Fine-grained analysis and mapping of urban flood susceptibility with interpretable machine learning: A case study of Hefei, China

Study region

Built-up area of Hefei City, China.

Study focus

Climate change has increased frequency of extreme rainfall events. Mapping the urban flood susceptibility and exploring the impact factors can enhance urban resilience. Existing methods often treat cities as uniform entities, making it challenging to capture the complexity of these localized characteristics. This paper proposes a novel approach combining interpretable machine learning and spatial autocorrelation. An ensemble learning model assesses susceptibility by incorporating terrain, urban construction, and precipitation factors. An improved spatial weight matrix is proposed to perform spatial autocorrelation for revealing spatial distribution of flood susceptibility, and the local factors are explained by LIME to provide a fine-grained analysis of different regions.

New hydrological insights for the region: (1)NDVI is the most influential factor emphasizing the importance of green spaces in urban flood regulation. (2)Micro-topography significantly affects urban flood susceptibility, and normalizing DSM based on micro-watersheds provides an accurate representation. (3)High flood susceptibility in Hefei, as revealed by spatial autocorrelation analysis, follows patterns similar to built-up areas and is influenced by major roads. Based on this, LIME analysis reveals distinct regional impact factors, such as NDVI, land use, distance to water bodies, and road density, supporting targeted flood management strategies. These findings can provide valuable insights for flood prevention and urban planning, contributing to the overall resilience of the urban environment.