Soil salinization is a major cause of land degradation and ecological deterioration. Traditional monitoring methods are spatially limited and inefficient. This study explores the use of multi-source remote sensing to improve the accuracy and timeliness of soil salinity inversion, providing support for precise management of saline-alkali land.The research was conducted in Lihe Township, Ningxia, China. We integrated Sentinel-1 SAR, Sentinel-2 multispectral data (including reflectance and derived salinity, vegetation, and texture indices), and SRTM DEM data. Feature selection was performed using Recursive Feature Elimination (RFE), Partial Least Squares (PLS), and Out-of-Bag (OOB) estimation methods. Soil salinity inversion models were developed for three maize growth stages (D1–D3) and a bare soil period (D4) using Random Forest (RF), Support Vector Regression (SVR), and Back-Propagation Neural Network (BPNN) algorithms.The RF model achieved the best and most stable performance across all periods, with an average coefficient of determination (R 2) > 0.77 on the test set. RFE yielded the most robust feature subsets. Sensitive features evolved with crop phenology: from vegetation and salinity indices in early stages, to topographic and texture features in middle stage, and finally to original bands and specific salinity indices in late and bare soil periods. Spatially, salinity was highest in the eastern and northern regions, with the west remaining stable. Temporally, soil salinity exhibited “summer-autumn leaching and winter-spring accumulation”.The multi-period, multi-model inversion framework developed in this study offers a reliable approach for stage-specific monitoring and management of regional soil salinization.