FEA-Net: Feature-Enhanced Attention-UNet for Hierarchical Classification and Grading of Postdisaster Building Images

With the deep integration of deep learning and high resolution remote sensing technologies, building damage assess ment methods based on semantic segmentation show significant application value in post-disaster emergency response. However, the existing methods face two core challenges: the imbalance problem of category distribution in the dataset, which leads to the model decision biased towards the majority of categories, and the progressive damage feature ambiguity problem, which leads to cross-category misjudgment. Therefore, this paper proposes a feature-enhanced Attention-UNet-based image classification method for disaster buildings to improve the reliability of post disaster damage assessment. The method adopts a ResNet34 en coder on the basis of the UNet framework and introduces multi scale difference feature extraction and enhancement modules (Multi-scale Contextual Boundary Module (MCBM) and Bound ary Dynamic Feature Module (BDF)). These modules enhance the model's ability to represent damaged regions at different scales, incorporate an attention mechanism to guide the model to focus on changing regions. Meanwhile, a joint loss function combining DICE loss and FocalLoss is used to optimize the model's learning ability in the case of category imbalance, and the method is validated on the xBD disaster building classification dataset. The experimental results show that the method achieves an F1-Score of 74.8% and an IoU of 60.4% on the xBD dataset, outperforming existing benchmark methods. It can more accurately identify the damage of disaster buildings and improve the reliability of post disaster assessment. Compared with generic U-Net variants, our model targets post-disaster building assessment with a lightweight attention–difference–fusion design that sharpens progressive damage boundaries and mitigates class imbalance. It delivers consistent gains on xBD and strong generalization on LEVIR CD under favorable accuracy–efficiency trade-offs.