High-resolution earthquake-induced landslide hazard assessment in Southwest China through frequency ratio analysis and LightGBM

Abstract

Earthquake-induced landslides cause extensive harm, necessitating accurate predictions for effective risk management. To tackle the dual challenges posed by inadequate model accuracy and the absence of frequency-derived landslide intensity as critical information, this paper presents a robust modelling method that generates high-resolution landslide hazard map for risk assessment. Firstly, the Frequency Ratio (FR) model is employed to quantitatively explore the correlation between geo-environmental factors and landslide occurrences, enabling the exclusion of less significant factors. Subsequently, these FR values are integrated into the Light Gradient Boosting Machine (LightGBM) model, resulting in the composite model, FR-LightGBM. To test the proposed model, data from the areas affected by the Luding earthquake based on the seismic landslide intensity model following the 2022 Ms 6.8 earthquake in Sichuan, China, is examined. Model validation, using the area under the curve (AUC) of the receiver operating characteristic curve, highlighted the superior predictive accuracy of the FR-LightGBM model, marking a 3.5 % improvement in AUC value compared to both the Convolutional Neural Network (CNN) and Logistic Regression (LR) models. The results have identified high-risk areas along the Dadu riverside and the Xianshuihe fault zone correspond well with the actual landslide distribution. Overall, the practical applicability of the FR-LightGBM model is significantly enhanced by its capacity to prioritize lithology, faults, and aspect as crucial factors. Additionally, based on an extended landslide inventory established using high spatial resolution (2 m) satellite imagery from the Chinese Gaofen satellite series, the proposed approach delivered a high-resolution (12.5 m) hazard map across the Luding earthquake area, contributing to relevant studies and risk management.

Keywords

• Earthquake-induced landslides
• Landslide hazard assessment
• Frequency ratio analysis
• LightGBM
• Gaofen-1/2