Sensing urban road inundation risk during extreme rainfall via a Vehicle-as-Flood-Gauge approach

Remote sensing techniques typically capture flood extent at discrete time points but cannot directly estimate flood depth. Social media, as an emerging source of volunteered geographic information (VGI), provides real-time visual evidence of fine-grained flood severity. However, this potential remains underexplored. In this study, we propose Social-VFG, an interpretable framework for assessing road inundation using vehicle wheels observed in images. To avoid training annotation-intensive wheel segmentation models, wheel contours above floodwater are efficiently extracted by coupling a trained wheel detector with the Segment Anything Model (SAM). Based on these contours, an interpretable rule-based indicator is proposed to characterize wheel inundation status as water levels rise. To mitigate the indicator’s sensitivity to segmentation noise and scale variations, four complementary features are further integrated into a random forest classifier to improve the robustness of inundation classification. Using the 2023 Zhuozhou flood as a case study, road inundation risk maps derived from Weibo images show strong spatial consistency with flood extents extracted from GF-3 SAR imagery. Importantly, social media observations provide fine-grained information on flood severity and effectively compensate for spatial gaps caused by satellite imagery limitations. The resulting road inundation risk maps can support flood-aware route planning under extreme rainfall conditions.