Shapley Explainable AI (XAI) to Assess the Potential of Satellite Remote Sensing Data for Estimating Active Layer Thickness (ALT)

Permafrost environments are increasingly threatened by warming temperatures. Active Layer Thickness (ALT) is a critical variable because it influences several environmental processes, particularly carbon storage and cycling. Remote Sensing (RS) techniques hold great potential for mapping ALT over remote, challenging, and inaccessible permafrost regions, especially in comparison to in situ field surveys. In this study, various RS sources (radar, optical, topographic) were assessed for ALT estimation. ALT samples were obtained from airborne radar imaging (P- and L-Band) through the NASA Arctic Boreal Vulnerability Experiment (ABoVE), and then upscaled to satellite observations using machine learning (ML) modelling. Using Explainable AI (XAI) methods, we provide interpretation of our ML modeling by assessing the relevance of 116 input RS features. The Shapley additive explanations (SHAP) XAI method was utilized for this interpretation task. Temperature and topography features provided the greatest explanation in ALT variability. Therefore, this study provides insight into the potential of satellite RS for estimating ALT, and the relative importance of RS features for this challenging task.