WHUL8-CDb, a block-level labelled cloud detection dataset collected from Landsat 8

WHUL8-CDb was collected for the training of weakly-supervised cloud detection methods. It includes 500 globally distributed Landsat 8 Level-1C images (Fig. 1) collected following the distribution of the L8 Biome dataset as it comes to biomes and dominant land cover types. The motivation is that clouds are similar with each other in different regions, while the backgrounds are very different. In order to increase the diversity of training samples, we selected images with more background pixels by setting the cloud percentage from 0 to 50 when searching images. To handle all cloud situations, 5 images full of clouds were collected and included in WHUL8-CDb.