多体制遥感卫星成像数据高精度处理新方法  

A new method for high precision processing of multi-system Earth observation satellite data

作　　者：付琨[1,2,3,4] 仇晓兰 韩冰[1,3] 孙显 FU Kun;QIU Xiaolan;HAN Bing;SUN Xian(Key Laboratory of Geo-information Processing and Application System(GIPAS),Chinese Academy of Sciences,Beijing 100190,China;Key Laboratory of Network Information System Technology(NIST),Chinese Academy of Sciences,Beijing 100190,China;Aerospace Information Research Institute,Chinese Academy of Sciences,Beijing 100094,China;University of Chinese Academy of Sciences,Beijing 100049,China)

机构地区：[1]中国科学院空间信息处理与应用系统技术重点实验室,北京100190 [2]中国科学院网络信息体系技术重点实验室,北京100190 [3]中国科学院空天信息创新研究院,北京100094 [4]中国科学院大学,北京100049

出　　处：《遥感学报》2023年第7期1511-1522,共12页NATIONAL REMOTE SENSING BULLETIN

基　　金：科技创新2030——“新一代人工智能”重大项目(编号:2022ZD0118402)。

摘　　要：光学和SAR等对地观测卫星需要经过成像、辐射/几何校正等处理和不断的时序积累,才能为计算机解译提供精度高、稳定性好、时间持续的数据和特征。传统中低分辨率对地观测卫星通常基于地理网格内地物目标电磁波散射特性简化为理想点目标的假设,进行逐像素处理。然而,高分宽幅、大斜视、多通道等新体制卫星的工作模式更加复杂,其数据处理对星地全链路各环节产生的误差非常敏感,对成像参数标定或估计的精度提出了更高要求,此时基于理想点目标假设来进行参数估计、成像及校正处理的方式已难以满足处理精度要求。并且,近年来多体制卫星组网式协同和融合应用的新发展,也使得当前的理想点目标假设难以表征和建模多源多时相数据特征。为此,本文提出了多体制遥感卫星成像数据高精度处理的新方法,首先创新提出了“超像素”的概念和表征理论框架,建立了基于超像素的精确成像模型,然后通过挖掘超像素稳定特征并借鉴生成对抗学习机制,实现了星地全链路高耦合成像参数的高精度估计和持续精化,有效提升了多体制遥感卫星成像数据产品的精度,为计算机解译提供了好的数据产品输入。Earth observation satellites,such as optical and SAR satellites,require processing such as imaging,radiometric/geometric correction,and continuous accumulation in order to provide high-precision,stable,and time continuous data and features for computer interpretation.Traditional medium and low resolution Earth observation satellites typically perform pixel-by-pixel processing based on the assumption of ideal point targets,which means that the ground object grid has a invariant time-frequency characteristic.However,the working modes of advanced satellite systems,such as high-resolution,wide-swath,large squint angle,and multi-channel,are more complex,and their data processing is very sensitive to the errors generated in the whole chain of the satellite to ground,which puts higher requirements on the accuracy of imaging parameter calibration or estimation.Hence,the method of assuming sensor pixels as ideal point targets for parameter estimation,imaging,and correction processing is no longer able to meet the processing accuracy requirements.Moreover,in recent years,the new development of multi-system satellite network collaboration and fusion applications has made it difficult to characterize and model the features of multi-source and multi-temporal data based on the current ideal point target assumption.To this end,this article proposes a new method for high-precision processing of multi-system remote sensing satellite imaging data.Firstly,the concept and characterization theory of“Hyper-pixel”are proposed,and an accurate imaging model based on hyper-pixels is established.Then,by mining stable features of hyper-pixels,and inspired by generative adversarial learning mechanisms,high-precision estimation and continuous refinement of high coupling imaging parameters are achieved.This effectively improves the accuracy of multi-system remote sensing satellite imaging data products,and provides better data input for computer interpretation.

关 键 词：遥感卫星 超像素 成像处理 辐射校正 几何校正 深度学习 

分 类 号：TP701[自动化与计算机技术—检测技术与自动化装置] P2[自动化与计算机技术—控制科学与工程]