Multigraph fusion neural network for predicting available cadmium levels in farmland soils

Abstract

Cadmium (Cd), a primary heavy metal pollutant in farmland soils, poses a significant threat to soil ecosystems and human health, developing accurate machine learning models to predict soil available Cd concentration is crucial for formulating effective remediation strategies. However, most existing models rely on the simple aggregation of multiple environmental and geographical variables to predict soil available Cd concentration, often neglecting the complex interrelationships among these variables and the spatial effects of geographic factors. In this work, a novel multigraph fusion neural network model based on the spatial relationships between sampling points and various geographic factors (elevation, mine, roads, and rivers) is proposed. By integrating multiple spatial relationship graphs, the model effectively captures the spatial effects of geographic factors on the farmland soil environment. The results demonstrate that the multigraph fusion model significantly outperforms the other models in predicting soil available Cd concentration, achieving R² value of 0.82, RMSE value of 0.0338 mg kg⁻¹, and MAE value of 0.0249 mg kg⁻¹. Compared with the single-graph models and baseline models, the multigraph fusion model provides lower residual distribution range and more stable prediction performance. These findings validate the feasibility of incorporating the spatial effects of geographic factors to increase the prediction performance of soil available Cd concentration models and offer valuable analysis tools into the environmental drivers underlying the spatial heterogeneity in heavy metal contamination in farmland soils.