Sub-Seasonal Precipitation Bias-Correction in Thailand Using Attention U-Net With Seasonal and Meteorological Effects

Tanatorn Faijaroenmongkol, Kanoksri Sarinnapakorn, Peerapon Vateekul

IEEE Access · 01 January 2023

In one sentence. Attention U-Net with multi-scale residual blocks bias-corrects CFSv2 sub-seasonal rainfall over Thailand, improving RMSE by 8.65% over linear downscaling.

Abstract

There have been many attempts to forecast sub-seasonal to seasonal (S2S) precipitation. One of them is the Climate Forecast System version 2 (CFSv2) model; however, a bias correction must be applied before CFSv2 data can be used in each local region. In this research, we aim to address the S2S precipitation forecasting using our new bias correction on CFSv2 data. Our model is based on the deep learning model: Attention U-Net having two proposed enhancements: (i) a multi-scale residual block to learn patterns on different scales and (ii) a combination of customized regression loss and classification loss. Further, we apply a log scaler to reduce the impact of skewness in the data. Finally, seasonal and meteorological effects are provided as additional input for our model. CFSv2 is employed as a dataset to be corrected in our study, while rainfall data from Thailand’s Hydro-Informatics Institute (HII) is served as the ground truth. In the result, it demonstrates that our model outperforms two baseline models: namely, a linear-downscaling technique (traditional approach) and a traditional Attention U-Net model. Our model’s root mean square error (RMSE) and temporal correlation coefficient (TCC) improve about 8.65% and 13.77% respectively, over the linear-downscaling technique. Besides the results of the Attention U-Net model reveal that our model’s RMSE and TCC improve by about 15.56% and 12.06%, respectively.

Labels

Labels and summary are generated by Claude and reviewed weekly. Wrong? Edit this entry.