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

Area of Science:

• Meteorology
• Artificial Intelligence
• Climate Science

Background:

• Artificial intelligence (AI) is increasingly used for meteorological data forecasting, but faces challenges like long training times and high computational costs.
• Applying existing AI models across different regions to reduce repetitive training is a significant issue.
• Super-resolution (SR) reconstruction models offer potential for efficient data processing.

Purpose of the Study:

• To explore the transfer learning capabilities of a super-resolution (SR) reconstruction model for meteorological data.
• To evaluate the SR reconstruction performance of a ResNet model using 2-meter temperature data from South China.
• To assess the model's ability to reconstruct other meteorological data (wind speed, atmospheric pressure) using transfer learning.

Main Methods:

• Utilized a ResNet model integrated with sub-pixel convolution modules for SR reconstruction.
• Evaluated transfer learning performance by applying the model to temperature data from different regions.
• Conducted 2x and 4x SR experiments for temperature and other meteorological data reconstruction.

Main Results:

• The ResNet model effectively captured data features for SR reconstruction.
• Transfer learning experiments across various regions showed favorable SR reconstruction performance.
• The transfer learning-based neural network model outperformed traditional interpolation methods in accuracy.
• Successful reconstruction of other meteorological data (wind speed, atmospheric pressure) was achieved.

Conclusions:

• Neural network models exhibit strong transfer learning capabilities applicable to meteorological data.
• Transfer learning significantly reduces computational costs and training times for AI in climate research.
• The study confirms the feasibility and high significance of transfer learning in meteorological data analysis and climate applications.