Correction of Monthly Average 2 m Temperature Prediction:A Method Based on Deep Learning
FANG Wei
WANG Binglun
Abstract：As a technique to reduce the error in short-term climate prediction,bias correction has become an important research direction.This study explored the application of deep learning techniques in bias correction,focusing on the U-Net model,which,despite its popularity,has inherent limitations.First,U-Net is based on a convolutional neural network,which has a limited receptive field,preventing it from fully capturing spatial features from a global perspective.Second,the subsampling operation in U-Net often leads to a loss of important image details.To address these issues,we implemented the following two measures.First,we integrated the original model with a Vision Transformer,which is capable of learning global features,thereby overcoming the limitation of the convolutional neural network's small receptive field.The second was to introduce the full-scale connection operation from the UNet 3+model and compensate for the image details lost in the original down-sampling process in the decoder.The improved model is called UNet-Former 3+.A correction experiment was carried out on the spring and winter datasets of the monthly average 2m temperature in CMIP6,with ERA5 as the experimental label.We compared its performance against six other correction methods:quantile mapping,ridge regression,U-Net,CU-Net,Dense-CUnet,and RA-UNet.The experimental results show that the average absolute error and root mean square error of this model were reduced by 49%and 57%,respectively,outperforming the other six methods.Overall,the UNet-Former 3+model demonstrated superior correction performance for both spring and winter seasons.
Keywords：short-term climate predictionbias correctionVision Transformerfull-scale connectionUNet-Former 3+
Publication Date：2024-12-28
Online Publishing Date：2025-08-15(First online date of this platform, not the publication date of the document)
Pages：12( 906-917 )