3DCNN landslide susceptibility considering spatial-factor features

Effective landslide disaster risk management contributes to sustainable development. A useful method for emergency management and landslide avoidance is Landslide Susceptibility Mapping (LSM). The statistical landslide susceptibility prediction model based on slope unit ignores the re-lationship between landslide triggering factors and spatial characteristics. It disregards the influence of adjacent image elements around the slope-unit element. Therefore, this paper proposes a hardwired kernels-3DCNN approach to LSMs considering spatial-factor features. This method effectively solved the problem of low dimensionality of 3D convolution in the hazard factor layer by combining Prewitt operators to enhance the generation of multi-level 3D cube input data sets. The susceptibility value of the target area was then calculated using a 3D convolution to extract spatial and multi-factor features between them. A geospatial dataset of 402 landslides in Xiangxi Tujia and Miao Autonomous Prefecture, Hunan Province, China, was created for this study. Nine landslide trigger factors, including topography and geomorphology, stratigraphic lithology, rainfall, and human influences, were employed in the LSM. The research area’s pixel points’ landslide probabilities were then estimated by the training model, yielding the sensitivity maps. According to the results of this study, the 3DCNN model performs better when spatial information are included and trigger variables are taken into account, as shown by the high values of the area under the receiver operating characteristic curve (AUC) and other quantitative metrics. The proposed model outperforms CNN and SVM in AUC by 4.3% and 5.9%, respectively. Thus, the 3DCNN model, with the addition of spatial attributes, effectively improves the prediction accuracy of LSM. At the same time, this paper found that the model performance of the proposed method is related to the actual space size of the landslide body by comparing the impact of input data of different scales on the proposed method.