Abstract: Deep learning has been proven to show remarkable performance in hydrological modeling; however, the spatial features of meteorological data are rarely incorporated in current deep learning hydrological models. In this study, we propose a spatiotemporal DL-based hydrological model by coupling principal component analysis (PCA) and long short-term memory (LSTM). PCA and LSTM were used to capture the spatial characteristics of meteorological data and understand long-length temporal dynamics, respectively. We used the source region of the Yellow River to test the PCA-LSTM model and compared the results with those of LSTM-only and THREW models. The Gaussian noise method was also used to evaluate the robustness of the PCA-LSTM model. The proposed PCA-LSTM model showed better performance than THREW and LSTM models, with Nash-Sutcliffe efficiency coefficients of 0.92, underlining the potential of the PCA-LSTM model for hydrological modeling and prediction.