Regional landslide susceptibility assessment based on improved semi-supervised clustering and deep learning

Data-driven model has been increasingly used in landslide susceptibility mapping (LSM). However, the traditional landslide susceptibility assessment lacks reasonable and effective criteria for the selection of negative (non-landslide) samples, which limit the modeling effect of machine learning models. This study proposes a semi-supervised fuzzy C-means clustering based on cosine Euclidean distance (C-SFCM) for reliable landslide negative sample selection, and the genetic particle swarm optimization (GAPSO) algorithm is used to optimize the initial clustering centers. Taking Yiliang County, Yunnan Province, China, as the study area, we selected nine factors, including topography, geological conditions, land cover and human activities as influencing factors for landslide susceptibility mapping, and construct the basic dataset with landslide catalogs. The C-SFCM is performed on the original data set containing a small number of landslide samples and unlabeled samples and to select the low affiliation data far away from the positive samples of landslides as reliable negative samples of landslides. Then, we construct three deep learning models, the 1D convolutional neural network gated recurrent unit (1D CNN-GRU), the 1D convolutional neural networks long short term memory (1D CNN-LSTM) and the deep neural network (DNN), and two traditional machine learning models, support vector machines (SVM) and logistic regression (LR) for training and testing, respectively, and used them for landslide susceptibility mapping. The results show that C-SFCM has a more reasonable negative sample area, while it can identify the spatial characteristics of landslide samples and negative samples effectively than SFCM under traditional Euclidean distance. Compared with the random selection of negative samples scheme, the classifier trained on the C-SFCM-selected dataset has higher accuracy and better prediction capability for potential landslides. Furthermore, 1D CNN-GRU achieved the highest prediction accuracy (AUC = 0.918) in the C-SFCM scenario, with a 9.55% improvement compared to the random selection scenario.