【Objective】Under the global context of climate change and anthropogenic impacts, soil salinization has become increasingly severe. However, traditional salinization monitoring suffers from being time-consuming, labor-intensive, and costly. Hyperspectral-based salinization monitoring often relies on single mathematical transformations and one-dimensional spectral information, while also exhibiting poor model interpretability. Research utilizing combined spectral transformations to construct spectral indices for salinization estimation urgently requires in-depth exploration. Thus, this study aims to fully exploit spectral information, enhance data sensitivity, and establish a high-precision, interpretable salinization inversion model based on spectral indices.【Method】Dongying City was selected as the study area, where hyperspectral datasets were collected through field surveys, and soil samples were analyzed in the laboratory for salinity determination. The samples were divided into training and testing sets in a 7:3 ratio based on salinity gradients. Spectral data were preprocessed using Savitzky-Golay (S-G) filtering and Multiplicative Scatter Correction (MSC). Four spectral transformations were applied: Reflectance (R), Reciprocal (1/R), Logarithm of Reciprocal (log(1/R)), and Continuum Removal (CR). The Fractional Order Derivative (FOD) transformation was subsequently performed on each form. Ten types of two-dimensional spectral indices were constructed from the combined transformed data at each derivative order. Optimal band combinations and differential orders were identified by assessing correlation coefficients with soil salt content (SSC). Using these spectral indices as features and measured salinity as the dependent variable, four machine learning models—Partial Least Squares Regression (PLSR), Convolutional Neural Network (CNN), eXtreme Gradient Boosting (XGBoost), and Support Vector Machine (SVM)—were constructed. The hyperparameters of all models were optimized using the Bayesian Optimization (BO) algorithm, which iteratively fitted a probabilistic surrogate model to guide the search for hyperparameters that minimize cross-validation error. Each model was trained and tuned via ten-fold cross-validation. Performance was evaluated using the Coefficient of Determination (R2), Root Mean Square Error (RMSE), and Residual Prediction Deviation (RPD). The best-performing model was further interpreted using SHapley Additive exPlanations (SHAP) to identify influential spectral features. 【Result】Results demonstrated that:(1) FOD effectively enhances spectral sensitivity by highlighting gradient information during spectral curve variations; (2) Mathematical transformations combined with FOD significantly improve correlations between spectral data and SSC; (3) The 2-order NDI index after CR treatment achieved the highest absolute correlation coefficient (|r|=0.91) with SSC; (4) The CR-FOD-XGBoost model delivered optimal accuracy (testing set: R2=0.94, RMSE=0.85 g·kg?1, RPD=4.33); (5) In the optimal model, GDI1 contributed most significantly while DI clusters adjacent to zero contributed minimally. 【Conclusion】Collectively, this study demonstrates that combining spectral transformations to construct indices with Bayesian-optimized XGBoost modeling effectively improves soil salinity inversion accuracy, providing scientific foundations for salinization control and ecological sustainability. Future research should focus on enhancing spectral sensitivity responsiveness to further improve model performance, thereby advancing theoretical frameworks for sustainable land-use and environmental conservation strategies.