Landslide susceptibility mapping using GIS-based machine learning algorithms for the Northeast Chongqing Area, China

This study examines the northeast part of Chongqing using four machine algorithms: logistic regression, multilayer perceptron, random forest and support vector machine. Eleven factors, including aspect, slope, distance from the road and rainfall, are selected, and 581 landslide points and equivalent safety points are used to establish a landslide sensitivity analysis model in Northeast Chongqing. The 407 landslide points in the sample (70%) are used for training, and 174 landslide points (30%) are used for verification to predict landslide sensitivity. The outcomes are used to generate a landslide sensitivity zoning map. The landslide grid distribution histogram and area under the curve (AUC) are also used to evaluate the prediction results. Results show that three types of machine learning exhibit a clear positive relationship between rainfall and the occurrence of landslides. The ratios of the area of landslide occurrence to the area of landslide sensitivity obtained by the four algorithms are 1.49, 1.23, 1.5 and 1.2, respectively. The success rates of the training samples under the four models in the AUC are 0.794, 0.816, 0.848 and 0.831, and the prediction rates of the test samples are 0.788, 0.802, 0.822 and 0.809. The analysis results show that the random forest prediction of the four algorithms is similar to the actual observation in Chongqing, the large-scale study area in the northeast is more accurate and the fit for disaster occurrence is more consistent. This research can provide an auxiliary decision-making basis for prevention and management of local landslide disasters.