Local-Scale Probabilistic Mapping of Spring Thaw Progression in Interior Alaska Using Satellite Radar and Physics-Augmented Deep Learning

Seasonal freeze-thaw (FT) transitions in the Northern high latitudes (NHL) exert a fundamental control on permafrost stability, vegetation phenology and productivity, soil respiration, surface and shallow subsurface hydrology, and carbon sink activity, while exhibiting complex variations in space and time which are not adequately captured by existing operational satellite FT products. To improve our understanding of local-scale (∼100 m) environmental changes over the NHL, probabilistic soil FT dynamics were mapped from Sentinel-1 radar backscatter images for Interior Alaska using a physics-augmented deep learning (DL) approach. A simplified radar backscatter model was developed and calibrated for simulating Sentinel-1 images over a variety of land cover and terrain conditions, and different component (e.g., vegetation and soil) FT states. The simulated backscatter images were used for training a U-Net DL model along with other ancillary geospatial data. The resulting probabilistic thaw maps effectively captured heterogeneous FT patterns during the spring thaw progression, demonstrating higher accuracy (91.0% to 92.6%) against in situ field measurements than an alternative threshold-based classification method (82.4% to 88.6%). Logistic regression analysis further showed a strong relationship between the DL estimated soil thaw probability and soil temperature (R2= 0.68), suggesting the potential of deriving unfrozen soil water content from FT probability estimates. Our findings enable enhanced delineation of FT patterns in highly heterogeneous boreal landscapes for improved understanding of climate driven FT controls on NHL ecosystems.