Effects of different division methods of landslide susceptibility levels on regional landslide susceptibility mapping

Abstract

Reasonable division methods of landslide susceptibility indexes (LSIs) are crucial for producing landslide susceptibility levels (LSLs), including very low, low, moderate, high, and very high levels. However, few studies have systematically compared division methods such as natural break, equal interval, quantile, geometric interval, and K-means. Moreover, these methods start from LSIs but ignore the nonlinear correlation between known landslides and LSIs. To address this, the natural break-frequency ratio (FR) method is proposed, combining the natural break method for LSLs division with the FR method. First, the five conventional methods divide LSIs predicted by three machine learning models in An’yuan County, China. Then, the natural break-FR method is proposed to divide the same LSIs and compared with these methods. The natural break-FR, equal interval and K-means method yielded the largest sum of landslide ratio in very high and high susceptibility level, showing these methods can use high and very high susceptibility levels to predict as many landslides as possible. Finally, statistical perspectives of known landslide identification, division area proportion, and landslide ratio are applied to discuss how to select a suitable division method. Results show different division methods have comparative effects on final LSLs. The landslide ratios of equal interval, K-means, and natural break methods at high and very high susceptibility levels are greater than the former methods. The natural break-FR method performs best with MLP and SVM, but in the more precise RF model, the equal interval method outperforms it, followed by the natural break-FR method.