Improvement of models to enhance the information delivery capability of AI-based flood early warning systems; [AI 기반홍수 예 · 경보 시스템의 정보 전달력 향상을 위한 모형 개선]

초록

The year 2023 recorded historically high temperatures and increased precipitation globally, with the Philippines particularly affected due to its vulnerability to tropical cyclones and typhoons. In the Eastern Visayas region, including Biliran Island, frequent floods and landslides have highlighted the limitations of existing flood warning systems, which rely primarily on observational equipment and expert judgment and often lack systematic coverage. This study develops an AI-based flood hazard classification and real-time water level prediction framework to enhance disaster preparedness and response. For flood hazard classification, machine learning models—Random Forest (RF) and Decision Tree (Tree)—were employed. The RF model demonstrated superior performance with an overall F1-Score of 0.83 across all hazard classes, whereas the Tree model achieved an F1-Score of 0.66, indicating relatively lower predictive accuracy. For real-time water level prediction, a deep learning Transformer model was applied, utilizing rainfall and water level time-series data to accurately forecast overall patterns and peak water levels. Results show that integrating machine learning and deep learning models can overcome the limitations of conventional flood management approaches, providing reliable hazard predictions and actionable information for early warning systems. The proposed AI framework enhances the timeliness and accuracy of flood alerts, supporting more effective disaster response. Future research should incorporate additional hydrological data and continuously refine the models to further improve predictive accuracy and strengthen local community resilience against floods. © 2025, Korean Wetlands Society. All rights reserved.

키워드