KAN- and CNN-Driven Snow Stratigraphy Retrieval Across Antarctic and Subarctic Environments

Accurate retrieval of snow density (ρ) and specific surface area (SSA) is critical for understanding snowpack evolution, energy balance, and hydrological processes. This study leverages the Kolmogorov–Arnold Network (KAN) relative to a conventional Convolutional Neural Network (CNN) in predicting vertical profiles of ρ and SSA across two contrasting environments: the stable, cold Antarctic plateau and the heterogeneous Canadian subarctic zones. Results demonstrate that both models accurately reproduce large-scale stratigraphic patterns, but KAN generally achieves higher depth-preserving skill for SSA in the more complex Canadian environment, while CNN yields more accurate ρ retrieval across two regions. Depth-resolved correlation patterns reveal performance fluctuations in mid-layers, attributed to compaction, vapor transport, and metamorphic processes not explicitly represented in purely data-driven models. SHAP analysis confirms the complementary sensitivity of low- (10 GHz, 19 GHz) and high-frequency (37 GHz, 89 GHz) microwave channels to deeper and near-surface snow layers, respectively. Under extreme snow conditions (high ρ or SSA), both models capture strong physical correlations between snow properties, temperature, measurement depth, and microwave signals, while also revealing saturation effects at high density and low SSA. The results highlight KAN's potential as a robust alternative to CNN for cryospheric remote sensing, especially in heterogeneous snow regimes where nonlinear feature interactions and depth-dependent relationships are critical. The demonstrated capability has implications for satellite retrieval algorithm design, avalanche hazard forecasting, and climate model parameterization of snow–atmosphere interactions.