Landslides pose a great threat to the safety of people’s lives and property within disaster areas. In this study, the Zigui to Badong section of the Three Gorges Reservoir is used as the study area, and the land use (LU), land use change (LUC) and band math (band) factors from 2016–2020 along with six selected commonly used factors are used to form a land use factor combination (LUFC), land use change factor combination (LUCFC) and band math factor combination (BMFC). An artificial neural network (ANN), a support vector machine (SVM) and a convolutional neural network (CNN) are chosen as the three models for landslide susceptibility mapping (LSM). The results show that the BMFC is generally better than the LUFC and the LUCFC. For the validation set, the highest simple ranking scores for the three models were obtained for the BMFC (37.2, 32.8 and 39.2), followed by the LUFC (28, 26.6 and 31.8) and the LUCFC (26.8, 28.6 and 20); that is, the band-based predictions are better than those based on the LU and LUC, and the CNN model provides the best prediction ability. According to the four groups of experimental results with ANNs, compared with LU and LUC, band is easier to access, yields higher predictive performance, and provides stronger stability. Thus, band can replace LU and LUC to a certain extent and provide support for automatic and real-time landslide monitoring.
Landslides are geological disasters affected by a variety of factors that have the characteristics of a strong destructive nature and rapid development and cause major harm to the safety of people’s lives and property within the scope of the disaster. Excessive landslide susceptibility mapping (LSM) factors can reduce the accuracy of LSM results and are not conducive to researchers finding the key LSM factors. In this study, with the Three Gorges Reservoir area to the Padang section as an example, the frequency ratio (FR), index of entropy (IOE), Relief-F algorithm, and weights-of-evidence (WOE) Bayesian model were used to sort and screen the importance of 20 LSM factors; then, the LSMs generated based on different factor sets modeled are evaluated and further scored. The results showed that the IOE screening factor was better than the FR, Relief-F, and WOE Bayesian models in the case of retaining no fewer than eight factors; the score for 20 factors without screening was 45 points, and the score for 12 factors screened based on the IOE was 44.8 points, indicating that there was an optimal retention number that had little effect on the LSM results when IOE screening was used. The core factor set obtained by the method for comparing the increase in scores and the increase in corresponding factors effectively improved the accuracy of the LSM results, thus verifying the effectiveness of the proposed method for ranking the importance of LSM factors. The method proposed in this study can effectively screen the key LSM factors and improve the accuracy and scientific soundness of LSM results.
19 March 2019 in International Journal of Pattern Recognition and Artificial Intelligence