Distilling Geospatial Foundation Model semantics into interpretable Sentinel-1/2 proxies for few-shot and cross-year crop type classification

Geospatial Foundation Models (GFMs) show strong generalization across Earth observation tasks, but their direct use in agricultural monitoring can be constrained when foundation-model embedding products are not available for a given year or observation window. We propose a physically grounded semantic distillation framework that transfers GFM semantics to standard satellite features. We hypothesize that selected GFM embedding dimensions can be approximated from interpretable physical proxies. Using Lasso regression, we reconstruct selected AlphaEarth Foundations (AEF) embedding dimensions from a sparse set of Sentinel-1/2 temporal features and identify the most influential proxy variables. On maize–soybean mapping in the U.S. Corn Belt, the distilled features closely approximate AEF’s discriminative performance while achieving high reconstruction fidelity. In cross-year transfer without target-year supervision, the distilled features consistently outperform the multi-modal satellite baseline across labeled-sample budgets; under the lowest sample budget, the proposed method achieves an overall accuracy of 93.33%, compared with 83.90% for the multi-modal satellite baseline. These results indicate that lightweight and interpretable proxy models can retain a substantial portion of the cross-year generalization advantage of foundation-model representations, providing a practical alternative for few-shot classification and cross-year monitoring.