Inversion Model of Coastal Soil Salinity Based on Hybrid Feature Selection and Machine Learning Algorithms

Abstract

Soil salinization in coastal regions poses a serious threat to sustainable agricultural development. Feature selection is critical for improving soil salinity inversion accuracy. To this end, this study developed a Pearson correlation coefficient (PCC)-based hybrid feature selection framework. First, the PCC was used for preliminary screening, followed by variable importance in projection (VIP), gradient boosting machine (GBM), and recursive feature elimination (RFE) for secondary selection, yielding four schemes: PCC, PCC-VIP, PCC-GBM, and PCC-RFE. Subsequently, five regression models, including backpropagation neural network (BPNN), adaptive boosting (AdaBoost), extreme gradient boosting (XGBoost), random forest (RF), and multiple linear regression (MLR), were constructed and evaluated based on these feature subsets. The best model was further interpreted using the SHapley Additive exPlanations (SHAPs) method and used to generate the soil salinity spatial distribution map. The results showed that: 1) the PCC-GBM-BPNN model achieved the best performance (R ${}^{2} =0.743$ , root-mean-square error (RMSE) = 4.081 g/kg, and ratio of performance to deviation (RPD) = 2.02), improving R2 by 0.209 compared with the PCC-BPNN model; 2) SHAP analysis revealed that the enhanced vegetation index (EVI) was the most influential feature for soil salinity inversion; and 3) the inversion results indicate that soil salinization in the study area was mainly characterized by mild and moderate levels, consistent with previous studies. These findings suggest that hybrid feature selection can effectively enhance soil salinity inversion performance, providing an effective approach for soil salinity monitoring in coastal regions.