融合多尺度视网膜图像增强的动态云检测

云检测是气象卫星各类定量遥感产品的基础,无论是以云图为基础的天气分析还是以去云为前提的各类大气和地表参数反演、沙尘火情等灾害检测,都需要对遥感影像中的云进行准确识别,尤其是薄云和云边缘等细节识别。针对静止气象卫星(以Himawari-8为例)精细化云检测,本文提出了一种基于多尺度视网膜图像增强的动态云检测算法。该算法基于云层与背景信息辐射特征不同的原理,构建可见光和红外波段的晴空辐射背景场,通过多尺度图像增强和最大类间差方法对辐射差值进行云细节信息的增强和提取。利用2021-2022年的75景MODIS云检测产品作为验证数据进行算法精度验证,整体上算法精度达到91.13%,召回率为94.02%,精确率为86.71%,有较强的适用性和稳健性,且已经较好地支撑了近两年的定量遥感产品业务化应用。

Dynamic cloud detection based on multi-scale retinal image enhancement

Cloud detection is the basis of various quantitative remote sensing products of meteorological satellites. Whether it is weather analysis based on cloud images, inversion of various atmospheric and surface parameters, sand, dust, fire and other disaster detection based on cloud removal, accurate cloud detection in images, especially details such as thin clouds and cloud edges is required. Aiming at the refined cloud detection of geostationary satellites, this paper proposes a dynamic cloud detection method based on multi-scale retinal image enhancement. The clear sky radiation background field is applied to enhance and extract the cloud detail information of the radiation difference through the multi-scale image enhancement and the maximum inter-class difference method. This paper uses 75 MODIS cloud detection products from 2021 to 2022 as the verification data to verify the algorithm accuracy. The overall algorithm accuracy reaches 91.13%, the recall rate is 94.02%, and the precision rate is 86.71%. Overall, the algorithm has strong applicability and robustness. It is excellent and has well supported the commercial application of quantitative remote sensing products in the past two years.