Intelligent recognition of building seismic resistance based on Res-Mda-Unet++

Conventional building seismic-resistance assessment methods are inefficient and expensive, thus limiting their large-scale application. To solve this problem, this study designed a Res-MDA-UNet++ recognition network integrating residual network and multidimensional detail attention (MDA) mechanism. Taking UNet++ as the basic framework and ResNet50 as the encoder backbone to enhance feature extraction, the model embeds the MDA module (fusing channel and spatial attention) into the residual structure to improve the representation of key featuresand the model's recognition accuracy. It also fuses field survey-based semantic grid data with remote-sensing imagery at the band level to compensate for single-source data limitations. The model was validated using the RGB-seismic factor dataset collected from Rong County, Sichuan, achieving an average F1-score of 96.91% and an average IoU of 94.04% in building seismicresistance recognition, significantly outperforming mainstream semantic segmentation models. It excels in image detail perception and fine-grained segmentation, providing a reliable technical approach for large-scale, efficient and accurate building seismic-resistance identification.