A Hybrid Physical and Machine Learning Model for Assessing Landslide Spatial Probability Caused by Raising of Ground Water Table and Earthquake in Atsuma, Japan — Case Study

Abstract

Landslides are catastrophic natural events primed and/or triggered by extreme rainfalls and strong earthquakes. Simultaneous occurrence of rainfall and seismic activity increases the likelihood of landslides. However, the researchers focused on this aspect are not much. In the present research, a hybrid model was developed to predict the landslide occurrences probability in Atsuma, Japan triggered by rainfalls and earthquakes. The proposed model is a combination of a physical and machine learning model for improving the accuracy of the landslide susceptibility mapping. The proposed model consisted of a physical module, a machine learning module and a matrix approach module. The physical module assessed the effects of rainfall and peak ground acceleration (PGA) on landslide occurrence probability based on a pseudo-static model. The machine learning module applied Multilayer Perceptron Neural Networks to assess landslide susceptibility, using 611 landslide events caused by strong earthquakes and extreme typhoons. The landslide susceptibility maps obtained from these two modules were then combined into final susceptibility map through a matrix approach. The final susceptibility map included five susceptible levels: very low, low, moderate, high, and very high. To evaluate the proposed model performance, the resulting models were assessed using the areas under the receiver operating characteristic curves. The areas under the success rate curves from the physical module, machine learning module and matrix-based approach showed 79.2%, 82.7% and 83.9% accuracy, respectively. Furthermore, the predicted rate curves showed that the areas under the curve for physical module, machine learning module and matrix-based approach were 78.4%, 82.3% and 83.4%, respectively. These results suggest that the proposed hybrid model improves the prediction capability compared to physically-based method or machine learning model and can be readily used to assess spatial probability of landslide.