A novel prediction approach driven by graph representation learning for heavy metal concentrations

Area of Science:

• Environmental Science
• Computational Chemistry
• Agricultural Science

Background:

• Heavy metal (HM) contamination in soil-rice systems poses significant public health risks.
• Current prediction models lack interpretability due to insufficient consideration of inter-factor correlations.
• Early prediction is crucial for mitigating HM accumulation.

Purpose of the Study:

• To develop an accurate and interpretable method for predicting HM concentrations in soil-rice systems.
• To leverage Graph Representation Learning (GraRL) for modeling complex environmental factor relationships.
• To enhance the efficiency and accuracy of HM accumulation prediction models.

Main Methods:

• Developed the GraRL-HM method, comprising two modules: PeTPG and GCN-HM.
• PeTPG module: Generated a graphic structure using graph representation and communitization to identify correlations and transmission paths of environmental factors.
• GCN-HM module: Employed a graph convolutional neural network (GCN) for predicting HM concentrations based on the refined graph structure.

Main Results:

• The PeTPG model reduced graph scale by 53.5%, simplifying 396 correlation paths to 184 by removing invalid ones.
• The refined graph structure enhanced learning efficiency and representation accuracy.
• The GCN-HM model outperformed four benchmark models, improving prediction accuracy (R²) by 36.1%.

Conclusions:

• The GraRL-HM method offers a novel, accurate, and interpretable approach for predicting HM accumulation in soil-rice systems.
• The study provides valuable insights for intelligent regulation and guidance in managing heavy metal pollution.
• This approach enhances the potential for precision agriculture and environmental risk assessment.