Identification of Soil Texture Classes Under Vegetation Cover Based on Sentinel-2 Data With SVM and SHAP Techniques

Understanding the spatial variability of soil texture classes is essential for agricultural management and environment sustainability. Sentinel-2 data offer valuable vegetation information as proxies properties inference. However, applications them in classification are still limited. This study investigated usefulness predicting class using an interpretable machine learning (ML) strategy. Specifically, multitemporal images were used to get exhaustive cover information. Basic digital elevation map (DEM) derivatives stratum extracted. Three support vector machines with different input parameters (purely DEM stratum, purely Sentinel-2, plus stratum) developed. Moreover, order improve transparency black box ML models, novel SHapley Additive exPlanations (SHAP) method was applied interpret outputs analyze importance individual variables. Results showed that model all variables provided desirable performance overall accuracy 0.8435, F1-score 0.835, kappa statistic 0.7642, precision 0.8388, recall 0.8355, area under curve 0.9451. The performed much better than solely stratum. contributions explain about 17%, 41%, 28% sandy, loamy clayey soils, respectively. SHAP visualized decision process indicated elevation, red-edge factors critical classes. offered much-needed insights into mapping ML-assisted tasks.