA, black
carbon and dust deposition from the Big Hump forest fire in 2012 was linked to the observed
increased glacier melt from that year compared to other years (Kaspari et al., 2015).
As forest fire aerosols can be transported over long distances, the impact of black carbon
has regional impacts beyond the sources of emission (Kim et al., 2005). On the Greenland ice
sheet, observations from shallow ice cores (Keegan et al., 2014) and snow pits (Thomas et al.,

12

2017) linked black carbon deposition to forest fire activity in North America. In the Himalayas,
black carbon concentration reached high levels (20-25 ng mL-1) in southeastern Tibetan glaciers
following episodes of wildfires, increasing melt (You et al., 2016). Glaciers in New Zealand
experienced unprecedented snow-darkening and melting from smoke plumes from the disastrous
forest fires in Australia in 2019 – 2020 (Pu et al., 2021).
Forest fire smoke has been linked to aerosol optical depth (AOD) (Barnaba et al., 2011),
providing a means for remote sensing detection of atmospheric black carbon flux. The AOD is a
parameter used to determine the aerosol load distributed within a column of air in the
atmosphere. In other words, it indicates how much light is blocked by aerosols in the
atmosphere. Thomas et al. (2017) used the Cloud-Aerosol Lidar with Orthogonal Polarization
(onboard CALIPSO) and Moderate Resolution Imaging Spectroradiometer (MODIS) Aqua
instrument to detect aerosols from the forest fires while moving from Canada to Greenland,
linking black carbon deposition on the Greenland Ice sheet to forest fires. Williamson and
Menounos (2021) used AOD at wavelengths 550 nm to quantify forest fire-generated aerosols
from the two MODIS sensors, Terra and Aqua, to assess the importance of black carbon in
altering glaciers’ albedo.
NASA’s latest reanalysis dataset, the Modern-Era Retrospective Analysis for Research
and Applications, version 2 (MERRA-2), uses the Goddard Earth Observing System, version 5
(GEOS-5), to assimilate field observation and modelled datasets to provide aerosols’ surface
mass concentration amongst other datasets (Randles et al., 2017). MERRA-2 assimilates biascorrected AOD from MODIS and the Advanced Very High-Resolution Radiometer (AVHRR)
instruments, and non-bias-corrected AOD retrievals from the Multiangle Imaging
SpectroRadiometer (MISR) over bright surfaces and ground-based Aerosol Robotic Network
13

(AERONET) observations. MERRA-2 then uses the GEOS-5 radiatively coupled to the Goddard
Chemistry, Aerosol, Radiation and Transport model (GOCART) to simulate five types of
aerosols, including dust, sea salt, sulphate and black and organic carbon. The black carbon
aerosol surface mass concentration dataset has been used to characterize the historical trends and
contributing sources of black carbon in Beijing, China (Qin et al., 2019). The MERRA-2 black
carbon dataset was also used to analyze the spatiotemporal variation of black carbon over
Northern Eurasia during the 2016 Siberian forest fires (Sitnov et al., 2020) and to assess and
quantify atmospheric black carbon contribution from forest fires in the circum-Artic in 2019 and
2020 (X. Chen et al., 2023).

1.7 Research Gap
Glaciers are critical sources of freshwater but are rapidly melting due to rising temperatures.
Several studies found that a reduction in albedo from LAPs, including dust, black carbon and
snow algae, accelerates melt (Di Mauro, 2020; T. H. Painter et al., 2013; Skiles et al., 2018).
While the impact of LAPs has been widely studied, as presented above, the interaction among
them remains poorly understood. In particular, the interaction between red snow algae and black
carbon from wildfire soot has yet to be studied. The chapters below seek to address this
knowledge gap through a field experiment and by leveraging remote sensing and machine
learning techniques.

14

2

Chapter 2: Quantifying the interaction between wood ash and snow algae
on Place Glacier, British Columbia

2.1 Introduction
Western Canada’s glaciers are retreating at an accelerated rate (Bevington & Menounos,
2022), having consequences on the regional hydrology, ecosystem (Moore et al., 2009b), and
hydroelectric power generations due to changes in runoff (Hock et al., 2019b). Air temperature
and albedo - an index measuring the proportion of reflected light at the Earth's surface - largely
influence glacier surface melt during the ablation season (Box et al., 2012). Albedo reduction
occurs due to changes in snow and ice grain sizes (Wiscombe & Warren, 1980) and the
accumulation of LAPs such as mineral dust (Di Mauro et al., 2015), black carbon (Flanner et al.,
2007), and microbial growth (Healy & Khan, 2023). Many studies investigated the impact of red
snow (e.g. Engstrom et al., 2022), mineral dust (e.g. Di Mauro et al., 2015), black carbon (e.g.
(Flanner et al., 2007; Hadley & Kirchstetter, 2012b) and the combination of dust and black
carbon (e.g. Nagorski et al., 2019) on snow and glacier albedo, but research focused on the
interaction between snow algae and black carbon is lacking.
Unlike inorganic LAPS, red snow algae forms perennial populations that resurface in the
spring and summer, following overwintering burial by snow when sufficient interstitial water is
available (Jones et al., 2001). While snow algae blooms can cover up to 65% of an individual
glacier surface, they display a high degree of interannual variability in intensity and spatial
extent (Engstrom & Quarmby, 2024), which is not fully understood. It is well known that snow
algae are both nutrient- and water-limited (e.g. Anesio & Laybourn-Parry, 2012; Ganey et al.,

15

2017; Jones et al., 2001); however, the provenance of these nutrients for Western Canadian
glaciers remains poorly documented.
Glaciers are oligotrophic environments, meaning that they have a low nutrient content.
Most nutrient input on glaciers comes from precipitation and airborne material (Hodson et al.,
2008; Anesio & Laybourn-Parry, 2012). On the ‘Dark Zone’ of the Greenland Ice Sheet, a
nutrient addition experiment indicates phosphorous from mineral dust from local lithologies to
be a limiting nutrient for ice algae blooms (McCutcheon et al., 2021). In the Arctic, snow algae
preferentially grow in the snow with high concentrations of dust, which is believed to provide
input of nutrients for the snow algae population (Lutz et al., 2015), and on North American
stratovolcanoes, increased carbon dioxide concentrations promoted snow algae growth
(Hamilton & Havig, 2020). A field experiment in Alaska demonstrated that adding aqueous
nitrogen, phosphorous and potassium nutrients increased snow algae concentration by fourfold
(Ganey et al., 2017). Similarly, snow algae in Antarctica bloomed in greater populations in
coastal regions than inland regions, likely due to the input of nutrients from local minerals and
marine fauna on the coast (Hodson et al., 2021). Red snow algae nutrient limitations generally
stem from local mineralogy and the environment, suggesting that those limitations are not global;
rather, they are driven by regional geology and environmental processes.
In Western North America, forest fires are common between spring and autumn, releasing
large amounts of aerosol and black carbon into the atmosphere (Kondo et al., 2011). Black
carbon is a light-absorbing particle and short-lived aerosol emitted from the incomplete
combustion of carbon-based fuels such as fossil fuels, biofuels, and wood. It absorbs solar
radiation across all wavelengths in the visible spectrum and impacts the radiation budget through
positive radiative forcing, increasing melt when deposited on snow and ice (Kang et al., 2020).
16

This enhanced radiative forcing is expected to increase liquid water at the glacier surface. Black
carbon also has the potential to nourish snow algae. In the Arctic Ocean, forest fire aerosol
deposition was linked to increased summertime phytoplankton bloom in 2014 (Ardyna et al.,
2022). The study suggests that aerosol from Siberian forest fires in the summer of 2014 led to Nenrichment of the sea surface, resulting in increased phytoplankton activity.
The deposition of black carbon on snow and ice can accelerate total glacier melt by up to
20% and reduce snow cover duration by several days (Kang et al., 2020). Following large forest
fires and intense forest fire seasons, several field and model observations found an increased
deposition of black carbon on snow and ice (e.g. Kaspari et al., 2015; Marshall & Miller, 2020;
Williamson & Menounos, 2021). As forest fire activity and aerosols intensify, increased snow
water and nutrients from the black carbon may enhance snow algae growth, further impacting
glacier albedo and melt. However, little is known about the interaction between snow algae and
black carbon, despite mentions made about this potential feedback (Aubry‐Wake et al., 2022).
In this study, I conducted a field experiment on Place Glacier, British Columbia, Canada
[50.42°, -122.6°], where I fertilized snow with wood ash and quantified the treatment effect on
snow algae growth and melt. I measured snow algae cell abundance, melt and surface spectral
reflectance in the field during three field visits conducted in the summer of 2023. I estimated the
radiative forcing and associated snowmelt potential of snow algae using the spectral
measurements taken on the first field visit and compared it to previous studies.

17

2.2. Methods
2.2.1 Study Site
I selected Place Glacier for the field experiment due to its ease of access via helicopter
from Pemberton and field observation of snow algae blooms in 2022. Place Glacier is a small (3
km2) glacier located in the southern Coast Mountain range (Fig. 2.1), northeast of Pemberton,
British Columbia, Canada (50.42°, -122.6°). The glacier extends from ∼2610 m a.s.l. (above sea
level) in the accumulation zone to ∼1900m a.s.l at the toe (Mukherjee et al., 2022). Winter
accumulation at Place Glacier results in a deep snowpack that covers the glacier ice late into the
summer ablation period.

A

B

Figure 2-1 A. Map showing the location of Place Glacier in the southern Coast Mountains of British Columbia,
Canada. Blue polygons are glaciers from the Randolph Glacier Inventory V. 7 (RGI Consortium, 2023). B.
Hyperspectral mosaic RGB of Place Glacier on 22 August 2022 produced by the Airborne Coastal Observatory. The
glacier boundary is outlined in red.

18

2.1.1 Snow algae and black carbon experiment
I conducted a replicated (10 blocks x 4 plots block -1 = 40 plots total), manipulative field
experiment to quantitatively measure the interaction between snow algae and black carbon from
wood ash (Fig 2-2). The experiment occurred on the mid-elevation bench of Place Glacier
(approx. 2000m a.s.l.) from 3 June to 2 July 2023. Each experimental block consists of four
circular plots, each plot being 2.0 m2 in area. In each block, I selected two plots with an
established bloom (colonized) and two plots with no visible bloom (uncolonized), spacing the
plots by at least 2 meters. One of the colonized and one of the uncolonized plots received a
treatment of 0.4 g of wood ash, with particles filtered to 50 µm, and the other two plots remained
as control. In the absence of a perfect experimental treatment for emissions from fire, wood ash
was used as a proxy. I replicated this plot setup ten times, spacing each block by at least 15 m.

19

Figure 2-2: Map of the experiment plots and their location on Place Glacier, BC, Canada. Location for
P8-2 and P8-1 missing.

I used three measurements to quantify the effect of wood ash on the snow algae: snow
melt, algae cell abundance and spectral reflectance. I measured snow melt and algae abundance
during our three field visits on 3-4 June, 11 June, and 2 July. I only collected optical properties
on 3-4 June for all plots and for four of the blocks on 11 June due to rapidly changing
atmospheric conditions. The field logistics on July 2 were not conducive to collecting optical
properties, as I accessed the glacier by foot. Elemental metal and total carbon and nitrogen
concentrations were determined for half of the treated and non-treated plots, randomly selected.
This allowed to understand the metal concentration present in the snow that did not receive
treatment and in the snow that received the ash treatment. Table 2-1 shows which plots on which
dates receive reflectance, cell count and nutrient measurements.

20

Table 2-1: Description of measurements conducted per dates on the experiment plots. Snowmelt is not
included, as each plot on each field visit was measured. No*: Data loss due to human error during
processing of the sample.
Test
ID
P1-1
P1-2
P1-3
P1-4
P2-1
P2-2
P2-3
P2-4
P3-1
P3-2
P3-3
P3-4
P4-1
P4-2
P4-3
P4-4
P5-1
P5-2
P5-3
P5-4
P6-1
P6-2
P6-3
P6-4
P7-1
P7-2
P7-3
P7-4
P8-1
P8-2
P8-3
P8-4
P9-1
P9-2
P9-3
P9-4
P10-1
P10-2
P10-3
P10-4

June 3
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
No
No
No
No
No
No
No
No
No
No
No
No
No
No
No
No
No
No
No
No

Reflectance
June 4
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
No
No
Yes
Yes
Yes
Yes
Yes
Yes
No
No
Yes
Yes
No
No
Yes
Yes
No
No
Yes
Yes
No
No

June 11
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
No
No
No
No
No
No
No
No
No
No
No
No
No
No
No
No
No
No
No
No
No
No
No
No
No
No
No

21

Cell Count
June 4
June 11
Yes
Yes
No
Yes
Yes
Yes
No
Yes
Yes
Yes
No
Yes
Yes
No*
No
Yes
Yes
Yes
No
Yes
Yes
Yes
No
No*
Yes
Yes
No
Yes
Yes
Yes
No
Yes
Yes
Yes
No
Yes
Yes
Yes
No
Yes
Yes
Yes
No
Yes
Yes
Yes
No
Yes
Yes
Yes
No
Yes
Yes
Yes
No
Yes
Yes
Yes
No
Yes
Yes
Yes
No
Yes
Yes
Yes
No
Yes
Yes
Yes
No
Yes
Yes
Yes
No
Yes
Yes
Yes
No
Yes

July 2
No
Yes
Yes
Yes
Yes
Yes
Yes
No
Yes
No
Yes
Yes
Yes
Yes
No
Yes
Yes
No
No
No
Yes
Yes
No
No
Yes
No
No
No
No
No
No
Yes
No
No
No
No
No
No
No
No

Nutrient
June 4
Yes
No
Yes
No
No
Yes
No
No
No
No
No
Yes
No
Yes
Yes
No
Yes
No
Yes
No
No
Yes
No
No
Yes
Yes
No
Yes
Yes
Yes
Yes
No
Yes
No
Yes
No
No
Yes
Yes
Yes

I measured snow melt in each plot using ablation stakes drilled in the centre of each plot.
I measured the length of the stake extending above the snow surface, with the depth of snow
calculated as the difference in the current and previous measured lengths. To avoid introducing
errors from snow cups and local melt around the stake, I measured the snow height from an ice
axe set perpendicularly to the base of the stake. On 3 June and 4 June, I also measured the depth
of the snowpack using an industrial gridded steel probe. Using 3 hours near-surface air
temperature for Place Glacier collected at 1924 m.a.s.l. shared from the Geological Survey of
Canada (appendix A), I calculated the melt factor using a simple degree day model as follows:
𝑀 = 𝐾𝐼 𝑃𝐷𝐷 + 𝐾𝑆 𝑃𝐷𝐷

(2.1)

Where M is the depth of snow melted in millimetres water equivalent w.e., PDD is the positive
degree day sum for the specific period and KI and KS are the degree day melt factors for ice and
snow respectively. For this study, I calculated the KS only and converted the height of snow melt
to mm w. e. using a snow density of 500 kg m-3. I then compared the melt factor (KS) to the
annual melt factor modeled by Shea et al. (2009) for Place Glacier from 1965 to 1995.
I sampled snow algae abundance from snow cookies in all plots three times during the
experiment. At the time of our last visit on 2 July, the snow had already melted out over the
location of the experimental blocks, but I sampled wet snow when present. I collected snow
cookies of 7.62 cm (3 inches) in diameter and approximately 2 cm thick in each plot using a
short segment of PVC tube labelled with a 2 cm depth line. I chose to sample the top 2 cm of the
snow based on previous studies observing the majority of the red snow algae located at or near
the snow surface (Engstrom et al., 2022; Ganey et al., 2017) and because the optical properties of
snow are mainly affected by the top 2 cm layer (Di Mauro et al., 2024). Using a clean spatula to
cut the snow at the base of the PVC tube, I carefully transferred the snow cookie into a Whirl22

Pak® bag upon collection. The samples were kept frozen on snow in the field and transported in
coolers from Pemberton to the University of Northern British Columbia in Prince George, where
they were stored at -20ºC.
2.2.3 Cell abundance
I conducted cell count at the University of Northern British Columbia using a Nikon FN1
microscope and a hemacytometer. The samples were thawed at room temperature, then preserved
in glutaraldehyde. The melted samples were homogenized, and a 10 mL aliquot subsample was
fixed with 50% glutaraldehyde to 1% final concentration. I measured algae abundance as cell
count per unit of volume via microscopy. I performed algal cell count with an Improved
Neubauer Hemacytometer cell counting chamber under a 40x objective. I counted individual
algal cells, measured cell diameter, and cell surface area using the NIS Elements software. All
nine 1x1 mm grids of the hemacytometer were used for the count. For each algae count, I subsampled 1 mL from the preserved preparation and performed four cell and dimension counts
over 0.1 mL each. The surface area and diameter of the cells were averaged for each count, and
the mean of the four counts was retained. All cells were considered in each count due to the low
population number per 0.1 mL. The averaged cells per volume were scaled per unit of area using
the following formulas to allow for comparison with other studies.
𝐴𝑣𝑒𝑟𝑎𝑔𝑒 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑐𝑒𝑙𝑙𝑠 ×104

𝑐𝑒𝑙𝑙𝑠

𝑁 𝑇𝑜𝑡𝑎𝑙 = 𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑠𝑞𝑢𝑎𝑟𝑒 𝑐𝑜𝑢𝑛𝑡𝑒𝑑 𝑜𝑛 ℎ𝑒𝑚𝑜𝑐𝑦𝑡𝑜𝑚𝑒𝑡𝑒𝑟 ( 𝑚𝐿 ) × 𝑠𝑎𝑚𝑝𝑙𝑒 𝑣𝑜𝑙𝑢𝑚𝑒
𝑁 𝑡𝑜𝑡𝑎𝑙

𝐶𝑒𝑙𝑙𝑠 𝑚−2 = 𝑆𝑛𝑜𝑤 𝑐𝑜𝑜𝑘𝑖𝑒 𝑎𝑟𝑒𝑎 (𝑚2 )

(2.2)
(2.3)

23

2.2.4 Empirical measurement of snow reflectance
I measured spectral reflectance from 350 to 2500 nm over the plots using the Spectral
Evolution RS-5400 portable spectrometer, which has a spectral resolution of 2.7 nm from 350 to
700 nm, 5.5 at 1500, and 5.8 at 2100 nm. I collected spectra with the bare optic pointing directly
downward, held 15 to 30 cm above the snow surface. I conducted the measurements at ±2h from
solar noon. Reference scans were taken approximately every 15 minutes over a 99% clean white
5x5 inch Spectralon board to account for changing atmospheric conditions.
2.2.5 Radiative Forcing and Red Snow Algae Melt Estimate
To compare with previous studies, I estimated the snow algae instantaneous radiative
forcing (IRF) from the reflectance data over the visible spectrum on 3 and 4 June for the
estimated snow grain size, following similar method used by Engstrom et al. (2022), Ganey et
al.(2017) and Healy & Khan (2023). I calculated the IRF of the snow algae over the visible
spectrum as follows:
IRF ≈ ∑850
350 𝐼( 𝜆) ( 𝛼𝑤ℎ𝑖𝑡𝑒 (𝜆) − 𝛼𝑟𝑒𝑑 (𝜆)) 𝛥𝜆

(2.4)

Where IRF is calculated as the sum of the spectral solar irradiance (I), multiplied by the
wavelength interval (𝛥𝜆) and the difference between clean snow albedo (𝛼𝑤ℎ𝑖𝑡𝑒 ) and the snow
algae reflectance (𝛼𝑟𝑒𝑑 ) over the wavelengths 350 to 850 nm. The field spectrometer
measurements (𝛼𝑟𝑒𝑑 ) were collected on flat grounds of < 5º slope, and thus, the error due to the
anisotropy factor is considered negligible (C. Donahue, personal communication, 2024),
allowing the use of the reflectance data without conversion. Downwelling irradiance was
measured at Place Glacier on 3 June and 4 June using a calibrated cosine diffuser attached to the

24

field spectrometer. I obtained clean snow albedo (𝛼𝑤ℎ𝑖𝑡𝑒 ) from SNICAR-ADv3 (Flanner et al.,
2021) using the following parameters. The solar zenith angle was estimated to be 34º for Place
Glacier on 3 June and 4 June 2023 at 18:30 UTC (11:30 local time) using the ‘pvlib’ open-source
Python tool (Jensen et al., 2023). The snowpack thickness was derived from the snowpack depth
measured on the field (~2m), and snowpack density was set to 500 Kg/m3. The snow grain shape
was set to spheroids. Multiple effective grain sizes were input based on the nearest 50 nm as
modelled from the measured spectra ice absorption feature centred at λ = 1030 nm (Nolin &
Dozier, 2000). I calculated the effective snow grain size from the field spectra based on the
continuum-removed depth rather than integrating the area to reduce uncertainty introduced by
wet snow, which can shift the location of the continuum minimum (Donahue et al., 2023). All
the other parameters in SNICAR were set as the default or zero. The model outputs the albedo at
10 nm intervals (𝛥𝜆).
From the IRF calculated for the red snow algae, I calculated a first-order estimate of the
glacier and snow melt rate by applying a simple energy analysis, similar to the method used in
Donahue et al. (2023), Kaspari et al. (2015) and Painter et al., (2013). For this calculation, I
assumed that the snow and ice are at the melting point (0°C) and that the additional energy is
used entirely for melting and is not contributing to raising the temperature of the snow and ice.
The assumption that isothermal snow and ice are at 0°C is reasonable, as from the interpretation
of the air temperature sensor at Place Glacier, the snow temperature is very close to 0°C when
the snow begins to melt and is isothermal between May 1st and May 15th, 2023, before the
sensor is snow free on May 17th (Appendix A, Fig. A1). Additionally, I did not consider the
amount of energy utilized for the photosynthesis process by the algae in this calculation, as
suggested by Cook et al. (2020), and assumed that all of the absorbed energy is used for melting.
25

I obtained hourly instantaneous radiative forcing (HRF) as described in equation 2.3, using
incoming irradiance simulated for Place Glacier on June 3rd, 2023 from the PVSystem solar
irradiance program (https://pvlighthouse.com.au) acquired at hourly intervals from sunrise to
sunset. The radiative forcing is assumed to remain constant during each 1h time step; therefore,
the total hourly energy supplied by the radiative forcing is equivalent to HRF multiplied by 3600
s h-1. To calculate the red snow algae contribution to melting, I divided the total hourly energy by
the latent heat of fusion of water at 0°C (0.334x106J kg-1), returning melt in terms of millimetres
water equivalent per hour (mm w.e hr-1). I then summed the hourly melt to obtain the estimated
daily melt caused by the red snow algae.
2.2.6 Total metals, carbon and nitrogen content
The Northern Analytic Laboratory Service (NALS) at the University of Prince George
tested the wood ash and randomized snow samples for total metals. The randomized snow
samples represented samples from ten control plots and ten treated plots with the wood ash, all
without snow algae as per visual inspection. The total metals test conducted by NALS followed
the United States Environment Protection Agency (USEPA) method 200.7 (U.S. Environmental
Protection Agency, 1994a) for assessing metals in the snow samples (liquid) and method 200.2
(U.S. Environmental Protection Agency, 1994b) for digesting the wood ash (solid) with an
Agilent 5100 ICP-OES on the wood ash digestates. The statistical difference in nutrients between
the treated plots and control plots was assessed using t-tests.
To identify which total metals most strongly influence algae growth, I applied a forward
stepwise feature selection model using linear regression. The forward stepwise feature selection
is a method that uses machine learning and statistics to build a model in which the most relevant

26

variables, nutrients in this case, are added iteratively until no further improvement is achieved or
a predetermined number of variables is reached. I used the mlextend Python library to perform
the algorithm and selected the top three nutrients. I then used the three selected nutrients in an
ordinary least squares (OLS) regression model to estimate their individual contributions to algae
cell count from June 11.
NALS performed a total carbon and nitrogen element test on the wood ash. They used a
Costech 4010 Elemental Analyzer to combust the samples via the standard sequential
combustion/reduction setup recommended by Costech with a flow rate of 100 mL min-1 of
helium 5.0 as the carrier gas. The combustion oven was set to 1000°C, the reduction oven to
650°C, and the GC oven to 70°C. Gases were then separated on a standard 3m column and
quantified with the built-in thermal conductivity detector (TCD).
2.2.7 Statistical Test
Statistical differences in the snowmelt and snow algae concentration and diameter among the
different groups were assessed using the ANOVA and Kruskal-Wallis test based on the
distribution of the dataset. I used the Shapiro-Wilk test to determine the normality of the
snowmelt and snow algae concentration and diameter data. The Shapiro-Wilk test quantifies how
well the data fit the normal distribution with the test statistic (W), ranging from 0 (deviation from
normality) to 1 (normally distributed), and provides a ρ-value that indicates whether the data
follows a normal distribution. A ρ-value < 0.05 indicates that the data do not follow a normal
distribution. Based on the Shapiro-Wilk test results (Table 2-2), all the snowmelt data were
normally distributed (ρ > 0.05), and only part of the cell concentration and diameter data
followed a normal distribution; therefore, I conducted the ANOVA test on the snowmelt data and
the Kruskal-Wallis on the cell concentration and diameter data. For test results where the
27

ANOVA or Kruskal-Wallis was significant (p-value < 0.05), I applied a post-hoc test to identify
which pairs of groups are significantly different through pairwise comparisons. Since only one of
the Kruskal-Wallis tests indicated a significant difference, I used Dunn’s test with Bonferroni
correction to determine which group pairs are significantly different. Due to the small sample
size on July 2 for the cell concentration (n=17) and diameter (n=15), I also applied the nonparametric Mood’s Median test to compare the cell concentration medians to verify for Type II
error in the Kruskal-Wallis test (failing to detect a real difference).

Table 2-2: Shapiro-Wilk Test results for the snowmelt and cell concentration data and sample size (n).

Snowmelt June
11
Snowmelt July 2
Cell
concentration
June 11
Cell
concentration
July 2
Cell diameter
June 11
Cell diameter
July 2

W-stats
ρ-Values
n
W-stats
ρ-Values
n
W-stats
ρ-Values
n
W-stats
ρ-Values
n
W-stats
ρ-Values
n
W-stats
ρ-Values
n

Colonized
Treated
0.843
0.313
10
0.914
0.313
10
0.859
0.0745
10
0.921
0.513
6
0.665
0.001
9
0.942
0.672
6

Uncolonized
Treated
0.953
0.279
10
0.91
0.279
10
0.804
0.0163
10
0.957
0.759
4
0.899
0.406
5
0.964
0.807
4

28

Colonized
Control
0.889
0.744
10
0.956
0.744
10
0.842
0.0607
10
0.75
0.000
3
0.818
0.033
9
NA
NA
1

Uncolonized
Control
0.959
0.249
10
0.905
0.249
10
0.791
0.0162
10
0.927
0.576
4
0.943
0.690
5
0.931
0.601
4

According to the Shapiro-Wilk test, the radiative forcing for clean snow (stat = 0.922, ρ = 0.334),
snow colonized with snow algae (stat = 0.881, ρ = 0.133), and snow darkened with ash (stat =
0.946, ρ = 0.674) are normally distributed. I assessed the statistical difference in radiative forcing
between clean snow with the ANOVA test and used the post hoc Dunn’s test to determine which
groups are significantly different.
2.3. Results & Discussion
2.3.1 Wood ash effect on snow algae abundance and melt
Snow algae blooms in the Coast Mountains typically appear in late June or early July and
can last late into the summer (Engstrom & Quarmby, 2024). Snow algae bloomed approximately
one month early in 2023, with observations in the Olympic Peninsula and on Place Glacier of
snow algae as early as late May. Temperatures in May 2023 were higher than normal, with
temperatures in the Coast Mountains near Place Glacier reaching 4-5°C above normal (Zhang et
al., 2023). Snow algae typically bloom when sufficient liquid water is available within the
snowpack. Thus, the higher-than-normal temperatures in May likely triggered the early bloom.
Compounded with snowpack that was 20% below normal for the area around Place Glacier (B.C
River Forecast Centre, 2023), early snow loss, as observed on 2 July, resulted in a reduced snow
algal growth season on the Place Glacier bench, providing a short window for the field
experiment.
Cell abundance and diameter
The statistical tests provide no evidence that wood ash had an impact on snow algae cell
concentration during the experiment (Fig. 2-3). The Kruskal-Wallis test performed on the cell
abundance on June 11 indicated a significant difference in the data (Stats = 16.93, ρ = 0.0007, n
29

= 40). The post-hoc Dunn’s test, however, showed that the significant differences in algae cell
median concentration were between the colonized and uncolonized plots, and no significant
differences were identified between the treated and control plots. The colonized treated cell
concentration was significantly higher than the uncolonized treated (ρ = 0.0029, n = 40) and the
uncolonized control (ρ = 0.031, n = 40), and the colonized control cell concentration was
significantly larger than the uncolonized treated (ρ = 0.0272, n = 40). These results suggest that
plots that were already colonized during the first field visit support a higher cell concentration
than the plots that had no visible red snow algae (uncolonized plots). This difference is likely
because the cells in the colonized plots had undergone cell reproduction during the seven days
between the two field visits, leading to a higher cell concentration. Whereas the algae cells in the
uncolonized plots were possibly just starting to emerge. The microscopy, however, provided no
evidence for this hypothesis. All cells observed in the microscopy were red, and no specific
observations were conducted to discriminate among the cell types, as presented in Figure 1-1. At
the end of the experiment, no significant difference was detected among the plots’ concentration
in the Kruskal-Wallis test (Stats= 3.033, ρ = 0.387, n = 17) and in the Mood's Median test (Stats
= 0.830, ρ = 0.841, n =17). The cell abundance results, therefore, show that experimental
additions of wood ash had no effect on cell abundance during the field experiment.

30

Figure 2-3: Mean cells per mL for the 4 plot settings from 4 June, 11 June and 2 July 2023 on Place
Glacier, British Columbia, Canada

Algal cells per square meter during the field experiment varied from a minimum of 4.9
× 106 to a maximum of 8.0 × 107 on June 4, and 2.9 × 106 to 8.8 × 107 on June 11. The algae
sample collected on 2 July consisted of firn and slushy snow, making it difficult to collect
consistent volume. For this reason, I did not convert the cells ml-1 to cells m-2. The maximum
algal cell concentrations are comparable to those previously observed in the Eastern Alaska
Mountains [5.1 × 107 (Takeuchi et al., 2013) and in the Sierra Nevada Mountains, California [1.3
× 107 ] (Painter et al., 2001).
The wood ash treatment did not impact cell diameter during the field experiment. The
Kruskal-Wallis test on June 11 (Stats = 2.878, ρ = 0.411, n = 28) and July 2 (Stats = 2.896, ρ =
31

0.408, n = 15) indicated no significant difference in the cell diameter among the plot types. The
Mood’s Median test indicated the same lack of statistical difference among the cell diameter on
July 2 (Stats = 2.946, ρ = 0.400, n = 15). The averaged cell diameter during the experiment
varied from 13 ± 4 µm to 41 ± 3 µm with a mean of 27 ± 9 µm. Cell diameter from algae
samples collected in the Southern Coast Mountains had similar dimensions, with a diameter
ranging from 20 to 30 µm (Engstrom et al., 2022).
Snowmelt over experimental plots
Snowmelt on Place Glacier over the field experiment from 3 June to 2 July, 2023
averaged 135.1 ± 7.1 cm or 675.5 ± 35.5 mm w.e (Fig. 2-4). I observed no significant difference
in the measured melt between the plot treatments and the colonized versus non-colonized plots
on June 11(ρanova = 0.18) and July 2 (ρanova = 0.23), as calculated by the ANOVA statistical test.
The melt factors for snow from June 3 to June 4, June 4 to June 11 and June 11 to July 2 are
respectively 3.5, 4.5 and 3.7 mm w.e +ºC-1 d-1, and is 3.8 mm w.e +ºC-1 d-1 for the entire
experiment period. In comparison, Shea et al. (2009) modelled a melt factor of 2.71 mm +ºC-1 d-1
for Place Glacier.

32

Figure 2-4: Snowmelt and temperature trends on Place Glacier, British Columbia, Canada. A. Daily
temperature, average measure melt and positive degree day melt model (PDM) from Shea et al. (2009) B.
Melt per plot setting from 4 June and 11 June 2023. C. Melt per plot setting from 11 June and 2 July
2023.

The average melt consistently trends higher for the treated colonized plots during the
experiment. However, the difference is not large enough to be significant and only varies by a
few centimetres from the other averages. I expected to observe higher snow melt in the treated
and colonized plots from the increased radiative forcing and reduced albedo from the algae and
ash. A few factors may explain why no significant difference was observed between the plot
settings. First, I recorded snowmelt as a point measurement taken in the middle of the plot, as
described in section 2.2.2. The algae blooms typically did not cover the entire plot area and were
rather concentrated in heterogeneous snow cup patches. Therefore, it is possible that the
measurement did not capture melt induced by the algae if they were not taken directly within the
33

immediate bloom area. Second, the melt caused by the algae during the 30-day experiment in
June was possibly too little to be detected by statistical tests. Engstrom et al. (2022) estimated 16
cm of melt from snow algae over a two-month period (July and August) for two glaciers in the
Purcell Mountains. Based on this estimate, the bloom observed on Place Glacier during our 30day field experiment would have produced approximately half the melt (8 cm). Compared to an
average total melt of 137 cm, the melt induced by the snow algae might have been too small to
be detected in statistical tests and may explain the lack of statistical differences between the
colonized and non-colonized plots.
In comparison to the Shea et al. (2009) melt model, the calculated melt factor for the field
experiment was 3.6 mm w.e +ºC-1 d-1, compared to 2.71 mm w.e +ºC-1 d-1. This discrepancy
could be explained by the fact that the melt factor calculated for the field experiment is over a
much shorter period compared to Shea et al. (2009), which was modelled for the entire ablation
seasons (May 15 to September 30) from 1965 to 1995. In contrast, the melt factor measured
during this experiment was calculated from June 3 to July 2, around peak insolation. The
difference between the two melt factors may also be due to surface albedo, as glaciers in the
decades have become darker (Williamson & Menounos, 2021). This difference supports the
concept that sensible heat alone is not the only factor that enhances melt on glaciers and that
radiation heat from aging snow grain and light absorbing particles contribute to glacier melt via
radiative forcing.
2.3.3 Wood ash Nutrients
Previous field studies and laboratory observations suggest that algae's response to
nutrients is study-dependent. For example, a study on North American volcanoes (Hamilton &

34

Havig, 2017) and in the Arctic (Lutz et al., 2015) found no correlation between nutrient and
snow algae abundance, whereas several other studies linked nutrients to algae abundance. In
Alaska, for example, a field addition of nitrogen-phosphorous-potassium fertilizer promoted
algae growth (Ganey et al., 2017). On Harding Glacier, Alaska, researchers observed more
prominent algae blooms on the continental side than on the coastal side of the glacier. They
speculated that the input of terrestrial nutrients from aerosols is higher on the continental side
than on the coastal side, promoting the algae bloom (Takeuchi et al., 2006). On the Greenland Ice
Sheet, McCutcheon et al. (2021) found that phosphorous from the local mineral dust increased
algal blooms, and Holland et al. (2019) observed a decrease in nutrient concentration in the
glacier meltwater surrounding areas of algal blooms.
In this study, nutrient levels of aluminium, barium, calcium, chromium, iron, potassium,
magnesium, manganese, nickel and phosphorous were significantly higher in the treated plots
with wood ash than in the control plots (Appendix B: Table B.1). The wood ash contains 0.05%
of total nitrogen and 5.7% of total carbon. The forward stepwise feature selection indicates that
Chromium, Iron and Lead have the most significant impact on snow algae abundance. Overall,
the OLS model is statistically significant (F statistic = 7.138, ρ = 0.0076), and explains about
58.6% of the variation, as per the adjusted R-squared, indicating a moderately strong model. All
three nutrients have a statistically significant influence on algae abundance. Iron (coefficient =
16 170, ρ = 0.001) has a positive impact on algae abundance and chromium (coefficient = 1631000, ρ = 0.033) and lead (coefficient=-1199000, ρ = 0.048) have a negative impact on algae
abundance. This suggests that red snow algae are likely inhibited by heavy metal contamination,
such as chromium and lead, and iron may be a limiting nutrient that supports algae growth. This

35

supports the relationship observed in Phillips-Lander et al. (2020) which observed that Fe-rich
minerals increased red snow algae growth.
2.3.4 Reflectance and Radiative Forcing
Spectral reflectance
Spectral reflectance from the different plot settings varied consistently according to the
plot composition, but the ash treatment did not impact the spectral reflectance (Fig 2-5. A-D).
The reflectance of the uncolonized plots (‘clean snow’ and ‘clean snow + ash’) is spectrally flat
in the visible wavelengths with values above 75% reflectance, as expected for white snow and
decreases in the near-infrared typical for snow spectra (Wiscombe & Warren, 1980). The
colonized plots (‘red snow’) reflectance varies between 40% and 80% in the visible wavelength,
in accordance with other observations of red snow algae reflectance (e.g. Khan et al., 2021;
Painter et al., 2001). The red snow spectra depict the uniquely biological chlorophyll absorption
feature located around 680 nm used by Painter et al. (2001) to map snow algae from
hyperspectral images, and the concave carotenoid absorption feature from 400 to 600 nm used in
the RGND to map snow algae extent from satellite multispectral sensors (e.g. Ganey et al., 2017;
Takeuchi et al., 2006).

36

Figure 2-5: Spectral reflectance measured on 3 and 4 June coloured according to the different plot
settings A. Uncolonized and untreated [Clean snow], B. Colonized and untreated [Red snow], C.
Uncolonized and treated [ Clean snow with wood ash], and D. colonized and treated [Red snow with
wood ash]. The green lines point out to the chlorophyll absorption feature (680 nm), and the red
rectangles point out to the carotenoid absorption feature (400-600 nm)

To assess the impact of ash and red snow algae on snow reflectance value, I computed the
statistical difference with the ANOVA and post hoc Dunn’s test among the radiative forcing of
the respective groups. The ANOVA test and Dunn’s test conducted on the radiative forcing
values for the clean snow, and treated snow (ash) reveal no significant difference between the
clean snow and the snow with ash (ρ = 1.000). The colonized and untreated snow radiative
forcing (red snow), however, was significantly different than the clean snow (ρ = 0.000094) and
the treated snow (clean snow + ash), ρ = 0.004684), with the red snow having a mean radiative
37

forcing statistically higher. Despite adding wood ash in amounts about a hundred-fold of what is
observed in snowpacks in North America following forest fires, I did not detect an increase in
radiative forcing pre- and post-treatment. The lack of difference can be attributed to black carbon
lowering snow albedo by approximately 2-6% only (Wang et al., 2020). Additionally, the actual
proportion of black carbon particulate matter contained in the wood ash from the bioenergy plant
is likely less than 50% of the total wood ash added to the plot, as characterization of wood ash
composition from small-scale residential wood combustion of mixed hard- and softwood
determined that black carbon comprised approximately 40% of the total particulate emission
(Meyer, 2012).
Radiative Forcing
I calculated instantaneous radiative forcing (IRF) using the field-measured reflectance
values collected under blue sky conditions on 3 and 4 June 2023 and modelled clean snow
albedo from SNICAR (Whicker et al., 2022). The average IRF for the ‘clean snow’ is 38 ± 2 W
m-2, 192 ± 12 W m-2 for the red snow algae and 54 ± 4 W m-2 for the snow with wood ash (Fig.
2.6). The maximum IRF is respectively 88.6 ± 5.4, 274.2 ± 16.7 and 111.8 ± 7.44 W m-2 for the
clean snow, red snow and snow with wood ash. The average estimated melt from the red snow
algae is 27.8 mm w.e d-1 as calculated for June 3rd, 2023. This melt assumed clear skies
conditions for the entire day and is calculated for the area measured directly under the field
spectrometer.

38

Figure 2-6: Instantaneous radiate,ve forcing or wood ash, clean snow and red snow as measured on June
4 and July 11 from the field spectrometer reflectance data.

In previous studies of snow algae, calculated radiative forcing varied from an average
IRF of 21.6 W m-2 in Alaska (Ganey et al., 2017) and 88.0 W m-2 in Antarctica (Khan et al.,
2021) to a peak bloom IRF of 235.56 W m-2 in the North Cascades, US (Healy & Khan, 2023)
and 295 W m-2 in the southern Coast Mountains, BC (Engstrom et al., 2022). The average
calculated IRF of snow algae in this study is comparable to the IRF calculated from peak algae
blooms, even though the spectra used in this study were obtained during the early bloom. The
IRF translate to a melt estimate of up to 27.8 mm w.e. d-1. While this value represents the
maximum potential red snow algae can melt under clear skies conditions, it shows that the snowdarkening effect from snow algae bloom, even during early blooms, can have great implications
on snowmelt. As explained above, no difference in melt was measured between the colonized
and non-colonized plots, which is in part attributed to the method. Nonetheless, those results

39

demonstrate that the potential melt contribution from snow algae on glaciers is non-negligible
and should be considered when modelling glacier melt.
The wood IRF was on average 54 ± 4 W m-2 compared to 38 ± 2 W m-2 for the clean
snow, but was not statistically different according to the ANOVA and Dunn’s test. Field
observations of black carbon in snow in the United States suggest radiative forcing can vary
from 1 W m-2 for low black carbon concentration to above 100 W m-2 (Kang et al., 2020).
Measured black carbon from surface snow samples in the Olympic Peninsula, US, post forest
fire, totalled 3120 µg/L, corresponding to a radiative forcing of 37 – 53 W m-2 (S. Kaspari et al.,
2015). On the Juneau Icefield, they estimated a median radiative forcing of 40 W m-2 in July
from black carbon only. Those values are comparable to the calculated IRF values of 54 ± 4 W
m-2 for the wood ash. However, compared to the calculated IRF for the ‘clean snow’ plot, the
increased IRF from the wood ash is lower than those observed in the previous study.
Additionally, a visual inspection of airborne imaging spectroscopy of Place Glacier from before
and after the wood ash was added failed to identify an increase in radiative forcing when
integrated over a 2m pixel size (Chris Donahue, personal conversation). Such a small response
was likely too weak to induce an increased melt or darkening effect and may explain why no
significant difference was observed in the measured snowmelt.
2.4 Conclusion
The field experiment is inconclusive regarding wood ash treatment effects on snow algae growth.
The nutrient analysis suggests, however, that iron may increase red snow algae growth while
heavy metals such as chromium and lead may inhibit their growth. Iron and chromium were both
found to be significantly more elevated on snow treated with wood ash. Due to limited data from

40

the last field visit, further studies are needed to assess the relationship between snow algae and
forest fire ash. As forest fires in Western North America are predicted to increase in frequency
and intensity, it is imperative to understand the feedback between snow algae and black carbon.
Here, I showed that snow algae have the potential to increase radiative forcing to up to 274 ± 16
W m-2 and 192 ± 12 W m-2 on average during the early bloom stage, translating to a melt
potential of 27.8 mm w.e. d-1. Those results support the need to include bio-albedo reduction in
predictive models of glacier change. In addition, this study suggests that higher-than-normal
spring temperatures can shift snow algae bloom timing, which can have an exacerbating impact
on seasonal snow cover on glaciated terrain and lead to early exposure of bare ice.

41

3

Chapter 3: Satellite and airborne remote sensing of snow algae in relation
to black carbon deposition from forest fires in Western Canada

3.1

Introduction
Glaciers represent important sources of freshwater in western North America (Bonsal et

al., 2020b). Yet, rates of glacier mass loss accelerated in the last decades (Bevington &
Menounos, 2022; Hugonnet et al., 2021), leading to important impacts on stream flow and
aquatic ecosystems due to changes in runoff (Huss & Hock, 2018). Seasonal and long-term
decreases in albedo, or glacier darkening, represent one factor that could explain the accelerated
mass loss from Western North America’s glaciers in the last two decades, as surface albedo is an
important factor controlling energy balance processes and melting at the surface of glaciers (Box
et al., 2012). The formation of snow algae blooms and black carbon deposition from wildfire
activity during the snowmelt season represent two processes that diminish glacier albedo.
However, the interaction between snow algae and black carbon from forest fires remains
uncertain.
Systematical mapping of snow algae blooms in North America from 2019 to 2022 shows
that blooms occur on 5% of the total glaciated area (Engstrom & Quarmby, 2023) and can cover
up to 60% of a glacier surface. Blooms of red snow algae appear on glaciers at the onset of
snowmelt and remain on the surface until the end of the melt season (e.g. Engstrom et al., 2022).
The algae produce carotenoid pigments to cope with the harsh glacial environment (LaybournParry et al., 2012). The photoprotective carotenoid pigment, primarily astaxanthin, colours the
snow red, reducing the albedo by up to 20% (Chevrollier et al., 2022; Ganey et al., 2017; Lutz et
al., 2014). The excess energy absorbed by the pigment increases snowmelt on glaciers, which

42

can lead to earlier exposure of bare ice, thus further decreasing glacier albedo as ice is darker
than snow. Previous studies used satellite imagery to study snow algae on snow and ice. Satellite
imagery from Sentinel-2, SPOT and Landsat OLI combined with band ratios, as explained in
section 2.1, have been widely used to map and quantify red snow algae bloom over glaciated
terrain (e.g. Takeuchi et al., 2006. Engstrom et al., 2022; Ganey et al., 2017; Khan et al., 2021).
Others have used Uncrewed Aerial Vehicles (UAV) surveys to analyze the impact of red snow
algae blooms (Healy & Khan, 2023) on snow radiative properties. These studies provide
evidence of the powerful potential of remote sensing techniques to study the distribution and
impacts of red snow algae blooms on glaciated terrains.
Black carbon deposition on glacier surfaces from forest fires decreases glacier albedo and
increases the absorption of shortwave radiation, accelerating melt. The excess energy absorbed
by black carbon further decreases the albedo due to increased snow grain size (Hadley &
Kirchstetter, 2012c). In the last two decades, forest fires in Western Canada increased in size,
severity, duration and frequency (Hanes et al., 2018), leading to the deposition of black carbon
over many glaciated areas downwind of the forest fires. In regions adjacent to frequent forest fire
activities in British Columbia and Alberta, decadal decreasing trends in the glacier surface albedo
were linked to changes in aerosol optical depth attributed to forest fire aerosols (Williamson &
Menounos, 2021). Likewise, the albedo at Haig Glaciers, located in the Canadian Rocky
Mountains, decreased to around 0.12 during years of heightened regional wildfire activity, like
those observed in 2003 and 2017 (Marshall & Miller, 2020). The analysis of an ice core extracted
from the South Cascade Glacier in Washington State, US, also indicated a high deposition of
black carbon during wildfire events (S. D. Kaspari et al., 2020). As the severity and frequency of
wildfires are expected to increase in western Canada (Coogan et al., 2019), the deposition of
43

black carbon over glaciers will also increase. While the distribution of light-absorbing particles
like mineral dust on glaciers has been studied from space, identifying black carbon deposition
over glaciers from multispectral satellites was deemed unlikely (Warren, 2013). The average
amount of black carbon found in remote areas of the Northern Hemisphere reduce the visible
albedo by approximately 1 to 6%, depending on the snow type (Warren & Wiscombe, 1980).
Satellite measurements typically have errors of a few percent making the reduced albedo signal
from black carbon challenging to detect from satellites (Warren, 2013). Moreover, black carbon
has the same spectral signature as thin snow, making it difficult to detect in patchy areas.
Therefore, atmospheric aerosol products such as the MERRA-2 black carbon flux provide a
useful alternative method to track black carbon over glaciated areas, assuming part of the
suspended black carbon gets deposited.
Recent studies combined available data from satellite imagery and data-driven models
such as machine learning algorithms to automate the mapping of light-absorbing particles and
snow and ice algae. Cook et al. (2020), for example, used a random forest classifier trained on
spectra collected in the field to classify ice surfaces and map the albedo from UAV and Sentinel2 images over the Greenland Ice sheet. Similarly, Engstrom & Quarmby (2024) developed a
random forest classifier to systematically map red snow algae using Sentinel-2 satellite images
over Western North America. Remote sensing observations and machine learning algorithms
helped to improve our understanding of the distribution of mineral dust and algae on snow and
ice and their impact on radiative forcing and albedo. However, despite the increasing interest in
surface darkening processes on glaciers, to date, I am aware of no studies that examined the
interaction between black carbon and snow algae using remote sensing data and machine
learning methods.
44

The objective of this paper is to investigate the interaction between snow algae and black
carbon from forest fires using remote sensing and machine learning. Since black carbon is not
only emitted from forest fires, an assumption is being made in this chapter that it dominates the
aerosol load, given the frequency and intensity of fires in the studied area. Here, I model the
occurrence of snow algae blooms on glaciers in the Canadian Cordillera below 60ºN using
satellite imagery from the Harmonized Landsat Sentinel 2.0 from 2015 to 2023 in conjunction
with black carbon flux and dust flux from MERRA-2, and meteorological data from ERA-5
Land. Modelling red snow algae using machine learning helped to understand how the different
variables affected the blooms and provided insight into interactive surface darkening processes.
3.2

Study area
The study area focused on glaciers in the Canadian Cordillera south of 60ºN and north of

49ºN, comprising all of Alberta and British Columbia glaciers, excluding the Insular mountains
from Vancouver Island (Fig. 3-1). This mountainous region covers roughly 1,000,000 km2 and
can be subdivided into three main systems: the Western, Interior, and Eastern (Bolch et al.,
2010). The Western System comprises the Insular, Coast and St. Elias Mountains. The Interior
systems encompass the Purcell, Selkirk, Cariboo and Monashee in the south of the region and the
Hazelton, Skeena, Cassiar and Omineca mountains in the north. The Canadian Rocky Mountains
define the Eastern System. For this study, I excluded the Insular mountains due to the limited
number and size of glaciated areas and subdivided the other mountain ranges into nine regions,
as suggested by Schiefer et al. (2007). Each of these regions defined in figure 3-1 has a similar
climate and glacier characteristics.

45

Figure 3-1: The glaciated region of the Canadian Cordillera below 60ºN is divided into nine regions. The
small glaciers of the Insular Mountains on Vancouver Island are omitted in this study. Details about the
regions are found in Table 3.1.

46

Table 3-1: Glaciated area, and elevation of the nine regions

Region
Number
Region Name
1 Southern Coast
Mountains
2 Central Coast
Mountains
3 Northern Coast
Mountains
4 St. Elias
5 Interior North
6 Interior South
7 Northern
Rockies
8 Central Rockies
9 Southern
Rockies

3.3

Glaciated
Area
(km2)

Mean
Elevation
(m)

Max
Elevation
(m)

Min
Elevation
(m)

6535.35

2065

3444

521

1698.27

1620

2550

411

8619.02
3199.29
535.58
1929.01

1650
1373
1884
2468

2940
4297
2411
3239

200
136
1167
1353

352.19
422.32

2249
2362

2819
3279

1604
1532

1350.75

2629

3438

1630

Data

3.3.1 Harmonized Landsat Sentinel
For this study, I use surface reflectance data from NASA’s Harmonized Landsat Sentinel
(HLS) version 2.0. The HLS dataset combines and harmonizes Landsat8/9 Operational Land
Imager (OLI) and Sentinel-2A/B Multi-Spectral Instrument (MSI) through a set of algorithms,
including atmospheric correction, cloud and cloud shadow masking, geographic co-registration
and common gridding, bidirectional reflectance distribution function (BRDF) normalization, and
bandpass adjustment (Claverie et al., 2018). The HLS produces a seamless surface reflectance
record with near-global coverage and revisits intervals every three days at the equator, increasing
in frequency with increasing latitude. HLS data consists of two products, S30 (30m spatial
resolution) derived from Sentinel-2 imagery and L30 (30m spatial resolution) derived from
47

Landsat 8-9/OLI, acquired since April 2013 (Landsat-8), November 2015 (Sentinal-2A), July
2017 (Sentinel-2B) and October 2021 (Landsat-9). In this study, I utilized L30 and S30 products
from June 1st to September 30th from 2015 to 2023 inclusive.
3.3.2 Aerosol Optical Depth
The black carbon and mineral dust used in this chapter originates from NASA’s Global
Modeling and Assimilation Office (GMAO) Modern-Era Retrospective analysis for Research
and Applications, Version 2 – MERRA-2 – reanalysis data (Gelaro et al., 2017). MERRA-2 is a
reanalysis dataset that assimilates aerosol optical depth (AOD) from ground- and space-borne
remote sensing platforms and meteorological observations from radiosondes and aircraft
(Randles et al., 2017). The model is based on the Goddard Earth Observing System Model
Version 5 (GEOS-5) and includes bias-corrected AOD datasets of Moderate Resolution Imaging
Spectrometer (MODIS), Advanced Very High-Resolution Radiometer (AVHRR) instrument,
AOD from Multi-Imaging Spectroradiometer (MSR) and direct measurements of the groundbased Aerosol Robotic Network (AERONET).
MERRA-2 models the spatial distribution of black carbon and mineral dust concentration
since 1980 at a spatial resolution of 0.5º×0.625º and time scales of 1-hourly, 3-hourly and
monthly. The data is available online through NASA’s Goddard Earth Sciences (GES) Data and
Information Services Center (DISC). I acquired all hourly black carbon and mineral dust grid
cells that covered the glaciated regions from 2015 to 2023. Based on the hourly dataset, I
obtained monthly mean and max from June 1st to September 30th of each year.

48

3.3.3 ERA5-Land – Climatic parameters
Weather and long-term monitoring stations are sporadic in mountainous landscapes and at
high elevations where glaciers exist in Western Canada, preventing the use of climatic variables
collected at discrete locations. Instead, I used the global climate reanalysis dataset of ERA5 Land
(Muñoz-Sabater et al., 2021) to obtain climatic parameters over the study area from 2015 to
2023. I used the 2 m air temperature, snow density, snowfall, snowmelt, temperature of snow
layer, and surface net solar radiation from ERA-5 Land hourly data produced at 9 km resolution
(https://cds.climate.copernicus.eu/cdsapp#!/dataset/reanalysis-era5-land?tab=overview)
resampled to a monthly average and monthly maximum. Previous studies suggest that climatic
parameters and snow’s physical conditions, such as snowmelt, snowfall and surface solar
radiation and liquid water, impact snow algae abundance (Ganey et al., 2017; Onuma et al.,
2016, 2022). Snow algae first appear on the snow surface when temperatures are above freezing
(≥ 0 ºC) (Jones et al., 2001). The ERA5 Lands grids that covered the tile areas as per the HLS
images from my glaciated region were aggregated into averages for each tile of the nine glacier
regions.
3.4

Data Acquisition and Methods
In this study, I developed a framework to assess the interaction between red snow algae

and black carbon. This framework consists of developing a dataset of red snow algae occurrence
over the study area, referred to as the algae ratio, along with the predictor variables. Once the
dataset was established, I assessed and preprocessed the data (section 3.3.4), then hypertuned and
trained the machine learning models consistent with top machine learning practices (section
3.3.6). Next, I evaluated the model outputs and finally identified the key variables controlling the
prediction of red snow algae (section 3.3.7). I explained the steps involved in acquiring and
49

preparing the red snow algae occurrence dataset in sections 3.3.1 to 3.3.3. The flow chart (Fig. 32) illustrates the process utilized to generate the red snow algae ratio dataset.

Figure 3-2: Flow chart of the workflow used to acquire the algae ratio from the HLS image collection.
The section mentioned in the workflow corresponds to the next three sections.

3.4.1 Image acquisition
I used the HLS surface reflectance data to create eight melt-season (June- September)
time series of algae blooms over the glaciated regions of British Columbia and Alberta from
2015 to 2023. I obtained all available HLS images filtered at 70% cloud coverage over the
glaciated regions from June 1st to September 30th, which I assume to be the snow algae
blooming season based on observations of blooms in Western North America (Engstrom &
Quarmby, 2024). I used a high cloud cover filter in this first step to include as many images as
possible that covered the glaciated regions. I processed entire HLS scenes, which can contain
50

less than 10% coverage of glaciers, and therefore, keeping a high cloud filter allowed me to
include more images that may contain partial cloud coverage over glaciated areas, which
otherwise would not have been included. From the HLS images collection, I loaded S30 B2, B3,
B4 and B11 bands and L30 B2, B3, B4 and B5 bands in memory corresponding to the blue,
green, red and SWIR bands. I used the glacier polygons from the BC Glacier Inventory 2021
(available here: https://catalogue.data.gov.bc.ca/dataset/glaciers) and Randolph Glacier Inventory
7.0 (RGI Consortium, 2023) Western Canada inventory (available here:
https://nsidc.org/data/nsidc-0770/versions/7) to constrain the workflow to glacier polygons of
British Columbia and Alberta only. The HLS bands were clipped to the glacier polygons,
excluding all the data outside of the polygons. I then scaled the bands to a 0.0001 factor to bring
the band values to a 0 to 1 scale as the HLS band data is provided in a 0 to 10000 scale.
3.4.2 Mask and threshold
Once the bands were loaded in memory, I applied a series of masks and thresholds to
clean and prepare the data for the algae time series. I masked pixels with B3 < 0.5 to exclude
bare ice. I used the green band to differentiate snow from ice rather than the Normalized
Difference Snow Index (NDSI), as the NDSI overestimated snow coverage by including bare-ice
pixels (A. Bevington, personal conversation). Since the images were already clipped to the
glacier polygon, the need to apply an index to eliminate non-snow and ice surfaces was not
necessary. I then applied the Green Blue Normalized Difference index (GBND) > 0.05 to mask
𝐵3−𝐵2

rock fall and mineral debris on snow. The GBND (𝐵3+𝐵2) is a newly developed index by Di
Mauro et al., (2024). The index is useful to differentiate mineral dust from algae blooms, as at
low cell concentrations, the RGND index used to map snow algae can confound mineral dust to
algal bloom (T. Painter et al., 2001).
51

3.4.3 Snow Algae Index
Finally, using the clean glacier image, I calculated the snow algae index for each pixel. I
used the Red Green Normalized Difference (RGND), as explained in section 1.4, as it considers
varying light intensity between the different images rather than the ratio between the red and
green bands, as proposed by Takeuchi et al. (2006). RGND values above 0.02 were assigned as
algal snow, and values below as algal-free snow. I chose to set a threshold of 0.02 as per my
observations from hyperspectral data for Place Glacier. The snow samples with RGND values
below 0.02 either had a very low concentration of cells or no cells. From the RGND pixel values,
I quantify the occurrence of the algae blooms using the ratio of algal snow pixels over the total
cloud-free pixels for a given glacier. Before calculating the ratio, I performed a second cloud
mask utilizing the SWIR band, as clouds, compared to snow and ice surfaces, have high
reflectance values in the SWIR band (A. Bevington, personal conversation). I applied a threshold
of 0.05 on the SWIR band to mask clouds. I kept images with cloud coverage below 20% only
for the analysis. The pixel ratio was determined for each tile based on the naming convention
from the HLS image collection.
3.4.4 Data aggregation
The data used in this study comes from three spatially gridded products, the HLS, MERRA-2
and ERA5-Land, which all have different resolutions. To account for the difference in grid size, I
spatially aggregated the data based on the HLS tile naming convention, where one tile is
approximate 109 Km by 109 Km. Each glacier region contains from 6 (St.-Elias) to 18 tiles
(Southern Coast Mountains), and on average, 9 tiles depending on the size of the region
(Appendix C, Table C1). I averaged the ERA5-Land and MERRA-2 values that covered each tile
when multiple grids covered the area of the HLS tile. Due to the MERRA-2 large grid size,
52

multiple HLS tiles were assigned to be the same value. Likewise, an HLS tile may be present in
multiple regions, depending on where the boundary of the region versus the tile fell. In that case,
the algae index contains a different value, but the ERA-5 Land and MERRA-2 values are the
same. To account for the difference in spatial resolution, I extracted monthly values. I created
two datasets for each HLS tile: the monthly mean and the monthly maximum values.
3.4.5 Trend analysis
I used ‘linregress’ from the Scipy stats library in Python to calculate the linear
relationship between the algae ratio (dependent variable Y) and black carbon flux (independent
variable X) for each mountain region. I calculated the Pearson correlation coefficients to assess
the variation in the linear relationship amongst the different regions of the study area. I
performed the linear regression calculation on the monthly max and monthly mean values. A
correlation score of > 0.5 indicates a moderately positive linear relationship, and < 0.3 indicates
a weak relationship. Furthermore, I obtained the ρ-value using a <0.05 confidence level for each
correlation to determine the significance of the trend.
Additionally, I obtained the correlation matrices of the monthly mean and max of the nine
predictors and one target variable to be used in the models. The correlation matrices provided
insight into the data structure and collinearity of the predictor variables and the algae ratio. I
utilized correlation matrices to identify multicollinearity within the predictor variables and guide
which of the nine predictors provided redundancy in the model.
3.4.6 Data Preprocessing
Before performing the models, I preprocessed the data to address inconsistency and
missing values in the datasets. The first preprocessing step involved cleaning the missing data

53

(NaNs) associated with the algae ratio. The data conversion to obtain the algae index from the
HLS bands created NaNs values when the band values were infinite, or both the red and green
band values were equal to zero. Additionally, the algae ratio calculation introduced NaNs if the
cloud coverage over the glacier was above 20% (step 3.4.3). In that case the values were dropped
due to the heterogenous distribution of algae bloom, making it difficult to use interpolation to
infill values where clouds masked most of the glacier. The regions near the coast contained the
most NaN values due to increased cloud coverage. Next, upon assessing the variable’s
distribution, I applied the log-scaling method to normalize variables with a skewed distribution.
The algae ratio, the black carbon flux and snowmelt were transformed to the logarithmic scale.
Finally, I conducted data normalization for the dependent and independent variables. Data
normalization is a technique that transforms values on a similar scale, improving the model’s
performance and training stability. I used the min-max normalization method on all of the
independent variables as follows:

𝑥′ =

𝑥−min (𝑥)
max(𝑥)−min (𝑥)

(2.1)

This method is preferred over the z-score for data that does not follow a Gaussian or normal
distribution. The data was then ready for the machine learning models.
3.4.7 Models
To assess how red snow occurrence responds to every variable, I applied a Random
Forest and XGBoost model.The predictor variables are black carbon flux, dust flux, longitude,
latitude, surface air temperature, surface net solar radiation, snowmelt, temperature of the snow
layer and snowfall. I then used permutation feature importance and Shapley Additive
exPlanations (SHAP) to explain the output of the models and assess the relationship between
54

each variable and red snow algae. I applied the model to the entire datasets, years separately and
using the previous year data for the black carbon to assess lag effect.
Random Forest Regressor
A Random Forest Regressor is an ensemble-based algorithm that combines the
performance of multiple decision trees to predict the outcome value (Breiman, 2001). Random
Forest Regressor is a non-parametric model, meaning it can deal with data that do not follow a
normal distribution and can detect higher-order relationships between the predictors and the
target data. The algorithm splits the data into multiple subsets of training data using the bootstrap
aggregating or bagging technique, in which multiple subsets of the original training data are
created. Decision trees have three main parts: root, internal and leaf nodes. The algorithm starts
at the root node (the training dataset), followed by the internal node, where decision-making
occurs. The leaf node is at the end and holds the final decision. At each node, a subset of the
dataset is drawn at random. When the decision tree is finalized, a simple regression is performed
on each leaf node to obtain the predicted dependent value. The Random Forest algorithm uses
the bagging approach to aggregate multiple decision trees, in which each tree runs in parallel,
meaning there are no interactions between individual trees. I utilized the random forest regressor
from the Scikit-Learn Python library (Pedregosa et al., 2011) to determine the influence of the
environmental and climatic variables on the monthly ratio of red snow algae from 2015 to 2023.
I tuned the hyperparameters with the GridSearchCV from scikit-learn (Table 3.2). GridSearchCV
uses all specified hyperparameter combinations to calculate and maximize the model
performance.

55

Table 3-2: Parameters description and best value used for the Random Forest Regressor models

Parameter
n_estimators
max_depth

Description
The number of decision trees in the forest
The maximum depth of the tree
The minimum number of samples needed to split an
min_samples_split internal node.
The minimum number of samples needed to create
min_samples_leaf new leaves.
The number of features to consider when seeking the
max_feature
best split.
Define if bootstrap samples are used when creating
bootstrap
the decision trees

Monthly Mean
100
20

Monthly Max

2

3

1

1

sqrt'

sqrt'

FALSE

FALSE

2500
50

Extreme Gradient Boosting
Extreme Gradient Boosting (XGBoost) is a powerful machine learning algorithm based
on decision trees introduced by Chen & Guestrin (2016). It is similar to the Random Forest
model as it is based on decision trees, although the XGBoost algorithm learns by boosting,
whereas the Random Forest learns by bagging. Boosting is an ensemble learning technique that
combines multiple weak learners in sequence, where each subsequent model tries to correct the
errors of its predecessor to improve the overall prediction accuracy. XGBoost utilizes the
gradient boosting ensemble technique (Friedman, 2001) in which each new model reduces the
residual error by optimizing a loss function using gradient descent, learning incrementally from
previous iterations. This technique results in a model with low variance and low bias. The
gradient boosting algorithm generated decision trees sequentially, learning from the last model
until no more improvements occur or a threshold is reached. The algorithm utilizes a learning
rate to scale the contribution of each tree. The learning rate leads to each tree having a more
minor impact, increasing the total number of trees to reach the same level of accuracy. Using
empirical evidence, Friedman (2002) showed that using multiple trees that point in the right
56

direction yields better prediction and lower variance than using a few decision trees with low
bias and high variance. XGBoost builds on the gradient boosting model, improving its
performance and speed. To improve computational efficiency, XGBoost enables parallel
processing. The algorithm utilizes regularization techniques to improve the overall performance,
constraining the model’s complexity by penalizing overfitted models. Common regularization
techniques include the Least Absolute Shrinkage and Selection Operator (Lasso/L1) and the
Ridge (L2).
XGboost also uses a more sophisticated pruning strategy based on a maximum depth of
tree and minimum loss reduction (gamma). The minimum loss reduction calculates the gain at
each split as follows:
1

𝐺𝑎𝑖𝑛 = [

𝐺𝐿2

2 𝐻𝐿 + 𝜆

+

2
𝐺𝑅

𝐻𝑅 + 𝜆

−

(𝐺𝐿+ 𝐺𝑅 )2
𝐻𝐿 +𝐻𝑅 + 𝜆

]− 𝛾

(2.2)

Where G and H are the gradients and second-order gradients, L and R define the left and right
child nodes, λ is the regularization term, and 𝛾 is the minimum loss reduction. If the gain
(reduction in the loss) is less than gamma, then the split is pruned (discarded). Using this
method, XGboost prunes the trees while in the building process rather than after the trees are
fully grown. The final prediction from the algorithm is a weighted sum of all the weak predictive
models.
I implemented XGBoost using the xgboost library in Python (Chen & Guestrin, 2016)
and used GridSearchVC from Skikit-learn to tune the parameters described above. The
parameters used in this study are outlined in Table 3.3. Once the model was tuned and trained, I
evaluated the feature importance and SHAP values (Lundberg & Lee, 2017), as explained in

57

section 3.4.3, to gain insight into which input variables yield the best estimate of the predicted
values.

Table 3-3: Parameters description and best value used for the XGBoost models

Parameter
max_depth

Description
The maximum depth of the tree

Monthly
Mean
9

Monthly Max

1

5

0

0

8

0.6
0.8

0.6
0.8

reg_alpha

The minimum weight needed to create a new node in the
decision tree. A large min_child_weigth leads to a more
conservation algorithm.
The minimum loss reduction required for a node to make
a split.
Specifies the ratio of observation to subsample for each
iteration.
The ratio of variables to be used.
L1 regularization term on weights. Can reduce the
number of features to create a more interpretable model.

0.01

0.1

learning_rate
n_estimators

Specifies the step size used to scale the contribution of
each new tree. A small learning rate reduces the impact
of each tree.
The number of decision trees in the forest

0.01
1000

0.01
1000

min_child_weight
gamma
subsample
colsample_bytree

I ran the Random Forest and XGBoost models on the monthly mean and monthly max
datasets. To assess whether years of high forest fire activity influenced the models’ performance,
I ran the models on the monthly mean datasets for each year. Due to a lack of data from the
monthly mean dataset in 2016, I ran 2016 with the monthly max dataset. Lastly, I applied the
model to the monthly mean dataset, using the black carbon flux from the previous year, such as
the 2016 dataset used the 2015 black carbon flux, to evaluate the lag-year effect from the black
carbon. All the models’ variable importance was assessed with the permutation feature
importance and the Shapley Additive Explanation as described below.

58

3.4.8 Evaluating Variables Importance
Once the models were tuned and trained, I evaluated the feature importance of each
variable using the permutation feature importance and SHapley Additive explanation (SHAP)
values (Lundberg & Lee, 2017) to gain insight into which input variables yield the best estimate
of the predicted values.
Permutation Feature Importance
The permutation feature importance uses a model inspection technique to measure how
much each feature contributes to the model’s statistical score. The model inspection is done by
randomly shuffling the values of a feature and measuring how much it decreases the resulting
model’s score. The advantage of the permutation feature importance technique is that it is modelagnostic and is particularly helpful when using non-linear predictors (Breiman, 2001). The one
downside of the permutation technique is when two features are strongly correlated. If one of the
correlated features is permuted, the model can still access the latter from its correlated feature,
resulting in a lower reported importance value for the two features than it might be. One way to
counter this issue is to cluster highly correlated features and retain only one feature from the
cluster (Scikit Learn, 2024). The permutation feature importance provides insight into which
features have the most impact on the model’s score.
Shapeley Additive Explanation
The SHAP (Shapley Additive exPlanation) is a method introduced by Lundberg and Lee
(2017) to interpret predictions from complex models, such as ensemble models like the one used
in this study. SHAP values are based on cooperative game theory in that they measure a player's
individual contribution (or variable) to the model's output for each set of observations. The
59

SHAP values are computed by evaluating the impact on the model’s prediction when including
or excluding each feature across all possible variable combinations. The larger the SHAP value,
the higher the contribution of a variable to the model’s outcome. The SHAP technique indicates
whether a predictor has a positive, negative or neutral relationship with the target value, as
visualized through the SHAP beeswarm plots.
3.5

Results

3.5.1 Correlation
Region Correlation
The correlation between black carbon and red snow algae occurrence is generally weak to
moderate, with interior and southern regions exhibiting stronger relationships than coastal and
northern regions. Overall, the correlation coefficient between black carbon and snow algae
indicates weak [0.1 to 0.3] to moderate [0.3 to 0.5] positive linear relationships, except for the St.
Elias subregion, which indicates a weak negative linear relationship (Fig. 3-3). The Central
Rockies subregion has the strongest relationship with R scores of 0.55 and 0.47 for the monthly
mean and monthly max values, followed by the Interior South with R scores of 0.37 and 0.43,
respectively. The correlation coefficient is generally higher for the interior regions than the
coastal regions, and stronger for the southern regions than the northern regions. The Northern
Coast Mountains, Northern Interior and St. Elias subregion depict the weakest relationship with
R-score varying from -0.12 to 0.17.

60

B

A

Figure 3-3: Pearson correlation coefficient between black carbon flux and red snow algae ratio for the nine subregions A.
Monthly mean correlation and B. Monthly max correlation

Variables Covariance
The correlation between red snow algae ratio and environmental variables varies in
strength and direction, with temperature, black carbon flux, and solar radiation showing
moderate positive relationships. The monthly mean temperature at 2 m (t2m), black carbon flux,
mean longitude, and monthly mean surface solar radiation (ssr) have the highest positive
correlation with the monthly mean algae ratio with R-scores of 0.361, 0.348, 0.336 and 0.305,
respectively, indicating a moderate positive correlation (Fig. 3-4). Both the monthly mean
latitude and dust flux have a weak negative correlation with the monthly mean algae ratio with rscore values of -0.214 and -0.131, respectively. Similarly, the max longitudes and monthly max
temperature at 2 m (t2m), black carbon flux and surface solar radiation (ssr) have the strongest
correlation coefficient with the monthly-max algae ratio, albeit at lower scores than the monthly
61

mean values. The r-scores are respectively 0.281, 0.259, 0.188 and 0.138. Likewise, the max
latitude [-0.172] and monthly max dust flux [-0.207] show a weak negative correlation with the
monthly max algae ratio. I calculate the p-value to assess the significance of the correlation (ρ ≤
0.05). The mean dataset has 2 non-significant correlations (sf – snow algae and rsn – tsn), and
the max dataset has four non-significant correlations (tp – lat, tp – algae ratio, tp, rsn and ssr –
dust flux) (Appendix E, Table E1 & E2). Based on the correlation matrix, I selected which
predictors to be used in the model.

Figure 3-4: Correlation matrices of the A. Monthly mean and B. Monthly max predictors and target
(algae ratio) values.

62

Correlation among the predictor variables ranges from weak to high. Predictors with high
correlation coefficients indicate collinearity and can lead to redundancy in the random forest and
XGboost model despite their robustness against multicollinearity. The surface solar radiation
(ssr) and temperature at 2 m (t2m) have a very high correlation for both the monthly mean
(0.866) and monthly max (0.851), as expected, indicating that areas with high surface solar
radiation tend to have high temperatures. Snow density (rsn) and snowmelt (smlt) show high
correlations with each other for the monthly mean (0.752) and monthly max (0.790), suggesting
that they capture related snow dynamic processes. Dense snow has reduced air content and
increased liquid water content, bringing it closer to the “ripe” state and making it more
susceptible to melting, as it requires less energy for melting to occur. Based on manual iterations
of the model with different combinations of the highly correlated pairs and their resulting
importance, I removed snow density (rsn) from the predictor list as it decreased the models’
performance. Removing surface solar radiation and/or temperature at 2m did not improve the
model's robustness; therefore, both predictors were kept in the final models. The latitudes and
longitudes had a strong negative correlation, likely because they were geographic coordinates,
and a shift in one typically correlated to a specific shift in the other. The final retained variables
for the model based on the correlation matrices were longitude, latitude, black carbon flux, dust
flux, surface air temperature, net surface solar radiation, snowmelt, the temperature of the snow
layer and snowfall.
3.5.2 The relation between Algae Ratio and the Environmental, Climatic and Geographic
Variables
I calculated the feature importance from the Random Forest and XGBoost model using
SHAP feature importance techniques and the permutation feature (Fig. 3-5 & 3-6). The different
63

models reveal similar patterns of feature importance. The most important variables for all model
iterations are the longitude, then black carbon flux, followed by the two-metre air temperature,
and the least important variable is snowfall.
SHAP features importance
The SHAP values range approximately from -0.2 to 0.3, implying that the variables
moderately impact the final predictions. Furthermore, the relatively small SHAP values indicate
that the prediction of the model outcome, the algae ratio, is not dominated by a single variable
but is rather influenced by all variables, with increased contribution from higher-rated variables.
The beeswarm plots (Fig. 3-5) order the variables by their importance, suggesting that
longitudes, black carbon flux and 2-metre air temperatures have the most impact on predicting
the algae ratio, as they are the top three variables in all four model iterations. The x-axis of the
beeswarm plots indicates the impact of a variable on the model output, and the colour map is
associated with the value of the variable, where blue indicates a high value and brown is a low
value. Dust flux, latitude and snowmelt typically follow in the middle of the order, and surface
solar radiation, the temperature of the snow layer and snowfall are generally the least impactful
variables in the models, sitting at the bottom of the list.
When assessing the relationship between the predictors and target variable, the SHAP
values for the Random Forest and XGBoost, using the monthly mean datasets, show that four
predictors depict a clear linear relationship between the predictor values and their impact on the
algae ratio (Fig. 3.5). Those are the longitude, 2 metres air temperature, black carbon flux, and
the snowmelt. The other predictors do not have a clear impact on the model outcome. High air
temperature values and black carbon flux are linked to a positive impact on the algae ratio value

64

and vice versa. Low longitude values indicate an eastward location or a large distance from the
coast and are associated with a high predicted algae ratio. Similarly, low surface melt values
result in a high model output. The model results from the monthly max datasets depict a slightly
different outcome. While the longitudes and 2-metre air temperature still show a relatively clear
linear relationship to the algae ratio, the monthly maximum black carbon flux does not indicate a
clear linear relationship with the predicted algae ratio. The latitude, on the other end, has a
positive linear impact on the algae ratio when using the maximum values.

65

Figure 3-5: Beeswarm plots for the SHAP variable importance for the algae predictors with the Random Forest and XGBoost models.

Permutation feature importance
The permutation feature importance calculates on the model predictors indicate their
contribution to the model outcome (Fig. 3-6). According to the permutation technique, the
longitude is the strongest predictor of red snow algae occurrence. The R2 for the Random Forest
and XGBoost models will decrease by respectively 45.6% and 37.0% for the monthly mean
values and 38.9% and 41.3% for the monthly max values when the longitude is not included in
the model. The second most important predictor is the black carbon flux. The Random Forest
scores will decrease by 38.4% and 35.2% for the monthly mean and monthly max and the
XGBoost, respectively, by 31.1% and 23.6%. The 2-metre air temperature has the third highest
permutation scores for the two models, contributing on average to 23.4% of the model outcome
values. The other predictors, the dust flux, surface solar radiation, the latitudes, the snow and the
temperature of the snow layer, contribute to less than 20% of the model and are not consistently
rated among the Random Forest and XGBoost iterations. The snowfall has the least impact on
the model, with a permutation score of 4.9% and 8.9% for the Random Forest monthly mean and
monthly max and a score of 2.4 and 5.7% for the XGBoost monthly mean and monthly max.

Figure 3-6: Permutation feature importance scores for the Random Forest and XGBoost for the A.
Monthly mean dataset and B. Monthly max dataset

68

3.5.3 Main Model Performances Evaluation
The Random Forest and XGBoost performed moderately well, with R-score values
ranging from 0.546±0.028 to 0.613±0.027 (Fig. 3-7). The Random Forest models have a mean
R-scores of 0.600, RMSE of 0.0453 and MAE of 0.0795 for the monthly mean and an R-scores
of 0.533, RMSE of 0.0861 and MAE of 0.131 for the monthly max. For the XGBoost model, the
average R2 are 0.604 and 0.564, the RMSE are 0.0455 and 0.0827, and the MAE are 0.0793 and
0.127 for the monthly mean and max. The Random Forest regressor and the XGBoost models
performed similarly, and they both performed slightly better when executed on the monthly mean
versus the monthly max values. All models have a slightly higher MAE than RMSE, indicating
that most of the model’s errors are small, with a few large errors.

69

A

B

C

Figure 3-7: Boxplot representing the A. Coefficient of determination (R2) B. Root Mean
Squared Error (RMSE), and C. Mean Absolute Error (MAE) for each of the models using the
mean and max dataset. The scores were calculated for 50 training and testing dataset
iterations.

70

Comparing the predicted value to the test datasets provide information on the fitting of
the models. The plots for the Random Forest and XGBoost for the monthly mean dataset (Fig. 38 A & C) depict a similar trend in that the main cluster of points follows the x = y line
moderately well. The monthly mean models show outliers on both sides of the x = y line, with a
minor trend towards underfitting the actual algae ratio for values above 0.2. The monthly max
plots (Fig. 3-8 B & D) show a slightly worse fit than the monthly mean plot, with larger residual
errors as implied by the larger RMSE. The cluster of points for the monthly max plots trend
lightly steeper than the x = y line, suggesting a light overfitting model. All four models indicate
that the bulk of the monthly algae ratio values are within 0.0 to 0.1 for the monthly mean and 0.0
to 0.2 for the monthly max.

71

Figure 3-8: Plots of actual vs predicted algae ratio value for A. Random Forest Monthly Mean, B.
Random Forest Monthly Max, C. XGBoost Monthly Mean and D. XGBoost Monthly Max. The dash line
is a 1:1 line and the red lines are linear regression model.

72

3.5.4 Annual Models Performance and Feature Importance
The annual models’ R2 score range from 0.043±0.239 to 0.703±0.094 for the random
forest and 0.064±0.160 to 0.706±0.118 for the XGBoost (Fig. 3-9). The random forest performed
slightly better than the XGBoost on the annual datasets. The 2019 model yield the lowest
performance in the models and 2018 the best performance. The 2015 XGBoost model has the
largest variation in model performance, with a standard deviation of 0.63. In general, the annual
model's R2 score is lower than the one obtained for the models run on the entire dataset, and
there are no differences in the model performance between the low fire seasons (2016, 2019,
2020 and 2022) and high fire seasons (2015, 2017, 2018, 2021 and 2023).

Figure 3-9: Random Forest (teal) and XGBoost (brown) R2 Score when applied to the model run
on a singular year. All models were run on the mean dataset except for 2016, which was run on
the max dataset.

73

Regarding the variable importance, the permutation feature value ranges from 0.007 to 0.370 for
the random forest and 0.011 to 0.487 for the XGBoost. The SHAP values range from 0.0009 to
0.0424 and 0.0008 to 0.0389 for the random forest and XGBoost, respectively. When averaged
across the annual models, the dust flux is the most important variable, followed by the longitude
(Fig. 3-10). While the dust flux is ranked first in the XGBoost, it has the largest value range
among the nine predictor variables. Although the black carbon flux and 2-metre air temperature
ranked in the top three variables in the overall model, they have a much lower impact on the
model output when assessed with the annual model. The permutation feature importance value is
less than 0.1 for the surface air temperature and black carbon and less than 0.01 for the SHAP
values. The variable importance amongst the nine annual models is inconsistent, as shown in the
permutation feature and SHAP plots (Appendix F & G, Fig F 1-18 & Fig G 1-18).

74

Figure 3-10: Feature importance for the model run annually. A. Random Forest Permutation Values, B.
XGBoost Permutaiton values, C. Random Forest Shap values and D. XGBoost Shap values.

75

3.5.5 Lag-year model performance and feature importance
The model R2-score performed with the lag-year black carbon flux dataset is 0.606 ±
0.043 for the random forest and 0.603 ± 0.049 for the XGBoost. The RMSE and MAE are
respectively 0.069 ± 0.006 and 0.039 ± 0.003 for the random forest and 0.070 ± 0.006 and
0.049 ± 0.003 for the XGBoost. The random forest XGBoost performed very similarly on the
lag-year dataset. According to the permutation feature importance and SHAP value (Fig. 3-11 &
3-12), longitude and dust flux are the most important variables when using the lag-year dataset.
The SHAP beeswarm plots (Fig. 3-12) indicate that low longitude value, thus eastward location,
and low dust flux value positively influence the red snow algae. The temperature of the snow
layer (tsn) is the third most important variable according to the permutation feature importance
and the SHAP value for the XGBoost. It ranked fourth for the random forest SHAP beeswarm
plot (Fig. 3-12.A.). The next most important variable is the surface air temperature (t2m). The
black carbon flux has less impact on the red snow algae ratio for the model ran on the lag year
dataset. The snowfall and latitude remain among the weakest variables for the lag-year models.

76

bc_flux

A.

bc_flux

B.

Figure 3-11: Permutation feature importance for the lag-year model for A. the random

forest and B. the XGBoost.

77

Figure 3-12: Beeswarm plots of SHAP values performed on the lag-year dataset for the A.

random forest and B. XGBoost model.
78

3.6

Discussion

3.6.1 Variables Influence from the Main Models
The models’ variable importance provides valuable knowledge on the influence of each
variable on the occurrence of red snow algae. The combination and comparison of the two
machine learning models consolidate the results. The section below explains the variables’
influence and implication with previous studies.
Influence of Longitude
The models’ feature importance indicates that the longitude is the most important
predictor explaining the monthly red snow algae ratio, with low longitude values (eastern
location) positively related to the bloom occurrence. In this study area, eastern longitudes
indicate glaciers located further away from the coast, such as the Interior and Canadian Rocky
Mountains. Previous studies observed a similar pattern in the distribution of red snow algae.
Takeuchi et al. (2006) found that red snow algae were more predominant on the inland side of
the Harding Icefield, Alaska than on the coastal side. They hypothesized that this spatial pattern
was due to geographic effect on snow chemistry, such as higher sea salt content inhibiting snow
algae propagation on the coastal side of the range and/or greater nutrient input from terrestrial
aerosols on the continental side than the coastal side promoting algae growth. Similarly, Enstrom
and Quarmy (2024) found that snow algae bloom preferentially grew on the lee side of the
southern Coast Mountains. They associated the high concentration of algae blooms on the
continental side with lower cloud cover and higher summer air temperatures than the coastal side
of the range.

79

The SHAP value indicates that the red snow algae ratio increases along the eastern
longitudes, suggesting that continental glaciers support a greater proportion of red snow algae in
this study. The Interior and Rocky Mountains are the easternmost regions of the study area. Both
regions depict a continental climate, with warm and dry summers and low cloud coverage.
Additionally, because westerlies are the prevailing wind direction in Western Canada, the Interior
and Rocky Mountains may receive greater amounts of terrestrial aerosols due to their inland and
downwind location. The permutation feature importance indicates that the models’ ability to
predict red snow algae occurrence can decrease by up to 40% when longitude is removed as a
predictor.
Influence of black carbon flux
The second most important variable in this study explaining red snow algae occurrence
on glaciers is the black carbon flux, with increased black carbon flux contributing positively to
the red snow algae ratio. Interestingly, the regions located directly downwind from the Interior
Plateau, such as the southern Interior, central Rocky and southern Rocky Mountain, where most
of the large forest fires occurred between 2015 and 2023, as depicted in Fig. 3-13, have the
highest correlation coefficient between the red snow algae and black carbon flux. Due to their
downwind location from frequent large forest fires, those three regions likely received high
concentrations of black carbon during years of high fire activity, as observed by other studies
(e.g. Marshall & Miller, 2020; Williamson & Menounos, 2021). Those three regions are also the
most eastward locations from the study area, suggesting that the black carbon flux and longitude
variables are co-related. In contrast, the eastern location may have the greatest influence on snow
algae occurrence because they are the most impacted by black carbon.

80

It is not currently known if black carbon directly induces red algae growth. The results
from this study indicate a positive relationship between the two. An explanation of the positive
relationship between black carbon flux and the ratio of red snow algae is that the decrease in
albedo from the black carbon increases liquid water content throughout the snowpack, which was
found to be a limiting growth factor for red snow algae (Ganey et al., 2017). An alternative
explanation is that black carbon provides a source of nutrients for the red snow algae, thus
promoting growth.

Figure 3-13: Burned area from 2015 to 2023 from MODIS Burned Area Monthly Global 500m.

81

Black carbon flux had a clear positive relationship with the red snow algae occurrence for
the model performed on the monthly mean dataset only. The monthly max plots indicate that the
lowest black carbon flux tends to have a negative impact on the red snow algae; however, the
relationship for the rest of the black carbon flux is unclear. Likewise, the correlation between
monthly max black carbon flux and red snow algae was generally lower than when calculated on
the monthly mean dataset. This could be explained by the fact that the maximum value may
introduce a bias in the dataset by representing a single event of rapid but intense duration
releasing a high amount of black carbon over a short period into the atmosphere. Such events
may not lead to significant black carbon deposition over glaciated terrain due to its brief
duration, explaining the unclear relationship between large black carbon flux and red snow algae
for the maximum dataset. In comparison, a large monthly mean black carbon flux likely
represents a consistent input of black carbon in the atmosphere over a month, which is more
likely to deposit black carbon on glaciers.
Influence of 2-metre air temperature
According to the Random Forest and XGBoost models, the 2-metre air temperature is an
important variable explaining red snow algae occurrence on glaciers. This finding accords with
previous studies that found red snow algae experience more growth or will expand under warm
conditions (e.g. Onuma et al., 2022; Healy & Khan, 2023). One of the challenges, however, is
the covariance between black carbon and surface air temperature, where years of high
temperatures tend to be years with high forest fire activity. The SHAP beeswarm plots (Fig. 3-5)
for the four models show that higher temperatures positively impact the red snow algae
occurrence. This positive relationship can be explained by the fact that red snow algae require
liquid water to bloom (Labourn Parry et al., 2000). Thus, warmer average air temperatures will
82

likely promote growth via increased meltwater and suitable areas for red snow algae. Cold
temperatures, on the other side, can prevent snow algae bloom, such as in the St-Elias
Mountains, where sparse algae blooms were linked to below-freezing average summer
temperatures (Engstrom & Quarmy, 2024) or at high elevations where snow remains dry during
the summer. Air temperature is also linked to other factors that may influence the occurrence of
red snow algae bloom, such as the duration of the growth period and the precipitation type.
Onuma et al. (2022) found that one of the main factors affecting snow algae growth was
interruption from new snow covers, which is linked to cold air temperatures. Overall, the result
suggests that warmer air temperature is positively related to the occurrence of red snow algae
Influence of dust flux
Dust flux contributes less than 20% of the model outcome. Its low explanation for the red
snow algae ratio in the models from this study indicates that atmospheric dust flux is not a good
predictor for red snow algae occurrence. These findings differ from previous studies that found
snow algae growth is promoted by mineral dust from the local lithology. For example, in
Greenland, phosphorous-rich mineral dust promotes ice algae growth (McCutcheon et al., 2021),
and on stratovolcano in the Pacific Northwest, US, snow algae utilized Fe, Mn and P leached
from local rocks (Hamilton & Havig, 2017). The dust flux used in the model was obtained from
MERRA-2, which mainly captures dues aerosol from large barren and arid land regions such as
the Sahara Desert and the North American desert, where strong winds transport soil particles into
the atmosphere (Global Modeling and Assimilation Office (GMAO), 2015). Therefore, it is
possible that global and regional dust flux patterns do not have a significant impact on snow
algae. However, localized mineral dust, which may not be captured in MERRA-2, still provides
the nutrients for snow algae. This process, however, was not captured in the model.
83

Predictors with low influence
The results show that surface net solar radiation, latitude, snowmelt, snow layer
temperature, and snowfall have the lowest impact on snow algae occurrence with the lowest
contribution to the model outcome. The model indicates that surface solar net radiation has
minimal impact on the algae ratio. While red snow algae are photosynthetic microbes, net solar
radiation does not negatively or positively impact growth in this study. Red snow algae blooms
occur in Antarctica (e.g. Huovinen et al., 2018), in the Arctic (e.g. Lutz et al., 2017), and in
alpine areas worldwide (e.g. Engstrom et al., 2020; Yoshimura et al.,1997). Yet, snow algae
blooms do not differ between areas that receive additional solar radiation during the summer,
such as the poles and at high elevations, versus lower latitudes and elevation areas where net
solar radiation is lower. To adapt to the change in solar radiation and other environmental factors,
studies observed a change in the red snow algae community structure with altitude (Takeuchi,
2013; Yoshimura et al., 1997) and locations (Engstrom et al., 2020; Lutz et al., 2016) instead.
Latitude was a variable of low importance overall. The model made no distinction
between the north and south distribution of snow algae. Red snow algae blooms have been
observed from the Sierra Nevada, California US (Painter et al., 2001) to the Eastern Alaska
Mountains. While those two locations extend from 37.55 °N to 64°N along the North American
Cordillera system, their maximum algae cell concentration in cells per meter square are both in
the order of 107, supporting that latitude does not impact snow algae bloom.
The results show that snowmelt has a weak negative relation with red snow algae. In a
previous study, the duration of snow cover throughout the summer season was noted to impact
the bloom extent (Engstrom & Quarmby, 2024). Years with summer snowpack that persisted late
into the season resulted in a large bloom extent. In contrast, in 2021, during the Pacific
84

Northwest heat wave, early snow loss over Western Canada and the Pacific Northwest US was
linked to a weak snow algae bloom. The snowfall, which was moderately to highly correlated
with snowmelt, is a variable with a low or no explanation for the red snow algae occurrence in
the models. In Onuma et al. (2022) snow algae model, snowfall events impacted the timing of
red snow algae appearance and could temporarily reduce snow algae growth. In this study,
snowfall is not an important variable for red snow algae. An explanation for the lack of
importance of snowfall can be related to the differences in the temporal resolution of the
datasets. Daily snowfall may impact red snow algae growth; however, when averaged over the
month, the snowfall effect is likely too small to be picked up by the model. The temperature of
the snow layer is also of low importance for the algae occurrence in the models. Red snow algae
blooms are restricted to periods when liquid water is available (Laybourn-Parry et al., 2012).
Liquid water content in the snowpack decreases rapidly when the snow temperature drops below
0°C (Pomeroy & Brun, 2010). In this study, snow layer temperature was not a good proxy for
liquid water availability.
3.6.2 Annual models and black carbon lag year models
Annual models
I ran the annual models to assess whether the years of high fire activity yield better
predictions. The models’ performance did not vary accordingly and explained about 40% of the
results except for 2018 and 2019. The 2018 model performed relatively well, with an R2 score of
0.7 and the 2019 performed poorly, with an average R2 of 0.1. The 2019 dataset, compared to the
other years, has more noise, outliers and extreme values, making it hard for the model to capture
meaningful relationships and thus create an accurate prediction. This explains why the 2019
model performed poorly. As for 2018 performance, it may be due to a higher correlation between
85

variables, greater variation in the predictor variables or clean patterns in the 2018 data compared
to the other datasets. The 2018 season was a high forest fire year in British Columbia, however, it
is the only year with high forest fire activity with an elevated R2 score. Therefore, it does not
provide an explanation for high forest fire season leading to better predictions.
The overall lower performance compared to the main model may be explained by the
small dataset used to run the annual model. The size of the dataset can also influence the
variables’ importance. The annual models were run on datasets with approximately 150 data
points. Such low data points can lead the model to rank irrelevant variables as important because
those variables are the best at explaining noise in the dataset. Small datasets can also influence
the ranking of variable importance due to the limited examples of certain relationships between
the different predictors and the target variable, leading to incorrect attributes of variable
importance. While most variables varied in importance among the nine annual models, the dust
flux and longitude consistently ranked as the top two variables with the most influence on the
algae ratio. The SHAP beeswarm plots (Appendix G, Fig G.1-18) indicate that low dust flux and
eastward longitude are linked to high algae ratio, agreeing with the findings from the main
models.
Lag year model
I perform the lag year model to assess whether a lag year relationship exists between
black carbon and algae, in which red snow algae are more predominant in years following high
forest fire season. The lag year model performed similarly to the main model. However, the
variable importance differs slightly compared to the main model. In the lag year model, black
carbon is not among the top predictors, and the dust flux and temperature of the snow layer
became more important. The black carbon is likely not an important predictor in the lag year
86

model as it refers to the monthly black carbon from the previous year. In the main model, black
carbon is one of the strongest predictors and is positively related to the red snow algae ratio.
Thus, using the monthly black carbon from the previous year likely leads to irrelevant patterns or
correlations with the current red snow algae, explaining why it is not rated as an important
predictor for that model. The dust flux, then, likely becomes more important to fill the gap
introduced by the lag year black carbon data points. The lag year model does not provide
evidence that years following high fire season lead to strong red snow algae bloom. The lag year
model likely performs similarly to the main model as both random forest and XGBoost rely on
aggregating multiple weak trees, making those models less sensitive to removing individual
variables. The model may also find some level of relationship between the lag year black carbon
and red snow algae, thus resulting in an unchanged performance score.
3.6.3 Models Performance and Associate Limitations
The performance of the random forest and XGBoost models was evaluated using
independent test datasets. The main models performed moderately well, with the best
performance attained on the monthly mean datasets in both models. This is likely because the
maximum datasets capture the monthly extreme, which may not provide a good proxy for
modelling the red snow algae occurrence, as the extremes are not representative of the general
trend over the month and can bias the model. While both models provide similar performance,
utilizing multiple machine-learning algorithms improves the robustness of the performance and
interpretability of the models’ results. While the models did not fully capture all the variables
that contribute toward the occurrence of red snow algae blooms, they contributed to improving
our understanding of the relevant drivers to model red snow algae bloom and suggest that black
carbon may positively influence red snow algae.
87

The annual models’ performed on average moderately weak. Those models were run on a
much smaller dataset, as explained above; therefore, it is not surprising to observe a lower
performance. Since small datasets contain fewer data points, the diversity and variability in the
data are generally lower than for large datasets. Moreover, since random forest and XGBoost rely
on the data to capture patterns and relationships, the small dataset may not provide enough
information to the model to result in accurate predictions or trends among the data may be too
weak for the model to detect. Small datasets can also lead to higher model bias because there is
less data to approximate the true relationship. This would explain why the main and lag-year
models performed better, as they were run on much larger datasets.
The uncertainties and biases associated with the predictor variables and target value
ultimately affect the models' performance. In terms of target value, although the RGND has been
widely used in previous studies to map and quantify snow algae (e.g. Ganey et al., 2017;
Huovinen et al., 2018), it can overestimate the extent of red snow algae bloom. For example,
Healy & Khan (2023) found that their optimized red/green band index applied to the MicaSense
bands tends to overestimate the extent of the algae bloom in the late season when debris-covered
snow is more common. However, they found that the red/green index had the most success at
mapping snow algae in their study compared to a machine learning method. To minimize the
uncertainty related to the index, I followed the approach suggested by (Di Mauro et al., 2024) to
mask rockfall and debris prior to applying the RGND index.
The use of meteorological reanalysis data such as ERA5-Land can result in bias,
especially in mountainous locations where ground stations are sporadic. Zhao & He (2022)
report through correlation and RMSE analysis that while ERA5-Land properly captured the
temperature trends over the Chinese Qilian Mountains, the value scale is not adequately
88

represented. Similarly, in a study evaluating ERA5-Land reanalysis as an input dataset for
surface energy balance modelling over mountain glaciers in Western Canada, Draeger et al.
(2024) found that ERA5-Land simulated the mean melt energy well if the temperature is
corrected through a lapse rate. While accurate magnitudes in temperature value are needed to
simulate daily melt energy, the monthly trend likely provides sufficient information for
modelling the occurrence of red snow algae as it is not based on energy balance calculation. As
for the snow variables, a study evaluating the snow cover properties of ERA5-Land found that
the snow cover extent is accurately represented compared to satellite-based datasets (Kouki et
al., 2023). Due to ERA5-Land being a relatively new product, evaluation of the net surface solar
radiation, snowmelt, and temperature of the snow layer has yet to be conducted.
The coarse resolution of the ERA5-Land dataset introduces some limitations as
meteorological values vary widely spatially and vertically. In terms of snow data in the model,
snowmelt and snow temperature vary with elevation, which is not captured in the current model.
The current grid size generally covers the entire glacier surface for the small-scale glaciers, or a
small aggregate of grid points covers the large icefields. Resolving this lack of resolution in the
dataset could help increase the accuracy of the model overall.
Another limitation of the models is related to the MERRA-2 black carbon aerosol
datasets and the assumption that most black carbon aerosols are emitted by forest fires rather
than anthropogenic sources. No studies have characterized the provenance of aerosols deposited
over glacier surfaces in western Canada. A study from the Pacific Northwest, US, however,
found that radiative forcing from black carbon over snow and ice was generally low except for
regions that are being impacted by forest fires (Skiles et al., 2018), suggesting that forest fire
aerosols likely dominate black carbon deposition in mountainous areas in Western Canada.
89

Moreover, previous studies investigating black carbon concentration produced during forest fires
in 2019 in Indonesia (Julian Sari et al., 2022) and from forest fires in Siberia between 2000 and
2016 (Sitnov et al., 2020) found a good positive correlation between BC aerosols from MERRA2 and forest fire hot spot.
3.7

Conclusions
Here, I assessed the interaction between black carbon, red snow algae and multiple other

environmental, climatic and geographical variables using decision-tree-based random forest
regressor and XGBoost machine learning models. The two main algorithms performed
moderately well at predicting the occurrence of red snow algae when compared to the testing
dataset. The models' performance suggests that the predictor variables explained part of the red
snow algae distribution but that other variables likely played a role in this complex phenomenon.
Due to the limited availability of widespread datasets of snow wetness covering all Western
Canada, the models did not include a predictor for snow liquid water content. Previous studies
identified snow wetness as a limiting factor controlling snow algae growth and appearance; thus,
integrating a value to quantify snow-liquid water would likely increase the model accuracy. In
this study, the data was constrained to the coarse spatial resolution from freely available satellite
imagery, meteorological reanalysis, and aerosol datasets, as well as the intermittent temporal
availability of multispectral satellite imagery. Applying the model to a finer spatial and temporal
resolution would capture subtle localized and seasonal effects that were missed in this model.
The most important variables that explained red snow algae occurrence in the models are
longitudes, black carbon and surface air temperatures. The positive relationship between snow
algae and air temperatures aligns with previous studies that suggest that the red snow algae range
is likely to expand with warming temperatures. Based on the geographical location of the study
90

area, the easternmost location had the strongest impact on the red snow algae. Those eastern
regions also have the strongest correlation between snow algae and black carbon, likely due to
their location downwind from the major forest fires that happened between 2015 and 2023.
Therefore, those results suggest that black carbon may induce snow algae growth, although
further studies observing the direct relationship are needed to confirm the interaction.
This study shows that leveraging machine learning can further our understanding of the
factors controlling the distribution of red snow algae blooms. It also shows that machine learning
is a powerful tool for investigating the interactions between glacier-darkening processes such as
snow algae and black carbon. Understanding the different factors influencing the distribution and
occurrence of red snow algae bloom enhances our ability to simulate glacier albedo and predict
glacier melt.

91

4

Chapter 4: Thesis Conclusions
This thesis investigated the relationship between red snow algae and black carbon via two

methodologies. First, I described and presented the findings from the field experiment I
conducted on Place Glacier during the summer of 2023, where I fertilized snow with wood ash
and compared snow algae growth and snow melt from a control and treated group. Second, I
applied random forest and XGBoost models using data from the Harmonized Landsat Sentinel
dataset, Merra-2 aerosol product and ERA5-Land climate reanalysis dataset to assess the
importance of various variables, including black carbon in modelling red snow algae occurrence.
Understanding interactions between glacier surface darkening agents such as red snow algae and
black carbon from forest fires provides insight into glacier albedo processes. Since, surface
glacier ablation is controlled mainly by surface air temperature and albedo, understanding the
interaction between surface darkening processes helps to improve our ability to predict glacier
albedo and surface melt. The significant findings of this study are as follows:
•

The field experiment presents no statistical evidence that wood ash enhances red snow
algae growth.

•

During the early bloom stage, the instantaneous radiative forcing of red snow algae in the
southern Coast Mountains of British Columbia I calculated was, on average, 192 ± 12 W
m-2 with a maximum of 274 ± 16 W m-2. This increase in radiative forcing translates to a
melt potential of 27.8 mm w.e. d-1. Those results support previous studies that found that
snowmelt induced by red snow algae is sufficient to increase glacier melt rates.

•

The random forest and XGBoost models suggest that black carbon flux from forest fire
positively influences red snow algae blooms. The models also imply that the occurrence
92

of red snow algae bloom is a complex phenomenon and that other factors likely influence
the distribution and occurrence of the blooms.
•

Red snow algae bloom in British Columbia and Alberta preferentially grows on
continental glaciers. Those regions are adjacent to the main area affected by large forest
fires in British Columbia supporting the idea that black carbon may promote red snow
algae growth.

A few main limitations inherently constrain the findings from this study. First, the temporal
resolution of the data from the field experiment and the remote sensing analysis limits the quality
of the results as the two analyses relied on temporally fixed data points. The field experiment
relied on three field visits, making a small dataset despite having multiple replicates, limiting the
statistical power of the analysis and the ability to capture the natural variability in the trends.
Moreover, with only three field visits, the dataset is more sensitive to human errors and
incidents. For instance, on the third field visit, the snow had melted in most of the plots,
significantly reducing the representativeness of the data and complicating trend analysis. To
mitigate these challenges, using automated instruments such as automated ablation stakes,
temperature sensors, and time-lapse cameras would help to supplement the in-person data
collection. On the other end, the remote sensing analysis was limited by the availability of
satellite imagery. Satellite imagery from the Harmonized Landsat Sentinel dataset is constrained
by the re-visit time and obstruction from clouds. Consequently, cloud-free images are not
available at regular time intervals, especially when acquired over a large study area as used in
this study, making it difficult to create time series. The use of monthly datasets instead of a
continuous time series limits the ability of the analysis to capture short-term environmental
variations or trends that could be significant for red snow algae growth. Moreover, identifying
93

precise timing and sequence of ephemeral factors driving algae bloom growth, such as summer
snowfall, becomes challenging when using monthly data.
Second, red snow algae are an intricate phenomenon, and while machine learning models
such as random forest and XGBoost are great tools for capturing complex patterns, the
performance and accuracy of the models are limited by the data inputs. Utilizing remote sensing
and reanalysis datasets allowed me to apply the model to a large study area; however, it limited
the variable inputs to readily available datasets. For example, the black carbon and dust flux
datasets refer to aerosol rather than deposited aerosols over glaciers, leading to the assumption
that high concentrations of atmospheric black carbon and dust lead to large deposition rates over
glaciers. Previous studies found that red snow algae are limited by the availability of liquid water
(eg. Ganey et al., 2017); however, identifying snow wetness over glaciers is challenging.
Therefore, the model did not include snow wetness even though it has been found to influence
snow algae distribution and growth.
To conclude, here are a few suggestions that can be undertaken from this thesis for future
research. First, the interaction between black carbon and red snow algae could be further
analyzed via a nutrient addition incubation experiment as conducted by McCutcheon et al.
(2021). While snow algae cultivation generally targets one species, such an experiment would
allow complete control of exterior factors and directly investigate the interaction between the
algae and the black carbon. This experiment, however, would only assess whether the algae
utilized the black carbon as a source of nutrients and would not evaluate if the increased liquid
water from the black carbon radiative forcing led to increased algae growth. A second
recommendation is to increase the temporal and spatial resolution of the dataset used in the
machine learning models by applying the model to multiple small regions, or target regions
94

adjacent to frequent forest fire activity. By doing so, the model could capture fine-scale
variations in environmental factors such as snowpack characteristics, microclimate or terrain
conditions that may influence snow algae blooms. Increasing the temporal resolution would help
to capture short-term changes in the weather and extremes such as snowfall and heatwave events
or rapidly changing snowpack conditions. Finally, glacier melt models should quantify albedo
change with respect to biological LAPs and inorganic LAPs. As Chapter two findings indicate,
red snow algae snowmelt contribution is substantial and needs to be considered in the glacier
melt rates. Considerations should also be taken for the interactions between red snow algae and
inorganic LAPs.

95

5

References

Anesio, A. M., & Laybourn-Parry, J. (2012). Glaciers and ice sheets as a biome. Trends in
Ecology & Evolution, 27(4), 219–225.
https://doi.org/https://doi.org/10.1016/j.tree.2011.09.012
Ardyna, M., Hamilton, D. S., Harmel, T., Lacour, L., Bernstein, D. N., Laliberté, J., Horvat, C.,
Laxenaire, R., Mills, M. M., van Dijken, G., Polyakov, I., Claustre, H., Mahowald, N., &
Arrigo, K. R. (2022). Wildfire aerosol deposition likely amplified a summertime Arctic
phytoplankton bloom. Communications Earth & Environment, 3(1), 201.
https://doi.org/10.1038/s43247-022-00511-9
Aubry‐Wake, C., Bertoncini, A., & Pomeroy, J. W. (2022). Fire and Ice: The Impact of Wildfire‐
Affected Albedo and Irradiance on Glacier Melt. Earth’s Future, 10(4), e2022EF002685.
https://doi.org/10.1029/2022EF002685
Barnaba, F., Angelini, F., Curci, G., & Gobbi, G. P. (2011). An important fingerprint of wildfires
on the European aerosol load. Atmospheric Chemistry and Physics, 11(20), 10487–10501.
https://doi.org/10.5194/acp-11-10487-2011
Bevington, A. R., & Menounos, B. (2022). Accelerated change in the glaciated environments of
western Canada revealed through trend analysis of optical satellite imagery. Remote Sensing
of Environment, 270, 112862. https://doi.org/https://doi.org/10.1016/j.rse.2021.112862
Bolch, T., Menounos, B., & Wheate, R. (2010). Landsat-based inventory of glaciers in western
Canada, 1985-2005. Remote Sensing of Environment, 114(1), 127–137.
https://doi.org/10.1016/j.rse.2009.08.015

Bonsal, B., Shrestha, R. R., Dibike, Y., Peters, D. L., Spence, C., Mudryk, L., & Yang, D.
(2020a). Western Canadian freshwater availability: current and future vulnerabilities.
Environmental Reviews, 28(4), 528–545. https://doi.org/10.1139/er-2020-0040
Bonsal, B., Shrestha, R. R., Dibike, Y., Peters, D. L., Spence, C., Mudryk, L., & Yang, D.
(2020b). Western Canadian freshwater availability: current and future vulnerabilities.
Environmental Reviews, 28(4), 528–545. https://doi.org/10.1139/er-2020-0040
Box, J. E., Fettweis, X., Stroeve, J. C., Tedesco, M., Hall, D. K., & Steffen, K. (2012). Greenland
ice sheet albedo feedback: thermodynamics and atmospheric drivers. The Cryosphere, 6(4),
821–839. https://doi.org/10.5194/tc-6-821-2012
Breiman, L. (2001). Random Forests (Vol. 45).
Chen, T., & Guestrin, C. (2016). XGBoost: A scalable tree boosting system. Proceedings of the
ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 13-17August-2016, 785–794. https://doi.org/10.1145/2939672.2939785
Chen, X., Kang, S., Yang, J., & Hu, Y. (2023). Contributions of biomass burning in 2019 and
2020 to Arctic black carbon and its transport pathways. Atmospheric Research, 296,
107069. https://doi.org/https://doi.org/10.1016/j.atmosres.2023.107069
Chevrollier, L.-A., Cook, J. M., Halbach, L., Jakobsen, H., Benning, L. G., Anesio, A. M., &
Tranter, M. (2022). Light absorption and albedo reduction by pigmented microalgae on
snow and ice. Journal of Glaciology, 1–9. https://doi.org/10.1017/JOG.2022.64

97

Clarke, G. K. C., Jarosch, A. H., Anslow, F. S., Radić, V., & Menounos, B. (2015). Projected
deglaciation of western Canada in the twenty-first century. Nature Geoscience 2014 8:5,
8(5), 372–377. https://doi.org/10.1038/ngeo2407
Claverie, M., Ju, J., Masek, J. G., Dungan, J. L., Vermote, E. F., Roger, J.-C., Skakun, S. V, &
Justice, C. (2018). The Harmonized Landsat and Sentinel-2 surface reflectance data set.
Remote Sensing of Environment, 219, 145–161.
https://doi.org/https://doi.org/10.1016/j.rse.2018.09.002
Coogan, S. C. P., Robinne, F.-N., Jain, P., & Flannigan, M. D. (2019). Scientists’ warning on
wildfire — a Canadian perspective. Canadian Journal of Forest Research, 49(9), 1015–
1023. https://doi.org/10.1139/cjfr-2019-0094
Cook, J. M., Hodson, A. J., Taggart, A. J., Mernild, S. H., & Tranter, M. (2017). A predictive
model for the spectral “bioalbedo” of snow. Journal of Geophysical Research: Earth
Surface, 122(1), 434–454. https://doi.org/https://doi.org/10.1002/2016JF003932
Cook, J. M., Tedstone, A. J., Williamson, C., McCutcheon, J., Hodson, A. J., Dayal, A., Skiles,
M., Hofer, S., Bryant, R., McAree, O., McGonigle, A., Ryan, J., Anesio, A. M., Irvine-Fynn,
T. D. L., Hubbard, A., Hanna, E., Flanner, M., Mayanna, S., Benning, L. G., … Tranter, M.
(2020). Glacier algae accelerate melt rates on the south-western Greenland Ice Sheet. The
Cryosphere, 14(1), 309–330. https://doi.org/10.5194/tc-14-309-2020
Di Mauro, B. (2020). A darker cryosphere in a warming world. Nature Climate Change, 10(11),
979–980. https://doi.org/10.1038/s41558-020-00911-9
Di Mauro, B., Fava, F., Ferrero, L., Garzonio, R., Baccolo, G., Delmonte, B., & Colombo, R.
(2015). Mineral dust impact on snow radiative properties in the European Alps combining
98

ground, UAV, and satellite observations. Journal of Geophysical Research: Atmospheres,
120(12), 6080–6097. https://doi.org/https://doi.org/10.1002/2015JD023287
Di Mauro, B., Garzonio, R., Ravasio, C., Orlandi, V., Baccolo, G., Gilardoni, S., Remias, D.,
Leoni, B., Rossini, M., & Colombo, R. (2024). Combined effect of algae and dust on snow
spectral and broadband albedo. Journal of Quantitative Spectroscopy and Radiative
Transfer, 316, 108906. https://doi.org/https://doi.org/10.1016/j.jqsrt.2024.108906
Donahue, C. P., Menounos, B., Viner, N., Skiles, S. M., Beffort, S., Denouden, T., Arriola, S. G.,
White, R., & Heathfield, D. (2023). Bridging the gap between airborne and spaceborne
imaging spectroscopy for mountain glacier surface property retrievals. Remote Sensing of
Environment, 299, 113849. https://doi.org/https://doi.org/10.1016/j.rse.2023.113849
Dumont, M., Gardelle, J., Sirguey, P., Guillot, A., Six, D., Rabatel, A., & Arnaud, Y. (2012).
Linking glacier annual mass balance and glacier albedo retrieved from MODIS data. The
Cryosphere, 6(6), 1527–1539. https://doi.org/10.5194/tc-6-1527-2012
Engstrom, C. B., & Quarmby, L. M. (2024). Satellite mapping of red snow on North American
glaciers. Science Advances, 9(47), eadi3268. https://doi.org/10.1126/sciadv.adi3268
Engstrom, C. B., Williamson, S. N., Gamon, J. A., & Quarmby, L. M. (2022). Seasonal
development and radiative forcing of red snow algal blooms on two glaciers in British
Columbia, Canada, summer 2020. Remote Sensing of Environment, 280, 113164.
https://doi.org/https://doi.org/10.1016/j.rse.2022.113164
Flanner, M. G., Arnheim, J. B., Cook, J. M., Dang, C., He, C., Huang, X., Singh, D., Skiles, S.
M., Whicker, C. A., & Zender, C. S. (2021). SNICAR-ADv3: a community tool for

99

modeling spectral snow albedo. Geoscientific Model Development, 14(12), 7673–7704.
https://doi.org/10.5194/gmd-14-7673-2021
Flanner, M. G., Zender, C. S., Randerson, J. T., & Rasch, P. J. (2007). Present-day climate
forcing and response from black carbon in snow. Journal of Geophysical Research:
Atmospheres, 112(D11). https://doi.org/https://doi.org/10.1029/2006JD008003
Friedman, J. H. (2001). Greedy function approximation: A gradient boosting machine. The
Annals of Statistics, 29(5), 1189–1232. https://doi.org/10.1214/aos/1013203451
Ganey, G. Q., Loso, M. G., Burgess, A. B., & Dial, R. J. (2017). The role of microbes in
snowmelt and radiative forcing on an Alaskan icefield. Nature Geoscience 2017 10:10,
10(10), 754–759. https://doi.org/10.1038/ngeo3027
Gelaro, R., McCarty, W., Suárez, M. J., Todling, R., Molod, A., Takacs, L., Randles, C. A.,
Darmenov, A., Bosilovich, M. G., Reichle, R., Wargan, K., Coy, L., Cullather, R., Draper,
C., Akella, S., Buchard, V., Conaty, A., da Silva, A. M., Gu, W., … Zhao, B. (2017). The
modern-era retrospective analysis for research and applications, version 2 (MERRA-2).
Journal of Climate, 30(14), 5419–5454. https://doi.org/10.1175/JCLI-D-16-0758.1
Government of British Columbia. (2024, April 26). Wildfire Averages.
Gray, A., Krolikowski, M., Fretwell, P., Convey, P., Peck, L. S., Mendelova, M., Smith, A. G., &
Davey, M. P. (2020). Remote sensing reveals Antarctic green snow algae as important
terrestrial carbon sink. Nature Communications, 11(1), 2527.
https://doi.org/10.1038/s41467-020-16018-w

100

Hadley, O. L., & Kirchstetter, T. W. (2012a). Black-carbon reduction of snow albedo. Nature
Climate Change, 2(6), 437–440. https://doi.org/10.1038/nclimate1433
Hadley, O. L., & Kirchstetter, T. W. (2012b). Black-carbon reduction of snow albedo. Nature
Climate Change, 2(6), 437–440. https://doi.org/10.1038/nclimate1433
Hadley, O. L., & Kirchstetter, T. W. (2012c). Black-carbon reduction of snow albedo. Nature
Climate Change, 2(6), 437–440. https://doi.org/10.1038/nclimate1433
Hamilton, T. L., & Havig, J. (2017). Primary productivity of snow algae communities on
stratovolcanoes of the Pacific Northwest. Geobiology, 15(2), 280–295.
https://doi.org/10.1111/GBI.12219
Hamilton, T. L., & Havig, J. R. (2020). Inorganic carbon addition stimulates snow algae primary
productivity. The ISME Journal, 14(3), 857–860. https://doi.org/10.1038/s41396-018-00486
Hanes, C. C., Wang, X., Jain, P., Parisien, M.-A., Little, J. M., & Flannigan, M. D. (2018). Fireregime changes in Canada over the last half century. Canadian Journal of Forest Research,
49(3), 256–269. https://doi.org/10.1139/cjfr-2018-0293
Healy, S. M., & Khan, A. L. (2023). Albedo change from snow algae blooms can contribute
substantially to snow melt in the North Cascades, USA. Communications Earth &
Environment, 4(1), 142. https://doi.org/10.1038/s43247-023-00768-8
Hock, R., Bliss, A., Marzeion, B. E. N., Giesen, R. H., Hirabayashi, Y., Huss, M., Radić, V., &
Slangen, A. B. A. (2019a). GlacierMIP – A model intercomparison of global-scale glacier

101

mass-balance models and projections. Journal of Glaciology, 65(251), 453–467.
https://doi.org/DOI: 10.1017/jog.2019.22
Hock, R., Bliss, A., Marzeion, B. E. N., Giesen, R. H., Hirabayashi, Y., Huss, M., Radić, V., &
Slangen, A. B. A. (2019b). GlacierMIP – A model intercomparison of global-scale glacier
mass-balance models and projections. Journal of Glaciology, 65(251), 453–467.
https://doi.org/DOI: 10.1017/jog.2019.22
Hodson, A. J., Sabacka, M., Dayal, A., Edwards, A., Cook, J., Convey, P., Redeker, K., & Pearce,
D. A. (2021). Marked Seasonal Changes in the Microbial Production, Community
Composition, and Biogeochemistry of Glacial Snowpack Ecosystems in the Maritime
Antarctic. Journal of Geophysical Research Biogeosciences, 126(7).
https://doi.org/10.1029/2020jg005706
Hugonnet, R., McNabb, R., Berthier, E., Menounos, B., Nuth, C., Girod, L., Farinotti, D., Huss,
M., Dussaillant, I., Brun, F., & Kääb, A. (2021). Accelerated global glacier mass loss in the
early twenty-first century. Nature, 592(7856), 726–731. https://doi.org/10.1038/s41586-02103436-z
Huovinen, P., Ramírez, J., & Gómez, I. (2018). Remote sensing of albedo-reducing snow algae
and impurities in the Maritime Antarctica. ISPRS Journal of Photogrammetry and Remote
Sensing, 146, 507–517. https://doi.org/https://doi.org/10.1016/j.isprsjprs.2018.10.015
Huss, M. (2011). Present and future contribution of glacier storage change to runoff from
macroscale drainage basins in Europe. Water Resources Research, 47(7).
https://doi.org/https://doi.org/10.1029/2010WR010299

102

Huss, M., & Hock, R. (2018). Global-scale hydrological response to future glacier mass loss.
Nature Climate Change, 8(2), 135–140. https://doi.org/10.1038/s41558-017-0049-x
Jensen, A. R., Anderson, K. S., Holmgren, W. F., Mikofski, M. A., Hansen, C. W., Boeman, L. J.,
& Loonen, R. (2023). pvlib iotools—Open-source Python functions for seamless access to
solar irradiance data. Solar Energy, 266, 112092.
https://doi.org/10.1016/j.solener.2023.112092
Jones, H. G., Pomeroy, J. W., Walker, D. A., & Hoham, R. W. (2001). Snow Ecology: An
Interdisciplinary Examination of Snow-Covered Ecosystems (1st ed.). Cambridge University
Press.
Jost, G., Moore, R. D., Menounos, B., & Wheate, R. (2012). Quantifying the contribution of
glacier runoff to streamflow in the upper Columbia River Basin, Canada. Hydrology and
Earth System Sciences, 16(3), 849–860. https://doi.org/10.5194/hess-16-849-2012
Kang, S., Zhang, Y., Qian, Y., & Wang, H. (2020). A review of black carbon in snow and ice and
its impact on the cryosphere. Earth-Science Reviews, 210, 103346.
https://doi.org/https://doi.org/10.1016/j.earscirev.2020.103346
Kaspari, S. D., Pittenger, D., Jenk, T. M., Morgenstern, U., Schwikowski, M., Buenning, N., &
Stott, L. (2020). Twentieth Century Black Carbon and Dust Deposition on South Cascade
Glacier, Washington State, USA, as Reconstructed From a 158-m-Long Ice Core. Journal of
Geophysical Research: Atmospheres, 125(11), e2019JD031126.
https://doi.org/https://doi.org/10.1029/2019JD031126
Kaspari, S., McKenzie Skiles, S., Delaney, I., Dixon, D., & Painter, T. H. (2015). Accelerated
glacier melt on Snow Dome, Mount Olympus, Washington, USA, due to deposition of black
103

carbon and mineral dust from wildfire. Journal of Geophysical Research: Atmospheres,
120(7), 2793–2807. https://doi.org/https://doi.org/10.1002/2014JD022676
Keegan, K. M., Albert, M. R., McConnell, J. R., & Baker, I. (2014). Climate change and forest
fires synergistically drive widespread melt events of the Greenland Ice Sheet. Proceedings
of the National Academy of Sciences, 111(22), 7964–7967.
https://doi.org/10.1073/pnas.1405397111
Khan, A. L., Dierssen, H. M., Scambos, T. A., Höfer, J., & Cordero, R. R. (2021). Spectral
characterization, radiative forcing and pigment content of coastal Antarctic snow algae:
approaches to spectrally discriminate red and green communities and their impact on
snowmelt. The Cryosphere, 15(1), 133–148. https://doi.org/10.5194/tc-15-133-2021
Kim, Y., Hatsushika, H., Muskett, R., & Yamazaki, K. (2005). Possible Effect of Black Carbon
Wildfire Soot on Arctic Sea Ice and Alaska Glaciers. AGU Fall Meeting Abstracts, 39.
Kondo, Y., Matsui, H., Moteki, N., Sahu, L., Takegawa, N., Kajino, M., Zhao, Y., Cubison, M. J.,
Jimenez, J. L., Vay, S., Diskin, G. S., Anderson, B., Wisthaler, A., Mikoviny, T., Fuelberg,
H. E., Blake, D. R., Huey, G., Weinheimer, A. J., Knapp, D. J., & Brune, W. H. (2011).
Emissions of black carbon, organic, and inorganic aerosols from biomass burning in North
America and Asia in 2008. Journal of Geophysical Research: Atmospheres, 116(D8).
https://doi.org/https://doi.org/10.1029/2010JD015152
L., W. A., G., H. H., R., C. D., & W., S. T. (2006). Warming and Earlier Spring Increase Western
U.S. Forest Wildfire Activity. Science, 313(5789), 940–943.
https://doi.org/10.1126/science.1128834

104

Laybourn-Parry, J., Tranter, M., & Hodson, A. J. (2012). The Ecology of Snow and Ice
Environments. Oxford University Press.
Lundberg, S. M., & Lee, S.-I. (2017). A Unified Approach to Interpreting Model Predictions. In
I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, & R.
Garnett (Eds.), Advances in Neural Information Processing Systems (Vol. 30). Curran
Associates, Inc.
https://proceedings.neurips.cc/paper_files/paper/2017/file/8a20a8621978632d76c43dfd28b6
7767-Paper.pdf
Lutz, S., Anesio, A. M., Edwards, A., & Benning, L. G. (2015). Microbial diversity on Icelandic
glaciers and ice caps. Frontiers in Microbiology, 6.
https://www.frontiersin.org/articles/10.3389/fmicb.2015.00307
Lutz, S., Anesio, A. M., Jorge Villar, S. E., & Benning, L. G. (2014). Variations of algal
communities cause darkening of a Greenland glacier. FEMS Microbiology Ecology, 89(2),
402–414. https://doi.org/10.1111/1574-6941.12351
Marshall, S. J., & Miller, K. (2020). Seasonal and interannual variability of melt-season albedo at
Haig Glacier, Canadian Rocky Mountains. The Cryosphere, 14(10), 3249–3267.
https://doi.org/10.5194/tc-14-3249-2020
Matsumoto, M., Hanneman, C., Camara, A. G., Krueger-Hadfield, S. A., Hamilton, T. L., &
Kodner, R. B. (2024). Hypothesized life cycle of the snow algae Chlainomonas sp.
(Chlamydomonadales, Chlorophyta) from the Cascade Mountains, USA. Journal of
Phycology, n/a(n/a). https://doi.org/https://doi.org/10.1111/jpy.13454

105

McCutcheon, J., Lutz, S., Williamson, C., Cook, J. M., Tedstone, A. J., Vanderstraeten, A.,
Wilson, S. A., Stockdale, A., Bonneville, S., Anesio, A. M., Yallop, M. L., McQuaid, J. B.,
Tranter, M., & Benning, L. G. (2021). Mineral phosphorus drives glacier algal blooms on
the Greenland Ice Sheet. Nature Communications 2021 12:1, 12(1), 1–11.
https://doi.org/10.1038/s41467-020-20627-w
Meyer, N. K. (2012). Particulate, black carbon and organic emissions from small-scale
residential wood combustion appliances in Switzerland. Biomass and Bioenergy, 36, 31–42.
https://doi.org/https://doi.org/10.1016/j.biombioe.2011.09.023
Moore, R. D., Fleming, S. W., Menounos, B., Wheate, R., Fountain, A., Stahl, K., Holm, K., &
Jakob, M. (2009a). Glacier change in western North America: influences on hydrology,
geomorphic hazards and water quality. Hydrological Processes, 23(1), 42–61.
https://doi.org/https://doi.org/10.1002/hyp.7162
Moore, R. D., Fleming, S. W., Menounos, B., Wheate, R., Fountain, A., Stahl, K., Holm, K., &
Jakob, M. (2009b). Glacier change in western North America: influences on hydrology,
geomorphic hazards and water quality. Hydrological Processes, 23(1), 42–61.
https://doi.org/https://doi.org/10.1002/hyp.7162
Moore, R. D., Pelto, B., Menounos, B., & Hutchinson, D. (2020). Detecting the Effects of
Sustained Glacier Wastage on Streamflow in Variably Glacierized Catchments. Frontiers in
Earth Science, 8. https://doi.org/10.3389/feart.2020.00136
Mukherjee, K., Menounos, B., Shea, J., Mortezapour, M., Ednie, M., & Demuth, M. N. (2022).
Evaluation of surface mass-balance records using geodetic data and physically-based

106

modelling, Place and Peyto glaciers, western Canada. Journal of Glaciology, 1–18.
https://doi.org/DOI: 10.1017/jog.2022.83
Müller, T., Bleiß, W., Martin, C.-D., Rogaschewski, S., & Fuhr, G. (1998). Snow algae from
northwest Svalbard: their identification, distribution, pigment and nutrient content. Polar
Biology, 20(1), 14–32. https://doi.org/10.1007/s003000050272
Muñoz-Sabater, J., Dutra, E., Agust\’\i-Panareda, A., Albergel, C., Arduini, G., Balsamo, G.,
Boussetta, S., Choulga, M., Harrigan, S., Hersbach, H., Martens, B., Miralles, D. G., Piles,
M., Rodr\’\iguez-Fernández, N. J., Zsoter, E., Buontempo, C., & Thépaut, J.-N. (2021).
ERA5-Land: a state-of-the-art global reanalysis dataset for land applications. Earth System
Science Data, 13(9), 4349–4383. https://doi.org/10.5194/essd-13-4349-2021
Nagorski, S. A., Kaspari, S. D., Hood, E., Fellman, J. B., & Skiles, S. M. (2019). Radiative
Forcing by Dust and Black Carbon on the Juneau Icefield, Alaska. Journal of Geophysical
Research: Atmospheres, 124(7), 3943–3959. https://doi.org/10.1029/2018JD029411
Onuma, Y., Takeuchi, N., & Takeuchi, Y. (2016). Temporal changes in snow algal abundance on
surface snow in Tohkamachi, Japan. Bulletin of Glaciological Research, 34, 21–31.
https://api.semanticscholar.org/CorpusID:132841299
Onuma, Y., Yoshimura, K., & Takeuchi, N. (2022). Global Simulation of Snow Algal Blooming
by Coupling a Land Surface and Newly Developed Snow Algae Models. Journal of
Geophysical Research: Biogeosciences, 127(2). https://doi.org/10.1029/2021JG006339
Painter, T., Duval, B., William, T. H., Mendez, M., Hentzelman, S., & Dozier, J. (2001).
Detection and Quantification of Snow Algae with an Airborne Imaging Spectrometer.

107

Applied and Environmental Microbiology, 67(11), 5267–5272.
https://doi.org/10.1128/AEM.67.11.5267-5272.2001
Painter, T. H., Seidel, F. C., Bryant, A. C., McKenzie Skiles, S., & Rittger, K. (2013). Imaging
spectroscopy of albedo and radiative forcing by light-absorbing impurities in mountain
snow. Journal of Geophysical Research: Atmospheres, 118(17), 9511–9523.
https://doi.org/https://doi.org/10.1002/jgrd.50520
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M.,
Prettenhofer, P., Weiss, R., Dubourg, V., Vanderpla, J., Passos, A., D., C., Brucher, M.,
Perrot, M., & Duchesnay, E. (2011). Scikit-learn: Machine Learning in Python. Journal of
Machine Learning Research, 12, 2825–2830.
Phillips-Lander, C. M., Harrold, Z., Hausrath, E. M., Lanzirotti, A., Newville, M., Adcock, C. T.,
Raymond, J. A., Huang, S., Tschauner, O., & Sanchez, A. (2020). Snow Algae Preferentially
Grow on Fe-containing Minerals and Contribute to the Formation of Fe Phases.
Geomicrobiology Journal, 37(6), 572–581. https://doi.org/10.1080/01490451.2020.1739176
Pu, W., Cui, J., Wu, D., Shi, T., Chen, Y., Xing, Y., Zhou, Y., & Wang, X. (2021). Unprecedented
snow darkening and melting in New Zealand due to 2019–2020 Australian wildfires.
Fundamental Research, 1(3), 224–231.
https://doi.org/https://doi.org/10.1016/j.fmre.2021.04.001
Qin, W., Zhang, Y., Chen, J., Yu, Q., Cheng, S., Li, W., Liu, X., & Tian, H. (2019). Variation,
sources and historical trend of black carbon in Beijing, China based on ground observation
and MERRA-2 reanalysis data. Environmental Pollution, 245, 853–863.
https://doi.org/https://doi.org/10.1016/j.envpol.2018.11.063
108

Randles, C. A., Da Silva, A. M., Buchard, V., Colarco, P. R., Darmenov, A., Govindaraju, R.,
Smirnov, A., Holben, B., Ferrare, R., Hair, J., Shinozuka, Y., & Flynn, C. J. (n.d.). The
MERRA-2 Aerosol Reanalysis, 1980 Onward. Part I: System Description and Data
Assimilation Evaluation. https://doi.org/10.1175/JCLI-D-16-0609.s1
RGI Consortium. (2023). Randolph Glacier Inventory - A Dataset of Global Glacier Outlines,
Version 7 [Data Set]. National Snow and Ice Data Center.
Schoen, E. R., Wipfli, M. S., Trammell, E. J., Rinella, D. J., Floyd, A. L., Grunblatt, J.,
McCarthy, M. D., Meyer, B. E., Morton, J. M., Powell, J. E., Prakash, A., Reimer, M. N.,
Stuefer, S. L., Toniolo, H., Wells, B. M., & Witmer, F. D. W. (2017). Future of Pacific
Salmon in the Face of Environmental Change: Lessons from One of the World’s Remaining
Productive Salmon Regions. Fisheries, 42(10), 538–553.
https://doi.org/https://doi.org/10.1080/03632415.2017.1374251
Scikit learn. (2024). Permutation feature importance.
Shea, J. M., Dan Moore, R., & Stahl, K. (2009). Derivation of melt factors from glacier massbalance records in western Canada. Journal of Glaciology, 55(189), 123–130.
https://doi.org/DOI: 10.3189/002214309788608886
Sitnov, S. A., Mokhov, I. I., & Likhosherstova, A. A. (2020). Exploring large-scale black-carbon
air pollution over Northern Eurasia in summer 2016 using MERRA-2 reanalysis data.
Atmospheric Research, 235, 104763.
https://doi.org/https://doi.org/10.1016/j.atmosres.2019.104763

109

Skiles, S. M., Flanner, M., Cook, J. M., Dumont, M., & Painter, T. H. (2018). Radiative forcing
by light-absorbing particles in snow. Nature Climate Change, 8(11), 964–971.
https://doi.org/10.1038/s41558-018-0296-5
Stahl, K., & Moore, R. D. (2006). Influence of watershed glacier coverage on summer
streamflow in British Columbia, Canada. Water Resources Research, 42(6).
https://doi.org/https://doi.org/10.1029/2006WR005022
Takeuchi, N. (2013). Seasonal and altitudinal variations in snow algal communities on an
Alaskan glacier (Gulkana glacier in the Alaska range). Environmental Research Letters,
8(3), 35002. https://doi.org/10.1088/1748-9326/8/3/035002
Takeuchi, N., Dial, R., Kohshima, S., Segawa, T., & Uetake, J. (2006). Spatial distribution and
abundance of red snow algae on the Harding Icefield, Alaska derived from a satellite image.
Geophysical Research Letters, 33(21). https://doi.org/10.1029/2006GL027819
Thomas, J. L., Polashenski, C. M., Soja, A. J., Marelle, L., Casey, K. A., Choi, H. D., Raut, J.-C.,
Wiedinmyer, C., Emmons, L. K., Fast, J. D., Pelon, J., Law, K. S., Flanner, M. G., & Dibb,
J. E. (2017). Quantifying black carbon deposition over the Greenland ice sheet from forest
fires in Canada. Geophysical Research Letters, 44(15), 7965–7974.
https://doi.org/https://doi.org/10.1002/2017GL073701
U.S. Environmental Protection Agency. (1994a). Determination of metals and trace elements in
water and wastes by inductively coupled plasma-atomic emission spectrometry .
U.S. Environmental Protection Agency. (1994b). Sample preparation procedure for
spectrochemical determination of total recoverable elements.

110

https://www.epa.gov/sites/default/files/2015-08/documents/method_200-2_rev_28_1994.pdf
Wang, X., Shi, T., Zhang, X., & Chen, Y. (2020). An Overview of Snow Albedo Sensitivity to
Black Carbon Contamination and Snow Grain Properties Based on Experimental Datasets
Across the Northern Hemisphere. Current Pollution Reports, 6(4), 368–379.
https://doi.org/10.1007/s40726-020-00157-1
Warren, S. G. (1982). Optical properties of snow. Reviews of Geophysics, 20(1), 67–89.
https://doi.org/https://doi.org/10.1029/RG020i001p00067
Warren, S. G. (2013). Can black carbon in snow be detected by remote sensing? Journal of
Geophysical Research: Atmospheres, 118(2), 779–786.
https://doi.org/https://doi.org/10.1029/2012JD018476
Warren, S. G., & Wiscombe, W. J. (1980). A Model for the Spectral Albedo of Snow. II: Snow
Containing Atmospheric Aerosols. Journal of Atmospheric Sciences, 37(12), 2734–2745.
https://doi.org/10.1175/1520-0469(1980)037<2734:AMFTSA>2.0.CO;2
Whicker, C. A., Flanner, M. G., Dang, C., Zender, C. S., Cook, J. M., & Gardner, A. S. (2022).
SNICAR-ADv4: a physically based radiative transfer model to represent the spectral albedo
of glacier ice. The Cryosphere, 16(4), 1197–1220. https://doi.org/10.5194/tc-16-1197-2022
Williamson, S. N., & Menounos, B. (2021). The influence of forest fire aerosol and air
temperature on glacier albedo, western North America. Remote Sensing of Environment,
267, 112732. https://doi.org/10.1016/j.rse.2021.112732

111

Wiscombe, W. J., & Warren, S. G. (1980). A Model for the Spectral Albedo of Snow. I: Pure
Snow. Journal of Atmospheric Sciences, 37(12), 2712–2733. https://doi.org/10.1175/15200469(1980)037<2712:AMFTSA>2.0.CO;2
You, C., Yao, T., Xu, B., Xu, C., Zhao, H., & Song, L. (2016). Effects of sources, transport, and
postdepositional processes on levoglucosan records in southeastern Tibetan glaciers.
Journal of Geophysical Research: Atmospheres, 121(14), 8701–8711.
https://doi.org/https://doi.org/10.1002/2016JD024904
Zhang, H.-M., Huang, B., Lawrimore, J., Menne, M., & Smith, T. M. (n.d.). NOAA Global
Surface Temperature Dataset (NOAAGlobalTemp), Version 5. NOAA National Centers for
Environmental Information. https://doi.org/doi:10.25921/9qth-2p70

112

6

Appendices

6.1

Appendix A: Spring surface air temperature at Place Glacier

Figure A.1: Surface air temperature at Place Glacier in April and May 2023 from HOBO temperature
sensor, elevation 1924 m. a.s.l. The red arrow indicates the approximate date when the air temperature
sensor is uncovered and above the snowpack/ice surface.

113

6.2

Appendix B: Elemental metal concentration of the wood ash, control and treated
samples.

Table B 1: Elemental metal concentration for the wood ash, control and treated samples. Rows in

bold represent elements with significantly higher concentrations in the treated samples.

Metals

Wood Ash
(mg Kg-1)

Control Sample
(mg L-1)

Treated Sample
(mg L-1)

Ρ - Value

Aluminum
Arsenic
Boron
Barium
Calcium
Cadmium
Cobalt
Chromium
Copper
Iron
Potassium
Magnesium
Manganese
Molybdenum
Sodium
Nickel
Phosphorus
Lead
Sulfur
Antimony
Selenium
Tin
Uranium
Vanadium
Zinc

10395
<20
321
3054
246384
19
9
66
110
12092
29355
25205
15000
8
6184
59
12298
13
10327
42
<20
79
35
25
3550

0.208
<0.02
<0.001
0.002
0.072
<0.0004
<0.001
<0.001
<0.002
0.228
0.132
0.093
0.004
<0.003
0.078
<0.001
0.009
<0.004
<0.08
<0.01
<0.02
<0.01
<0.01
<0.001
0.064

0.311
<0.02
<0.001
0.033
2.064
<0.0004
<0.001
0.001
<0.002
0.345
0.271
0.374
0.140
<0.003
0.098
0.001
0.108
0.001
<0.08
<0.01
<0.02
<0.01
<0.01
<0.001
0.059

0.00579
NaN
0.329
0.0
0.0
NaN
NaN
0.0
NaN
0.00328
0.00118
0.0
0.0
NaN
0.391
0.00455
0.0
0.155
NaN
NaN
NaN
NaN
NaN
NaN
0.0106

114

6.3

Appendix C: HLS tile names per Mountain Regions

Table C 1: HLS tile names per Mountain Regions
Region Number

Region Name

HLS Tile Name

1

Southern Coast
Mountains

2

Central Coast
Mountains

3

Northern Coast
Mountains

4

St. Elias

T09UXS
T09UXT
T09UXU
T09UYR
T09UYS
T09UYT
T09UYT
T09UYU
T10UCA
T10UCB
T10UCC
T10UCD
T10UDA
T10UDB
T10UDC
T10UEA
T10UEB
T10UEV
T09UVA
T09UVB
T09UVV
T09UWA
T09UWB
T09UWU
T09UWV
T09UXU
T09UXV
T08VML
T08VMM
T08VNK
T08VNL
T08VNM
T08VPH
T09UVB
T09VUC
T09VVC
T09VVD
T07VEG

115

5

Interior North

6

Interior South

7

Northern
Rockies

8

Central
Rockies
116

T07VFF
T07VFG
T08VLL
T08VLM
T08VMM
T+C45:C6009UWA
T09UWB
T09UXA
T09UXB
T09VVC
T09VVD
T09VVE
T09VWC
T09VWD
T09VWE
T09VWF
T09VXC
T09VXD
T09VXE
T10VCH
T10VCJ
T10UFD
T10UFE
T10UGB
T10UGC
T10UGD
T11ULS
T11ULT
T11ULU
T11UMR
T11UMS
T11UMT
T11UNR
T11UNS
T09VXE
T09VXF
T10VCJ
T10VCK
T10VC
T10VDJ
T10VDK
T10UEF
T10UFE

9

Southern
Rockies

117

T10UFF
T10UGD
T10UGE
T11ULU
T11ULV
T10UGD
T10UGE
T11ULT
T11ULU
T11ULV
T11UMT
T11UMU
T11UMV
T11UNS
T11UNT
T11UNU
T11UPR
T11UPS

6.4

Appendix D: Parameters used for Random Forest and the XGBoost models for the
Annual and lag year model

Table D 1: Random forest parameters used for the annual models and the lag model. The ‘F’ and “T’
stands for True and False. Table 3.2 describes the parameters.

Parameters
n_estimators
max_depth
min_samples_splits
min_samples_leaf
max_feature
Bootstrap

2015
1000
60
3
1
sqrt'
F

2016
200
40
2
5
log2'
T

2017
100
40
3
3
sqrt'
T

2018 2019
100 2500
100
40
3
5
3
2
sqrt'
sqrt'
F
F

2020
500
60
3
1
log2'
T

2021
500
40
2
1
log2'
T

2022
300
100
3
1
log2'
T

2023 Lag Year
200
100
200
80
3
5
1
2
log2'
log2'
F
F

Table D 2: XGBoost parameters used for the annual models and the lag model. Table 3.3 describes the
parameters.

Parameters
learning_rate
n_estimator
max_depth
min_child_weigth
gamma
subsamble
colsample_bytree
reg_alpha

2015
0.2
300
4
5
0
0.6
0.8
0

2016
0.01
100
6
1
0
0.6
0.8
0

2017
0.01
100
9
1
0
0.9
0.8
1E-05

2018 2019
0.2 0.01
300 100
7
6
2
3
0
0
0.8
0.6
0.8
0.7
0
0

118

2020
0.01
100
5
4
0
0.9
0.8
0

2021
0.1
100
5
1
0
0.9
0.8
0.01

2022
0.01
300
3
1
0
0.6
0.9
1

2023
0.01
100
7
1
0
0.7
0.8
0

Lag Year
0.01
1000
7
4
0
0.9
0.6
0.1

6.5

Appendix E: Correlation matrix significance values

Figure E 1: Mean variables correlation’s significance for the correlation matrix. The green indicates a
correlation with ρ< 0.05, thus significant.

119

Figure E 2: Max variables correlation’s significance for the correlation matrix. The green indicates a
correlation with ρ< 0.05, thus significant.

120

Appendix F: Annual model permutation feature importance

bc_flux

6.6

bc_flux

Figure F 1:Permutation feature importance for the random forest on the annual model 2015

Figure F.2: Permutation feature importance for the XGBoost on the annual model 2015

121

bc_flux
bc_flux

Figure F.3 : Permutation feature importance for the random forest on the annual model 2016

Figure F.4: Permutation feature importance for the XGBoost on the annual model 2016

122

bc_flux
bc_flux

Figure F.5: Permutation feature importance for the random forest on the annual model 2017

Figure F.6: Permutation feature importance for the XGBoost on the annual model 2017

123

bc_flux
bc_flux

Figure F.7: Permutation feature importance for the random forest on the annual model 2018

Figure F.8 : Permutation feature importance for the XGBoost on the annual model 2018

124

bc_flux
bc_flux

Figure F.9: Permutation feature importance for the random forest on the annual model 2019

Figure F.10: Permutation feature importance for the XGBoost on the annual model 2019

125

bc_flux
bc_flux

Figure F.11: Permutation feature importance for the random forest on the annual model 2020

Figure F.12: Permutation feature importance for the XGBoost on the annual model 2020

126

bc_flux
bc_flux

Figure F.13: Permutation feature importance for the random forest on the annual model 2021

Figure F.14: Permutation feature importance for the XGBoost on the annual model 2021

127

bc_flux
bc_flux

Figure F.15: Permutation feature importance for the random forest on the annual model 2022

Figure F.16: Permutation feature importance for the XGBoost on the annual model 2022

128

bc_flux
bc_flux

Figure F.17: Permutation feature importance for the random forest on the annual model 2023

Figure F.18: Permutation feature importance for the XGBoost on the annual model 2023

129

6.7

Appendix G: Annual model SHAP Value

Figure G 1: SHAP Values for the random forest on the annual model 2015

Figure G 2: SHAP Values for the XGBOOST on the annual model 2015

130

Figure G 3: SHAP Values for the random forest on the annual model 2016

Figure G 4: SHAP Values for the XGBOOST on the annual model 2016

131

Figure G 5: SHAP Values for the random forest on the annual model 2017

Figure G 6: SHAP Values for the XGBOOST on the annual model 2017

132

Figure G 7: SHAP Values for the random forest on the annual model 2018

Figure G 8: SHAP Values for the XGBOOST on the annual model 2018

133

Figure G 9: SHAP Values for the random forest on the annual model 2019

Figure G 10: SHAP Values for the XGBOOST on the annual model 2019

134

Figure G 11: SHAP Values for the random forest on the annual model 2020

Figure G 12: SHAP Values for the XGBOOST on the annual model 2020

135

Figure G 13: SHAP Values for the random forest on the annual model 2021

Figure G 14: SHAP Values for the random forest on the annual model 2021

136

Figure G 15: SHAP Values for the random forest on the annual model 2022

Figure G 16: SHAP Values for the XGBOOST on the annual model 2022

137

Figure G 17: SHAP Values for the random forest on the annual model 2015

Figure G 18: SHAP Values for the random forest on the annual model 2015

138