Prediction of Soil Salinity Parameters in the Songnen Plain Using FOD Processing and Machine Learning from Measured Hyperspectral Reflectance Under Different Surface Conditions

Soil salinization severely restricts ecosystem stability and the sustainable development of agricultural productivity. However, current understanding of the spectral–salinity quantitative relationships under the influence of surface cracking still remains limited. To address this gap, this study collected hyperspectral reflectance data (350–2500 nm) from salt-affected soil in both cracked and uncracked surface conditions across the Songnen Plain, and applied fractional-order differentiation (FOD) processing with orders ranging from 0 to 2 and a step size of 0.1. Based on this, 14 types of FOD spectral indices were constructed, incorporating one-dimensional (1D), two-dimensional (2D), and three-dimensional (3D) structures. For each spectral index, the optimal fractional order and corresponding band combinations were first selected through Pearson correlation analysis for pH and EC under both surface conditions; subsequently, feature selection was performed using XGBoost-SHAP explainable analysis among the 14 optimal indices across different dimensions. Furthermore, the predictive performance of four modeling methods, including partial least squares regression (PLSR), Gaussian process regression (GPR), support vector regression (SVR), and random forest regression (RFR), was evaluated. The results showed that FOD transformations significantly enhanced correlations with EC and pH compared to raw reflectance. All prediction models demonstrated higher prediction accuracy under cracked surface conditions than uncracked surface conditions, indicating that desiccation cracks positively modulate spectral signals to enhance salinity information expression. Across different surface states, model performance generally followed the ranking: PLSR > GPR > SVR > RFR, with PLSR achieving the best predictions for EC and pH under cracked surfaces (R2 of 0.88 and 0.76, RMSE of 0.29 dS/m and 0.35). This study not only deepens the understanding of fractional-order spectral response mechanisms in saline–alkali soils but also provides methodological support for regional monitoring of soil salinization.