1. Introduction
In recent years, human activities have had a huge impact on the geological environment, which has caused many engineering disturbance disasters around the world [
1,
2,
3,
4,
5,
6,
7]. Under the coupling effect of the earth’s internal and external dynamics, the characteristics of natural slope rock and soil mass and its stress state are constantly adjusted and changed over time. During this process, the stress concentration of the slope body and the weakening of rock-soil properties caused by engineering disturbances will lead to geological disasters such as landslides, collapses, and debris flows, thereby stabilizing the slope. This is an inevitable phenomenon in the evolution of slope landforms [
8,
9]. The engineering disturbance disasters studied in this paper refer to slope geological hazards caused by engineering construction and operation, which have a similar formation mechanism to natural geological disasters but the inducing factors are different [
10]. The main triggering factors of natural geological disasters are rainfall and earthquakes, while engineering disturbance disasters are stress redistribution caused by engineering construction.
Susceptibility evaluation is a method for quantitatively evaluating the temporal and spatial scope of geological disasters under basic geological and climatic conditions. It can provide data basis and technical support for disaster prevention and control. Therefore, to reduce the negative impact of engineering disturbance disasters, it is necessary to implement regional susceptibility evaluation [
10,
11]. According to the prediction principle, the susceptibility evaluation methods can be divided into four categories: physical model method, heuristic model, statistical model, and machine learning model [
12,
13]. The physical model requires detailed geotechnical parameters and is mainly aimed at the stability analysis of a single slope. Heuristic models, such as AHP, mainly rely on the experience of experts and are easily affected by the subjective opinions of experts [
14,
15]. Machine learning models can be divided into statistical machine learning and deep learning. The internal operation mode of deep learning is a black box rule, and the update and iteration of its parameters are unknown [
11,
16,
17,
18]. Compared with other models, the statistical model can clearly show the corresponding relationship between the index factors and the distribution of engineering disturbance disasters [
19,
20,
21]. Typical statistical models include logistic regression (LR) model, weight of evidence method, information value model, certainty coefficient model, gray relational degree model, etc. [
3,
18,
22,
23,
24,
25,
26,
27,
28].
Because the information value model has the advantages of simple operation and good algorithm stability, it can better avoid subjective judgment and can objectively reflect the evaluation results, so it is used by some researchers. Song et al. [
12] used the information value model to evaluate the engineering disturbance disaster susceptibility of the China–Nepal railway, and established the regional engineering disturbance disaster evaluation factor selection principle and susceptibility evaluation system; Ba et al. [
3] used an improved information model to evaluate the susceptibility of disasters in the Artvin area, and the accuracy rate reached 85.4%. Although the information model shows high accuracy in evaluating susceptibility, the information model can only determine the weight of the secondary classification of each evaluation factor, without considering the contribution of different evaluation factors in the occurrence of geological disasters, which is inconsistent with the actual situation. Because the occurrence of regional engineering disturbance disasters is affected by both control factors and impact factors, the contributions of the two to slope catastrophe are different, and the control factors of geological disasters in different regions are also different; those such as rainfall, lithology, and slope should have greater weight in the evaluation. Domestic and foreign scholars have conducted preliminary research in this area: for example, S. Lee et al. [
29] and N. N. Vasu et al. [
30] took the Yongin area and Mt Woomyeon area of Korea as the research area, respectively, and the susceptibility evaluation results obtained by using all index factors and only retaining the index factors with larger weights were compared; E. Yesilnacar and T. Topal [
31] used statistical analysis and machine learning models to calculate the weights of index factors in the same region, and compared the results of the two methods; Cao et al. [
13] obtained the weight of each index through principal component analysis (PCA) and fuzzy analytic hierarchy process on the basis of the information value model, and optimized the information value model, which improved the accuracy of susceptibility evaluation by 5~8%. However, the research on the optimization of the engineering disturbance disaster susceptibility model is not enough, and at the same time, it is necessary to conduct in-depth research on the calculation method of the weight coefficient of the influencing factors [
12]. The LR model can determine the contribution of evaluation factors to geological disasters based on the relationship between the historical occurrence of geological disasters and evaluation factors [
32]; PCA can eliminate the influence of correlation between evaluation indicators, and after extracting principal components, combined with entropy method, the weight of each evaluation factor can be calculated [
33]. In order to establish an evaluation system for engineering disturbance disaster susceptibility, this paper uses LR and PCA to optimize the information model to evaluate regional engineering disturbance disaster susceptibility, and establishes an engineering disturbance disaster risk management system based on the evaluation results.
Based on the above analysis, this paper selects the Himalayan alpine valley region in southeastern Tibet as the research area. The specific research steps and contents include: (1) Based on the analysis results of mathematical statistics, establish the selection principles of evaluation indicators; (2) Calculate the weight coefficient of the evaluation factor by using the PCA method and the LR method; (3) Use the information value model and the optimized information value model to calculate the susceptibility of regional engineering disturbance disasters; (4) Compare the prediction results of the two optimization models and the single information model, verify the improvement effect of each method on the IV model, and provide a basis and reference for optimizing the evaluation of regional engineering disturbance disaster susceptibility.
In previous studies, researchers have mostly considered the optimization effects of different evaluation methods, and there has been less research on weight calculation methods. Therefore, in order to improve the accuracy and rationality of the susceptibility assessment of regional engineering disturbance disasters, this paper determines the weight calculation method suitable for the study area by comparing the optimization effect of the weight calculation methods such as PCA and LR on the susceptibility assessment, which provides methods and ideas for the optimization of the susceptibility assessment methods later.
5. Discussion
Previous studies mainly focused on the influence of different susceptibility evaluation models on the evaluation results, for example, decision tree model, convolutional neural network model, and so on. However, there were few studies on the influence of evaluation factor weight coefficients [
13]. To study the influence of different weight calculation methods on the accuracy of the susceptibility evaluation model, this paper takes the Himalayan alpine valley area as the research object, based on the GIS platform and SPSS statistical analysis software, selecting elevation, lithology, slope, undulation, slope aspect, PGA, and rainfall as evaluation factors. The information model and two weighted information models were used to evaluate the regional susceptibility of the study area. The LR model shows that slope, aspect, rainfall, and elevation are the main influencing factors of engineering disturbance disasters, which is consistent with the research results of Zou et al. [
19]. The LR weight calculation model performs well, which may be related to the fact that the independent variable does not require a normal distribution and can be both discrete and continuous. Statistical analysis shows that engineering disturbance disasters are concentrated in areas with slope >60° and elevation between 2500 m and 3500 m. To evaluate the accuracy of different susceptibility evaluation models, the RPOE value and ROC curve were introduced. The study found that the accuracy of the LR-IV model is higher than that of the PCA-IV and IV-Only models, but the accuracy of the PCA-IV model is lower than that of the IV-Only model. It shows that a reasonable method for calculating the weight of evaluation factors can improve the accuracy of susceptibility evaluation models, but some weight calculation methods can have the opposite effect. Therefore, establishing a universal evaluation factor weight calculation method has become a future research direction.
6. Conclusions
Many experts and scholars have conducted more research on different types of evaluation models, but less on the weight coefficients of evaluation factors. They generally adopt the method of equal weight coefficients, which is inconsistent with the actual situation. Therefore, this article adopts different weight calculation methods. By comparing the evaluation results of different methods, it is proven that a reasonable weight calculation method can improve the accuracy of regional susceptibility evaluation. The detailed conclusion is as follows:
- (1)
PCA and LR weight calculation methods of evaluation factors show that lithology, elevation, slope, undulation, and rainfall play a major role in the occurrence of engineering disturbance disasters in alpine and canyon areas.
- (2)
The RPOE value and ROC curve evaluation results show that the prediction results of the LR-IV model are better than those of the PCA-IV model and the IV model, and the LR-IV model has higher susceptibility evaluation accuracy, but the PCA-IV model it is lower than the IV-Only model, which shows that only a reasonable weight calculation method can improve the accuracy of the susceptibility model.
- (3)
According to the susceptibility evaluation results of the LR-IV model, the study area was divided into four categories: low susceptibility area (26.28%), medium susceptibility area (37.40%), high susceptibility area (27.65%), and extremely high susceptibility area (8.66%), providing a data basis for disaster monitoring and prevention.
- (4)
The research results provide an important direction for improving the susceptibility evaluation model of engineering disturbance disaster, and provide a reference for subsequent researchers to improve the accuracy of susceptibility evaluation by optimizing the index weight calculation method.
The research results of this paper can be used for regional susceptibility evaluation of geological disasters in high mountain and canyon areas, providing a data foundation for disaster prevention and reduction. In addition, this paper supplements the shortcomings of research on weight calculation methods for evaluation factors, providing a reference for future research. However, this article only compares the evaluation results of three weight calculation methods, and more weight calculation methods should be studied in the future.