Assessment of landslide susceptibility along mountain highways based on different machine learning algorithms and mapping units by hybrid factors screening and sample optimization

To develop a better spatial prediction model of landslide susceptibility along mountain highways, this study compared assessment models of landslide susceptibility along mountain highways based on different machine-learning-algorithms (random forest (RF), support vector machine (SVM)), and mapping units (grid and slope units) using factor screening and sample optimization. Using a typical mountainous county, Chengkou, as a case study, a landslide inventory was prepared based on field investigations, satellite images, and historical records. In total, 334 landslides were identified. Twenty landslide conditioning factors influenced by topography, geology, environmental conditions, and human activities were initially identified. Less important conditioning factors were identified and eliminated using a geographical detector. Two common non-linear machine-learning models (RF and SVM) were used under the grid and slope units, and on all cells (including landslides and non-landslides); 10-fold cross-validation was utilized to select the training and validation datasets for these models. Subsequently, the trained models were utilized for landslide susceptibility mapping in the entire study area, divided into five classes. Subsequently, the performance of the models under different units was evaluated using a confusion matrix and receiver operating characteristic curve. Finally, partial dependency plots (PDP) and local interpretable modelagnostic explanations (LIME) were used to study the interpretability of landslide susceptibility model. The research results show that after removing the noise-generating factors, the model trained by the remaining 11 landslide-conditioning factors can deliver ideal forecast accuracy. In this study, the evaluation results of the slope unit for the RF or SVM model were more accurate and reasonable than those of the grid unit. In this case, the slope unit was found to be ideal. Under both the slope and grid units, the performance of the RF model was better than that of the SVM model. This study provides a reference for the spatial prediction of landslide susceptibility along mountain highways. (c) 2022 International Association for Gondwana Research. Published by Elsevier B.V. All rights reserved.

Automated summary

This study compared assessment models of landslide susceptibility along mountain highways using different machine learning algorithms and mapping units. The results showed that after removing noise-generating factors, the model trained with the remaining landslide-conditioning factors achieved ideal forecast accuracy. The slope unit provided more accurate and reasonable evaluation results, and the RF model outperformed the SVM model.