Accurately and rapidly assessing soil salinity is crucial for land quality evaluation and agricultural development. Hyperspectral remote sensing， as an effective monitoring technology， offers new avenues. Optimizing hyperspectral data processing to enhance features is critical for accurately estimating soil parameters. However， the impact mechanisms of different spectral combination treatments on soil salinity estimation need further study. This study aims to investigate the potential of hyperspectral data for estimating soil salinity under nonlinear transformation and fractional derivative combination treatments. Based on 60 typical soil samples from the Yulin area， the spectral characteristics of saline soils under different spectral transformation combinations were analyzed. The Competitive Adaptive Reweighted Sampling method （CARS） and Successive Projections Algorithm （SPA） were used to select characteristic bands as input variables. Partial Least Squares Regression （PLSR） and Random Forest （RF） models were established to compare and analyze the estimation capabilities of different spectral processing and modeling strategies for soil salinity. The results indicate that fractional derivatives， compared to integer derivatives， better reflect the absorption characteristics of saline soil spectra. Reciprocal and logarithmic transformations effectively enhance the correlation between spectral data and soil salinity information， especially in the 1905–2078 nm range. The PLSR model generally outperforms the RF model， with the CARS-PLSR achieving optimal modeling accuracy （R2=0.966）. These findings can provide a theoretical basis for the implementation of saline soil prediction and precision agriculture.