Global Downstream Flood Risk Assessment Using Vision-Language Models and Multi-Source Environmental Data

Abstract Submission: Communities downstream of major dams are increasingly vulnerable to severe flooding. This increased vulnerability is largely attributed to the complex and dynamic interactions between hydrological and environmental systems, exacerbated by the cascading impacts of anthropogenic climate change. While existing flood risk models have significantly advanced our understanding of flooding, improvement in their precision, scalability, and interpretability is still needed, particularly in high-risk and data-scarce regions. By combining vision and language modeling techniques, our approach integrates satellite imagery with diverse environmental and climatic datasets to enhance the predictive accuracy of flood risk assessments for communities downstream of dams. The model provides interpretable insights through visually grounded explanations with accompanying textual descriptions, ensuring transparency in risk assessment and supporting effective decision-making. We hypothesize that integrating real-time data with advanced modeling techniques will enhance the accuracy and timeliness of flood risk predictions, significantly reducing response times for disaster management teams. This data-driven approach equips policymakers, environmental agencies, and disaster response teams with critical insights, enabling informed decisions that mitigate flood risks and improve preparedness for climate-induced disasters. By bridging the gap between advanced modeling techniques and practical implementation, our research represents a significant advancement in flood risk management, contributing to developing proactive, sustainable strategies to combat the escalating threats posed by climate change.

Learning Objectives/Expected Outcome (Optional) :