Secondary-factor division optimization based on ET-GA to improve the accuracy of LSA obtained using SSA-DBN compared to the original division method: a case study in southern Sichuan, China

Landslide susceptibility assessment (LSA) is an essential method to prevent and mitigate landslides. Many studies thus involve the LSA and improve the accuracy of LSA using various methods; however, most of the studies optimize the LSA by improving the evaluation models and neglect the influence of factors. The present study initially selects fifteen factors and combines the optimization algorithms, namely, the genetic algorithm (GA), with entropy theory (ET) to divide the factors. The division results are compared with that of the natural breakpoint method (NBM) and equidistant empirical method (EEM) which are the most commonly used methods. The division results are employed to select factors based on the chi-square test and multicollinearity test, and the optimal number of factors involved in the evaluation and the factor systems are determined by the recursive feature elimination method based on the out-of-bag error. The LSAs of ET-GA, NBM, and EEM are then obtained using improved deep belief networks (DBN), and the distribution patterns of the landslide susceptibility maps are similar. The landslide density of very high susceptibility using the three methods is more than three times that of low susceptibility, which illustrates that the three methods are applicable to LSA in the study area. The assessment results are evaluated by the area under the receiver operating characteristic curve (AUC). The ET-GA showed the highest AUC, proving that the optimization algorithms can improve the accuracy of LSA and the ET-GA is more applicable for LSA in the study area.