A Flood Segmentation Model Enhanced by Residual Squeeze-and-Excitation (R-SE) Blocks

Abstract

Timely flood detection and segmentation play an important role in disaster management and recovery processes by enabling effective interventions. In this study, a segmentation model with ResNet-50 infrastructure is proposed, which provides high accuracy. Our proposed model achieves high segmentation accuracy on the Flood dataset by reducing computational cost compared to standard convolutions and by using depth-separable convolutions. In order to make learning more flexible and reduce overfitting, our model uses the PReLU (Parametric ReLU) activation function, which allows learning from negative values. Squeeze-and-Excitation (SE) blocks, which strengthen feature learning by highlighting important information, are also integrated into the model. In addition, Feature Enhancement Blocks contribute to the production of more accurate and detailed segmentation maps. The effectiveness of the model is verified using the Flood dataset, where it is evaluated with various measurement metrics. The results show that the proposed model outperforms popular segmentation architectures such as FCN, SegNet and U-Net. In particular, the average IoU value obtained is 85.37%, demonstrating the high overall accuracy of the model. This work provides a valuable contribution to the field of flood detection and provides a solid foundation for future systems that aim to improve segmentation accuracy for real-time disaster response.