A Comparison of Class Imbalance Techniques for Real-World Landslide Predictions

Landslides cause lots of damage to life and property world over. There has been research in machine-learning that aims to predict landslides based on the statistical analysis of historical landslide events and its triggering factors. However, prediction of landslides suffers from a class-imbalance problem as landslides and land-movement are very rare events. In this paper, we apply state-of-the-art techniques to correct the class imbalance in landslide datasets. More specifically, to overcome the class-imbalance problem, we use different synthetic and oversampling techniques to a real-world landslide data collected from the Chandigarh - Manali highway. Also, we apply several machine-learning algorithms to the landslide data set for predicting landslides and evaluating our algorithms. Different algorithms have been assessed using techniques like the area under the ROC curve (AUC) and sensitivity index (d'). Results suggested that random forest algorithm performed better compared to other classification techniques like neural networks, logistic regression, support vector machines, and decision trees. Furthermore, among class-imbalance methods, the Synthetic Minority Oversampling Technique with iterative partitioning filter (SMOTE-IPF) performed better than other techniques. We highlight the implications of our results and methods for predicting landslides in the real world.

I. Introduction

The Himalayan Region has been prone to landslide hazards mainly due to its precarious topographic characteristics and tectonic dynamism [14]. Landslide causes damage to property, blocks roads for days and kills and traps people. For places at very high altitude where everything from food to clothing is imported from cities, blocking roads is a huge issue. These reasons become very supporting evidence for combating this problem at the earliest.