Evaluation on landslide susceptibility based on self-organizing feature map network and random forest model：A case study of Dayu County of Jiangxi Province

Abstract

In order to further explore the influence of evaluation units and non-landslide sample selection methods on landslide susceptibility prediction, a landslide susceptibility evaluation model is established based on self-organizing feature map network and random forest model in this paper. According to the relationship between grid units and slope units, an optimized calculation method of landslide susceptibility index is proposed. Aiming at the deficiencies of grid units and slope units in the evaluation of landslide susceptibility, this model proposes an optimized calculation method for landslide susceptibility index based on the relationship between grid cells and slope cells. On this basis, a landslide susceptibility evaluation model was established based on the random forest Tree Bagger classifier. By comparing and analyzing the influence of self-organizing feature map network and random non-landslide sample selection methods on the evaluation results, the effectiveness of the three evaluation models of self-organizing feature map network, random forest and self-organizing feature map network -random forest were discussed. The evaluation model has been applied to the landslide susceptibility evaluation in Dayu County. The results show that the prediction accuracy of random forest and self-organizing feature map network-random forest is higher, reaching 91.19% and 94.94% respectively, and the AUC of success rate curve was 0.822 and 0.849 respectively. It shows that self-organizing feature map network-random forest has higher prediction rate and success rate，although the prediction accuracy of self-organizing feature map network clustering is limited, it can effectively improve the evaluation accuracy of random forest model as the basis for selecting non landslide samples.

Keywords

• slope units
• landslide susceptibility
• self-organizing feature map network
• random forest
• non-landslide sample