Graph Learning-Based Spatial Structural Identification of Drought Regions

Abstract. Under the backdrop of global climate change, the frequency and severity of drought events are continuously increasing, posing signif-icant challenges to human society, ecosystems, and economic development. Traditional drought simulation methods often overlook the interactions among meteorological, hydrological, and geographic information. Complex network theory offers a new perspective for exploring these interconnections. Graph Neural Networks (GNNs), as a deep learning technique capable of handling geospatial data and complex structures, have advantages in capturing geographic correlation information and network topology. Therefore, combining complex networks and GNNs for drought simulation is of great significance. This study proposes a framework that integrates complex networks and graph learning to identify the spatial structure of drought regions. Using latitude-longitude grids as nodes, topological indicators such as degree, betweenness centrality, and clustering coefficient describe node features, which are combined with SPEI time series statistical features to form multidimensional vectors. These are input into a Graph Convolutional Network (GCN) to obtain low-dimensional embeddings, and clustering is used to divide the space into subregions. Results show that the clustering based on multi-feature combinations exhibits stronger spatial continuity and clearer boundaries. Regions with high degree and betweenness centrality and low clustering coefficient serve as network hubs and information bridges, while medium-feature regions are intermediate connecting zones, and regions with low feature values are peripheral isolated areas. This method offers a novel approach to analyzing drought system structures and regional risk management.