PSO-WELLSVM: an Integrated Method and Its Application in Urban Waterlogging Susceptibility Assessment in the Central Wuhan, China
Wenying Du,Yue Gong,NengCheng Chen
DOI: https://doi.org/10.1016/j.cageo.2022.105079
IF: 5.168
2022-01-01
Computers & Geosciences
Abstract:Urban waterlogging susceptibility assessment can be used in preventing urban waterlogging disaster, which can cause serious damage. An integrated method based on particle swarm optimization (PSO) and weakly labeled support vector machine (WELLSVM), is presented to assess urban waterlogging susceptibility for a certain rainstorm. This method incorporated twelve explanatory factors, including daily precipitation, 3-day cumulated rainfall prior to flood occurrences, elevation, slope, curvature, aspect, topographic wetness index, stream power index, distance to river, leaf area index, impervious surface percentage, distance to road, to perform waterlogging susceptibility analysis. The rainstorm on July 6, 2016 in the main districts of Wuhan, China was used as the scenario to test its feasibility. Cohen's kappa index, accuracy, precision, and recall were used to evaluate the performance of the proposed model. The accuracy of the proposed model (93.6% for training data and 90.1% for testing data) is higher than WELLSVM, support vector machine, and logistic regression, demonstrating the advantages of utilizing unlabeled data and optimized parameter selection. The proposed model can also well identify the waterlogging susceptibility zones. The highest waterlogging susceptibility areas are located in the area with high impervious surface percent, intersections, culverts and overpasses, lakeshore, and riverbank. Elevation and precipitation factors are the most influential factors to waterlogging susceptibility. The proposed model was also tested by the other two storms on July 7, 2013 and June 21, 2019, proving the validity of it. The proposed model is helpful for instant waterlogging susceptibility analysis and can help decision-making of urban waterlogging control.