Multimodal traffic prediction is a critical endeavor within smart traffic management. An accurate traffic prediction can guide vehicle dispatching, improve vehicle utilization, and alleviate traffic congestion. However, this task is challenging due to the heterogeneous nature of data across different inter-modes and the non-stationary nature of short-term traffic flow data. Existing researches primarily focus on leveraging single-modal data and incorporating external factors like weather conditions for prediction. While they overlook the correlation among different inter-models hampers prediction accuracy. In this paper, we propose the Spatiotemporal Multimodal Graph Attention Temporal Convolutional Network (SMGATCN), a deep learning model for multi-modal traffic prediction. Firstly, we devise a cross-modal feature fusion mechanism to integrate and learn dynamic cross-modal associations. We utilize a dual-channel pooling layer to enhance the model’s generalization capabilities. Secondly, we propose a Graph Attention Temporal Convolution Network (GATCN), which utilizes spatiotemporal correlation features between the inter- and intra-model to alleviate traffic flow instability. Finally, extensive experiments on two real-world datasets show that SMGATCN outperforms existing methods across various metrics, highlighting its efficacy in multimodal traffic prediction scenarios.