The Dunhuang murals face severe salt-induced deterioration, yet traditional salinity detection methods are often invasive and inefficient. Hyperspectral remote sensing offers a non-destructive alternative, but current inversion models lack sufficient accuracy. This study proposes a multi-level optimization framework integrating Fractional Order Differentiation (FOD), correlation analysis, and various feature selection strategies alongside Partial Least Squares Regression (PLSR) to develop a robust salinity inversion model. Experimental results demonstrate that the prediction model jointly optimized using FOD spectral transformation and LASSO feature selection achieves a cross-validated coefficient of determination (R²) of 0.908. This represents a 15.96% improvement in accuracy compared to models relying solely on FOD-transformed spectra. The findings confirm that FOD significantly enhances subtle spectral responses associated with salinity features in hyperspectral data. Furthermore, the LASSO algorithm improves model generalizability through its sparse feature selection mechanism. Collectively, combining FOD spectral transformation with judicious feature selection substantially enhances the precision and reliability of salt damage detection. This approach provides a scientific, efficient, and non-destructive technological foundation for mural preservation and restoration.