Global flood detection using Sentinel-2A-MSI by combining histogram-based and regional methods compared with an automated RandomForest approach

Remote sensing can be a crucial source of up-to-date crisis information such as flood extent. Since the Intergovernmental Panel on Climate Change (IPCC) expects an increase of heavy precipitation events for the mid-latitude and tropical regions until the end of this century, it is likely that the number of flood events will increase as well. Therefore new services embedded in the Global Monitoring for Environment and Security (GMES) program like rapid flood mapping need to be developed to account for an increase of events. The study presents a global flood detection method based on multispectral Sentinel-2A image data using common indices for water delineation, empirical and adaptive image histogram thresholding, morphological and regional operators. In addition, one of the most promising ensemble classifier, the RandomForest (RF) classifier is evaluated. Since manual selection of training data is time-consuming, an automated sampling design based on the introduced Water Rank Layer (WRL) is presented. The thematic accuracy of the proposed methods was assessed at three test sites. The histogram based approach showed overall good results with Overall Accuracy (OA) ranging from 0.875 to 0.983 and Cohen’s kappa coefficients ranging from 0.705 to 0.901. A measurable increase of Overall Accuracy (OA) could be observed for the RF approach with accuracies ranging from 0.929 to 0.995 and kappa coefficients ranging from 0.679 to 0.967