Soil salinization is a critical environmental issue that threatens global ecological security and sustainable agricultural development. This study focuses on the Ebinur Lake Wetland Nature Reserve in Xinjiang, where 51 spectral indices were derived from Landsat-9 multispectral imagery. By integrating the Bootstrap Soft Shrinkage (BOSS) feature selection algorithm with the Particle Swarm Optimization-Random Forest (PSO-RF) model, we achieved high-precision prediction and spatial mapping of total soil salinity (TSS) and eight major water-soluble salt ions (Na++K+, Ca2+, Mg2+, Cl-, SO42-, HCO3-, and CO32-). The BOSS algorithm was employed to filter high-dimensional spectral features, effectively reducing redundant information while preserving key sensitive bands. Meanwhile, the PSO-RF model leveraged a nonlinear ensemble learning mechanism and global parameter optimization to capture the complex interactions between spectral features and salt ion concentrations. The results demonstrated that the PSO-RF model achieved a coefficient of determination (R2) exceeding 0.6 for TSS and all salt ions, with the highest prediction accuracy observed for TSS (R2 = 0.78), Mg2+ (R2 = 0.71), and SO42-(R2 = 0.71). Compared with the traditional Random Forest (RF) model, the PSO-RF model improved prediction accuracy by 5.4%-19.8%. Spatial mapping revealed that high salinity areas were concentrated along the edges of the Ebinur Lake wetland, with Na++K+ and Cl-distributions closely matching total salinity patterns. In contrast, HCO3-and CO32-exhibited characteristics of alkaline salinization in the northern region. The proposed BOSS-PSO-RF framework offers an efficient technical solution for soil salinization monitoring in arid regions. The multi-ion synergistic mapping results provide a scientific basis for zonal management and ecological restoration.