Explaining the spatial heterogeneity in soil-rice yield interactions via a hybrid interpretable machine learning and...

Rice (Oryza sativa L.) constitutes the main food for more than 50% of the global population, and its production significantly influences food security and sustainable development. Using various high-resolution soil and remote sensing datasets, a hybrid random forest (RF)-Shapley additive explanations (SHAP)-bivariate local spatial autocorrelation (BI-LISA) framework was proposed and applied to the city of Jiaxing, which is located in the lower reaches of the Yangtze River Delta, to elucidate the spatially heterogeneous relationships between environmental variables (especially soil properties) and rice yield. The results revealed that the soil conditions in the study area were generally suitable, with mean soil pH, organic matter content, bulk density and cation exchange capacity values of 6.47, 27.98 g/kg, 1.17 g/cm³ and 18.01 cmol(+)/kg, respectively. The mean rice yield was 8437 kg/ha, and yields were higher in the eastern, central and western regions than in other regions. RF and SHAP results revealed that soil physical and chemical factors contributed the most to rice yields, with a total relative importance of approximately 66.8%. BI-LISA revealed the following spatial interactions between soil properties and rice yields: 1) in the southwest, high soil bulkiness, the presence of sandy and acidic soil, and low nutrient levels were the main influencing factors; 2) in the north, low soil bulkiness, the presence of clay and acidic soil, and imbalanced nutrient levels were the predominant influencing factors; and 3) in the eastern coastal region, the major influencing factors were low soil bulkiness, the presence of sandy and alkaline soil, and low nutrient levels. The proposed hybrid framework can offer a valuable reference for precision agriculture, facilitating spatially explicit nutrient management to mitigate soil-specific constraints and increase yields.