Abstract: Objective: Reservoir-bank landslide, such as Xinpu landslide in Three Gorges Reservoir, is strongly influenced by external hydrological factors such as rainfall and water level, and they often exhibit visible periodic change. While traditional prediction approaches struggle to capture abrupt turning points in these periodic components, this study proposes a novel method for landslide periodic deformation velocity prediction combined with hydrological feature optimization, incorporating rainfall and reservoir water-level information. Methods: Meteorological data including temperature and wind speed were comprehensively utilized to refine the API estimation method, and the reservoir water level window variation feature data were constructed to enhance the interpretability of hydrological factors driving reservoir bank landslide deformation. Subsequently, a convolutional neural network long short-term memory model (CNN-LSTM) landslide periodic velocity prediction network was developed, incorporating both types of hydrological features as input. Results: Experimental results demonstrate a significant improvement in the correlation between optimized hydrological features and periodic deformation velocity. Specifically, the correlation coefficient between optimized API and landslide periodic velocity increased from 0.54 to 0.85. Furthermore, the RMSE for predicting the deformation rate of the periodic component of landslides decreased to 0.54 mm/d, representing a 47.06% improvement over traditional method. The R² value increased from 0.69 to 0.94. Conclusions: The results indicate that the proposed method can effectively improve the accuracy of landslide periodic velocity prediction by enhancing the interpretability of external hydrological factors. The related methods and results can provide reference material for early warning of reservoir-bank landslides.