单日的晴空合成图像,对于日常的地表监测等业务应用具有重要意义,针对风云四号B星快速扫描成像仪的1分钟连续成像序列数据,本文提出了一种基于二元高斯混合模型的晴空图像合成算法。算法假设固定地点反射率时序数据由云-晴空两种类型数据构成,分别满足高斯分布；在初始化云-晴空二元高斯分布参数后,贯序处理时间序列图像,判识新影像像素分属的云-晴空类型,并更新该地点云-晴空两种类型的平均值、标准差等参数；贯序处理全部日内影像数据后,以晴空类型反射率的平均值作为该地点晴空合成结果的反射率估计值。方法具有线性的时间和内存空间复杂度,晴空合成图像的有效晴空像素比例和图像信息熵逐渐增加；与典型晴空算法相比,具有更高的区分云-晴空的稳健特性和对云边缘阴影的过滤能力。较高频次的单日晴空合成图像可应用于植被、水体监测等日常业务的生态遥感应用方向。

Objective The clear-sky image synthesis in a single day is of great significance for daily water body recognition and other business applications. This paper proposes a clear sky image synthesis algorithm based on a binary Gaussian mixture model for the 1-minute continuous imaging sequence data of Geostationary High-speed Imager (GHI) of Fengyun-4B satellite. The GHI is the world"s first quantitative remote sensing instrument for day and night high-frequency imaging of geostationary orbit. It is designed for short-term approaching forecasting needs, with a total of 7 visible and infrared channels, providing Continuous observation of multiple spectral bands at 1-minute intervals in a 2000 × 2000km area. Method, in general, except for ice and snow, the reflectivity of clouds is higher than that of the underlying surface; Moreover, for the same location, the change in reflectivity is relatively small, while the change in reflectivity of passing clouds is significant. That is to say, the distribution of surface reflectance shows low mean and small variance (with relatively concentrated samples), while the distribution of cloud reflectance passing through shows high mean and large variance (with scattered samples). Therefore, the algorithm in this article assumes that the reflectance sequence samples of a single pixel within a single day are composed of clear sky reflectance samples and cloud reflectance samples, which respectively satisfy Gaussian distributions. The problem of synthesizing clear sky images is transformed into estimating the distribution parameters of clear sky reflectance samples. The algorithm is divided into three main steps: Initial Guess Value (Step I), Pixel Classification (Step C), and Parameter Update (Step U). In the initial parameter estimation, use the simple threshold method to initialize the cloud clear sky binary Gaussian distribution parameters; Sequential processing of time series images, using Gaussian distribution function to identify the cloud clear sky type to which new image pixels belong; Update the average, standard deviation, and other parameters of the two types of cloud clear sky at the location based on the new identification results. Finally, when the sequential processing of all intraday image data is completed, the average reflectance of the clear sky type is used as the estimated reflectance of the clear sky composite result for that location. Result the method has linear time and memory space complexity, and the effective clear sky pixel ratio and image information entropy of the clear sky composite image gradually increase; Compared with typical clear sky algorithms, it has higher robustness in distinguishing cloud clear sky and filtering ability for cloud edge shadows. Conclusion, high frequency single day clear sky composite images can be applied in ecological remote sensing applications such as vegetation, water environment, and water monitoring.