Optimizing Machine Learning Models with Multi-Source Variables for Soil Salinity Prediction in an Arid Oasis: Implications for Spatial Management

Soil salinization disrupts ecosystem material and energy cycling, leading to agricultural resource wastage, ecological degradation, and economic losses. This study developed soil salinity prediction models using 1247 samples from China’s Kashgar Oasis through five machine learning methods—Ordinary Least Squares (OLS), Backpropagation Neural Network (BP), Random Forest (RF), Support Vector Machine (SVM), and Extreme Learning Machine (ELM)—integrated with multi-source environmental variables. Dominant ions (Cl–, Na+, $${\text{SO}}_{4}^{{2 - }}$$ ) classified regional salinity as chloride-sulfate type, exhibiting significant spatial heterogeneity (variation coefficients: 0.37–4.04). The SVM model achieved optimal performance (R2 = 0.6885, lowest RMSE/MAE/MBE), demonstrating robust generalization and anti-overfitting capabilities via its Gaussian kernel. Spatial predictions from RF, BP, and OLS exhibited scattering, whereas SVM and ELM generated coherent patchy-zonal distributions aligned with field measurements. Variable optimization (EVI, NDVI, EC, TDS, DEM, TL, MAP) enhanced SVM’s RMSE by 23%, highlighting its superiority for soil salinity mapping in arid oasis ecosystems. The study provides critical insights into nonlinear drivers of salinization and informs targeted land management strategies.