Week-ahead shallow landslide displacement prediction using chaotic models and robust LSTM

Abstract

Time-series monitoring of landslide displacement is crucial for controlling the geo-risk associated with sudden landslide occurrence and slope failure. Accurate prediction is valuable for geohazard mitigation in advance of short-term displacement. In this research, a novel chaotic modeling framework is proposed to predict landslide displacement using a robust long short-term memory (LSTM) network. To facilitate the prediction framework, daily instant displacement is measured in three dimensions at 19 monitoring locations. Then, the chaotic characteristics are computed for data reconstruction purposes, and the reconstructed data are selected as inputs in the prediction model. Next, LSTM is applied as the prediction algorithm and is trained using reconstructed field data. A generic LSTM is often trained to minimize the mean square error (MSE) loss, which can be oversensitive to a few outliers. In this research, the pseudo-Huber loss is adopted as the loss function and is integrated with LSTM as an improvement over the MSE loss. The effectiveness and efficiency of the proposed framework have been validated by the benchmark LSTM and other machine learning algorithms. The computational results show that the proposed approach performed better than conventional LSTM and other machine learning algorithms. This framework may be valuable for engineers for practical landslide hazard estimation or rapid preliminary screening of slope stability.