基于CGAFNet的卫星云图临近预报研究  

Satellite Cloud Image Nowcasting Based on CGAFNet

作　　者：康奇秀 杜东升 陈立福[1] 欧小锋 叶成志 KANG Qixiu;DU Dongsheng;CHEN Lifu;OU Xiaofeng;YE Chengzhi(School of Electrical&Information Engineering,Changsha University of Science and Technology,Changsha 410004,China;Hunan Institute of Meteorological Sciences,Changsha 410118,China;Hunan Key Laboratory of Meteorological Disaster Prevention and Reduction,Changsha 410118,China)

机构地区：[1]长沙理工大学电气与信息工程学院,湖南长沙410004 [2]湖南省气象科学研究所,湖南长沙410118 [3]气象防灾减灾湖南省重点实验室,湖南长沙410118

出　　处：《热带气象学报》2024年第6期1074-1084,共11页Journal of Tropical Meteorology

基　　金：国家自然科学基金联合基金项目(U2242201);湖南省自然科学基金重大项目(2021JC0009)共同资助。

摘　　要：卫星云图外推技术能及时掌握云团的运动轨迹和变化情况,为临近预报和灾害性天气的监测提供重要参考。然而,现有的云图预测方法存在难以捕捉小尺度云团发展、云图细节特征不清晰、预测结果逐渐模糊等问题,导致最终的预报效果不理想。为了有效提取卫星云图的时空信息,预报中小尺度云团的发展,利用FY-4A红外云图,以湖南区域为中心的中东部地区作为研究对象,从时空序列预测的角度出发,提出了一种卷积门控循环注意力融合网络(ConvGRU Attention Fusion Network,CGAFNet),并提出了主副损失(Primary and Secondary Loss,PaSLoss)作为模型的损失函数,构建了编-解码结构,更好地提取了卫星云图的时空信息。为验证网络框架的有效性,与三个典型网络进行对比实验,结果表明,CGAFNet在云图外推任务中均方根误差为10.00 K,结构相似性为0.74,峰值信噪比为31.43,该模型能准确预测云团的生消演变过程,在各项指标上均优于其它网络,证明该方法能获得更准确的预测精度,且具备良好的泛化能力。Satellite cloud image extrapolation technology enables timely tracking of the movement and changes of cloud clusters,providing important references for nowcasting and severe weather monitoring.However,existing cloud image prediction methods face challenges such as difficulty in capturing the development of small-scale cloud clusters,unclear details in cloud images,and gradually blurred prediction results,leading to suboptimal forecasting performance.To effectively extract spatiotemporal information from satellite cloud images and forecast the development of mesoscale cloud clusters,this study utilized FY-4A infrared cloud images,focusing on the central and eastern regions of China with Hunan as the center.From the perspective of spatiotemporal sequence prediction,we proposed a convolutional gated recurrent attention fusion network(CGAFNet)and introduced primary and secondary loss(PaSLoss)as the model’s loss function.An encoder-decoder structure was constructed to better extract spatiotemporal information from satellite cloud images.To validate the effectiveness of the network framework,we conducted comparative experiments with three typical networks.The results show that CGAFNet achieved a root mean squared error of 10.00 K,a structural similarity index of 0.74,and a peak signal-to-noise ratio of 31.43 in the cloud image extrapolation task.Outperforming other networks across various metrics,the model accurately predicted the evolution of cloud clusters,demonstrating that this method can achieve more accurate prediction accuracy and possesses good generalization ability.

关 键 词：卫星云图 临近预报 时空序列预测 融合网络 注意力机制 

分 类 号：P456.1[天文地球—大气科学及气象学]