Presenter
Mr. Ze Wang, Dalian University of Technology
Co-author(s)
Dr. Heng Lyu, Dalian University of Technology
Prof. Chi Zhang, Dalian University of Technology
Sub-theme
3. Building Resilience for Disaster Prevention and Mitigation
Topic
3-7. Management of water risks induced by extreme weather and climate events
Body
Urban flooding is a frequent, dangerous, and costly natural disaster that causes significant damage to infrastructure and affects the livelihoods of millions of people every year. To effectively implement adaptive management measures for flood preparation and mitigation, it is essential to accurately identify areas susceptible to flood. With high accuracy and computational efficiency, machine learning is increasingly used for identifying high flood susceptible areas. However, previous machine learning models were typically used for the predictive purpose only, thus, lacking interpretability. They handled the input and output data from a statistical perspective, and their parameters and hidden states cannot contribute to understanding the mechanism of flood formation. To address the interpretability of the machine learning model, this study introduced graph attention network (GAT), an interpretable semi-supervised machine learning model, for flood susceptibility mapping and further interpreting the flooding process. The GAT formalizes the flood susceptibility mapping problem into a flooded-nonflooded classification task. The model uses nodes and edges to represent spatial units of the study area and their relative spatial relationships, respectively. Based on an attention mechanism that guides an individual unit to assign attention to its neighbors, the GAT can explicitly interpret the hydrological responses between adjacent units. In this study, the metropolitan area of Dalian, China, was used to validate the application of GAT in flood susceptibility mapping. Compared with two state-of-the-art machine learning models, a non-biased artificial neural network (ANN) and a convolutional neural network (CNN), GAT has shown superiority in a series of classification evaluation indicators and the distribution pattern of flood susceptibility. Meanwhile, the self-attention weight, weight entropy, and accumulated attended weight, derived from the attention mechanism, enable the GAT to serve in the cause analysis of floods and identify critical areas in the flooding process. The introduction of GAT moves a step to an interpretable machine learning model and provides an effective tool for urban flood management.