Abstract:

In order to address the challenges of inaccurate region segmentation and insufficient detail restoration in flood disaster recognition models within complex urban environments, AttResU-Net, an enhanced U-Net semantic segmentation model integrating residual networks and a self-attention mechanism was proposed. Building upon the classical U-Net architecture, the model employed a deep residual network as the encoder to strengthen feature representation. Simultaneously, self-attention mechanisms were incorporated into the decoder to enhance response capability on key flood-related regions. A comprehensive training and testing pipeline was established. The improved AttResU-Net was trained and evaluated on the FloodNet dataset, which contains diverse and complex urban environmental categories. Quantitative metrics and qualitative visual results demonstrate the model's superior performance, achieving a mean pixel accuracy (mPA) of 79.75%, pixel accuracy (PA) of 90.01%, and mean precision (mPrecision) of 81.78%. Comparative experiments against state-of-the-art models reveal that AttResU-Net attains significantly higher segmentation accuracy and global recognition capability, particularly for urban features such as trees, water bodies, roads, and buildings.

Key words: U-Net, flood disaster, image recognition, image segmentation, self-attention mechanism, residual network