Flood Inundation Range Extraction and Comparative Analysis Based on Segformer Model

[Objective] With the development of deep learning and computer vision technologies, the application of machine vision in flood monitoring has gradually become a research hotspot. This study aims to overcome the limitations of traditional manual and satellite remote sensing methods in flood monitoring and early warning, including insufficient accuracy and high costs, and to explore the advantages and limitations of the Segformer model in extracting flood boundaries, while proposing future research directions and improvement strategies. [Methods] Using deep learning and machine vision techniques, a specialized “RiverDataset” for flood monitoring was constructed, and the performance of the Segformer model in extracting flood inundation range was evaluated based on this dataset. Additionally, the Segformer model was compared with the U-Net models based on ResNet50 and VGG16 to assess their performance in water body segmentation tasks. Taking Shashi District, Jingzhou City, Hubei Province as an example, UAV remote sensing imagery was used to accurately extract the flood contours in the region. [Results] The U-Net (VGG16) model demonstrated excellent performance on the training set but was slightly inferior to the Segformer model on the validation and test sets. The Segformer model achieved superior performance across most indicators, particularly outperforming the U-Net (ResNet50) model in complex scenarios. Although the U-Net (ResNet50) model achieved a slightly higher IoU, its higher loss value and lower mIoU and mAP indicated that its overall performance was inferior to that of the Segformer model and the U-Net model (VGG16). Consequently, the U-Net model based on VGG16 performed well across all evaluation indicators, demonstrating strong fitting capability during training. However, when processing complex water bodies, this model struggled to capture contextual information in narrow regions due to its limited receptive field, leading to frequent occurrences of information voids. Furthermore, the U-Net model failed to effectively eliminate gridding effects, compromising the local consistency of feature information. In contrast, the Segformer model did not exhibit this issue. This difference was mainly due to the relatively small receptive field of the convolutional kernels used in the U-Net model, which restricted its ability to interpret contextual information within narrow regions and hindered the establishment of long-range information continuity. The Segformer model, not constrained by the limited receptive field of conventional convolutional kernels, could better capture broader contextual information in the image. Additionally, the U-Net model failed to effectively eliminate the impact of the grid effect on feature information, resulting in extracted features that disrupted the local consistency of information. The absence of structures for perceiving local regional information made the model incapable of effectively eliminating the influence of interfering objects in complex aquatic environments, particularly in scenarios requiring precise boundary extraction. [Conclusions] In complex aquatic environments, the Segformer model demonstrates superior segmentation performance and robustness. This study validates the efficiency of the Segformer model in extracting flood inundation range, highlighting its potential for practical water body monitoring applications. Future research should further optimize the model, expand the dataset, and explore its potential for real-time application to enhance the efficiency and accuracy of flood early warning systems.