Machine Learning Feature Selection Methods for Landslide Susceptibility Mapping

Abstract

This paper explores the use of adaptive support vector machines, random forests and AdaBoost for landslide susceptibility mapping in three separated regions of Canton Vaud, Switzerland, based on a set of geological, hydrological and morphological features. The feature selection properties of the three algorithms are studied to analyze the relevance of features in controlling the spatial distribution of landslides. The elimination of irrelevant features gives simpler, lower dimensional models while keeping the classification performance high. An object-based sampling procedure is considered to reduce the spatial autocorrelation of data and to estimate more reliably generalization skills when applying the model to predict the occurrence of new unknown landslides. The accuracy of the models, the relevance of features and the quality of landslide susceptibility maps were found to be high in the regions characterized by shallow landslides and low in the ones with deep-seated landslides. Despite providing similar skill, random forests and AdaBoost were found to be more efficient in performing feature selection than adaptive support vector machines. The results of this study reveal the strengths of the classification algorithms, but evidence: (1) the need for relying on more than one method for the identification of relevant variables;

Landslide susceptibility mapping using random forest, boosted regression tree, classification and regression tree, and general linear models and comparison of their performance at Wadi Tayyah Basin, Asir Region, Saudi Arabia

“…Using these guidelines, the number of trees in RF has been fixed to 1000 after a primary analysis and the m sampled at each node has been selected to be 3 to analyze the joint contribution of subsets of features while keeping a fast convergence during iterations. No calibration set is needed to regulate the parameters (Micheletti et al 2014). Two types of error were assessed: mean decrease in accuracy and mean decrease in node impurity (mean decrease Gini) (Calle and Urrea 2010).…”

A Comparative Assessment Between Three Machine Learning Models and Their Performance Comparison by Bivariate and Multivariate Statistical Methods in Groundwater Potential Mapping

“…RFs are very powerful and flexible ensemble classifiers based upon decision trees, the first developed by Breiman (2001) (Catani et al 2013;Micheletti et al 2014). RF consists of a combination of many trees, where each tree is generated by boot-strap samples, leaving about a third of the overall sample for validation (the out-of-bag predictions-OOB) (Oliveira et al 2012).…”

Extreme Learning Machines for spatial environmental data

“…Usually this space is constructed using expert and science-based knowledge and can be either incomplete or redundant. Therefore, feature selection is an important task (Micheletti et al, 2014). In the next sections, by following the complete methodology, we concentrate the presentation only on the new and the most relevant properties of ELM and corresponding results.…”

“…Machine learning algorithms, principally based on statistical learning theory (Hastie et al, 2009;Vapnik, 1998), being a universal non-linear modelling tools, play an important role in the modelling of environmental spatial data (Cracknell and Reading, 2014;Hsieh, 2009;Kanevski et al, 2004Kanevski et al, , 2009Melchiorre et al, 2011;Micheletti et al, 2014;Nefeslioglu et al, 2008). Recently, a new approach in machine learning, Extreme Learning Machine (ELM) (Huang et al, 2006), has gained a great popularity in the computer science community.…”