Enhancing Graph Neural Networks Using Proximity Graphs for Dust Source Emission Forecasting

Predicting dust source emissions is a complex spatiotemporal challenge, as these events depend on dynamic environmental factors like wind speed, soil moisture, and vegetation cover. Traditional machine learning models often struggle to capture the intricate spatial dependencies between these variables. This paper investigates whether constructing proximity graphs—which explicitly model spatial relationships—can enhance the performance of Graph Neural Networks (GNNs) in forecasting dust source activity.

The researchers treat dust emission forecasting as a binary classification problem, identifying whether a specific location will be an active dust source. Because the raw data is tabular, the authors first transform it into a graph structure. They construct spatial proximity graphs for each month using methods such as Delaunay triangulation, Gabriel graphs, k-Nearest Neighbor (k-NN) graphs, and Yao graphs. To capture temporal dynamics, they connect these monthly graphs by linking nodes to their nearest spatial neighbors in the subsequent month. They then apply three standard GNN architectures—GraphSAGE, Graph Convolutional Networks (GCN), and Graph Attention Networks (GAT)—to these structures to learn and predict emission patterns.

Findings

The study demonstrates that proximity-based graph structures are superior to random graph connections for message passing in GNNs. By explicitly encoding spatial proximity, the models better capture the underlying environmental dynamics of the Tigris and Euphrates river basin. The results show that GNNs utilizing these proximity graphs significantly outperform both GNNs built on random graphs and traditional Long Short-Term Memory (LSTM) models, providing a more robust framework for proactive dust storm mitigation.

Dust storms cause significant environmental and health hazards, yet forecasting them remains difficult due to the limitations of remote sensing and traditional modeling. This research provides a novel, effective framework for integrating spatial geometry into deep learning models. By leveraging proximity graphs, researchers can better model the complex, interconnected nature of dust sources, leading to more reliable predictions that can assist policymakers and public health officials in planning mitigation strategies.