With the rapid pace of urbanization, traffic congestion has become a critical global problem, especially for major cities. Timely planning and management of traffic flows can help avoid traffic jams, necessitating advanced solutions for efficient traffic prediction. Although many existing machine learning-based methods have achieved excellent performance in traffic prediction, they often fail to fully consider the temporal characteristics of traffic flow, as well as the significance of real-world spatial relationships. In this paper, we propose an improved Meta-Graph Learning model for traffic prediction, named the Geospatial-Enabled Meta-Graph Convolutional Recurrent Network (GEM-GCRN). The GEM-GCRN model introduces a new fusion module that dynamically fuses temporal and spatial information by integrating the metagraph and similarity matrix through trainable coefficients. Experimental results on the METR-LA and PEMS-BAY datasets demonstrate the superior predictive accuracy of the proposed GEM-GCRN model, particularly in long-term forecasting scenarios.