Landslide susceptibility assessment based on the SOM-I-SVM model

Abstract

When using machine learning models for landslide susceptibility evaluation, the non-landslide sample points are usually selected randomly outside the landslide influence area, leading to a certain error. To improve the accuracy of landslide susceptibility evaluation, this paper couples the self-organizing map (SOM) neural network, information (I) model, and support vector machine (SVM) model, and proposes a SOM-I-SVM model-based method of landslide susceptibility evaluation, comparing with K-means clustering to verify the reliability of this model. The Maojian District of the city of Shiyan is taken as an example, and seven factors of the distance from water system, slope, rainfall, distance from structure, relative height difference, distance from road, stratigraphic lithology are selected by correlation and importance analyses of environmental factors to establish a landslide susceptibility evaluation system. Based on these, the graded information values of each factor are calculated and used as input variables for landslide susceptibility evaluation. The SOM neural network and K-means clustering are used to select non-landslide samples, and the sample data set is substituted into the I-SVM model to predict landslide susceptibility. The prediction accuracies of the four models, SVM, I-SVM, KMeans-I-SVM and SOM-I-SVM, are compared, and the area under the ROC curve (AUC values) are 0.82, 0.88, 0.90 and 0.91, indicating that the SOM-I-SVM model can effectively improve the accuracy of landslide susceptibility prediction.

Keywords

• landslide; susceptibility assessment
• information model
• svm
• som