Self-Attention Driven Decoder for SAR Image-based Semantic Flood Zone Segmentation

Floods are destructive natural calamities that endanger people's lives, infrastructure, and the environment. Flood detection that is timely and accurate can help with disaster management and save lives. Flood semantic segmentation from remote sensing data such as SAR images has gained popularity due to recent advances in computer and memory capacity. In this context, encoder-decoder based CNN architectures are widely adopted. However, the inter-class feature sharing in these images makes distinguishing flood-prone zones challenging. To properly extract features and decode class labels, robust encoders and decoders are necessary. A common strategy that dramatically upsamples the decoder's feature maps, in particular, typically causes information loss and gives subpar segmentation results. In this context, the current work proposes a novel decoder that uses a self-attention layer to improve the feature maps before assigning class labels. The proposed method has been statistically and qualitatively verified using publicly available dataset.