IGWRABoost

Traditional global machine learning models face certain limitations, such as challenges in capturing the spatial non-stationarity of soil texture and insufficient interpretability. To address these issues, this study proposes a modified interpretability-oriented geographically weighted regression adaptive boosting (IGWRABoost) method and applies it to the spatial prediction of soil texture in the Yongqiao District of Suzhou City in Anhui Province, China, a major wheat-producing area on the Huang-Huai-Hai Plain. AdaBoost submodels were constructed by incorporating geographical weighting in local neighborhoods. Subsequently, multiple local models underwent Gaussian weighted fusion during the prediction phase to capture the spatial non-stationary relationships between texture and environmental factors. Results indicate that the IGWRABoost method achieved the highest R² and the lowest root mean square error (RMSE) and mean absolute error (MAE) values in soil texture prediction. SHAP-based global and local interpretations further revealed that the temperature vegetation dryness index (TVDI) was the joint dominant factor for sand and clay, normalized difference vegetation index (NDVI) prominently contributed to the prediction of sand content, and the effects and strengths of the various covariates differed significantly across different spatial locations. The IGWRABoost-SHAP framework not only improved the accuracy of soil texture prediction but also enhanced model interpretability. Therefore, this study serves as a methodological reference for digital soil texture mapping and refined management of agricultural water resources in the Huang-Huai-Hai Plain and other similar agricultural regions.