Predicting Freeze-Thaw States in Alaska Permafrost Landscapes using climate Reanalysis and Machine Learning

Freezing and thawing (FT) processes in the soil active layer play a critical role in high-latitude ecosystems, influencing carbon cycling, hydrology, and infrastructure stability. Accurate prediction of FT states is essential for assessing permafrost dynamics under a changing climate. This study introduces a comprehensive framework for FT state prediction that leverages MERRA-2 reanalysis climate data, ensemble machine learning, and in-situ soil temperature measurements. In addition, Genetic algorithms are employed to identify the most influential geospatial features affecting FT transitions across three shallow active layer depths. The framework is validated in two distinct regions on Alaska’s North Slope (Deadhorse and Toolik Lake) over four prediction horizons (+0, +7, +30, and +90 days). Furthermore, the study evaluates the impact of four training approaches (location-specific, cross-location, location-agnostic, and depth-agnostic) on model performance, addressing the challenge of using prediction methods in real-world scenarios.