A hybrid ensemble-based deep-learning framework for landslide susceptibility mapping

• Published in: International journal of applied earth observation and geoinformation Vol. 108; p. 102713
• Main Authors: Lv, Liang, Chen, Tao, Dou, Jie, Plaza, Antonio
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.04.2022
• Subjects: data collection; Deep learning; Heterogeneous ensemble learning; humans; Landslide susceptibility mapping; landslides; neural networks; quantitative analysis; spatial data; Stacking; Deep learning; Landslide susceptibility mapping; Stacking; Heterogeneous ensemble learning
• ISSN: 1569-8432; 1872-826X
• DOI: 10.1016/j.jag.2022.102713

[Display omitted] •Fully used the deep learning model as the base model for heterogeneous ensemble.•Stacking ensemble method improved the performance of the base model effectively.•Ensemble learning methods reduced the uncertainty of the single models’ results.•The importance and correlation of all LCFs were analyzed before training. Landslides are highly hazardous geological disasters that can potentially threaten the safety of human life and property. As a result, landslide susceptibility mapping (LSM) plays an important role in the landslide prevention system. Recently, many deep learning (DL) models have been adopted for LSM, but they also face problems such as sensitivity to overfitting and lower mapping accuracy. In this paper, a novel hybrid LSM framework is proposed based on four heterogeneous ensemble learning (HEL) methods with three single DL models: deep belief network (DBN), convolutional neural network (CNN) and deep residual network (ResNet). The proposed model is tested at the Three Gorges Reservoir area, China. 202 historical landslides and ten conditioning factors were selected to construct a geospatial dataset for LSM. The conditioning factors with high-correlation and low importance were removed from the dataset by using the Spearman Correlation Index and random forests. The geospatial dataset was then divided into two subsets: 70% for training and 30% for testing. Then LSM results was carried out by the single and proposed HEL-based models. The quantitative evaluation of the results showed that the proposed HEL-based models improved the LSM accuracy, and outperformed the single DL LSM models. Stacking model achieved the highest AUC value (0.984), highest Kappa (86.95%), highest overall accuracy (94.17%), highest precision (88.87%), highest Matthews correlation coefficient (87.03%) and highest F1-score (91.34%) among all of models for the testing dataset, while the Boosting model obtained the highest Recall value (96.02%). At the same time, HEL-based models proposed in this study also show better stability and can avoid the overfitting effectively. In addition, the Gini index showed that elevation factor contributes most in LSM in the study area. In general, the proposed framework has promising applicability in improving LSM accuracy.