Flood disasters severely threaten socio-economic development and the safety of lives and property, making the accurate acquisition of disaster information crucial for effective response. Traditional satellite remote sensing is limited by spatial resolution and revisit frequency, and often fails to capture dynamic flood details. Unmanned Aerial Vehicles (UAVs), leveraging high mobility, high-resolution imaging, and flexible deployment, can quickly access disaster-affected areas to acquire fine-grained imagery, compensating for the shortcomings of traditional methods. However, existing UAV image datasets for flood research generally suffer from limited sample sizes and lack of scenario diversity. This study focuses on the typical flood event in Jiujiang, Jiangxi Province in 2022. UAV field surveys were conducted to collect 412 high-resolution images during the disaster event. Through a rigorous data processing workflow, 12,080 training-ready samples with a resolution of 512×512 pixels were generated, covering six semantic categories: water, building, road, flooded building, flooded road, and background. To evaluate dataset quality, multiple mainstream semantic segmentation models were tested. The experimental results fully validate the reliability of the proposed dataset, as well as its challenging nature in evaluating segmentation algorithms for complex scenes. This dataset provides crucial data support for scientific decision-making by local disaster management authorities and holds significant value for enhancing flood prevention and mitigation capabilities in affected regions.
Li et al. (Sun,) studied this question.