基于朴素贝叶斯的FY-4A/AGRI云检测方法

光学遥感云检测是定量遥感和遥感应用的基础, 尝试将朴素贝叶斯的机器学习方法应用于风云四号气象卫星A星(FY-4A)搭载的先进静止轨道辐射成像仪(AGRI)红外通道数据云检测。因辐射物理方法的云检测采用可见光通道导致存在日夜不连续现象，仅选取FY-4A/AGRI载荷7个红外通道的光谱数据，构建10种特征分类，利用正交偏振云-气溶胶激光雷达(CALIOP)与FY-4A/AGRI时空匹配数据，对不同地表类型和不同季节的数据集进行分类训练和验证。与CALIOP数据交叉验证显示除积雪上空云识别准确率约为81%，深海、浅水、陆地和荒漠上空的云识别准确率均高于92%，误判率基本低于10%，总体云识别精度达到90%；与2021年10月和2022年1，4，7月MODIS 2级云检测产品比对，深海、浅水云识别准确率均在88%以上，误判率分别低于3%和10%，夏季云识别效果最佳，总体云识别准确率高达90%。云检测结果不仅得到云、可能云、可能晴空和晴空4种分类结果，还得到每种特征和综合特征云检测分类器的不确定性概率值，这为云和地表相关检测产品提供重要参考。

FY-4A/AGRI Cloud Detection Method Based on Naive Bayesian Algorithm

Optical remote sensing cloud detection is the foundation for subsequent quantitative remote sensing and applications. A cloud detection method based on naive Bayesian algorithm is studied and applied to the advanced geostationary orbital radioimager (AGRI) on Fengyun-4A satellite. Cloud detection method considering radiation physics of visible light channels is discontinuous between day and night. To avoid the direct impact of solar radiation, the spectral data of 7 infrared channels loaded by AGRI are analyzed to construct 10 cloud detection feature classifiers. Using cloud polarized lidar with orthogonal polarization (CALIOP) data as the true value of cloud detection, and using its spatiotemporal matching data with AGRI, classification training and validation are conducted for datasets of different surface types and different seasons. The cloud detection results and CALIOP data cross-verification show that the cloud recognition accuracy over snow is about 81%, the cloud recognition accuracy rate over the deep sea, shallow water, land and desert is higher than 92%, the false positive rate is basically less than 10%, and the overall cloud recognition accuracy reaches 90%. Compared with MODIS level 2 cloud detection products in October of 2021 and January, April and July of 2022, the recognition accuracy rate of deep-sea and shallow water clouds is above 88%, and the false positive rate is lower than 3% and 10%, respectively. The overall cloud recognition accuracy rate in four seasons is more than 86%, of which the summer cloud recognition effect is the best, and the overall cloud recognition accuracy rate is as high as 90%. The recognition effects of the method are good during both day and night, ensuring not only the accuracy of day and night cloud detection, but also the continuity of cloud detection in the morning and evening transition zone. Due to the use of dynamic surface type files and sufficient training sample sizes for deep and shallow waters, the overall cloud recognition accuracy of the method is relatively ideal in four seasons, with the best performance in summer and autumn. The cloud recognition accuracy of deep and shallow water is generally high, but there are still omissions and misjudgments. The method can output classification results of cloud including probable cloud, probable clear sky, and clear sky, and it also outputs the uncertainty probability value of each feature and a comprehensive feature cloud detection classifier, which can provide important reference for cloud and surface related detection products.