AI Insights for Wind Speed Retrieval From GNSS Reflectometry

Artificial intelligence (AI) models developed for Global Navigation Satellite System Reflectometry (GNSS-R) observations have demonstrated competitive performance in estimating geophysical parameters, especially ocean surface wind speeds. However, the transition from transparent physical scattering models to complex deep learning architectures raises concerns regarding reduced model transparency and trust. Understanding the decision-making processes of these “black-box” models is essential for assessing model behavior, detecting anomalies, and ensuring reliability in AI-based Earth observations. In this study, we investigate the role of explainable artificial intelligence (XAI) in addressing the transparency gap for hybrid deep learning models designed for GNSS-R observations. Focusing on ocean wind speed retrieval as a well-characterized benchmark, our study is structured around three primary objectives: first, assessing the robustness and efficiency of XAI explainers, second, interpreting a benchmark hybrid model trained using a manually selected feature set with Shapley additive explanations (SHAP) and gradient-weighted class activation mapping (Grad-CAM), which provide quantitative branchwise attribution and qualitative spatial saliency, and third, proposing an XAI-based feature selection pipeline that leverages SHAP-based ranking and exclusion, comparing its efficacy against conventional statistical methods. The results demonstrate that SHAP is effective not only for model interpretation but also for supporting computationally efficient feature selection and model debugging. Meanwhile, Grad-CAM offers complementary spatial interpretability by highlighting salient regions in the delay–Doppler map inputs. This study demonstrated the potential of integrating XAI as a diagnostic and validation tool into the model development cycle, enabling more transparent, robust, and trustworthy AI models for upcoming GNSS-R missions and future applications.