Remote sensing estimation and analysis of Net Primary Production (NPP) have important roles for global ecosystem conservation. The dominant information of ground object landscape in remote sensing images of different scales is different. The multi-scale analysis of modern ecological research is crucial. This study area is Haibei Tibetan Autonomous Prefecture in Qinghai Province. We used Landsat 8 OLI remote sensing image, Wide-band Imaging Spectrometer (WIS) of Tiangong-2, fusion image (Landsat 8 OLI images and WIS of Tiangong-2 fusion), and MODIS image as input parameters of CASA model. We aim to explore the spatial distribution of NPP at different scales in study area, and to comparatively analyze different source data in estimating the accuracy of the NPP. The results showed that:(1) the NPP value using Landsat 8 OLI data was 150-200 g C m^-2 a^-1 and the proportion was the highest. The NPP value of the WIS of Tiangong-2 data and the fusion image were 50-100 g C m^-2 a^-1, which accounted for the highest proportion. The NPP value of MODIS data's proportion of less than 50 g C m^-2 a^-1 was the highest. (2) The root mean square error (RMSE), MAE, and MAPE of the WIS of Tiangong-2 data were the smallest. The correlation between the WIS of Tiangong-2 image's grassland NPP value inversed by Landsat 8 OLI image was the highest. It indicated that the accuracy of the WIS of Tiangong-2 data for describing grassland NPP was higher than that of fusion image and MODIS image. The MODIS image has a large error value and the lowest accuracy in describing the grassland NPP in this region.