Deep learning based cloud detection for fy-4a/agri snow mapping considering cloud and snow spectral characteristics

Cloud detection is the first step in remote sensing surface parameter retrieval. Due to the similar spectral properties of cloud and snow, cloud products commonly used in snow monitoring sensors have a certain degree of cloud and snow misjudgment problem. With the launch of a new generation of geostationary satellite (such as China's FY-4A), its time resolution is 15 minutes, and high-frequency observations make accurate cloud and snow identification possible. This study utilizes the high-frequency and multispectral observation characteristics of FY4A, combining the multi band threshold method with deep learning algorithm, to fully explore the spectral and texture differences of cloud and snow, as well as the characteristics of rapid cloud changes, and achieve high-precision cloud detection. Then, the CALIPSO data is used to evaluate the accuracy of the new algorithm's detection results. From the results, the new algorithm for cloud and snow recognition is more accurate and consistent with CALIPSO observations. In terms of specific accuracy indicators, the cloud hit rate(CHR) increase 1.07% and the false alarm rate(FAR) decrease 5.15%. At the same time, both in terms of single scene or daily composite results, the proportion of cloud cover has decreased about 20%, and the proportion of snow cover can increase by up to about 15%. This laid a solid foundation for high-precision fractional snow cover retrieval and spatiotemporal reconstruction in the future.