In this study, we optimized the traditional Carnegie-Ames-Stanford Approach (CASA) model based on China Ecosystem Research Network (CERN) datasets, and compared the estimation accuracy of the two-leaf model and the optimized CASA model at the site scale and pixel scale at eight CERN sites covering major ecosystem types. The optimized CASA model with better performance at the pixel scale combined with China Land Cover Data (ChinaCover) were employed for mapping and monitoring the spatio-temporal changes of terrestrial vegetation net primary production (NPP) in China from 2000 to 2019. The results show that:(1) the optimization for model input parameter of photosynthetically active radiation based on FY2D PAR can effectively avoid the uncertainty caused by the spatial interpolation, and significantly improve the accuracy of PAR estimation; (2) The two-leaf model shows higher NPP estimation accuracy at the site scale, while the optimized CASA model performs better for the NPP estimation at the pixel scale; (3) At the national scale, the CASA model with optimized maximum light energy use efficiency, water stress coefficient and photosynthetically active radiation can better simulate China's terrestrial vegetation NPP. The estimated total NPP in Chinese terrestrial vegetation ranges from 2.703 PgC/a to 2.882 PgC/a in the past 20 years and indicates a fluctuated and slow increasing trend. The spatial distribution of the NPP in China shows a general pattern of gradually increasing from northwest to southeast.