Soil salinization and alkalization pose major threats to agricultural sustainability and ecological security in arid and semi-arid regions. Under the combined influence of climate change and human activities, the spatiotemporal evolution of cropland soil salinity content (SSC) and its driving mechanisms remain inadequately understood, particularly regarding the nonlinear threshold responses to multiple factors and causal pathways. Therefore, this study utilized field soil measurements and remote sensing imagery to estimate cropland SSC in the Manas River Basin-a typical irrigated agricultural region in China-from 2013 to 2022. A Random Forest (RF) model was constructed using features optimized by the Normalized Difference Phenology Index (NDPI), and SHapley Additive exPlanations (SHAP) combined with Piecewise Structural Equation Modeling (PSEM) were applied to quantitatively assess the effects of climate, drought, topography, and other factors on SSC. The results indicate that: (1) The RF model incorporating NDPI features achieved reliable predictions (R2 = 0.854, RMSE = 1.649 g/ kg, RPD = 2.284). Between 2013 and 2022, overall cropland SSC in the basin showed improvement, with the area of non-salinized soils increasing by 18%, exhibiting a spatial heterogeneity pattern of "low in oasis cores and high at the edges." (2) Human activities exert dual effects: cropland retirement facilitates the withdrawal of highsalinity areas, whereas unreasonable cropland expansion can exacerbate local salinization risks. (3) Topography and hydrothermal conditions jointly dominate the spatial differentiation of SSC. Potential evapotranspiration (PET) and soil moisture (SM) have positive direct effects on SSC (0.144 and 0.077), while precipitation (PRE) and the Standardized Precipitation Evapotranspiration Index (SPEI) show negative direct effects (-0.287 and -0.096). Elevation exhibits the largest direct effect on SSC (0.329) and generates significant indirect effects (-0.255) by modulating hydroclimatic processes (PRE, SM, and SPEI), with all drivers displaying nonlinear threshold responses. These findings provide scientific evidence for monitoring, managing, and mitigating cropland salinization in arid regions.