Exploring thawing patterns: data-driven modelling of permafrost dynamics using Random Forest and Neural Networks

ABSTRACT: This study examines permafrost thaw dynamics in the Northwest Territories' Mackenzie River Delta (MRD) using machine learning (ML) techniques, specifically Random Forest (RF) and Neural Networks (NNs). Utilizing a 10-year dataset from 79 observation sites, we predicted the spatial distribution of two key permafrost parameters: Mean Annual Ground Temperature (MAGT) and Active Layer Thickness (ALT). Both RF and NNs contributed complementary strengths — while NNs yielded lower RMSE and higher R² values overall, RF showed enhanced performance in areas with sparse data coverage. The strong agreement between both models underlines the robustness of our results. Beyond confirming the expected north-to-south warming gradient, our study reveals its fine-scale spatial variability linked to vegetation, topography, and hydrological conditions.