Improved DeepLabv3+ Based Flood Water Body Extraction Model for SAR Imagery

In recent years, with the intensification of global climate change and the increasing frequency of natural disasters, flooding has become a widespread concern worldwide. Deep learning-based techniques for extracting floods from remote sensing images are crucial in disaster analysis, emergency response, and early warning systems. In this study, we innovatively developed a high-precision model for extracting floodwater bodies based on DeepLabv3+. To enhance the network’s comprehension of inter-feature dependencies, we introduced the feature-dense connectivity operation. Furthermore, to prioritize essential channel and spatial information, particularly at water body edges, we seamlessly integrated the Convolutional Block Attention Module (CBAM) mechanism into our model. This strategic combination of DeepLabv3+, feature-dense connectivity, and CBAM not only showcases our novel approach but also significantly enhances the model’s overall effectiveness in floodwater body extraction. The experiments prove that our method outperforms models such as DeepLabv3+, SegFormer, SETR and FastFCN, performing the best in all metrics on both pre-disaster and post-disaster data from the DaLi flood dataset. Additionally, the model’s mean Intersection over Union (mIoU) results improved by 0.3% and 4.1% compared to the unimproved DeepLabV3+ on pre-disaster and post-disaster data, respectively. This study contributes to the timely extraction of flood hazards and facilitates emergency response.