Impact of Extreme Class-Imbalance on Landslide-Risk Prediction and Mitigation Using Two-Stage Deep Neural Network

Abstract

The classification of landslides is the one of the most challenging topics because of the complexity of the relationships of various dynamic and uncertain factors and the physical gaining data processes. Landslide incidents frequently occur in the upper northern region of Thailand due to its topography. The landslide classification method is proposed to capture the significant features from an extreme case of class-imbalance dataset. The imbalanced nature of some real-world data is one of the current challenges for machine learning researchers. A study area of this research covers an area of 25 square kilometers at Chiang Rai, in Thailand which contains 30,408 non-landslide data and 1,077 landslide data. The percentage of landslide data is 3.54% of the total data. This paper proposed the solution of landslide classification given by extreme class-imbalance dataset. The proposed method has two main steps i.e., firstly, mitigating class-imbalance dataset. and secondly two-stage learning. Performance of the proposed method benchmarks the baseline, the logistic regression (LR), the random forest classifier (RFC) methods given by enhanced dataset. In the case of imbalance dataset, the one-class method is assessed against the proposed method along with the LR and the RFC methods. Experimental results demonstrate that the proposed method has improved the landslide-risk prediction performance over the baseline, the LR, the RFC and the one-class SVM methods in terms of an average area under the curve scores by 0.48, 0.48, 0.03, and 0.06, respectively, in both enhanced dataset and imbalance dataset.