Landslide Susceptibility Prediction Considering Spatio-Temporal Division Principle of Training/Testing Datasets in Machine Learning Models

In most of the landslide susceptibility prediction (LSP) models, the landslide-non landslide spatial datasets are divided into training/testing datasets according to the principle of spatial random, however, this spatial randomness division inevitably introduces uncertainties into LSP modelling. Theoretically, LSP modelling is based on past landslide inventories to predict the spatial probability of future landslides, which has significant time series characteristics rather than only spatial random characteristics. Therefore, we believe that it is necessary to divide spatial datasets into the model training/testing datasets based on the time series of landslide occurrence. Taking Wencheng County in China as an example, 11 types of environmental factors and 128 time-accurate landslides are obtained; Then, the landslide and non-landslide samples connected with environmental factors are divided into two different types of training/testing datasets according to the principles of landslide time series and spatial random, respectively. The division ratios of training/testing datasets are set as 9∶1, 8∶2, 7∶3, 6∶4 and 5∶5, respectively, to avoid the influences of different ratios on the LSP results. Thus, the training/testing datasets under 10 combined working conditions are obtained. Finally, several typical machine learning models, such as Support Vector Machine (SVM), Multi-Layer Perceptron (MLP) and Random Forest (RF), are then trained and tested to perform LSP and analyze their uncertainties. Results show that: (1) The LSP uncertainties performed by the time series-based SVM, MLP and RF models are slightly lower than those by spatial random-based models, which verifies the feasibility of dividing by time series; (2) The time series division of training/testing datasets is actually a "deterministic" case among the spatial random division, which is more consistent with the actual situation of landslides. Of course, it is also feasible to carry out spatial random division for training and testing datasets when lacking landslide occurrence time.