MSflood: A Multi-Sources Segmentation For Remote Sensing Flood Images

Flood detection requires the identification of areas that have been affected by flood events. This is essential for monitoring natural disasters, which is a crucial aspect of building sustainable cities. In this paper, we propose a novel segmentation method to identify flood segments using Synthetic Aperture Radar data, Digital Elevation model in conjunction with hydrology maps. The classical deep learning semantic segmentation methods face a challenge when dealing with multiple data sources. To address this, the encoder-decoder architecture has been updated to enhance the recognition of multi-sources patterns by combining high-level abstract features with low-level spatial features. This work introduces a novel approach called MSFlood for semantic segmentation, employing a multi-sources encoder architecture to extract features from various input sources. To achieve this, a Fusion Block with attention for multi-sources data is added to merge the extracted features. The performance of our method is evaluated using the MMFlood dataset, and it surpasses the state-of-arts segmentation methods.