Global Navigation Satellite System (GNSS) is one of the popular approaches for zenith tropospheric delay (ZTD) retrieval due to its advantages of high precision and high temporal resolution. However, obtaining ZTD maps with high spatial resolution and high precision from sparse GNSS stations is greatly challenging, and related studies for areas with large height difference areas are never investigated. Here we present a real-time high-precision ZTD (RHZ) model for generating high-resolution regional ZTD maps from sparse GNSS stations. The model is divided into three modules: real-time calibration, co-estimation, and SCHA-based fitting. The real-time calibration module is mainly responsible for enriching the spatial information of the modeling data through the corrected GTrop model. The output virtual ZTD obtained from real-time calibration module and the GNSS-derived ZTD are then used for modeling by spherical cap harmonic analysis (SCHA, i.e., SCHA-based fitting module) after adaptively determining the weighting, which is performed in the co-estimation module. The North America region (30°N–49°N, 125°W–101°W) with complex terrain and dense GNSS stations was selected as the experiment area, and only 229 of 1833 GNSS stations were selected to simulate sparse station conditions (about 140 km station spacing). The numerical results show that the ZTD series derived from the RHZ model is consistent with that from GNSS at different elevation groups, with an average root mean square (RMS) and Bias of 11.49 and 0.82 mm, respectively. In addition, the ZTD maps derived from RHZ model have great spatial performance, and the comparison results with ERA5 show the average RMS and Bias of 14.98 and − 9.94 mm, respectively.