基于生成对抗网络的三维云构建方法  

3D Cloud Construction Method Based on Generative Adversarial Network

作　　者：程文聪 王志刚[1] 张文军[1] 史小康[1] CHENG Wen-cong;WANG Zhi-gang;ZHANG Wen-jun;SHI Xiao-kang(Beijing Aviation Meteorological Institute,Beijing 100085,China)

机构地区：[1]北京航空气象研究所,北京100085

出　　处：《计算机仿真》2022年第8期33-40,179,共9页Computer Simulation

基　　金：高分对地观测专项项目(GFZX0402180102)。

摘　　要：针对航空气象服务保障中的云信息需求,提出一种基于深度学习的空域三维云构建方法。利用深度生成对抗网络模型从卫星云图通道数据中推算出关注区域内各个地理网格单元上的云顶高、云底高和云量参数数据,基于网格单元区域的云顶高和云底高参数在三维显示平台中构建出云的几何立体轮廓,再使用粒子系统根据云量对云立体轮廓内部进行填充,从而构建具有较高仿真度的三维云以支持航空气象服务保障工作。基于欧洲中期天气预报中心数值模式分析场产品和风云4A气象卫星数据进行了相关实验,实验结果以及演示验证软件均表明,所提方法可以有效构建实际空域的三维云仿真模型并用以支持航空气象服务保障工作。Driven by the requirement of the cloud information in aviation meteorological service and support,to create the 3D cloud of the airspace,a method based on deep learning is proposed.We can inference the cloud top height,cloud bottom height and cloud amount in each geographical grid cell of the concerned area by using the deep generative adversarial network model.The geometry shape of cloud in each grid cell was constructed in the 3D display platform based on cloud top height and cloud base height.Then we filled the interior geometry of the cloud with a particle system according to the cloud amount,so as to construct the 3D cloud with high fidelity to support the work of aviation meteorology.We conducted experiments based on the re-analysis products of the European Centrefor Medium-Range Weather Forecastsand FY-4A satellite cloud date.The experiments and the demonstration software show the proposed method can be effective to build a 3D cloud of the airspace and support the work of the aviation meteorology.

关 键 词：深度学习 生成对抗网络 气象 云参数 三维仿真 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]