Machine learning-driven multi-sensor and cross-frequency SAR fusion for high-resolution soil moisture retrieval: integrating LSTM networks, cross-sensor calibration and multi-source data synergy

Surface soil moisture (SSM) is significant for hydrological, agricultural, and climate system , yet retrieval from remote sensing is constrained by vegetation-induced signal saturation, neglect of temporal persistence, and the absence of firm cross-frequency calibration for non-coincident Synthetic Aperture Radar (SAR) missions. This thesis develops and evaluates a machine-learning- driven framework which integrates multi-frequency SAR (SAOCOM L-band, Sentinel-1, and RCM C-band), optical (Sentinel-2 NDVI), and meteorological measurements to address these constraints and enable high-resolution, scalable SSM retrieval. The methodology is structured around three pillars. First, feature-level fusion of Sentinel-1/2 and meteorological data in Italian croplands established a baseline (Random Forest, R² ≈ 0.86, RMSE=4.36%) while exposing limitations of static models. Second, Bayesian-optimized LSTM models applied in Argentina explicitly exploited temporal memory and multi-frequency synergy, achieving R² ≈ 0.84 (RMSE ≈ 0.022 m³/m³). In this context, non-SAR contribution is defined as the relative explanatory power of ancillary optical and meteorological variables specifically precipitation, air temperature, and NDVI quantified through feature-ablation and importance analyses within the trained LSTM framework. These analyses indicated that precipitation accounted for approximately 11.9% of the total model contribution, temperature for 9.5%, and NDVI for 3.6%, complementing the SAR-derived information. Third, cross-frequency calibration between non-coincident SAOCOM and RCM observations in Canada (Ensemble Learning Regression (ELR), R ≈ 0.72) and kernel-based domain adaptation in Kenya (ensemble R² ≈ 0.44) demonstrated that asynchronous SAR missions can be harmonized and transferred across agro-ecological zones. Results confirmed that dual-frequency SAR fusion improves retrieval accuracy by ~0.20 R relative to single-frequency models, LSTMs were able to capture hysteresis in soil moisture development, and ancillary optical–meteorological inputs reduced errors by >11.7%. Aside from enhanced performance, the approach delineated a mission-agnostic pipeline that can generate soil moisture map with spatial resolution of 10–30 m for irrigation planning, flood/drought monitoring, and crop stress analysis. By bridging physical scattering physics with interpretable machine learning (ablation) and cross- validation across Italy, Argentina, Canada, and Kenya, this thesis transposes SSM retrieval from sensor-domain science to operational and transferable solutions.