Table 1 General statistics of the training, validation and test splits of the WorldFloods  dataset. Since raw images from S2 can be many megapixels in size, we tile each image into 256-pixel square patches. The training set distribution has a higher percentage of cloudy pixels compared with the validation and test datasets; this is because we were interested in distinguishing water/flood pixels whereas detecting clouds is a byproduct of the model.

From: Towards global flood mapping onboard low cost satellites with machine learning