基于U-Net改进的日平均2 m气温订正方法  

Improved Daily Average 2 m Temperature Correction Method Based on U-Net

作　　者：王冰轮 方巍[1,2,3] WANG Binglun;FANG Wei(School of Computer,Nanjing University of Information Science&Technology,Nanjing 210044,Jiangsu,China;Key Laboratory of Transportation Meteorology of China Meteorological Administration,Nanjing Joint Institute for Atmospheric Sciences,Nanjing 210041,Jiangsu,China;Provincial Key Laboratory for Computer Information Processing Technology,Soochow University,Suzhou 215000,Jiangsu,China)

机构地区：[1]南京信息工程大学计算机学院,江苏南京210044 [2]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 [3]苏州大学江苏省计算机信息处理技术重点实验室,江苏苏州215000

出　　处：《应用科学学报》2025年第1期51-65,共15页Journal of Applied Sciences

基　　金：国家自然科学基金(No.42075007,No.42475149);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);中国气象局交通气象重点开放实验室开放研究基金项目(北极阁基金项目)(No.BJG202306)资助。

摘　　要：针对数据订正常用的深度学习模型U-Net中不能充分学习空间特征以及图像细节信息丢失的问题,提出了S-CUnet 3+模型。S-CUnet 3+采取以下两个措施对U-Net进行改进:一是将原模型与能够学习图片全局特征的Swin Transformer有机结合起来;二是引入多尺度连接操作。模型还采用了预训练与微调的训练策略针对多个预报步长同时订正。7个预报步长的日平均2 m气温预报值订正的实验结果表明,S-CUnet 3+模型对所有预报步长的预报都有明显的订正效果,其中24 h预报步长的订正效果最好,平均绝对误差和均方根误差分别下降了50.64%和49.25%,且相比于基于历史资料的模式距平积分预报订正、分位数回归、岭回归、U-Net、CU-Net、Dense-CUnet和RA-UNet这7种订正方法,S-CUnet 3+取得了更好的订正效果。In response to the limitations of the widely utilized deep learning model U-Net,which is unable to adequately learn spatial features and suffers from the loss of image detail information,the S-CUnet 3+model has been proposed.S-CUnet 3+enhances U-Net in two ways:firstly,it integrates the original model with the Swin Transformer,enabling it to learn global features of images,and secondly,it introduces multi-scale connection operations.The model also adopts pre-training and fine-tuning strategies to correct multiple forecast lead times simultaneously.Experimental results of correcting daily average 2 m temperature forecasts across seven lead times show that the S-CUnet 3+model has a significant correction effect for all lead times,with the best correction effect at the 24-hour lead time.The mean absolute error and root mean square error are reduced by 50.64%and 49.25%,respectively.Moreover,S-CUnet 3+outperforms seven existing correction methods:anomaly numerical-correction with observations,quantile regression,ridge regression,U-Net,CU-Net,Dense-CUnet,and RA-UNet.

关 键 词：数据订正 深度学习 Swin Transformer 预训练 微调 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]