Enhanced deep learning-based water area segmentation for flood detection and monitoring

Abstract

This study presents a hybrid architecture tailored for semantic segmentation challenges, mainly targeting the water area extraction for flood detection and monitoring. The model integrates an efficient transformer-based encoder, utilizing an efficient multi-head self-attention module for capturing hierarchical feature maps through a ‘downsample-upsample’ strategy. The proposed decoder architecture comprises one feature refinement head block and three CNN-based dual-branch context blocks. The convolutional block attention module is employed within the feature refinement head block to refine feature representation. The depth-wise separable atrous spatial pyramid pooling module is central to this architecture, facilitating efficient multi-scale contextual information capture. Compared to the state-of-the-art models, our model and the PSPNet model obtained the highest precision, recall, and F1-scores of above 80%, and mIoU surpassing 70%. The proposed method outperformed PSPNet in recall, F1-score, mIoU, and pixel accuracy, albeit with a slight deficit in precision. In terms of scale and efficiency, compared to the PSPNet model, our model has lower complexity and slightly higher inference speed, highlighting its effectiveness and efficiency in the water area segmentation for flood detection. The source code is available at https://github.com/manhhv87/mmsegmentation.git.