Dynamic prediction of landslide life expectancy using ensemble system incorporating classical prediction models and machine learning

Abstract

With the development of landslide monitoring system, many attempts have been made to predict landslide failure-time utilizing monitoring data of displacements. Classical models (e.g., Verhulst, GM (1,1), and Saito models) that consider the characteristics of landslide displacement to determine the failure-time have been investigated extensively. In practice, monitoring is continuously implemented with monitoring data-set updated, meaning that the predicted landslide life expectancy (i.e., the lag between the predicted failure-time and time node at each instant of conducting the prediction) should be re-evaluated with time. This manner is termed “dynamic prediction”. However, the performances of the classical models have not been discussed in the context of the dynamic prediction yet. In this study, such performances are investigated firstly, and disadvantages of the classical models are then reported, incorporating the monitoring data from four real landslides. Subsequently, a more qualified ensemble model is proposed, where the individual classical models are integrated by machine learning (ML)-based meta-model. To evaluate the quality of the models under the dynamic prediction, a novel indicator termed “discredit index (β)” is proposed, and a higher value of β indicates lower prediction quality. It is found that Verhulst and Saito models would produce predicted results with significantly higher β, while GM (1,1) model would indicate results with the highest mean absolute error. Meanwhile, the ensemble models are found to be more accurate and qualified than the classical models. Here, the performance of decision tree regression-based ensemble model is the best among the various ML-based ensemble models.

Keywords

• Dynamic prediction
• Landslide life expectancy
• Machine learning
• Ensemble system