Discussion on the tree-based machine learning model in the study of landslide susceptibility

This study reported an application of the tree-based models to landslide susceptibility. The landslide inventory and ten conditioning factors were first constructed, based on data availability and climate. Subsequently, three tree-based models, decision tree (DT), DT-Boosting, and random forest (RF)...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Natural hazards (Dordrecht) 2022-09, Vol.113 (2), p.887-911
Hauptverfasser: Liu, Qiang, Tang, Aiping, Huang, Ziyuan, Sun, Lixin, Han, Xiaosheng
Format: Artikel
Sprache:eng
Schlagworte:
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:This study reported an application of the tree-based models to landslide susceptibility. The landslide inventory and ten conditioning factors were first constructed, based on data availability and climate. Subsequently, three tree-based models, decision tree (DT), DT-Boosting, and random forest (RF), were established and compared with the support vector machine (SVM) to analyze the difference in model prediction. Finally, the effect and causes of tree-based algorithms on prediction results were explored based on the working mechanism of the susceptibility model. Results show that there is no multicollinearity among the conditioning factors. The predicted results produced by the tree-based model display the discontinuous distribution compared with the SVM, not only presented in the point-based prediction but the surface-based heterogeneity. Moreover, heterogeneity on the susceptibility map relates to the tree-based algorithm and factor grading, especially the classification of important factors. Besides, DT-Boosting appears the highest numerical features, with large values of AUC (0.981), specificity (0.960), sensitivity (0.956) and accuracy (0.958) in the training phase, and high prediction of AUC (0.862), specificity (0.759), sensitivity (0.843) and accuracy (0.801) in the validation phase. In terms of fluctuation, the RF is smaller than that of DT-Boosting. Further, the susceptibility map generated by RF, with the largest D -value of 7.81, can well capture the difference in landslide susceptibility. This study provides a deep understanding for the application of tree-based machine learning models to landslide susceptibility.
ISSN:0921-030X
1573-0840
DOI:10.1007/s11069-022-05329-4