Graph Attention Networks Predict Soil Microplastics and Organic Matter
A study using Graph Attention Networks (GATs) achieved high accuracy in predicting soil microplastics and organic matter from 91 georeferenced samples, but cross-validation revealed limited generalization due to small sample size and sparse graph structure. The two-layer GAT architecture incorporated spatial coordinates, soil properties, and land use data, yielding RMSEs of 625.06 (R²=0.87) for microplastics and 0.43 (R²=0.91) for organic matter. The research underscores the need for denser datasets and improved graph connectivity for spatial soil prediction.
Key facts
- Graph Attention Networks used for spatial prediction of soil microplastics and organic matter
- 91 georeferenced soil samples analyzed
- Two-layer GAT architecture incorporated spatial coordinates, soil properties, and land use data
- RMSE of 625.06 (R²=0.87) for microplastics
- RMSE of 0.43 (R²=0.91) for organic matter
- Cross-validation showed limited generalization due to small sample size and sparse graph structure
- Study demonstrates potential of GATs for spatial soil prediction
- Need for dense datasets and improved graph connectivity highlighted
Entities
Institutions
- arXiv