Encoder Decoder with multi-head Attention layers for multi-channel Climate Downscaling (EDA)

An Encoder-Decoder model of Climate Downscaling of ERA5

training data: ERA5 in Taiwan region with the shape of (14,9)
auxiliary data: ERA5 10 meter high wind field vector u,v near Taiwan region with the shape of (14,9) and low-resolution(downsampled) label data
label data: TCCIP grid daily precipitation with the shape of (70,45) (resized)

Model Architecture

Encoder involves Multi-head Attention Layer and Fully Connected Layer (FC).
Decoder involves Upsampling Layer with Efficiency Sub-pixel method (of ESPCN) and convolutional layers.

Model Prediction

Prediction of precipitation on 2019.08.15
left-top: ERA5 Reanalysis Data of Precipitation
left-bottom: TCCIP gridded observation
right-top: MAE error heatmap
right-bottom: model prediction

Metrics

Evaluation on metrics of pixel-wise error: mean absolute error (MAE) and root mean square error (RMSE).
And relationship: Pearson Correletaion (Corr) and Structural Similarity Index (SSIM).

Compared to traditional statistical downscaling method, BCSD (Bias-Corrected Spatial Disaggregation):

Environment

tensorflow==2.14.0+nv23.11
numpy==1.24.4
pandas==1.5.3
(last updated: 2024/02/21)