A sequential multi-model machine learning framework integrating environmental and management factors for mapping exchangeable potassium in rice soils

Exchangeable potassium is a key determinant of rice productivity, yet its spatial distribution is difficult to predict because it is influenced by both pedogenic factors and intensive agricultural management. This study developed a staged multi-model machine learning framework to map exchangeable potassium in rice systems by progressively integrating terrain, spectral, geological, soil, climate, and management predictors. A total of 318 soil samples were collected across irrigated and rainfed rice fields, and exchangeable potassium was measured using ammonium acetate extraction. Six machine learning algorithms were evaluated across five feature-integration stages. Model performance improved consistently as predictors were added, with R2 increasing from 0.38 to 0.68 under random validation (0.16–0.41 under spatial block cross-validation). Variable importance and SHAP analyses showed a progressive shift in the most influential predictor groups, from terrain and hydrological factors to geological and soil properties, and finally to management variables. Irrigation, cropping intensity, and planting schedule were the most influential predictors in the final model. Spectral indices showed weaker relationships than soil and management variables. These findings demonstrate that integrating environmental and management predictors is essential for accurate potassium mapping and provide a framework whose transferability to other regions and periods remains to be tested.