Exploring cross-regional and cross-variable transferability of a ResNet-based super-resolution method for the ERA5 data

Abstract

Abstract With the development of artificial intelligence, diverse datasets can be assisted in refined operations AI. In recent years, AI has been applied to meteorological data forecasting. However, using AI presents challenges such as long training times and high computational costs. Applying similar meteorological data models across different regions to reduce repetitive training costs remains a significant issue to address. This study explores the transfer learning capabilities of a super-resolution (SR) reconstruction model using 2-meter temperature data from SouthChina. The ResNet, integrated with sub-pixel convolution modules, effectively captures data features. By leveraging similar temperature data across different regions, the model’s SR reconstruction performance is evaluated. Experiments compare the model’s transfer learning abilities across various regions. Additionally, given the correlation of meteorological features within the same region, the study attempts to reconstruct other meteorological data (e.g. wind speed, atmospheric pressure, etc.) with SR. Both 2x and 4x SR experiments for the two tasks yield favorable results. Compared to traditional interpolation methods, the transfer learning-based neural network model produces more accurate outcomes. The findings indicate that neural network models possess strong transfer learning capabilities, which are highly significant for climate research and related applications and confirm the feasibility of transfer learning in meteorological data.