Comparison of Machine Learning Methods for Grid-Based Flood Risk Assessment and Mapping

This study aims to create a flood risk assessment and mapping of Ulsan Metropolitan City, South Korea targeting typhoon 'Chaba' that occurred in 2016. Five influencing factors were used to estimate the risk of flooding. As static data, 'Area of buildings', 'Population', and 'Landcover' were used, and as dynamic data, at the time of Typhoon 'Chaba' (October 5, 2016) 'Precipitation' and 'River stage' data in 10 minutes units were used. The influence factors were constructed in the form of a 10m grid size based on the 'National Point Number' of South Korea. The value of the influencing factor corresponding to each grid was normalized to make the scale the same, and this was applied to the risk model algorithm to calculate the risk level (grade 5) for each grid. The risk grade values for each grid calculated through the process were used as dependent variables when constructing the machine learning model, and the values of five influence factors were used as independent variables. About 15,000 sample datasets were constructed, and training, validation, and test datasets were randomly classified in a 6:2:2 ratio. We compared the performance of three machine learning techniques to derive an optimal model for risk rating calculation; Random Forest (RF), Support Vector Machine (SVM), and K-Nearest Neighbor (KNN). The built high-performance model can be applied in connection with real-time meteorological and river stage data in the future. This allows real-time identification of areas with a high risk of flooding in the event of disasters such as typhoons and floods, allowing rapid disaster response, on-site management, and restoration work. In addition, it is judged that it can help the national and local government officials to make more efficient decisions in disaster situations.