Abstract: The time series deformation signal of synthetic aperture radar interferometry (InSAR) usually contains trend
deformation and random fluctuation. As a noise component,the latter is easy to interfere with the extraction of key features by
the model,which affects the accuracy of landslide prediction. Therefore,a deformation decomposition prediction framework
based on time-series InSAR data is proposed. The original deformation signal is decomposed into trend term and random term
by variational mode decomposition (VMD) method. Then,the autoregressive moving average model (ARMA) and the improved
Transformer-LSTM hybrid model are used to predict and synthesize the final deformation prediction results. The landslide
deformation prediction test of a reservoir in Xinjiang shows that the prediction performance of this method is better than
that of traditional models such as long short-term memory(LSTM),and the goodness of fit (R2 )is higher than 0. 95. The root
mean square error (RMSE)and mean absolute error (MAE)of representative points were significantly reduced. The study reveals
that the improved Transformer-LSTM model can effectively capture the characteristics of sudden deformation fluctuations
in random terms. Combined with the advantages of ARMA model in stationary time series modeling,it can effectively improve
the prediction performance of complex landslide deformation sequences,and has important reference value for improving the
risk assessment and prevention and control of reservoir bank landslide disasters.