Spatial prediction of soil cadmium concentration: a multi-model prediction system with novel evaluation metrics

Abstract

This study proposes a modular, automated Multi-Model Prediction System (MMPS) to overcome nonlinear relationships, data integration difficulties, and model instability in soil cadmium (Cd) spatial prediction under small sample conditions. We further introduce two novel evaluation metrics — Global Performance Metrics (GPM) to assess robustness across runs, and Global Mean Local Standard Deviation (GMLSD) to quantify spatial stability of prediction maps. Our system achieved high predictive accuracy, with the LGBM model demonstrating superior performance (Global R² = 0.614). Crucially, GPM yielded substantially more consistent evaluations, exhibiting up to an 88.6% reduction in performance standard deviation across random seeds compared to traditional metrics. The MMPS, which integrates geostatistical, deterministic, and machine learning algorithms, was tested in a small mining watershed in Hunan, China. Using multi-source data (remote sensing, topography, soil properties) from 149 samples, we compared 12 models. Results showed that machine learning models achieved higher prediction accuracy but tended to have lower spatial prediction stability (higher GMLSD). Topographic data was the most influential input, boosting performance by 53.5%, whereas spectral or soil data alone showed limited impact. Key driving factors were identified as distance to river, soil pH, B2 band, and organic matter. The MMPS provides a high-precision, low-cost intelligent tool for pollution management. Together with our robust evaluation framework, this research offers a new paradigm for environmental spatial prediction, enabling smarter, scalable, and more reliable pollution control strategies.