Assessing the Consistency of Landslide Susceptibility Models: Frequency Ratio, Logistic Regression, Artificial Neural Network, and Random Forest

Abstract

Abstract Several models have been proposed to analyze landslide susceptibility, including physically-based models, index-based models, statistical models, and machine-learning algorithms. Previous studies focusing on model comparison mainly determined a better model based on predicting accuracies. In this study, we suggest a better model should not only produce results with higher accuracy but also be consistent in its performance. This study aims to provide a new perspective to assess the model performance, by analyzing the consistency of modeling outcomes. This requires multiple modeling trails. Specifically, four commonly used models were selected and tested, including frequency ratio (FR), logistic regression (LR), artificial neural network (ANN), and random forest (RF). The study area is the Thu Lum basin, located in the mountainous range of Lai Chau Province, Viet Nam. This study applied 13 predisposing factors, and the model training and testing procedures were randomly performed multiple times, from 5 times to 50 times for each model. Seven accuracy indexes were used to summarize and assess model consistency. We also explored the consistency of each factor’s contribution in different models. The result shows that 10 independent modeling trials are acceptable to reveal the model consistency, and among the four models, the RF model is considered the best one because it consistently produces higher accuracies. However, it also shows inconsistent importance rank of predisposing factors in different trials, which leads to higher uncertainty in explaining the landslide environment. To address this issue, we suggest finding consensus from multiple modeling outcomes could be a more reliable approach. We expect this study can be a useful reference for determining a suitable model for analyzing landslide susceptibility in a given area.