Improving predictive accuracy for soil cadmium distribution in highly heterogeneous region based on a synergistic dual-stage feature selection strategy.

Cadmium (Cd) contamination in soil poses a significant threat to crop yield and public health. Accurately predicting its spatial distribution in highly heterogeneous regions remains challenging due to environmental complexity and high-dimensional data noise. This study proposes a novel dual-stage variable selection framework, OPGD-RFE, which integrates an optimal parameter-based Geodetector (OPGD) with recursive feature elimination (RFE) to identify key drivers and reduce data redundancy. In a case study conducted in a complex industrial-mining region (Zhangjiakou City), we systematically collected four types of multi-source environmental variables: geology, soil, climate, and socio-economic activities. Four machine learning models were trained on datasets selected by different methods, and their performance was compared. Results indicated that the machine learning model trained on the dataset selected by OPGD-RFE significantly outperformed others: the optimal random forest (RF) model achieved a test set coefficient of determination (R2) of 0.67 and root mean square error (RMSE) of 0.27 mg kg-1, representing a 39.58% improvement in R2 over OPGD and a 76.32% improvement over RFE. The optimized model revealed a distinct spatial distribution pattern of cadmium and successfully delineated four high-risk areas. Geographically weighted regression (GWR) analysis further elucidated the varying local driving mechanisms within these sub-regions, providing essential information for targeted risk management.