<p>Landslides occurring on the banks of reservoirs pose significant threats to the safety of hydropower stations, nearby infrastructure, and human lives. It is challenging to accurately predict the displacement of landslides due to the complex geological conditions and the coupling effects of rainfall and reservoir level fluctuations. This study proposes a deep-learning-based model for spatiotemporal prediction of landslide displacement by introducing the spatiotemporal heterogeneity of landslide response to rainfall and reservoir level fluctuations. Utilizing InSAR time-series data and the maximum information coefficient, we reveal and quantify the spatiotemporally heterogeneous responses of the Cheyiping landslide to triggering factors. A data fusion unit is designed to integrate the response characteristics of the landslide into the spatiotemporal prediction framework. The spatiotemporal heterogeneity analysis indicates that the tension cracks caused by reservoir water level fluctuations are responsible for larger and faster displacements in the lower and middle parts of the landslide. We also observe a previously overlooked area with significant response and suggest increased attention should be given during the period of reservoir water level variations. Furthermore, the proposed model outperforms other models in predicting the entire displacement field of the landslide and remains robust under different geological conditions. This study elucidates the spatiotemporal patterns of landslide response, offering a predictive framework that contributes to the precise localization and prevention of landslide hazards.</p>