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

Area of Science:

• Environmental Science
• Soil Science
• Machine Learning

Background:

• Potentially toxic elements (PTEs) in cultivated soils threaten food security.
• Existing risk assessments lack characterization of dynamic interplay between PTEs and environmental drivers.

Purpose of the Study:

• To develop an innovative heuristic graph convolutional network (GCN) model for enhanced ecological risk prediction.
• To improve the characterization of the dynamic interplay between environmental drivers and PTEs in soils.

Main Methods:

• Developed a heuristic GCN model integrating adaptive graph topology and directional feedback optimization.
• Utilized 466 soil samples and 28 environmental drivers from the Yangtze River Basin.
• Employed a three-phase heuristic algorithm to prune graph edges and quantify feedback.

Main Results:

• The heuristic GCN model achieved 23.1% higher predictive accuracy than traditional methods.
• Identified key regulators of ecological risk: pH, base saturation, calcium carbonate, exchangeable bases, and soil organic carbon.
• Successfully linked soil chemistry with machine learning for causal inference.

Conclusions:

• The developed model provides simplified, quantified, and adaptive ecological risk prediction.
• Enables efficient screening of risk management pathways for agroecosystems.
• Offers precise and integrated strategies for ecological risk control.