SegNet-ATT: Cross-Channel and Spatial Attention-Enhanced U-Net for Semantic Segmentation of Flood Affected Areas

Pranav Sunil, Soumen Sinha, Rahul Roy

Abstract: Rapid and accurate detection of flood-affected areas is crucial for effective disaster response and relief operations. Traditional methods often fail to provide the necessary speed and precision, underscoring the need for advanced technological solutions. This research introduces SegNet-ATT, an innovative approach utilizing an attention-augmented U-Net architecture to semantic segment flood-affected areas in aerial imagery. The SegNet-ATT model incorporates self-attention, cross-attention, channel attention, and spatial attention mechanisms to enhance feature extraction and contextual understanding. These attention mechanisms enhance the model’s ability to discern subtle distinctions between water and land, which is critical for effective flood segmentation. These regions are then utilized to determine drop zones for potential rescue operations. The results demonstrate the effectiveness of the SegNet-ATT model, which achieved an impressive accuracy of 89.46%. Comparing this with other models like DeepLab v3, we identify that although they demonstrate marginally better performance, they have steeper computational requirements leading to longer runtimes. This research demonstrates the potential of leveraging attention-enhanced deep learning models for quicker and more effective disaster response.