Introducing a Pattern-Based Approach for Landslide Susceptibility Prediction

Machine learning (ML) models are extensively used in data-driven landslide susceptibility prediction (LSP). Dataset used in LSP with ML containing positive (landslide) samples and negative (non-landslide) samples, while the spatial biases of non-landslide samples for LSP are frequently ignored. The main objective of this study is to develop a pattern-based approach that properly tackles the spatial biases of non-landslide samples combining two models, i.e. balanced iterative reducing and clustering using hierarchies (BIRCH) and Random Forest (RF). In this study, BIRCH is employed to select four types of non-landslide samples, representing four spatial patterns. In the meanwhile, another set of non-landslide samples is randomly selected to serve as control. RF model is trained to calculate the susceptibility index using these five types of non-landslide samples along with landslide samples derived from landslide inventory, producing five LSP scenarios. Results indicate that the pattern-based approach offers an effective way to find the non-landslide samples and provide sufficient and reliable spatial patterns, and therefore proves itself as a better solution to the LSP.

Keywords: Landslide susceptibility, Machine learning, Risk, Remote sensing.