Verification of the threshold rainfall based on meteorological and climate data using SNS inundation video data

Urban inundation is becoming increasingly severe due to the effects of climate change and more frequent occurrences of intense rainfall. As cities grow denser and infrastructure becomes more vulnerable, the risk of urban inundation continues to rise. Traditional monitoring systems—such as closed-circuit televisions (CCTVs), infrared (IR) sensors, and depth cameras—are commonly used to estimate threshold rainfall, which is the point at which inundation begins. However, these systems often fall short in accurately validating real-time flood depth, especially during rapidly evolving flood events. To address this issue, this study proposes a deep learning-based model designed to detect and validate inundation levels using both CCTV videos and social media videos. Social media (SNS) videos provide real-time, accessible, and user-rich content, which allows for more dynamic inundation monitoring without requiring manual input or direct human intervention. The proposed model integrates YOLO model for real-time object detection and U-Net model for semantic segmentation to estimate flood depth, flow velocity, direction, and overall severity. Furthermore, the model detects signboards in videos, enabling automatic extraction of key information such as time and location. To train and evaluate the model, datasets were collected from Kaggle and various publicly available social media platforms. Model performance was assessed using receiver operating characteristic (ROC) curves and mean squared error (MSE) metrics. The proposed system was applied to verify inundation depth against pre-identified threshold rainfall values and demonstrated both accuracy and reliability. This approach provides valuable data for early warning systems and supports proactive disaster risk management. By combining AI with user-generated content, this model offers a scalable and cost-effective solution for real-time urban inundation monitoring and response.