Landslide susceptibility analysis can provide theoretical support for landslide risk management. However, some susceptibility analyses are not sufficiently interpretable. Moreover, the accuracy of many research methods needs to be improved. Therefore, this study can supplement these deficiencies. This study aims to research the evaluation effects of random forest (RF) and extreme gradient boosting (XGBoost) classifier models on landslide susceptibility, and to compare their applicability in Fengjie County, Chongqing, a typical landslide‐prone area in southwest of China. Firstly, 1624 landslides information from 1980 to 2020 were obtained through field investigation, and a geospatial database of 16 conditional factors had been constructed. Secondly, non‐landslide points were selected to form a complete data set and RF and XGBoost models were established. Finally, the area under the ROC curve (AUC) value, accuracy, and F‐score were used to compare the two models. The results show that even though both classifiers have a highly accurate evaluation of landslide susceptibility, the RF model performs better. In comparison, the RF model has a higher AUC value of 0.866, and its accuracy and F‐score are approximately 2% higher than XGBoost. The land use, elevation, and lithology of Fengjie County contribute to the occurrence of landslides. This is due to human engineering activities (such as land reclamation, and housing construction) resulting in low slope stability and landslides in widely distributed sandstone, siltstone, and mudstone layers owing to their low permeability and planes of weakness.

Landslide susceptibility mapping (LSM) has been widely used as an important reference for development and construction planning to mitigate the potential social‐eco impact caused by landslides. Originally, most of those maps were generated by the judgements of experts, which is time‐consuming and laborious, and whose accuracy is difficult to be quantified because of the subjective effects. With the development of machine learning algorithms and the methods of data collection, big data and artificial intelligence have now been popularized in this field, significantly improving mapping accuracy and efficiency. Various machine learning‐based methods, mainly including conventional machine learning, deep learning, and transfer learning have been applied and compared in LSM in different areas by previous researchers. Nevertheless, none of them can be effective in all cases. Although deep learning‐based methods were proven more accurate than conventional machine learning‐based methods in most data‐rich situations, the latter is sometimes more popularly used in LSM, as there is not that much data in this field to train a deep learning network perfectly. In a more rigorous situation when there is very limited data, transfer learning‐based approaches are applied by several researchers, which have contributed to improve the workability and the accuracy of LSM in data‐limited areas. Such technical explosion has promoted the application of landslide susceptibility maps, thus contributing to mitigating the social‐eco impact associated with landslides. This paper comprehensively reviews the whole process of generating landslide susceptibility maps based on machine learning methods, introduces and compares the commonly used machine learning methods, and discusses the topics for future research.

Machine learning-based methods are commonly used for landslide susceptibility mapping. Most of the recent publications focused on quantitative analysis, i.e., improving data processing methods, comparing and perfecting the data-driven model itself, but rarely taking the qualitative aspects of the local landslide occurrences into consideration and the further analysis of the key features was always lacking. This study aims to combine qualitative and quantitative analysis and examine its effect on mapping accuracy; based on the feature importance ranks and the related literature, the key features for identifying landslide/non-landslide points of different sub-zones were further analyzed. Before modeling, the study area Yunyang County, Chongqing City, China, was manually divided into four sub-zones based on the information from geological hazards exploration in Chongqing, including the mechanism of landslide formation and sliding failure and geomorphic unit characteristics. Upon the qualitative analysis basis, five grid searches tuned random forest models (one for the whole region and four for the sub-zones independently) were established by 1654 data points and 20 conditioning features. Compared with the conventional data-driven method, the integrated quantitative evaluation based on the qualitative analysis results showed higher reliability, which not only improved the mapping accuracy but also increased the AUC values of all four sub-models, which were 8.8%, 2.3%, 1.9% and 9.1% higher than that of the parent model. Moreover, the quantitative evaluation based on the qualitative analysis revealed the key factors affecting local landslide formation. Therefore, qualitative analysis is recommended in future landslide susceptibility modeling with the additional combination of data-driven methods.

No abstract