Flood susceptibility assessment based on a novel random Naive Bayes method: A comparison between different factor discretization methods

摘要

Random Naive Bayes (RNB) is a machine learning method that uses the Random Forest (RF) structure to optimize Naive Bayes (NB). It is interesting to see whether RNB could optimize NB and achieve satisfied assessment results like RF in the flood susceptibility assessment study. RNB has rarely been used in study of using machine learning methods to spatially analyze natural disasters, and thus it was selected as the analysis method. Based on the data feasibility, 12 spatial factors that affect the occurrence and spatial distribution of floods were selected. To avoid the influence of subjective equal-interval classification method, natural breaks and quantile method were used to discretize factors with continuous values, respectively. Here, a recently proposed repeatedly random sampling method was adopted to select negative samples for RNB to generate a most accurate classifier (MAC) that was employed to compute the probability of flood occurrence in the study area. Consequently, this paper adopted the integrated framework of GIS and RNB to spatially assess the flood susceptibility using the Wanan County in China as an instance. The results demonstrated that when integrated with the repeatedly random sampling method, the MAC-based flood susceptibility maps corresponding to different factor discretization methods were similar, meaning this framework can effectively avoid the effects caused by different factor discretization methods. Also, to testify the classification performance of RNB, RF and NB were chosen to compare the classification performance with it. The results indicated the classification performance in the order of RF > RNB > NB. This means RNB is able to achieve better classification performance than NB, but it exists limitations when compared with traditional strong classifiers like RF. The findings of this paper proved that RNB is a feasible approach for natural hazard susceptibility assessment.