Landslide Risk Prediction and Regional Dependence Evaluation Based on Disaster History Using Machine Learning and Deep Learning

Around the world, many people were killed and injured by landslides, and thousands of houses and buildings were destroyed. Therefore, the evaluation and analysis of landslide hazards are very important for human, environmental, cultural, economic and social sustainability. This paper discusses the susceptibility of landslides and the regional dependence, in other words differences in characteristics between regions, based on ArcGIS and AI technologies such as machine learning and deep learning. Considering previous studies and various kind of data, 11 factors related to historical landslide: elevation, slope, aspect, curvature, lithology, soil, land-use, precipitation, distance from rivers, faults, and roads are selected in this research. The above characteristic data are aggregated and processed through ArcGIS to train the models and verify their performance, and finally check their versatility by applying them to the other regions. In machine learning, KNN, DT and MLP (ANN) models outperformed LR and RF models. In deep learning, the performance of CNN depends on the structure of the model, i.e., layer depth, input window size, and training strategy. It is necessary to optimize the deep learning model structure to show the better performance than machine learning. A model that is optimized in a specific region may not necessarily be applicable in another region as it is observed that the LR model and the DT model in the HKD area performed far less than the results for the other area (regional dependence). In this study for the three regions, the KNN model and the MLP model comprehensively outperform the other models from the aspect of the regional dependence.