Prediction of Landslide Displacement Based on Long Term Monitoring Data and LSTM Network

Landslide displacement change is a direct representation of hazard. Displacement prediction is crucial to disaster prevention and mitigation. Taking the Bazimen landslide as an example， landslide displacement prediction was performed based on ten years of monitoring data and neural network models （LSTM， RNN）. The total displacement is decomposed into a trend term and a periodic term using the one-time moving average method. The trend term is predicted by segmental fitting using a cubic polynomial function， and the relationship between the periodic term and the characteristic factor is established and predicted by the neural network model. Among them， the characteristic factors of the periodic term are initially selected according to the displacement influencing factors， and then the irrelevant factors are eliminated by Pearson correlation analysis. The predicted trend and period terms are summed to the total displacement prediction， and the error analysis is performed between the predicted and observed values. The absolute errors were 10 mm （LSTM） and 24 mm （RNN）. The correlation coefficients R² are 0.9715 （LSTM） and 0.6675 （RNN）. The results show that LSTM performs better in the face of long-time sequences. The prediction results can provide theoretical reference for disaster prevention and mitigation of the Bazimen landslide.