A GNN-based station-level downscaling method for 2-m air temperature by considering region terrain

• Integrating multi-dimensional information, propose a T2m downscaling method. • Propose a Geospatial-Perceived Graph Neural Network (GeoGNN). • Propose a feature extracting method for region terrain. • Experiment results verify the effectiveness. Existing deep learning-based downscaling (DS) methods for meteorological variables treat downscaling as a super-resolution task in computer vision, heavily relying on grid-based data. In response to the demand for real-time, high-resolution 2-m air temperature (T2m) in meteorology, we propose an innovative Graph Neural Network-based Station-Level 2-m air Temperature Downscaling method (GNN-SLTD), which takes in-situ observation data from irregularly scattered meteorological stations and digital elevation model (DEM) in a region as input, then downscales T2m to any locations within the research area. Taking Fujian Province, China, a topographically complex region, as the research area, the experiment results show that GNN-SLTD improves by at least 12.6%(RMSE) and 12.4%(MAE) in stations-to-stations downscaling task. In comparison to mountainous areas, plain areas exhibit smoother and more continuous atmospheric circulation patterns, making the downscaling easier. The distribution of backbone stations is also affecting the performance. In grid-to-stations downscaling tasks, GNN-SLTD overcomes the problem of poor accuracy of ERA5-Land in coastal areas and achieves great improvement in T2m downscaling. In addition, the ablation study verifies the indispensability of covariates and region terrain features in improving T2m downscaling performance. GNN-SLTD is significant for downscaling models based on deep learning and provides a new way to integrate multi-dimensional information.