[3G1-OS-14a] OS-14

10:15 AM - 10:30 AM JST(1:15 AM - 1:30 AM UTC)

[3G1-OS-14a-06] GNN-based Bias Correction and Super-Resolution of Global Reanalysis Data: Application to AMeDAS

〇Yuki Aoyama¹, Kyohei Atarashi¹, Hisashi Kashima¹, Koh Takeuchi¹ (1. Kyoto University)

Keywords:

Neural Network,Spatial-temporal data,Machine Learning

Although data-driven global weather forecasting models, such as GraphCast, have improved rapidly, they still suffer from representativeness errors, or differences between coarse grid outputs and weather station observations. Many current bias-correction methods struggle to reflect wind-driven information flow because they lack sufficient physical constraints. To address these issues, we employ E(n)-equivariant graph neural networks (EGNNs) for bias correction and introduce an Advective EGNN (AdvEGNN) that explicitly incorporates essential meteorological principles. Using ERA5 reanalysis and AMeDAS observations over Japan, we tested the ability of these models to predict future local weather. Overall, EGNNs achieve high accuracy. Specifically, the AdvEGNN reduces temperature prediction errors by approximately 7.0\%, outperforming standard graph convolutional networks (GCNs). These results demonstrate that incorporating physical principles into geometric deep learning can enhance the accuracy of meteorological forecasts.