FloodFusionNet: a multi-input multi-output neural network for flood-water detection and mapping using SAR and polarization data

Abstract

Flooding poses significant challenges worldwide, resulting in substantial economic losses and endangering human lives. Traditional waterbody detection methods often rely on passive sensing and optical images, which are susceptible to weather conditions and limited temporal availability. In contrast, Synthetic Aperture Radar (SAR) data from satellites, such as Sentinel-2, enable continuous and reliable flood analysis. This paper introduces FloodFusionNet (FFN), a novel multi-input multi-output network specifically designed for flood-water detection and mapping. Our approach integrates active sensing imagery and polarization channels, leveraging a newly developed augmentation module, Neural Patch Augmentation (NAP), to ensure class balance, preserve high-dimensional features, and generate robust training patches. The FFN encoder-decoder architecture fuses spatial and spectral information while employing pixel-wise convolutions, skip connections, and custom down/up sampling blocks to enhance feature representation. Experimental evaluations demonstrate that FFN achieves a Mean Intersection over Union (MeanIoU) above 0.95 on challenging test datasets, outperforming conventional deep learning approaches. Although the model occasionally struggles with small waterbody delineation, its overall accuracy, adaptability, and performance in heterogeneous and complex scenarios mark a significant advancement in remote sensing-based flood monitoring.