Abstract: Tropospheric delays act as one of the main error sources affecting the precision positioning of the GNSS. In recent years, machine learning has been widely used for modeling tropospheric delays. Based on the GNSS tropospheric delay and normalized difference vegetation index (NDVI) in China, a novel NDVI-constrained regional tropospheric delay model (NZTD) is established by considering the topographic factors represented by NDVI. The results show that the root mean square error (RMSE) and the mean bias (Bias) of the NZTD model in China are 7.91 mm and 1.43 mm, showing 70% and 65% decreases compared with that of the GPT3 model, respectively. Meanwhile, the NZTD model possesses stronger anti-interference to seasonal changes than the GPT3 model. In addition, the NZTD model demonstrates significantly improved accuracy performance compared with the regional tropospheric delay model without NDVI constraint, evidenced by the average enhancement of 8%. It shows that the NZTD model can refine the environmental characteristics of different geographical regions and reflect the fine changes of tropospheric delay, with better adaptability and accuracy.