Landslide Susceptibility Assessment Through TrAdaBoost Transfer Learning Models Using Two Landslide Inventories
Zhiyong Fu,Changdong Li,Wenmin Yao
DOI: https://doi.org/10.1016/j.catena.2022.106799
IF: 6.367
2023-01-01
CATENA
Abstract:The prediction performance of conventional landslide susceptibility (LS) models is generally limited by the size of samples in the landslide inventory, as unsatisfying performance may be yielded with landslide samples less than certain thresholds. Increasing landslide samples is the most reliable approach for solving this problem, but it can be expensive, time-consuming, and even unable to conduct. In this study, a novel method is presented that improves the performance of the LS models using the TrAdaBoost transfer learning algorithm. The proposed method transfers useful knowledge from one landslide inventory to another to improve the performance of LS models through which the effort to recollect landslide data is reduced. A database involving 373 historical landslide locations and 5 landslide influencing factors (LIFs, the slope angle, slope aspect, altitude, lithology, and curvature) in the study area (Zigui Basin, China) and 4,120 historical landslide locations and the corresponding LIFs in the source area (Wenchuan-Yingxiu area, China) were used for demonstration. The frequency ratio method was used to tackle problems of feature dissimilarities of LIFs between the study area and the source area. And then, with these quantified influencing factors as inputs, the three TrAdaBoost models (with decision trees (DT), support vector machine (SVM), and random forest (RF) as basic learners, namely TrAdaBoost-DT, TrAdaBoost-SVM, and TrAdaBoost-RF, respectively), and three conventional machine learning models (DT, SVM, RF) were adopted for showing the performance of the TrAdaBoost in improving the LS models. The area under the receiver operating characteristic curve (AUC) and the existing landslides were used to evaluate the performance of the LS models. The calculated results show that the AUC values of the DT, SVM, RF, TrAdaBoost-DT, TrAdaBoost-SVM, and TrAdaBoost-RF are 0.73, 0.82, 0.83, 0.80, 0.85, and 0.85, respectively; the landslide prediction accuracies of these models are 69%, 77%, 71%, 74%, 87% and 75%, respectively. Compared to the aforementioned results, when using the landslide inventory in the Wenchuan-Yingxiu area, the AUCs of the DT, SVM and RF increase by 0.07, 0.03 and 0.02, respectively, and the landslide prediction accuracies increase by 5%, 10% and 4%, respectively. The results of this study present that in the uncomplete landslide inventory environment, using the TrAdaBoost model to improve the performance of LS models is promising due to its low costs and the great improvement to LS models.