Impact of multi-source remote sensing fusion schemes on soil salinity inversion accuracy in tidal flat environments

Abstract

Multi-source remote sensing fusion is widely assumed to improve soil salinity (SSC) inversion accuracy. However, its actual effectiveness in complex tidal flat environments remains insufficiently validated. This study evaluated five fusion schemes (single-source Sentinel-2, single-source GF-2, S2+S1, GF-2+S1, and all-source fusion) in the Yangkou tidal flat, Laizhou Bay, using 139 field samples and five machine learning models (RF, XGBoost, CatBoost, LightGBM, Stacking). A three-step feature selection pipeline reduced 103 candidate features to 7 optimal subsets (max VIF: 155.8 → 4.2). Results show that the single-source Sentinel-2 Random Forest model achieved the best performance (R²=0.78, RMSE=1.95 g/kg, RPD=2.14), while no multi-source scheme surpassed it. SAR polarization features exhibited near-zero Spearman correlations with SSC, and GF-2 texture features lost effectiveness after resampling, together explaining fusion failure. These findings suggest that data source applicability assessment should precede any fusion scheme design.