Mapping homogeneous regions for flash floods using machine learning: A case study in Jiangxi province, China

• Published in: International journal of applied earth observation and geoinformation 2022-04, Vol.108, p.102717, Article 102717
• Main Authors: Zhang, Ruojing, Chen, Yuehong, Zhang, Xiaoxiang, Ma, Qiang, Ren, Liliang
• Format: Article
• Language: English
• Subjects: China; Clustering algorithm; Clustering validity indices; Flash floods; Jiangxi province; Regionalization
• ISSN: 1569-8432; 1872-826X
• DOI: 10.1016/j.jag.2022.102717

•A completely quantitative flash flood regionalization approach through key factor selection, clustering and post-processing using a combination of machine learning algorithms is developed.•Plentiful flash flood relevant geospatial data, including historical flash flood events, rainfall data, topographic data and 10–50 km2 small watersheds, are used to delineate the homogeneous regions of flash floods.•A regionalization map of flash floods in the Jiangxi province of China was generated for local government in future flash flood mitigation and prevention. Regionalization of flash floods aims to partition a geographical space into homogeneous regions in which flash floods have similar generation mechanism. In this paper, we present a flash flood regionalization approach using machine learning algorithms to generate the flash flood regionalization map. First, the random forest algorithm is used to identify thirteen key factors of flash floods from a series of rainfall and topographic factors that have great potential to drive the occurrence of flash floods. Second, the two-stage hybrid self-organizing-map-based clustering algorithm is built to delineate the homogeneous regions of flash floods according to the identified thirteen key factors. Third, the best clustering result is selected by the clustering validity indices and it is regarded as the initial regionalization map. Last, the post-processing is implemented to obtain the final regionalization map of flash floods. A case study in the Jiangxi province of China was conducted to generate the flash flood regionalization map with eighteen homogeneous regions. The regionalization map clearly divides the historical flash flood events with different densities into different regions and it is visually satisfied. The quantitative evaluation method further confirmed that the flash flood regionalization map can provide the determinant power of 77.31% for the spatial distribution of historical flash flood events. Hence, it is a valuable option to perform the flash flood regionalization of watersheds and it is beneficial for local government in the Jiangxi province of China in future flash flood mitigation and prevention.