Assessment of urban flood susceptibility based on a novel integrated machine learning method.
Journal:
Environmental monitoring and assessment
PMID:
39636366
Abstract
Flood susceptibility assessment is the premise and foundation to prevent flood disaster events effectively. To accurately assess urban flood susceptibility (UFS), this study first analyzes the advantages and disadvantages of multi-layer perceptron (MLP), and light gradient boosting machine (LightGBM), and designs a new integrated machine learning method by combining logistic regression (LR) method, i.e., LG-MLP-LR. Then, we verify the performance of LG-MLP-LR by taking the flood disaster events in Fuzhou from 2013 to 2016 as example and analyze the contribution of flood conditioning factors by calculating the SHapley Additive exPlanations values. Finally, the assessment results are compared with MLP, LightGBM, XG-MLP-LR, and CB-MLP-LR. The results show that (1) the selected flood conditioning factors can accurately depict the UFS of the study area; (2) compared with MLP, LightGBM, XG-MLP-LR, and CB-MLP-LR, the assessment results by LG-MLP-LR have higher average accuracy (94.950%) and higher average AUC (98.813%); (3) the factors affecting the occurrence and damage degree of flood disaster events in Fuzhou from 2013 to 2016 were elevation, topographic wetness index, maximum one-day rainfall, and stream power index, respectively. This study provides a new idea and method for the effective prevention and control of flood disasters in cities.