Urban waterlogging susceptibility assessment based on a PSO-SVM method using a novel repeatedly random sampling idea to select negative samples

Urban waterlogging occurs frequently and often causes considerable damage that seriously affects human activities and the economy. Effectively assessing waterlogging susceptibility can reduce or even avoid the damage caused by such disasters. Here, a Support Vector Machine (SVM) was chosen for waterlogging susceptibility assessment due to its simplicity, objectivity, and understandability. The Particle Swarm Optimization method was used to compute parameters of the SVM. When selecting negative samples for machine learning methods, the methods of subjective selection and single random selection used in previous studies made it easy to select improper negative samples, and thus affected the classification accuracy and generalization capacity of the trained classifiers. To overcome these shortcomings, we proposed a repeatedly random sampling and verifying model to select negative samples for an SVM. As such, this study adopted the spatial framework of integrating GIS and an SVM to assess waterlogging susceptibility using the primary urban area of Guangzhou as an example. The results demonstrate that the waterlogging susceptibility map derived from the most accurate classifier (MAC) can reflect the real occurrence and spatial distribution of waterlogging. Moreover, we randomly generated 100,000 groups of samples to test the classification accuracy and generalization capacity of the MAC; the results show that in 82% of the samples, the area under the curve value of the MAC was higher than that of the randomly generated classifier. This demonstrated that the sampling and verifying model can allow the selection of an MAC with a relatively high and stable classification accuracy. The proposed sampling method can overcome the shortcomings of negative sample selection method employed in previous studies, which makes the machine learning results more accurate and reliable. Furthermore, the method requires less data, which can be helpful in developing countries where the availability of long-term intensive hydrologic monitoring data is limited.