Enhanced multi-temporal cloud detection algorithm for optical remote sensing images

Abstract

Remote sensing images have been a crucial data source for land cover mapping and other applications. However, optical remote sensing images are frequently contaminated with clouds. Clouds have caused several limitations in remote sensing applications through optical satellite. Although several approaches have been conducted for cloud detection, they still fail to distinguish bright surfaces, snow, and clouds, especially for seasonally snow-covered images. Therefore, we aim to develop a fast and universal cloud detection method, which can accurately detect clouds in complex areas. Considering that many sensors do not have a thermal infrared band, we only use the visible, near infrared, and short-wave infrared bands to detect clouds. The proposed method is expected to be used for a variety of satellite data. In this study, a multi-temporal cloud detection method was proposed for optical images. Given that snow and clouds have a big difference in short-wave infrared bands, we first developed an Enhanced Cloud Index (ECI) based on the spectral properties of the bands to distinguish them. Then, we proposed an Enhanced Multi-Temporal Cloud Detection (EMTCD) algorithm based on the ECI index and multi-temporal images to extract cloudy pixels. Finally, we tested and compared the algorithm with three classical cloud detection algorithms, namely, Function of mask (Fmask), Cloud Cover Assessment (CCA), and Multi-Temporal Cloud Detection (MTCD) algorithms, to verify the accuracy of the proposed algorithm. Landsat-8 images were used as the data source in this study. Given that many operational cloud detection methods had failed in complex areas, we selected four Landsat-8 OLI scenes in two test areas with typical seasonal snow cover and complicated land covers as our test data. The images were all obtained from 2015. The test areas were the northeast and southwest of China. Test results indicated that the ECI index can effectively distinguish snow and clouds. The ECI index of snow was higher than that of clouds. The EMTCD method performed well in cloud detection, which had the best cloud detection result with an overall accuracy of 93.2% compared with that of 81.85%, 66.14%, and 86.3% for the classic Fmask, MTCD, and CCA cloud detection methods, respectively. The ECI index is effective in distinguishing clouds and snow. The EMTCD algorithm can provide a good performance in cloud detection without using the thermal infrared band, even for seasonally snow-covered regions with complicated high brightness ground surface, which is always challenging for traditional cloud detection algorithms. However, the method is developed based on multiple images. Compared with single temporal methods, the proposed method still has some limitations.