Machine learning-based flood susceptibility assessment for railway networks in Chiang Mai: A random forest approach

Abstract

Flooding is among the most damaging natural disasters, threatening public safety, causing loss of life and economic disruption, and degrading the resilience of heavy civil infrastructure, including urban railway networks. This study aims to develop a machine-learning-based assessment framework for the flood susceptibility of railway network corridors in the Chiang Mai Comprehensive Plan Area, Thailand. By integrating the Random Forest (RF) algorithm with a Geographic Information System (GIS), the research analyzed five spatial parameters: land use, land cover, elevation, slope, and distance to rivers to identify flood-susceptible sections along the railway network. A probability thresholding technique was applied with a threshold of 0.70 to address prediction errors arising from imbalanced data effectively. The evaluation results demonstrated that the RF model performed effectively, achieving an overall accuracy of 88.2% and an Area Under the Curve (AUC) of 0.863. Feature importance analysis revealed that distance to rivers and land use were the most influential factors driving railway susceptibility. The generated susceptibility map aligns closely with historical flood records, highlighting, from the 2024 event, high-risk clusters in river plains and the city center. Ultimately, these findings can assist engineers and urban planners in flood risk management and the sustainable development and resiliency of railway infrastructure.