Spatiotemporal Prediction of Rainfall-induced Landslides Using Machine Learning Techniques

Abstract Landslides cause significant damage to infrastructure and loss of life. Landslide susceptibility map (LSM), as an important reference for landslide hazard assessment, indicates prone areas of landslides. However, conventional LSM only predicts the spatial distribution of potential landslides, which cannot fully explain the occurrence of landslides at different times. In the present study, a spatiotemporal LSM is conducted to predict landslides both in space and time. A landslide database containing 223 recorded landslide events in southwestern Pennsylvania is used. Fourteen topographic spatial factors and eight rainfall temporal factors are used for machine learning (ML). Four ML models are applied in the study, including Logistic regression (LR), Support vector machine (SVM), Random forest (RF), and Gradient boosting machine (GBM). The results show that through cross-validation, RF outperforms the other algorithms with a value 0.86 of the area under the receiver operating characteristic curve (AUC score). The optimal model is used to generate spatiotemporal LSMs. It is concluded that by introducing spatial and temporal information simultaneously, ML models have the capability of learning the pattern of landslide occurrence both in space and time, providing an effective assessment tool to reduce catastrophic loss of landslide hazards.