基于生成对抗网络和卫星数据的云图临近预报  

Nowcasting of Cloud Images Based on Generative Adversarial Network and Satellite Data

作　　者：肖海霞 张峰 王亚强[4] 唐飞 郑玉 Xiao Haixia;Zhang Feng;Wang Yaqiang;Tang Fei;Zheng Yu(Key Laboratory of Transportation Meteorology of China Meteorological Administration,Nanjing Joint Institute for Atmospheric Sciences,Nanjing 210041;Shanghai Qi Zhi Institute,Shanghai 200232;Department of Atmospheric and Oceanic Sciences,Intitute of Atmospheric Sciences,Fudan University,Shanghai 200438;Chinese Academy of Meteorological Sciences,Beijing 100081)

机构地区：[1]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,南京210041 [2]上海期智研究院,上海200232 [3]复旦大学大气与海洋科学系/大气科学研究院,上海200438 [4]中国气象科学研究院,北京100081

出　　处：《应用气象学报》2023年第2期220-233,共14页Journal of Applied Meteorological Science

基　　金：国家重点研发计划(2021YFB3900400);中国气象科学研究院基本科研业务费专项资金(2020Z011,2021Y010);江苏省气象局青年项目(KQ202115)。

摘　　要：利用风云四号气象卫星A星(FY-4A)红外云图,基于生成对抗网络方法,提出了红外云图外推预报模型,实现了华东区域未来3 h的云图预报,预报的时空分辨率分别为1 h和4 km。结果表明:该外推模型预报的云图可较好描述云团移动、发展和减弱趋势,对研究区域内云团的强度、位置和形态得到较为理想的预报效果。为了验证提出的云图外推模型的有效性,将其与光流法和轨迹门控循环单元模型进行比较。对比试验结果表明:该云图外推模型具有最优的预报效果,说明使用生成对抗网络进行云图外推具有较好的可行性,能有效应用于气象业务中监测云团的生消和移动并提前预警灾害性天气的发生,为天气预报提供重要的参考依据。Satellite cloud images contain abundant information,which can reflect daily weather conditions.Nowcasting based on cloud images can strengthen the application of satellite data in the early warning and forecasting of severe weather.At present,the cloud images predicted by most nowcasting methods based on artificial intelligence are not accurate enough and the lead time is limited.Thus,it’ s necessary to improve the accuracy and period validity of cloud images in nowcasting.Using the infrared cloud image data of Fengyun-4A(FY-4A) and the generative adversarial network(GAN) method,an infrared cloud image extrapolation nowcasting model is proposed.The cloud images in the next 3 hours in East China are predicted by the proposed model,and the spatial resolution of predicted cloud images is 4 km and the temporal resolution is 1 hour.The results show that the evaluation values of SSIM(structural similarity),PSNR(peak signal to noise ratio) and RMSE(root mean square error) predicted by the proposed GAN-based cloud images extrapolation model are 0.75,20.92 and 10.00 K,respectively.In addition,the MAE(mean absolute error),MSE(mean squared error),and SSIM are chosen as loss function and analyzed,aiming to verify the rationality of the loss function in the generator.Comparative experiments of different loss functions show that it is reasonable and effective to choose SSIM combined with MAE as the loss function.Furthermore,to verify the effectiveness of the proposed GAN-based model,the prediction results are compared with those of the optical flow method and the TrajGRU model with the GAN-based model.The experimental results show that the cloud image extrapolation model based on GAN has the superior prediction performance,with the highest SSIM and PSNR,and the lowest RMSE within 1-3 h of cloud images nowcasting.The observational examples show that the cloud images predicted by the proposed model can well describe the movement,development and dissipation trend of clouds.Meanwhile,the experiments obtain accurate prediction pe

关 键 词：卫星云图 临近预报 深度学习 生成对抗网络 

分 类 号：P412.27[天文地球—大气科学及气象学] P45