Interpretable salinization estimation model for Dongying City based on integrated multi-dimensional spectral indices with XGBoost-driven transformations

Against the backdrop of intensifying global climate change and anthropogenic activities, soil salinization is becoming increasingly severe, posing a significant threat to agricultural production and sustainable development. This study collected sample data from Dongying City, Shandong Province as the research area, focusing on investigating the effectiveness of combining mixed-scale transformations and high-dimensional spectral indices in model inversion. The research process involved spectral transformations, upon which hybrid transformations were constructed. Subsequently, sensitive spectral bands were identified, leading to the development of two-dimensional and three-dimensional spectral indices. These indices served as feature variables to construct three types of models: Extreme Gradient Boosting (XGBoost), Partial Least Squares Regression (PLSR), and Convolutional Neural Network (CNN). The optimal model was evaluated using SHAP interpretability analysis. The results demonstrate that the proposed processing methods effectively enhance band sensitivity, improve model accuracy and generalization capability. Index operations conducted across different spectral bands can enhance spectral sensitivity characteristics or suppress the influence of noise to some extent. The outcomes provide a theoretical framework for optimizing spectral indices. The analysis reveals the salt-response mechanisms and noise resistance differences among various indices, facilitating the advancement of salinization monitoring towards greater precision, intelligence, and cost-effectiveness.