Spatiotemporal deep learning (DL) has emerged as a promising paradigm for hydrological simulation compared with lumped models using basin-averaged inputs. However, existing research primarily focuses on either Euclidean data, characterized by regular structures such as grid-like meteorological forcing, or non-Euclidean data, which features irregular topological connectivity such as downstream information transfer. Here, a novel spatiotemporal DL-based approach is developed to integrate both types of spatial information for streamflow modeling. Our proposed model utilizes convolutional neural network (CNN)- and graph neural network (GNN)-based modules to extract the spatial features of Euclidean and non-Euclidean data, respectively. Spatial-attention mechanisms are incorporated into these modules to automatically focus on important features. Long Short-Term Memory (LSTM) networks are then employed to capture temporal dependencies. The adaptability and response of the spatiotemporal DL-based model to different Euclidean and non-Euclidean inputs are investigated. The proposed model is applied to simulate multi-step and multi-gauge streamflow in the mid-lower Hanjiang River basin (ML-HRB), one of the largest sub-basins of the Yangtze River. Results show that when integrating Euclidean and non-Euclidean spatial data, the proposed model outperforms comparative models temporally, spatially and spatiotemporally. The average model improvements across all steps and gauges are of 17.75 % for the Nash-Sutcliffe Efficiency coefficient (NSE), 9.02 % for the correlation coefficient (CC), and 14.79 % for the Kling–Gupta efficiency metric (KGE), compared with the commonly-used lumped LSTM model. The simulations of the tributary stream gauges are particularly sensitive to model refinements and spatial input selection. This study sheds light on the potential of integrating different kinds of spatial information for spatiotemporal DL-based simulation.
