Flood-MATE: A Flood Segmentation Model in Urban Regions through Adaptation of Mean Teacher and Ensemble Approach

Abstract

Flood disasters remain one of the most recurring natural phenomena worldwide, resulting from excessive water flow submerging land for an extended period of time. The escalating occurrences of floods, particularly in urban areas, can be attributed to climate change, extreme weather patterns, uncontrolled urbanization, and complex geographical conditions. To mitigate the destructive impacts, such as loss of life and economic ramifications, automatic flood analysis and remote-sensing imagery segmentation offer valuable decision-making insights. However, the segmentation process for flood detection faces challenges due to the scarcity of labelled data and diverse resolutions, including medium resolution data. In response, the authors propose Flood-MATE, a novel semi-supervised learning approach based on the mean-teacher model. Our approach leverages the deep learning architecture and introduces a new loss function scenario for training. The dataset utilized in this study comprises SAR images of Sentinel-1 C-band that have undergone thorough processing. Promisingly, the results demonstrate a 4% improvement in the IoU metric compared to the baseline method employing pseudo-labelling.