Evaluation of Susceptibility under Different Landslide Sample Points and Polygonal Expression Modes

Landslide cataloging modes are usually points and polygons. The location of landslide points and the sampling range of polygons will affect the results of landslide susceptibility evaluation. In order to study the differences in the susceptibility results of different points and polygonal landslide sample sampling strategies, taking Ningnan County, Sichuan Province as an example, landslide polygons and landslide steep sill buffer zones were used to compare the susceptibility evaluation of different polygon expression patterns. The influence of landslide sill point and landslide mass center point was used to compare the influence of different point expression patterns on susceptibility evaluation, and three evaluation models were selected, namely, support vector machine (SVM), random forest (RF) and artificial neural network (ANN). Landslide susceptibility modeling was performed, and differences in modeling were analyzed using ROC curve, mean, and standard deviation. The results are as follows: (1) When the landslide samples are in the polygonal expression mode, the evaluation effect of the steep sill buffer zone is better than that of the landslide polygon. When the landslide sample is in a point expression mode, the evaluation effect of the landslide mass center point is better than that of the landslide steep point. (2) The susceptibility evaluation effect of the RF model is better under different sampling strategies, and the susceptibility results based on the RF model under different sampling strategies are also less different, and have better generalization ability than the SVM and ANN models. (3) The discrete factor is the main factor leading to the difference in the susceptibility results of the sampling strategy under the point expression pattern. Compared with the landslide polygon, the sampling strategy of the steep sill buffer preserves the spatial information of discrete environmental factors such as rock formations, so the evaluation effect is better. It can be seen that using refined terrain features such as landslide steep ridge areas as landslide sampling methods at the county scale can improve the accuracy of susceptibility evaluation.