Urban vehicle emission pollution has become the main factor affecting urban air quality. How to achieve fine-grained emission prediction is of great significance to vehicle emission supervision. Existing works mainly considered the emission prediction as a time sequence forecasting task, and extracted the temporal and spatial features of emission sequence based on road network prior information. However, road emission patterns are usually influenced by diverse environmental factors like weather and traffic conditions. These urban multi-source data have different properties, how to utilize cross-domain factors to assist emission prediction remains to be studied. To this end, a multi-source fusion spatiotemporal network is devised to simultaneously capture external factors’ impact on emission prediction. Specifically, the proposed model adopt an attention adaptive fusion module to achieve multi-source heterogeneous data fusion. Experiment results on Beijing emission dataset have indicated that our method surpass the existing baselines.