Abstract: In China,there exists an imbalance and inadequacy in the distribution of meteorological radar,with particularly significant gaps in coverage in the southwestern region and over the sea.While geostationary meteorological satellites possess a broader range of observation and high temporal and spatial resolution multispectral observational results,utilizing satellite data for radar reflectivity inversion can to some degree compensate for deficiencies in radar networks.However,the current inversion process faces challenges in effectively extracting the complex,multichannel characteristics of satellite data in both the channel and spatial dimensions during data processing.In light of these issues,the business process of radar reflectivity inversion utilizing geostationary satellites was explored.The high-level semantic features of satellite data was extracted through ResNet,and then spatial and channel attention modules was utilized to aggregate features pertinent to radar echoes in both spatial and channel dimensions and eliminate interference features from non-precipitation clouds.Through a comparison experiment utilizing multiple indicators,the research in this paper had been verified to have a notable improvement in the precision of echo inversion across various sizes,highlighting the algorithm's superiority.

Key words: business process, radar echoes, remote sensing, deep learning, Hamawari-8

摘要： 我国气象雷达分布存在着不均衡,不充分问题,在西南地区以及海上区域存在着很大的雷达组网覆盖盲区。而静止轨道气象卫星具有更大的观测范围以及高时空分辨率的多光谱观测结果,通过开展基于卫星数据的雷达反射率反演业务过程可以在一定程度上能够弥补雷达组网空缺的问题。但现有的反演业务过程在数据处理阶段很难在通道以及空间维度上有效的提取复杂的多通道卫星数据特征。针对上述问题,探究了基于静止轨道卫星的雷达反射率反演业务过程,首先通过ResNet提取卫星数据的高层次语义特征,然后使用空间注意力模块、通道注意力模块在空间、通道两个维度上聚合与雷达回波相关的特征,并消除非降水云团的干扰特征。最后,通过多指标的对比实验验证了本文研究在不同大小的回波反演精度上均有一定提升。

关键词: 业务过程, 雷达回波, 遥感, 深度学习, 葵花8号卫星