An alternative approach for the landslide prediction using an interpretable machine learning method

Keywords: interpretable machine learning, landslide modeling, Shapley additive explanations, tree-based machine learning, InSAR
Abstract Type: Paper Abstract

Abstract

Since landslides can be a serious threat to people and their properties, identifying landslide-susceptible areas is one of the crucial phases of the hazard mitigation. Machine Learning approaches are becoming more popular in landslide modeling due to their ability to handle non-linearity and capture the interaction effects between feature variables. Tree-based ensemble methods have been widely adopted in terms of the level of prediction. By integrating poor-performing simple regression trees, tree-based ensemble can ensure a better prediction accuracy. Despite the favor of machine learning approaches in modeling landslides, understandable explanation for the model is not often provided, making the approach black-box. This paper seeks to a better landslide prediction using a tree-based ML algorithm with boosting. Particularly, we employ an Interpretable Machine Learning (IML) to improve unknown black-box explanations. More specifically, we use a gradient-boosting regression tree method (GBM) to examine and model landslides to enhance prediction accuracy with the interpretability of the model. We also utilize Shapley Additive Explanations (SHAP) method to show how and what decisions the model is making. For robust explanatory variables of landslides, remote sensing data derived from InSAR (Interferometric Synthetic-Aperture Radar), NDVI (Normalized Difference Vegetation Index), and other geo-environmental elements are used in the landslide prediction model. We believe that our results could be alternative by showing the adequate explainability and transparency of the black-box approaches based on machine learning methods, especially for the landslide modeling.