Soil moisture plays a vital role in the water-energy cycle, climate dynamics, and agricultural practices. Conventional approaches for estimating surface soil moisture (SSM) with single-frequency Synthetic Aperture Radar (SAR) often face challenges in accuracy and adaptability due to differences in sensor properties. This study proposes a novel cross-frequency calibration framework for SSM retrieval using non-coincident multi-frequency SAR data, specifically leveraging SAOCOM (L-band) and RADARSAT Constellation Mission (RCM) (C-band) datasets. A domain adaptation approach is employed to transfer knowledge from SAOCOM to RCM, enabling robust SSM estimation without requiring coincident acquisitions. Ensemble Learning Regression (ELR) and Gaussian Process Regression (GPR) models were developed and optimized using Bayesian hyperparameter tuning and k-fold cross-validation. The adapted models achieved RMSE of 0.039 m³/m³ (ELR) and 0.047 m³/m³ (GPR), with corresponding coefficients of determination R2 = 0.81 and R2 = 0.72, respectively. Minimal bias (0.001) and strong agreement between predicted and observed values were confirmed through Bland-Altman and Q-Q plots. This framework enhances the scalability and generalizability of SSM retrieval, offering practical applications in drought monitoring, flood prediction, and irrigation management. The study highlights the effectiveness of machine learning and domain adaptation techniques in advancing large-scale soil moisture mapping.
Ondieki et al. (Fri,) studied this question.