Scale matching based remote sensing image cloud detection in southwest mountainous areas

Abstract

Sample quality is the key to automated cloud detection from regional remote sensing images, and scale is one of the major impediments to sample quality control. In this paper, we select the southwest mountainous area in China, which is fragmented, cloudy, and rainy, as the study area. We proposed a method for constructing a cloud detection dataset based on the idea of downscaling and the spectral characteristics of vegetation. Finally, we validated the dataset by the U-Net+ deep learning model. The experimental results show that the cloud detection accuracy reaches 95.11% when using the dataset constructed in this paper, which is approximately 40% higher than the cloud detection accuracy with large-scale samples. Additionally, it reduced the workload of masking a large number of samples for a specific region and realizing the possibility of efficient cloud detection in the region.