Heuristic Topological Graph Convolutional Network for Risk Prediction of Potentially Toxic Elements in Cultivated Soils

Abstract

Contamination of cultivated soils with potentially toxic elements (PTEs) poses a growing threat to global food security. Although existing risk assessments have examined the accumulation and toxicity of PTEs, their dynamic interplay with multidimensional drivers has remained inadequately characterized. Here, an innovative heuristic graph convolutional network (GCN) model is introduced by integrating adaptive graph topology with quantified directional feedback optimization to improve ecological risk prediction. Leveraging 466 spatially resolved soil samples and 28 environmental drivers of a typical rice production area Yangtze River Basin in China, the heuristic GCN model outperformed traditional approaches by 23.1% in predictive accuracy. A three-phase heuristic algorithm pruned 85.5% of spurious edges in the topological graph, and GCN adaptively quantified the directional feedback between environmental drivers and ecological risk. Topological networks and feature importance analysis jointly identified pH, base saturation, calcium carbonate, exchangeable bases, and soil organic carbon as pivotal regulators acting alongside geological factors. By linking mechanistic soil chemistry with machine-learning-based causal inference, our model supports streamlinedly simplified, directionally quantified, and dynamically adapted ecological risk prediction. This enables the screening of the most efficient pathway of risk management and provides more precise and integrated strategies for ecological risk control in agroecosystems.