Global products of remote sensing Normalized Difference Vegetation Index (NDVI) are critical to assessing the vegetation dynamic and its impacts and feedbacks on climate change from local to global scales. The previous versions of the Global Inventory Modelling and Mapping Studies (GIMMS) NDVI product derived from the Advanced Very High Resolution Radiometer (AVHRR) provide global biweekly NDVI data starting from the 1980s, being a reliable long-term NDVI time series that has been widely applied in Earth and environmental sciences. However, the GIMMS NDVI products have several limitations (e.g., orbital drift and sensor degradation) and cannot provide continuous data for the future. In this study, we presented a machine learning model that employed massive high-quality and global-wide Landsat NDVI samples and a data consolidation method to generate a new version of the GIMMS NDVI product, i.e., PKU GIMMS NDVI (1982−2020), based on AVHRR and Moderate-Resolution Imaging Spectroradiometer (MODIS) data. A total of 4.9 million Landsat NDVI samples that well spread across the globe were extracted for vegetation biomes in all seasons. The PKU GIMMS NDVI exhibited higher accuracy than its predecessor (GIMMS NDVI3g) in terms of R2 (0.975 over 0.942), mean absolute error (0.033 over 0.074), and mean absolute percentage error (9% over 20%). Notably, PKU GIMMS NDVI effectively eliminates the evident orbital drift and sensor degradation effects in tropical areas. The consolidated PKU GIMMS NDVI has a high temporal consistency with MODIS NDVI in describing vegetation trends. The PKU GIMMS NDVI product can potentially provide a more solid data basis for global change studies. The theoretical framework that employs Landsat data samples can facilitate the generation of remote sensing products for other land surface parameters.