Traditional noise prediction models, reliant on on-site monitoring, are hindered by data and computational constraints. This research addresses this challenge by introducing Generative Adversarial Networks (GAN) in conjunction with satellite maps. Based on the inherent interconnectedness between traffic noise and urban morphology elements, the research proposes a GAN model-based framework capable of generating noise heat maps from high-resolution satellite maps, offering a cost-effective and efficient alternative. This research also examines how model performance is influenced by input images through qualitative and quantitative methods. Using New York City as a case study, the proposed GAN-based models demonstrate accuracy in predicting noise distributions. Three parameters of input images likely to be influential in noise prediction accuracy were proposed. We also compare the model performance in different urban contexts. The study presents a valuable tool for architects and urban planners, enabling optimized urban planning and design strategies.