인공지능 기반 수치모델 초해상화 연구

A Review of Artificial Intelligence based Numerical Model with Super-resolution

Since high-resolution meteorological data can be used for precise research or useful services in real life, the need for high-resolution data production is increasing. Considering the local phenomenon and complex geographical characteristics of the Korean Peninsula, the installation of a large number of observatories is ideal, but this is limited by practical problems such as budget. Numerical models can be a good alternative to this. In this paper, a study was conducted to super-resolution the GDPS data of the 12 km unit grid into the LDPS data of the 1.2 km unit grid through the FSRCNN and VDSR techniques. As the performance evaluation index of the super-resolution model, MAE, which means the average difference between the super-resolution data and the LDPS data, and PSNR, which means the loss rate of the data, were used. As a result, the VDSR model was excellent for temperature and the FSRCNN model was excellent for U and V components for wind.