BaguanHR: Data Scaling for High-Resolution Weather Forecasting

BaguanHR, a novel framework, suggests transitioning from model transfer to data transfer to address the challenges faced in high-resolution weather forecasting. The creation of global weather forecasting models at a 0.1-degree resolution using machine learning is limited due to the scarcity of high-resolution data, as historical reanalysis exists only at a 0.25-degree resolution. Current methods attempt to refine 0.25-degree models with sparse 0.1-degree samples, but this process suffers from irreversible information loss associated with coarse-resolution forecasts. In contrast, BaguanHR utilizes super-resolution (SR) for more effective resolution transfer, benefiting from lower conditional entropy and input amplification. By employing variable-wise SR, the framework generates comprehensive 0.1-degree data from ERA5. BaguanHR outperforms both ML-based techniques and IFS-HR, the high-resolution variant of the Integrated Forecasting System. The research is accessible on arXiv with the identifier 2608.14652.