• Introduces ESCAPE, a self-calibrated, physics-informed ensemble for SAR SM retrieval. • Retrieves soil moisture and roughness from ALOS-2/PALSAR-2 without in situ data. • Embeds Oh2004 and WCM physics to stabilize ill-posed SAR inversion. • Ensemble PINNs reduce uncertainty and filter non-physical solutions. • Validated across Austria, Spain, and Japan without in-situ training labels. Soil moisture (SM) is a key variable governing land–atmosphere interactions and terrestrial hydrological processes. Recent satellite missions provide global SM observations, among which ALOS-2/PALSAR-2 offers full polarimetric L-band SAR measurements at spatial resolutions as fine as 6 m. Retrieving SM from SAR data requires disentangling scattering contributions from vegetation and surface roughness (e.g., root mean square height; H rms ). Conventional approaches typically rely on collocated in-situ measurements of SM and H rms for calibration, but such data are often sparse. In addition, the simultaneous estimation of multiple scattering parameters from limited SAR observables can lead to ill-posed inversion and unstable retrievals. To address these challenges, we propose ESCAPE (Ensemble-based Self-Calibrated Autoencoder with Physics-informed Estimation), a self-calibrating framework that integrates polarimetric ALOS-2/PALSAR-2 observations within a 10-member physics-informed neural network (PINN) ensemble. ESCAPE embeds physically based scattering models into an autoencoder architecture and estimates SM and H rms without using in-situ SM measurements as direct training targets. The ensemble strategy mitigates uncertainty arising from ill-posed physical constraints and gradient-based optimization. Spatial evaluation over Spain and Austria demonstrates robust performance, with a correlation coefficient R = 0.701, an unbiased root mean squared difference (ubRMSD) of 0.089 m 3 m −3 , and a bias of 0.006 m 3 m −3 . Temporal evaluation at an independent site in Japan yields R = 0.568, ubRMSD of 0.06 m 3 m −3 , and a bias of − 0.13 m 3 m −3 . Ensemble analysis further reveals that aggregating multiple PINN realizations improves robustness by reducing the influence of poorly converged members. Compared with conventional approaches trained directly on in-situ data, ESCAPE exhibits improved generalization across heterogeneous environments. These results highlight the potential of ESCAPE as a self-calibrated, satellite-only framework for high-resolution estimation of geophysical parameters, demonstrating the value of combining physical principles with ensemble-based deep learning for SAR remote sensing applications.
Lee et al. (Thu,) studied this question.