An Explainable AutoML Framework for Soil Salinity Mapping Using Multi-Source Data in an Arid Irrigated District, China

Soil salinization limits agricultural production and the management of soil and water resources in arid irrigated regions. Regional prediction remains difficult, while soil salinity is influenced by hydrology, topography, and land cover. This study developed an explainable automated machine learning (AutoML) framework to map surface soil salinity in the Qingtongxia Irrigation District, Ningxia, China. The analysis combined 108 soil samples (0–10 cm), collected in March-April 2024, with 36 candidate input features from Sentinel-2 spectral indices, topography, soil texture, groundwater depth, and climate. Pearson correlation analysis and recursive feature elimination with cross-validation (RFECV) selected 10 key input features. Among the tested models, AutoML achieved the best validation performance, with R2 = 0.78, MAE = 1.94 g/kg, and RMSE = 2.65 g/kg. The resulting 30 m prediction map captured broad regional patterns of soil salinity, with predicted values from 0.34 to 22.91 g/kg and higher salinity in northern and northeastern areas. Shapley additive explanations (SHAP) analysis highlighted brightness index, groundwater depth, clay content, elevation, and salinity index I as influential features. These findings suggest that explainable AutoML can help identify regional salinity risk and guide future sampling and model refinement.