Machine learning based classification of early seral vegetation in cut-blocks in the interior of northern British Columbia

Globally forests provide a wide range of essential services such as lumber for construction, tourism value, and habitat for animals. In many regions forest management is performed to maximize the utilization of these services and to promote sustainable forest ecosystems. Effective management requires detailed information on the current state of forests, how the forest is projected to develop through time, and knowledge about the provisioning of desired forest services, such as forage for wildlife species. Historically this information has been acquired using traditional field surveys, which is both costly and limited in the extent of area that can be sampled. The use of Remotely Piloted Aircraft Systems (RPAS) combined with machine learning potentially allows for more scalable methods of gathering information on forest inventories. In this thesis, I evaluate and advance the use of multispectral imagery collected from RPAS for the classification of early seral vegetation. This specific type of vegetation is both a key indicator of forest regeneration and habitat suitability for ungulates. However, accurate identification and classification of early seral vegetation is particularly challenging due to its small size, the fact that individuals are highly variable, and the fact that individuals can overlap and not exhibit distinct boundaries. The process of image classification is broken down into two major components: the segmentation of collected imagery into discrete units of vegetation and then the classification of those units into their specific species. These two components are presented as an overall framework for classification. I also provide operational recommendations to achieve successful results. The algorithms used in the segmentation of images are highly configurable and can be tuned to the input data to yield high quality results; however, what is more challenging is determining what a high-quality result is, and applying suitable metrics that allow the accuracy of the segmentation process to be evaluated. In this research I propose a method for scoring the quality of segmentation quality applied to forest imagery, in a format that can be easily integrated into a larger framework that will integrate with the classification of results. In the second component of my thesis, I evaluate various common classification algorithms and assessed their accuracy. This analysis considered both overall accuracy of classification, as well as only the classification accuracy of species of interest. I also explore under what circumstances this type of classification be feasible and provide recommendations on what variables are most important to control during the collection of training data, and best practice for capture of new datasets for classification with already trained models. My research demonstrates both the benefits and limitations of using RPAS imagery for segmentation and classification of early seral vegetation and suggests best practices that can be used when applying this framework.