Climate Downscaling with an Encoder-Resolver Model

Training data: ERA5 reanalysis data (precipitation, 10 meter wind field)

Labels (ground truth data): TCCIP daily gridded precipitation data

Github: https://github.com/AugChiang/EnRe3_Climate_Downscaling

Model Architecture

composed of multi-head attention layer, fully connected layer, etc. in Encoder part (transformer) and convolutional layer in Resolver part for climate downscaling.

Results

metrics, compared to traditional climate downscaling method of Bias Corrected Spatial Disaggregation (BCSD)