[A-HW22] River Channel Morphology, Water Resource Management, and Advanced Techniques

[AHW22-P16] Using Synthetic Information from Numerical Models as Inputs for Artificial Intelligence to produce Detailed Flood Map: A Preliminary Study

*HUNG-YING CHEN¹, CHUN-LIANG WU¹, Tsunhua Yang¹ (1.National Yang Ming Chiao Tung University)

Keywords:Numerical Models, Artificial Intelligence, Flood Map, TUFLOW, Diffusion Model

Imagery is a vital tool for visualizing the extent of floods and assessing their impact, as floods are among the most devastating disasters in the world. Tools such as numerical modeling and aerial photography can be used to create "flood maps" at any stage of a disaster's life cycle, serving purposes of precaution, response, and recovery. However, numerical modeling requires trained personnel and considerable preparation time. Additionally, environmental conditions can affect the quality and availability of aerial photography, limiting its effectiveness. The rise of artificial intelligence offers an alternative solution to these challenges, provided that sufficient training data is available. This study proposed a method that involves two main steps: first, using a numerical model called TUFLOW to simulate flooding scenarios and generate flood maps; second, feeding these synthetic images into an artificial intelligence (AI) technique known as diffusion model to examine the causal relationships between rainfall intensity, topography, and the resulting flood maps. A key distinction of this study compared to others is its approach of utilizing imagery to train the AI, rather than relying on quantified data, with the aim of fully leveraging the advantages of diffusion model in generating visuals. A study area of 26 km² in Tainan, Taiwan, was selected, and 182 different rainfall patterns were applied to a 24-hour rainfall event, resulting in accumulations of 24 mm to 720 mm. The findings confirmed that the proposed method can serve as an effective alternative for generating flood maps in the future.