Machine learning–based spatiotemporal mapping and risk zoning of soil salinization in an arid inland river basin: A case study of the Weigan River Basin, northwestern China

Soil salinization severely constrains agricultural sustainability and ecological security in arid inland river basins. This study establishes an integrated “inversion–evolution–zoning” framework to elucidate the spatiotemporal dynamics and risk structure of soil salinization in the Weigan River Basin from 2000 to 2024. Using multi-temporal Landsat imagery and field-measured salinity data (275 sites, 0–30 cm depth), three machine-learning models—Decision Tree, Gradient Boosting Decision Tree, and Random Forest—were compared. Random Forest achieved the highest accuracy (R² = 0.81, RMSE = 9.08 g/kg), with NDVI, ENDSI, and NDSI as the predominant contributors. Results reveal pronounced spatial heterogeneity: low salinity predominates in the upper alluvial fan, whereas high salinity concentrates in the lower fan and along the Tarim River, forming corridor-like belts and island-like patches. Although the proportion of non-saline land increased from 14% to 30% and saline land decreased from 61% to 47%, the evolutionary process was highly dynamic, characterized by bidirectional transitions among moderate, severe and saline classes, particularly during 2010–2020—a pattern driven by irrigation expansion and groundwater dynamics. Geo-informatic Tupu analysis identified 17 evolution trajectories, among which persistently stable (41.6%) and risk-reduction (32.5%) patterns dominate. Unstable areas are predominantly located in low-lying zones and desert–oasis ecotones. Integrating baseline salinity, evolutionary pathways, and transition frequency, the basin was delineated into four zones: stable (40.0%), improvement (27.4%), warning (22.6%), and high-risk (10.0%). This framework provides spatially explicit guidance for precision irrigation and groundwater control, and is transferable to other arid and semi‑arid inland river basins.