Paper Title
FLOOD MONITOR: A MULTI-TASK DEEP LEARNING FRAMEWORK FOR REAL-TIME FLOOD VICTIM DETECTION AND RISK ASSESSMENT USING SYNTHETIC VIDEO DATA

Abstract
Flood disasters continue to pose a serious threat to human life and demand rapid situational awareness for effective rescue efforts. However, the lack of large annotated flood video datasets limits the development of reliable computer vision-based monitoring systems. This paper introduces Flood Monitor, a multitask deep learning framework that integrates synthetic video generation with an end-to-end neural architecture to address this challenge. The system performs four key tasks: victim detection and counting, water depth estimation using human body reference cues, flood severity classification, and risk-based priority ranking. To overcome data scarcity, state-of-the-art text-to-video generative models are used to create a diverse synthetic dataset of 500 flood video clips. Experimental results demonstrate strong performance, achieving 87.3% accuracy in person detection, 82.1% in relative depth estimation, and 89.5% in severity classification. Domain adaptation experiments on real flood footage further validate the system’s real-world applicability, achieving 73.8% detection accuracy beyond synthetic training data. Additionally, qualitative analysis highlights the effectiveness of the priority ranking module in identifying high-risk victims and supporting rescue decision-making. Overall, the proposed framework provides a scalable and efficient pipeline that trans-forms raw flood video data into actionable insights for improved real-time disaster response and management. Keywords - Synthetic Video Data, Flood Surveillance, Victim Detection, Water Depth Estimation, Multi-Task Learning, Generative AI, Disaster Response, Domain Adaptation