Automatic segmentation of urban flood extent in video image with DSS-YOLOv8n

Abstract

With the impact of global climate change, floods triggered by extreme rainfall events seriously threaten the operation of urban systems in recent years. Real time and accurate urban flood inundation information is critical for disaster management and emergency response. With emergent technologies and citizen sensing, video image has become a core data source of urban system and exhibits great potential for flood management. Some research studies have made progress in video image flood extent extraction, but still face challenges such as non-universal datasets, outdated technology, and lack of support for multi-terminal deployment. In this study, an advanced approach is proposed to address the common challenges facing previous video image-based flood segmentation, by compiling a specialized dataset and training an enhanced flood segmentation model. Initially, a flood inundation dataset containing 2819 samples and 6048 labeled water instances is compiled based on urban flood video images searched from public platforms. Subsequently, Distributed Shift Convolution (DSConv) is introduced to enhance the performance of You Only Look Once version 8 for segmentation (YOLOv8n-seg) model for flood segmentation, and an optimal model is obtained, called DSS-YOLOv8n. Various cases prove that DSS-YOLOv8n has superior performance in flood extent segmentation. Compared to the baseline YOLOv8n-seg, the DSS-YOLOv8n has advanced performance with the Box mAP50 (mean Average Precision at 50 % Recall) value of 77.5 % (a 1.6 % enhancement), the Mask mAP50 value of 76.5 % (a 1.7 % enhancement), and a reduction of the floating-point operations by 0.6 G. Besides, the behavior of DSS-YOLOv8n for flood segmentation in complex scenarios and the comparison results with typical flood segmentation systems demonstrate its robustness and generality in urban flood segmentation. In brief, this study successfully demonstrates the advancement of the proposed approach in urban flood segmentation and further promotes the use of video images for urban flood management.