Estimation of major soil water-soluble salt ions from Sentinel-2 imagery: Integrating spectral and texture features

Soil water-soluble salt ions are key indicators for diagnosing both the type and severity of soil salinization and represent major constraints on crop growth. In the Alar Irrigation District of Xinjiang, severe soil salinization poses a substantial threat to agricultural sustainability, highlighting the need for fine-scale monitoring with high-resolution Sentinel-2 imagery. Although integrating spectral and texture features has been demonstrated to enhance soil property estimation, the effectiveness in monitoring soil water-soluble salt ions remains insufficiently validated. This study aimed to determine the optimal window size for texture feature extraction using the Gray-Level Co-occurrence Matrix (GLCM) method and to integrate spectral and texture features to improve the estimation accuracy of topsoil water-soluble salt ions (0–20 cm). Key features were identified using the Boruta algorithm, after which three machine learning models—Gradient Boosted Regression Tree (GBRT), Extreme Gradient Boosting (XGBoost), and Random Forest (RF)—were trained and evaluated. The results indicate that a 5 × 5 pixel window size yields the highest prediction accuracy, with contrast, entropy, auto-correlation, and dissimilarity identified as the most important texture features. Integrating spectral and texture features improved model performance by 4.71 % to 67.55 %. The RF model exhibited the highest predictive performance for Na + , Cl − , HCO 3 − , K + , SO 4 2− , and Mg 2+ , whereas GBRT performed best for Ca 2+ . Overall, the RF model demonstrated superior accuracy ( R 2 = 0.48–0.93, RMSE = 1.62–5.39 g⋅kg −1 , RPD = 1.39–3.26). Thus, this study establishes a comprehensive framework for monitoring soil salinity, providing valuable theoretical and practical insights for precision agriculture and sustainable land resource management.