Soil salinity accumulation in arid regions severely constrains agricultural productivity, highlighting the critical need for a robust remote sensing-based monitoring framework. However, multispectral remote sensing frequently suffers from limited sensitivity when isolating salinity signals within complex surface conditions. Focusing on the Aral Irrigation District in Xinjiang, this study systematically evaluated the effects of first- and second-order derivative (FD and SD) preprocessing on soil salinity inversion accuracy, and further analyzed the efficacy of two-dimensional (2D) and three-dimensional (3D) spectral indices constructed via discrete gradient transformations. The results demonstrate that FD preprocessing notably amplifies spectral slope variations and suppresses background noise. This transformation strengthens the physical contrast between the visible and shortwave infrared bands, which are associated with surface salt crystallization and moisture absorption, respectively, thereby substantially enhancing the sensitivity of multispectral predictors. Comparative analysis reveals that while FD-enhanced 3D indices consistently outperform their 2D counterparts, their band combination strategies must be rigorously optimized under derivative-based enhancement to achieve an optimal balance between index sensitivity and noise suppression. Feature optimization utilizing the Boruta algorithm effectively eliminated redundant variables, yielding an 8.54% improvement in predictive accuracy. The Random Forest (RF) model integrating FD preprocessing and the Boruta algorithm achieved the highest predictive performance (R2 = 0.75, RMSE = 0.86 dS & sdot;m- 1, RPD = 2.00, and RPIQ = 2.20) and demonstrated robust temporal stability during independent cross-temporal validation. This optimized framework enabled the high-accuracy spatial mapping of soil salinity grades, establishing a reliable monitoring approach to support sustainable regional land management.