The complementary nature of multi-source urban data (including remote sensing images, mobile phone data, and other spatial datasets) underscores the critical importance of data fusion in spatial attribute mapping. However, exponentially growing data coupled with decentralized data management has rendered traditional centralized data analytics and learning models increasingly inefficient at scale. To address these challenges, we propose an alternative urban attribute mapping framework based on federated learning which resolves data silo issues in multi-source collaborative modeling. Unlike traditional direct data sharing, the presented framework shares encrypted handcrafted features derived from multi-source data between dataholders. Therefore, it preserves the privacy of original urban data. With popular machine learning methods, the framework enables accurate inference of diverse urban attributes such as population, economic development, urban mobility, land use, and air quality. Experimental results from a case study in Shenzhen demonstrate that the presented framework successfully facilitates multi-source collaborative mapping solely through the exchange of model parameters and structures. The framework exhibits exceptional performance across five urban attribute mapping tasks: non-resident population, GDP, taxi travels, land use mix, and PM2.5 concentration. These results validate the effectiveness of multi-source data collaboration in data-rich environments. The primary contribution of this research lies in the development of a distributed multi-source federated mapping framework for urban attributes, offering an alternative solution to overcome urban data silos while establishing practical foundations for expanding urban mapping and cross-regional applications.