【作者】 刘波；

【导师】 高峰； 成巍；

【作者基本信息】 哈尔滨工程大学 ， 电子信息（专业学位）， 2023， 硕士

【摘要】 降水是全球水循环的基本组成部分,它是气象学、气候学和水文学中水循环的关键水文变量。但降水研究存在着严重的时空不确定性以及区域和海拔等挑战,并且地面降水观测受人类活动和地形条件的影响,存在着分布不均甚至人口稀疏和海上地区无分布的问题。地球同步卫星可提供全天候、区域性观测,因此引入卫星观测进行降水反演的任务显得尤为重要。针对高质量卫星降水反演数据集的需求,本文结合FY-4A卫星的先进的静止轨道辐射成像仪AGRI(Advanced Geostationary Radiation Imager)观测和GPMIMERG半小时累计降水数据,将端到端的深度学习算法应用到降水反演的研究中。端到端的深度学习算法显示了作为一种可全球应用的操作技术的潜力,为FY-4A气象卫星的应用提供了一个新的思路。本文的研究内容和主要工作如下:(1)对国内外卫星遥感反演降水技术进行总结与分析,考虑如今的反演手段主要基于卫星红外通道,没有利用到可见光通道。为了充分利用卫星多通道观测的优势,本文结合可见光和红外通道数据,在研究中区分夜间和白天场景,将端到端的改进U-Net网络应用于降水反演任务中。结合降水数据的分布特点确定研究方案,将研究内容分为降水识别和降水量反演两部分。(2)确定研究区域,并对原始数据进行预处理,以建立用于降水识别和降水量反演的训练样本库。然后,提出基于皮尔逊相关性方法进行通道选取,区分夜间和白天场景挑选AGRI数据中与降水相关性较高的通道组合作为通道方案,为后文模型方案的设计奠定基础。(3)结合降水量分布的特点,对基础U-Net网络进行改进,设计了基于改进U-Net网络的降水识别模型。该模型基于AGRI观测利用改进U-Net网络进行降水识别,根据不同通道方案区分夜间和白天场景,分场景进行实验。实验结果基于GPM-IMERG降水数据进行指标分析,并将加和后的结果与PERSIANN-CCS产品相比较,以验证方案的有效性。相比于PERSIANN-CCS产品,改进U-Net模型在夜间和白天场景反演结果的相关性、临界成功指数、命中率和准确度更高,并且虚警率更低。(4)在降水量月份分析的基础上,设计了基于改进U-Net网络的降水量反演模型。模型基于AGRI观测利用改进U-Net网络反演降水量,实验结果基于GPM-IMERG降水数据进行分析,并将加和后的结果与PERSIANN-CCS产品相比较。相比于PERSIANNCCS产品,改进U-Net模型在夜间和白天场景反演结果的相关性、临界成功指数更高,并且MAE和RMSE指标更低,白天场景表现优于夜晚场景。更多还原

【Abstract】 Precipitation is a fundamental component of the global water cycle,and it is a key hydrological variable of the water cycle in meteorology,climatology and hydrology.However,there are serious spatio-temporal uncertainties and regional and altitude challenges in precipitation research,and surface precipitation observation is affected by human activities and topographic conditions,and there are problems of uneven distribution or even sparse population and no distribution in maritime areas.Geostationary satellites can provide all-weather and regional observation,so it is particularly important to introduce satellite observation for precipitation inversion.In response to the demand for high-quality satellite precipitation inversion datasets,this article combines the Advanced Geostationary Radiation Imager(AGRI)observation of FY-4A satellite and GPM-IMERG half hour cumulative precipitation data to apply end-to-end deep learning algorithms to the research of precipitation inversion.The endto-end deep learning algorithm shows its potential as an operational technology that can be applied globally,and provides a new idea for the application of FY-4A meteorological satellite.The research content and main work of this paper are as follows:(1)Summarize and analyze the satellite remote sensing retrieval precipitation technology at home and abroad,considering that today’s retrieval methods are mainly based on the satellite infrared channel,and do not use the visible light channel.In order to make full use of the advantages of satellite multi-channel observations,this paper combines visible light and infrared channel data to distinguish night and day scenes in the study,and applies the end-toend improved U-Net network to precipitation retrieval tasks.Based on the distribution characteristics of precipitation data,the research program is determined,and the research content is divided into two parts: precipitation identification and precipitation inversion.(2)Determine the research area and preprocess the original data to establish a training sample library for precipitation identification and precipitation inversion.Then,the channel selection based on the Pearson correlation method is proposed,and the channel combination with high correlation with precipitation in the AGRI data is selected as the channel scheme by distinguishing night and daytime scenes,which lays the foundation for the design of the model scheme in the following.(3)Combined with the characteristics of precipitation distribution,the basic U-Net network was improved,and the precipitation identification model based on the improved UNet network was designed.This model is based on AGRI observation and uses improved UNet network to identify precipitation.Night and day scenes are distinguished according to different channel schemes,and experiments are carried out in different scenes.The results were analyzed based on GPM-IMERG precipitation data,and the results were compared with PERSIANN-CCS product to verify the effectiveness of the scheme.Compared with PERSIANN-CCS,the improved U-Net model has higher correlation,critical success index,hit rate and accuracy of night and day scene inversion results,and lower false alarm rate.(4)Based on the analysis of precipitation month,the precipitation inversion model based on the improved U-Net network is designed.The model was based on AGRI observation and the improved U-Net network was used for precipitation inversion.The experimental results were analyzed based on GPM-IMERG precipitation data,and the added results were compared with PERSIANN-CCS products.Compared with PERSIANN-CCS,the improved U-Net model has higher correlation and critical success index of night and day scene inversion results,and lower MAE and RMSE indexes,and the performance of day scene is better than night scene.更多还原