Can different machine learning methods have consistent interpretations of DEM-based factors in shallow landslide susceptibility assessments?

Abstract

Research on the application of machine learning (ML) models to landslide susceptibility assessments has gained popularity in recent years, with a focus primarily on topographic factors derived from digital elevation models (DEMs). However, few studies have focused on the explanatory effects of these factors on different models, i.e. whether DEM-based factors affect different models in the same way. This study investigated whether different ML models could yield consistent interpretations of DEM-based factors using explanatory algorithms. Six ML models, including a support vector machine, a neural network, extreme gradient boosting, a random forest, linear regression, and K -nearest neighbors, were trained and evaluated on five geospatial datasets derived from different DEMs. Each dataset contained eight DEM-based and six non-DEM-based factors from 8912 landslide samples. Model performance was assessed using accuracy, precision, recall rate, F1-score, kappa coefficient, and receiver operating characteristic curves. Explanatory analyses, including Shapley additive explanations and partial dependence plots, were also employed to investigate the effects of topographic factors on landslide susceptibility. The results indicate that DEM-based factors consistently influenced different ML models across the datasets. Furthermore, tree-based models outperformed the other models in almost all datasets, while the most suitable DEMs were obtained from Copernicus and TanDEM-X. In addition, the concave surface without potholes on steep slopes are ideal topographic conditions for landslide formation in the study area. This study can benefit the wider landslide research community by clarifying how topographic factors affect ML models.

Topics & Concepts