Comparative assessment of ensemble and machine learning models for susceptibility mapping of 2024 Wushi earthquake-triggered landslides
Vergleichende Bewertung von Ensemble- und Machine-Learning-Modellen für die Suszeptibilitätskartierung von durch das Erdbeben 2024 in Wushi ausgelösten Hangrutschen
PLOS One · 10.11.2026 · 21 (11) · e0362380
DOI: 10.1371/journal.pone.0362380 ↗Auf einen Blick
Für 904 durch das Erdbeben 2024 in Wushi ausgelöste Hangrutsche wurde ein LightGBM-Modell mit ANN, DT und NB verglichen. LightGBM zeigte die beste Vorhersageleistung im Validierungsdatensatz.
Fragestellung
Wie gut sagt LightGBM die Suszeptibilität für durch das Wushi-Erdbeben 2024 ausgelöste Hangrutsche im Vergleich zu ANN, DT und NB vorher?
Methode
904 Rutschungen wurden in Trainings- (70 %) und Validierungsdaten (30 %) geteilt. Verglichen wurden LightGBM, ANN, DT und NB; bewertet mit AUC, Accuracy und F1-Score.
Ergebnisse
LightGBM erreichte AUC 0,961, Accuracy 0,890 und F1-Score 0,895; ANN AUC 0,957, DT 0,947, NB 0,919. Wichtige Faktoren waren Höhe, PGA und Distanz zu Straßen.
Schlussfolgerung der Autor:innen
Die Ergebnisse zeigen, dass LightGBM die beste Gesamtvorhersageleistung und eine gute Generalisierungsfähigkeit aufwies.
Kernaussagen
- 904 Rutschungen aus dem Biedieli-Flussbecken wurden analysiert
- LightGBM hatte die beste Validierungsleistung: AUC 0,961
- Wichtige Faktoren: Höhe, PGA und Distanz zu Straßen
Deutsche Zusammenfassung: KI-gestützt aus dem Original-Abstract erstellt; alle Zahlen wurden automatisch mit dem Original abgeglichen. Maßgeblich ist der Original-Abstract.
Kurz gefasst
Vergleichsstudie in der Zeitschrift PLOS One, online erschienen am 10.11.2026. Verfasst von Yutong Wu und 5 weiteren Autor:innen. Beteiligte Einrichtung: School of Geology and Mining Engineering, Xinjiang University, Urumqi, China. Hauptthemen laut MEDLINE: Earthquakes, Machine Learning, Landslides. Die Arbeit nennt 48 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National College Students Innovation and Entrepreneurship Training Program, Natural Science Foundation of Xinjiang Uygur Autonomous Region, Science and Technology Department of Xinjiang Uygur Autonomous Region.
Abstract (Originalsprache)
Coseismic landslide susceptibility mapping plays a critical role in disaster risk mitigation, infrastructure protection, and territorial spatial planning, particularly in tectonically active mountain belts such as the Tianshan Mountains. Conventional machine learning models often suffer from overfitting and inadequate generalization capability due to their homogeneous structures and inflexible modeling strategies. In this study, targeting the coseismic landslides triggered by the 2024 Ms 7.1 Wushi earthquake in the Biedieli River basin, a landslide inventory comprising 904 events was compiled and randomly partitioned into training (70%) and validation (30%) datasets. Nine conditioning factors pertaining to topography, geology, environment, and seismicity were selected, including elevation, slope angle, slope aspect, curvature, lithology, distance to faults, distance to rivers, distance to roads, and peak ground acceleration (PGA). The predictive performance of an ensemble learning model, LightGBM, was compared against three benchmark machine learning models, i.e., artificial neural network (ANN), decision tree (DT), and naive Bayes (NB). Multicollinearity diagnosis and factor importance ranking were conducted using variance inflation factor (VIF) and correlation attribute evaluation (CAE), respectively. The model performance was validated and compared via the area under the receiver operating characteristic curve (AUC) and several statistical metrics including accuracy and F 1 -score. The results indicate that elevation (AM = 0.6296), PGA (AM = 0.5810), and distance to roads (AM = 0.4934) are the primary determinants of landslide occurrence. The LightGBM model achieved the best overall predictive performance on the validation dataset, with an AUC of 0.961 and the highest accuracy (0.890) and F 1 -score (0.895), followed by ANN (AUC = 0.957, accuracy = 0.876, F 1 -score = 0.876), DT (AUC = 0.947, accuracy = 0.856, F 1 -score = 0.860), and NB (AUC = 0.919, accuracy = 0.862, F 1 -score = 0.869). From the training dataset to the validation dataset, the AUC decay of LightGBM was merely 2.6%, indicating good generalization capability. The high and very high susceptibility zones account for 10.51% to 19.43% of the study area across the four models, predominantly concentrated in the mid-elevation valley zones (2036.5–2959 m), steep slopes exceeding 40°, areas with PGA greater than 0.48 g, and the vicinity of the Maidan fault zone. The susceptibility maps produced in this study can provide a scientific basis for post-earthquake emergency response, hazard risk assessment, and disaster mitigation in the Tianshan seismic belt.
Schlagwörter
MeSH-Schlagwörter (MEDLINE)
Autor:innen und Institutionen
- School of Geology and Mining Engineering, Xinjiang University, Urumqi, China.
- Key Laboratory of Green and Efficient Mining and Ecological Restoration in High-Altitude Arid Regions of Xinjiang, Urumqi, China.
ORCID-Profile: Junpeng Huang
Förderung
- National College Students Innovation and Entrepreneurship Training Program
- Natural Science Foundation of Xinjiang Uygur Autonomous Region
- Science and Technology Department of Xinjiang Uygur Autonomous Region
- Major Science and Technology Project of Xinjiang Uygur Autonomous Region
- Xinjiang Uygur Autonomous Region Tianchi Talent Introduction Plan
Zitieren
APA
Wu, Y., Huang, J., Li, W., Jia, X., Li, Y., & Wang, Z. (2026). Comparative assessment of ensemble and machine learning models for susceptibility mapping of 2024 Wushi earthquake-triggered landslides. PLOS One, 21(11), e0362380. https://doi.org/10.1371/journal.pone.0362380
Vancouver
Wu Y, Huang J, Li W, Jia X, Li Y, Wang Z. Comparative assessment of ensemble and machine learning models for susceptibility mapping of 2024 Wushi earthquake-triggered landslides. PLOS One. 2026;21(11):e0362380. doi:10.1371/journal.pone.0362380
BibTeX
@article{Wu2026,
author = {Wu, Yutong and Huang, Junpeng and Li, Wenqian and Jia, Xiangfeng and Li, Yaning and Wang, Zhiyi},
title = {Comparative assessment of ensemble and machine learning models for susceptibility mapping of 2024 Wushi earthquake-triggered landslides},
journal = {PLOS One},
year = {2026},
volume = {21},
number = {11},
pages = {e0362380},
doi = {10.1371/journal.pone.0362380}
}
EUVIMED