Predicting Landslide Susceptibility by Using Logistic Regression and Random Forest at Different Spatial Resolution in Taiwan
Shyue-Cherng Liaw1, Uen-Hao Wang2 and Wan-Jiun Chen3

1 Professor, Department of Geography, National Taiwan Normal University, Taiwan
2 Assistant Researcher, Forest Management Division, Taiwan Forest Research Institute, Taiwan
3 Professor, Institute of Natural Resource and Environmental Management, National Taipei University, Taiwan

Abstract

This study aims to compare the effectiveness of logistic regression and random forest methods in predicting landslide susceptibility, using four spatial resolutions for analysis. The research site is located in the Lioukuei Experimental Forest, Kaohsiung City. Landslide susceptibility models were developed and analyzed based on land cover maps from August 2009 (post-Typhoon Morakot) and January 2024. Variable selection for potential landslide influencers included both variables identified in previous studies, while this research uniquely incorporates canopy structural indices derived from airborne LiDAR, resulting in a total of 14 variables used for model construction. Results indicate that the random forest model outperforms the logistic regression model across all spatial scales, with the 10m model achieving optimal performance. The 10m model validation shows an AUC of 0.929 and an accuracy of 85.74%, demonstrating excellent predictive discrimination. Furthermore, in the 10m random forest model, four canopy structure indicators (Canopy Cover Ratio, Mean Top-of-the-Canopy, Variance of Canopy Height, and Canopy Volume Ratio) rank as the top variables, confirming their significant contribution to landslide susceptibility prediction by effectively integrating forest structure characteristics. This study provides valuable insights into spatial scale selection and LiDAR data application for landslide susceptibility modeling and offers a scientific basis for forest management and disaster prevention strategies.

Keywords: machine learning, airborne LiDAR, canopy structure indices, Lioukuei Experimental Forest.

Topic: Topic D: Geospatial Data Integration