A feature selection-based hybrid model for accurate spatial prediction of potentially toxic elements in sediments: Application to Poyang Lake, China

Abstract

Accurately predicting spatial distribution of potentially toxic elements (PTEs) in sediments is crucial for protecting aquatic ecosystem but remains challenging due to complex interactions of environmental variables. This study developed an integrated framework by combining optimal machine learning (ML) with ordinary kriging (OK) and feature selection to improve prediction accuracy of PTEs in Poyang Lake sediments. Three ML models—random forest (RF), extreme gradient boosting (XGBoost), and support vector machine (SVM)—were evaluated to identify the optimal approach for predicting PTE concentrations. Feature selection techniques including relative importance analysis and recursive feature elimination were applied to identify suitable predictors for each PTE. RF model outperformed the others across all PTEs (R²>0.70). Distinct sets of predictors were identified for each PTE, further refining RFOK model. Incorporating selected predictors, RFOK significantly enhanced prediction accuracy, increasing R² by 37.5% (Cr) to 421% (Cd) relative to OK, and by 133% (Cr) to 457% (Pb) relative to inverse distance weighting, effectively capturing fine-scale spatial variability. The findings highlight the effectiveness of the RFOK hybridization and the importance of feature selection in enhancing prediction accuracy within complex multifactor aquatic environments, providing scientific supports for designing targeted protection strategies against PTE pollution in aquatic ecosystem.