Dust Source Emission Forecasting using GraphSAGE and Delaunay Graph

Our work proposes a method for forecasting dust source emission (DSE) using a graph neural network (GNN). Dust storms pose a significant environmental and health hazard, and accurately forecasting their origin is crucial for mitigation strategies. Traditional methods like direct field measurements are limited in coverage or resolution. This work introduces a framework that constructs a novel connected graph, Delaunay graph from historical dust hotspot events, incorporating spatial and temporal relationships between environmental parameters obtained from remote sensing data. A state-of-the-art GNN model, inductive GraphSAGE, is applied to this graph to predict future active dust source locations. The proposed method outperforms a baseline random forest model, demonstrating its ability to capture temporal relationships throughout months. The AUC of GraphSAGE for dust source emission forecasting was approximately 73% which is 17% better than a simple random forest.