Multiscale semantic segmentation and damage assessment of village houses in post-flood scenarios using an enhanced DeepLabv3 with dual attention mechanism
Luyuan Wu, Yi Cheng, Aifang Su, Boyang Zhang, Zifa Wang, Jingbo Tong, Meng Li, Anqi Zhang
Abstract
Floods can cause varying degrees of damage to village houses, and rapid assessment of houses damage is crucial for post-disaster safety evaluations. To address this issue, the paper proposes a DeepLabv3-Dual Attention (DDA) model incorporating improved ResNet-50, Atrous Spatial Pyramid Pooling (ASPP), Selective Kernel(SK)-Attention, and Self-Attention mechanisms. In the DDA model, the ASPP structure is enhanced with Self-Attention, while the backbone network employs SK-Attention improved ResNet-50, enabling more effective feature extraction across both spatial and channel dimensions. Validated on the “7.20” heavy rain disaster dataset from Zhengzhou, DDA achieves optimal performance (learning rate: 0.01, batch_size: 16, epoch: 88) with 87.5% accuracy, 98.5% global accuracy, 77.8% MIoU, and 86.5% F1-score, outperforming DeepLabv3+, Swin-Unet, U-Net, YOLO baseline models. Ablation studies confirm dual-attention synergy improves MIoU by 2.5%, precision by 2.5% over single mechanisms. Further, a damage quantification framework is developed using OpenCV and Canny edge detection to extract contours, coupled with a maximum tangent circle algorithm for pixel-level width measurement and skeletonization-based pixel-level length calculation. The results of damage quantification model show that the proportion of images with damage width quantification error ⩽ 20% are 85.29%. This study serves as a reference for intelligent assessment and risk classification of structural damage to buildings affected by flooding.