Spaceborne GNSS reflectometry for soil moisture: from physics-based modeling to hybrid deep learning

Soil moisture is an Essential Climate Variable (ECV) that plays a critical role in the global water and energy cycles. While L-band radiometry has established a standard for monitoring soil water content, current missions are limited by coarse spatial resolution. Global Navigation Satellite System Reflectometry (GNSS-R) has emerged as a disruptive remote sensing technique capable of bridging this gap by providing high-spatiotemporal resolution data using signals of opportunity. However, the operational retrieval of soil moisture from GNSS-R data currently faces a fundamental dilemma: physics-based models offer interpretability but struggle with surface heterogeneity, while data-driven approaches provide high accuracy but lack physical domain knowledge consistency. This thesis addresses this challenge by developing a unified benchmarking framework that systematically evaluates and bridges the gap between physics-based and data-driven retrieval strategies. The research begins by establishing a rigorous comparison between a semi-empirical (SE) inversion model and an Artificial Neural Network (ANN) approach within the context of the European Space Agency’s upcoming HydroGNSS mission. To ensure robust validation, a stratified sampling scheme is implemented across diverse climate zones and land-cover types, leveraging data from the NASA’s Cyclone GNSS (CYGNSS) constellation alongside auxiliary inputs from multiple sources. Building upon this benchmark, the thesis proposes an advanced Hybrid Deep Learning (HDL) architecture. This novel model integrates the feature-extraction power of Convolutional Neural Networks (CNN) with the regression capabilities of standard neural networks, leveraging physical domain knowledge to guide the learning process. The results demonstrate a clear hierarchy in performance. The ANN-based approach achieves an overall Root-Mean-Square Error (RMSE) of approximately 0.047 m³/m³ (correlation R ≈ 0.90), significantly outperforming the semi-empirical model across most regimes, particularly under relaxed data-quality filtering. The proposed HDL model yields the highest accuracy, reducing the RMSE to 0.038 m³/m³ (R ≈ 0.93). Notably, the inclusion of climate-specific stratification and ancillary inputs results in error reductions of 44–47% in complex scattering environments. These findings indicate that while semi-empirical models remain valuable for their interpretability in data-sparse regions, deep learning architectures offer the flexibility and precision required for next-generation Earth observation. Consequently, a hybrid retrieval strategy is recommended to maximize the exploitation of future datasets. This work provides a foundational roadmap for the HydroGNSS mission, suggesting that future improvements should focus on expanding training coverage in underrepresented high-latitude regions to fully realize the potential of spaceborne GNSS-R hydrology.