Pan-Tropical Land Surface Modulation of CYGNSS GNSS-R DDM Coherence Based on Explainable Machine Learning

Quantifying how land surface properties modulate Global Navigation Satellite System Reflectometry (GNSS-R) coherence is essential for assessing the performance and limitations of GNSS-R land applications. However, these modulations remain poorly characterized at the pan-tropical scale. We develop an explainable machine-learning framework to quantify the pan-tropical land surface modulation of GNSS-R delay–Doppler map (DDM) coherence, including topography, surface water, soil moisture, and vegetation. We represent the coherence with a proxy derived from Cyclone Global Navigation Satellite System (CYGNSS) observations and model it with a Light Gradient Boosting Machine (LightGBM). Under spatial cross-validation, the model explains about 60% of the spatial variance in the coherence proxy and demonstrates reasonable generalization performance under spatial and regional holdout validation. Shapley additive explanations (SHAP) and accumulated local effects (ALE) show that GNSS-R DDM coherence is enhanced over smooth, wet, and moderately vegetated surfaces, but reduced over rough, steep, and densely vegetated terrain. Topographic factors, particularly terrain roughness, emerge as the dominant controls, while surface water occurrence and soil moisture are associated with enhanced coherence across identifiable empirical transition ranges. A dominant-control analysis further reveals distinct regional patterns in which topography, surface water, soil moisture, or vegetation serves as the leading control. Topography dominates about 68% of land, surface water and vegetation each about 14%, and soil moisture about 4%. These results provide a physically meaningful account of where and why land DDM coherence varies, together with an interpretable empirical basis for understanding how land-surface conditions influence DDM coherence and for informing the design of future GNSS-R land applications.