Geographically informed graph neural networks

Graph neural networks (GNNs) have been introduced to spatial statistical tasks due to their mechanisms of simulating spatial interactions and processes among geographical neighbours using graph structures. However, previous methods ignore quantifying differences in attributes among adjacent spatial characteristics. Considering this spatial characteristic by fitting the spatial statistic trinity (SST) framework may help improve models’ accuracy and robustness. Thus, we introduce the geographically informed graph neural network (GIGNN) by considering the additional geospatial feature: closer geographical entities may interact less when spatial disparities are captured. When setting up the model, GIGNN leverages differences of attributes by spatial stratified heterogeneity, quantifies connections between geographical entities, and inherits k-order neighbour attribute aggregation and message-passing mechanisms from GNNs. GIGNN is applied to an urbanization analysis study in the Greater Perth Area, Australia, showing higher accuracy than the existing machine learning models and other GNNs in simulation and prediction accuracy. GIGNN achieved an accuracy of 84.1% for simulation and an accuracy of 81% for prediction. Incorporating spatial characteristics into GNNs enhances simulation and prediction accuracy in geoscientific applications, highlighting the importance of spatially aware models in solving complex problems by capturing geographical data dependencies.