Artificial Intelligence and Data Science
Online ISSN : 2435-9262
Machine learning-based assessment of earthquake-induced landslide occurrence probability incorporating probability calibration and spatial error evaluation
Keigo KAWARA, Yuka ISHIKAWA, Hiroki ISHIBASHI
Author information
JOURNAL OPEN ACCESS

2026 Volume 7 Issue 3 Pages 14-25

Details
Abstract

Estimating the occurrence probability of earthquake-induced landslides is important for promoting efficient disaster management. This study developed an integrated framework combining machine learning-based probability estimation, probability calibration and performance evaluation considering spatial prediction errors. The framework was applied to landslides triggered by the 2024 Noto Peninsula earthquake. Logistic regression and LightGBM models were constructed, and their predictive performance was compared. The results showed that LightGBM achieved higher predictive performance than logistic regression. In addition, distance-based performance metrics were proposed to evaluate spatial errors between predicted and observed landslide locations. The results indicated that the LightGBM model had practical predictive performance when spatial prediction errors of approximately 100 m were allowed. Furthermore, probability calibration using isotonic regression improved the underestimation of landslide occurrence probability in the high-probability ranges. The calibrated probabilities showed greater consistency with the observed frequency of landslide occurrence. These results demonstrate that the proposed framework can support wide-area prioritization of landslide disaster countermeasures while accounting for allowable spatial prediction error in practical applications.

Content from these authors
© 2026 Japan Society of Civil Engineers
Previous article Next article
feedback
Top