Accurate displacement prediction plays an important role in landslide early warning. However, the majority of the existing data-driven models focus on single-point modeling based on time series data which cannot consider the deformation correlation in the whole slope. To overcome this drawback, this study proposed a spatial-temporal attention (STA) mechanism-based deep learning model which combined the convolutional neural network (CNN) with the long short-term memory (LSTM) neural network. A convolutional block attention module (CBAM) combined with CNN was developed to extract the spatial deformation characteristics of the slope. A temporal attention module and LSTM model were used to learn the significant historical information from the input external conditions time series data. The model also allowed to output the tempo-spatial attention weights to reveal the tempo-spatial characteristics of landslide deformation. The Paotongwan landslide with step-like behavior displacement in the Three Gorges Reservoir Area (TGRA) of China was used to validate the model performance. The results show that, the root mean square error (RMSE) and the mean average percentage error (MAPE) of the STA-CNN-LSTM model decreased 9.28% and 13.88%, respectively, compared with grey wolf optimization optimized support vector machine (GWO-SVM). The attention weight results calculated by STA-CNN-LSTM demonstrate that rainfall had a larger impact on the deformation of the Paotongwan landslide at the beginning of the monitoring while the influence of reservoir water level decreased with ongoing of the monitoring.