AgriROSE-L 2025 Case Study: Soil Moisture Estimation from L-Band PolSAR Data with Physical Autoencoders

Abstract

High-resolution soil moisture estimation is an active research area with important applications in hydrology and agriculture. In this study, we investigate the estimation from fully polarimetric L-band data using an F-SAR dataset recently acquired by the German Aerospace Center (DLR) during the AgriROSE-L 2025 airborne campaign. We compare a purely supervised machine learning (ML) model with a physical autoencoder (AE) that combines an ML encoder with a physical decoder into a single architecture. The autoencoder demonstrates improved robustness and prediction accuracy, particularly in scenarios with limited or no training data, thanks to the physical modeling. In addition, the autoencoder provides a fully explainable physical latent space and predicts additional geophysical parameters useful for target characterization. Combining physical modeling with machine learning is a promising approach for estimating soil moisture, helping to address the limitations of small labeled datasets. The source code is available at https://github.com/nbasargin/nb2026eusar.