基于深度学习的月平均2 m气温订正方法  

Correction of Monthly Average 2 m Temperature Prediction:A Method Based on Deep Learning

作　　者：方巍[1,2,3,4] 王冰轮 FANG Wei;WANG Binglun(School of Computer&Software/Engineering Research Center of Digital Forensics of Ministry of Education,Nanjing University of Information Science&Technology,Nanjing 210044,China;China Meteorological Administration Basin Heavy Rainfall Key Laboratory/Hubei Key Laboratory for Heavy Rain Monitoring and Warning Research,Institute of Heavy Rain,China Meteorological Administration,Wuhan 430205,China;Key Laboratory of Transportation Meteorology of China Meteorological Administration,Nanjing Joint Institute for Atmospheric Sciences,Nanjing 210041,China;Jiangsu Collaborative Innovation Center of Atmospheric Environment and Equipment Technology(CICAEET),Nanjing University of Information Science&Technology,Nanjing 210044,China)

机构地区：[1]南京信息工程大学计算机学院/数字取证教育部工程研究中心,江苏南京210044 [2]中国气象局流域强降水重点开放实验室/暴雨监测预警湖北省重点实验室/中国气象局武汉暴雨研究所,湖北武汉430205 [3]南京气象科技创新研究院/中国气象局交通气象重点开放实验室,江苏南京210041 [4]南京信息工程大学江苏省大气环境与装备技术协同创新中心,江苏南京210044

出　　处：《热带气象学报》2024年第6期906-917,共12页Journal of Tropical Meteorology

基　　金：国家自然科学基金项目(42075007、42475149);灾害天气国家重点实验室开放项目(2021LASWB19);江苏省研究生科研创新计划项目(KYCX22_1218);中国气象局交通气象重点开放实验室开放研究基金项目(BJG202306);中国气象局流域强降水重点开放实验室开放研究基金(No.2023BHR-Y14)共同资助。

摘　　要：作为减少短期气候预测误差的技术,数据订正成为了重要的研究方向。而深度学习作为一种新兴方法已经应用到数据订正技术中,其中常用的模型是U-Net,但它存在不可避免的缺陷。第一,U-Net基于卷积神经网络,但是受限于卷积神经网络的小感受野,这导致U-Net不能从全局的角度学习空间特征;第二,U-Net的下采样操作容易丢失图像细节信息。这两点都影响了该模型的订正性能。因此采取以下两个措施进行改进,一是将原模型与能够学习图片全局特征的Vision Transformer有机结合起来,使其能够从全局的角度学习空间特征;二是引入UNet 3+模型中的全尺度连接操作,弥补原下采样中丢失的图像细节信息。改进之后的模型称为UNet-Former 3+,在CMIP6中月平均2 m气温的春季和冬季数据集上进行订正实验,ERA5为实验标签。模型会与分位数映射、岭回归、U-Net、CU-Net、Dense-CUnet和RA-UNet这六种订正方法进行对比。实验结果表明,此模型在两个季节的平均绝对误差都下降49%,均方根误差都下降57%,两者都低于上述六种方法。总之,UNet-Former 3+在春季和冬季的订正效果优于上述六种方法。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,UNet 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,UNet,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.

关 键 词：短期气候预测 数据订正 Vision Transformer 全尺度连接 UNet-Former 3+ 

分 类 号：P456.5[天文地球—大气科学及气象学]