Interpretable Salinization Estimation Model for Dongying City Based on Integrated Multi-Dimensional Spectral Indices with XGBoos

The construction of three-dimensional indices and combinations of different dimensional indices based on hyperspectral data for salinization estimation is rarely reported, and the effects of inversions under different transformations need further exploration. In this study, Dongying City was chosen as the research area. Hyperspectral data from sample points were preprocessed using Savitzky-Golay (S-G) filtering and Multiplicative Scatter Correction (MSC), followed by conventional transformations (log(1\/R), Continuum Removal (CR), Standard Normal Variate (SNV)), fractional-order derivative (FOD), and Continuous Wavelet Transform (CWT). Additionally, mixed-order differential and mixed-wavelet transformations were constructed. The Competitive Adaptive Reweighted Sampling (CARS) algorithm was employed to select sensitive bands, and two-dimensional and three-dimensional spectral indices were created. The optimal band combinations were selected based on their correlation with soil salinity. Two-dimensional, three-dimensional, and combined two- and three-dimensional indices were used as feature variables to construct models, including XGBoost (Extreme Gradient Boosting), Partial Least Squares Regression (PLSR), and Convolutional Neural Network (CNN), to predict salinity content. The SHapley Additive exPlanations (SHAP) method was used to perform interpretability analysis on the best model. The results indicated the following: (1) The CARS algorithm effectively selects [1]sensitive bands, and the mixed-order differential and mixed-wavelet transformations can significantly enhance band sensitivity. (2) Compared to two-dimensional spectral indices, three-dimensional spectral indices can effectively improve inversion accuracy. In the PLSR, CNN, and XGBoost models, R\u00b2 of the prediction set improved by 0.11, 0.16, and 0.02, respectively, RMSE decreased by 0.44, 0.82, and 0.72, and RPD increased by 0.86, 0.79, and 1.31. Except for the CNN model, the combination of two-dimensional and three-dimensional spectral indices further improved inversion accuracy, with R2 improving by 0.01 and 0.01, RMSE decreasing by 0.1 and 0.16, and RPD increasing by 0.31 and 1.21 in the PLSR and XGBoost models. (3) Wavelet and fractional-order derivative transformations exhibited high accuracy under different models and effectively enhanced model precision and generalization. (4) XGBoost outperformed PLSR and CNN models in prediction performance. In the prediction set, compared to PLSR and CNN, R2 improved by 0.16 and 0.28, RMSE decreased by 0.23 and 1.94, and RPD increased by 4.36 and 4.81. (5) Among the two-dimensional spectral indices, OSI exhibited a strong positive driving effect. In the three-dimensional spectral indices, TDI 2 demonstrated a significant positive effect on model predictions. In the combined two-dimensional and three-dimensional indices, TDI 7 showed a mostly negative effect, while TDI 9 exhibited a higher positive effect.