Attention Mechanisms in Semantic Convolutional Neural Networks for Flood Area Segmentation

Flood area segmentation remains a significant challenge in remote sensing, primarily due to complex textures, lighting variations, and shadow effects in flood imagery. This study introduces Attention-SCNN, a novel semantic convolutional neural network framework enhanced with spatial and channel attention mechanisms. These modules refine feature maps by prioritizing relevant regions while suppressing noise, addressing critical segmentation challenges. Using the Flood Segmentation Area dataset, the proposed model achieved superior performance metrics, including an IoU of 67.04%, Dice Coefficient of 80.26%, and Accuracy of 84.89%, outperforming state-of-the-art models such as Fast-SCNN, Mobile Fast-SCNN, and PSPNet. Comparative analyses revealed that AttentionSCNN effectively delineates flood boundaries, reduces false positives, and demonstrates robust segmentation capabilities under adverse conditions. Furthermore, its lightweight architecture ensures computational efficiency, enabling realtime flood monitoring in resource-constrained environments. These findings underscore the potential of Attention-SCNN for advancing disaster management systems and highlight directions for future research, including the integration of multi-modal data for enhanced segmentation performance.