Application of convolutional neural network fused with machine learning modeling framework for geospatial comparative analysis of landslide susceptibility

Landslides in mountain settlements are among the most widespread and dangerous geohazards. In this study, we aimed to assess landslide susceptibility using Wenchuan, southwest China, as a case. For this purpose, we constructed an optimization method that combines a convolutional neural network with the machine learning algorithm of support vector machines, quadratic discriminant analysis, Bayesian optimized gradient boosting tree, and Bayesian optimized random forest. The model inputs were 13,886 historical seismic-induced landslide events interpreted from remote sensing imagery and ten evaluation features: elevation, slope angle, slope aspect, plan curvature, profile curvature, distance to roads, distance to rivers, distance to faults, land use pattern, and soil texture. The output was the probability of landslide occurrence for each prediction unit. Finally, we evaluated the assessed outcomes using both the receiver operating characteristic curve and 1074 latest recorded landslide dataset (2013–2020). The calculations showed that the overall susceptibility values to landslides in the high–very high interval produced by the hybrid convolutional neural networks was 9.95%–16.91%, which is close to the actual landslide susceptibility of the region. The receiver operating characteristic curve and statistical analysis of the latest landslide event outcomes demonstrated that the hybrid Bayesian optimized gradient boosting tree model had a higher classification accuracy than the other classifiers presented in this study. The research findings are available to local governments and disaster management authorities in guiding disaster prevention, mitigation policy formulation, and land use and provide reference value for evaluating landslide susceptibility in other mountainous areas.