AUTHOR=Ling Sixiang , Zhao Siyuan , Huang Junpeng , Zhang Xuantu TITLE=Landslide susceptibility assessment using statistical and machine learning techniques: A case study in the upper reaches of the Minjiang River, southwestern China JOURNAL=Frontiers in Earth Science VOLUME=10 YEAR=2022 URL=https://www.frontiersin.org/journals/earth-science/articles/10.3389/feart.2022.986172 DOI=10.3389/feart.2022.986172 ISSN=2296-6463 ABSTRACT=

Landslides have frequently occurred in deeply incised valleys in the upper reaches of the Minjiang River. Long-term interactions between rock uplift and river undercutting developed widely distributed landslides in this catchment, which recorded the typical tectonic geomorphology in the eastern margin of the Tibetan Plateau. In this study, we examined the landslides in the Minjiang catchment and aimed to compare the prediction ability of the statistical and machine learning (ML) models in landslide susceptibility assessment. We adopted the statistical models of the frequency ratio (FR) and information value (IV) models, and the ML models represented by a logistic model tree (LMT) and radial basis function classifier (RBFC) for landslide prediction. An inventory map of 668 landslides was compiled, and the landslides were randomly divided into training (80%) and validation (20%) datasets. Furthermore, 11 control factors of landslides based on topography, geology, hydrology, and other environments were applied for the analysis. The comprehensive performance of the four models was validated and compared using accuracy and area under the receiver operating characteristic curve (AUC). The results indicated that both sides of the valley along the Mingjiang and Heishuihe Rivers are in the high and very high susceptibility zones; in particular, the river segment from Wenchuan to Maoxian County has the highest susceptibility. The AUC values of the FR, IV, LMT, and RBFC models with the training data were 0.842, 0.862, 0.898, and 0.894, respectively, while the validation dataset illustrated the highest AUC value of 0.879 in the LMT model, followed by the RBFC (0.871), IV (0.869), and FR (0.839) models. Moreover, the LMT and RBFC models had higher accuracy values than the FR and IV models. This suggests that the ML models are superior to the statistical models in generating adequate landslide susceptibility maps, and the LMT model is the most efficient one for landslide prediction in the study region. This study provides a typical case in a landslide-prone region in the plateau margin to advance the understanding of landslide susceptibility assessment.