Displacement prediction of water-induced landslides using a recurrent deep learning model

Abstract

Displacement prediction is a direct and effective method for mitigating geohazards. Due to the influence of rainfall and reservoir water level variations, landslides often display step-like deformations with an increasing trend and periodic fluctuation, indicating long-term memory in displacement time series. Traditional data-driven methods are mostly suitable for short-term prediction, and extra data processing is applied to solve this problem. This paper proposes a novel deep learning-based displacement prediction method using long short-term memory (LSTM) networks. Based on open-source frameworks for deep learning, namely, Keras and TensorFlow, a detailed implementation of displacement prediction is proposed and illustrated. The Baishuihe landslide, a typical landslide with long-term monitoring, is taken as a case study, and both single-factor and multi-factor predictions are performed. The results indicate that multi-factor prediction can reduce overfitting and improve accuracy. Compared with the existing method, the multi-factor deep-learning model displays better performance. This study indicates that the LSTM-based deep-learning model is suitable and convenient for displacement prediction and has broad prospects in safety monitoring of water-induced landslides.