TDBOSegNet: A Semantic Segmentation Model for Typical Disaster-Bearing Objects in UAV Images of Flood Disasters

In recent years, the intensifying effects of climate change have led to an increase in natural disasters, posing significant challenges to ecosystems. Unmanned Aerial Vehicles (UAVs), with their rapid deployment and high flexibility, play a crucial role in capturing high-resolution, multi-angle imagery of flood-affected areas. The ability to quickly and accurately identify typical disaster-bearing objects (such as inundated roads and buildings) from UAV images is essential for effective disaster assessment and rescue operations. In this paper, we introduce TDBOSegNet, a semantic segmentation model designed specifically to identify typical disaster-bearing objects in UAV-captured flood images and determine whether they are inundated. The model employs an encoder-decoder architecture, enhanced by a Multi-branch Dilated Residual Block (MDRB), a Segment Anything Model (SAM) block, and an Atrous Spatial Pyramid Pooling (ASPP) module, which together improve feature extraction and segmentation performance. Experimental results on the FloodNet dataset show that TDBOSegNet achieves a mean IoU of 81.81%, precision of 90.45%, recall of 89.52%, and an F1 score of 0.8992, outperforming five classic and two state-of-the-art semantic segmentation models used for comparison. These results demonstrate the impressive prowess of TDBOSegNet in accurately segmenting disaster-bearing objects in UAV imagery, providing significant support for disaster assessment and emergency response efforts.