Regional-Scale Permafrost Mapping by Scalable Machine Learning

Abstract

This thesis presents an integrated approach to mapping permafrost at high-resolution using scalable machine learning (ML) methods and is composed of two interrelated components. The first component is a research paper outlining the development and validation of a ML framework and model to predict permafrost presence and ice content across west‐central Yukon, Canada (Chapter 2). The second component is a published, open-source R package developed in support of the aforementioned permafrost prediction model (Appendix), created with the goal of facilitating future applications of ML techniques to spatial predictions of permafrost or other spatial phenomena with built in support for model optimization.

In Chapter 2, permafrost distribution and relative ice content is modelled across west-central Yukon using as a training set a detailed field map of permafrost conditions across a proposed mine site (the Coffee Gold Project). Permafrost distribution was constructed using ML and a series of spatial predictors derived primarily from an open-access Canadian high-resolution (16 m) digital elevation model (e.g., elevation, slope, aspect, curvature, solar radiation). Using random forest (RF) classification, we predicted permafrost distribution across the training area with a balanced accuracy of approximately 0.86. We also validated the model's performance by accurately reproducing a permafrost distribution map produced at a second, 35 km distant mining project (Casino Mine Project) to a 73 % agreement. This high accuracy underscores the potential for ML to bridge the scale gap between detailed local field observations and broader regional permafrost mapping efforts and substantially improves upon the previously existing permafrost models over Yukon in terms of resolution and accuracy.

Complementing the research paper, the SAiVE R package (Appendix) offers a robust, reproducible, and user-friendly tool to streamline the creation of optimized ML models predicting spatially dependent variables. The package streamlines tasks ranging from predictor variable extraction and spatial sampling, model training, testing, and comparison and selection, and (if desired) the extrapolation of predictions over extensive geographic areas. By integrating functionalities for hyperparameter tuning and variable importance ranking, SAiVE not only facilitates advanced spatial modelling for permafrost studies but also provides a flexible framework adaptable to other environmental applications.

Together, these contributions offer a scalable solution for permafrost mapping that supports environmental management, infrastructure planning, and climate adaptation strategies in rapidly changing Arctic and sub-Arctic regions. I expect that the modelling approach presented here will lead to improved representations of permafrost in Yukon and beyond, with direct applications for research, hydrology, mining exploration, and infrastructure projects looking to extrapolate local permafrost measurements to broader scales.