Abstract: The data production of MOD17A3 NPP in northwest China is seriously deficient, with consequences on further research on vegetation net primary productivity (NPP) in this region. Based on the meteorological data, DEM, NDVI, and MOD17A3 NPP with high quality, the BP neural network model was constructed to simulate vegetation NPP and fill the areas without NPP data in northwest China from 2000 to 2014. Its spatio-temporal variations and relationships with meteorological factors were analyzed using unitary linear regression, R/S analysis, and partial correlation analysis. The results showed that: (1) Coefficients of determination (R²), mean absolute error (MAE), mean relative error (MRE) and root mean square error (RMSE) between MODIS NPP and simulated NPP were 0.833-0.906, 25.84-40.1, 0.16-0.23, 34.57-59.36, respectively, which well met the accuracy requirements. The BP neural network model was suitable for vegetation NPP simulation in northwest China. (2) The annual mean NPP had a strong spatial variation, showing a gradual decline from southeast to northwest and a high-value block area in the northwest of Xinjiang. (3) The annual mean NPP ranged between 106.64 and 156.17 g C·m^-2·a^-1 from 2000 to 2014, with a slightly fluctuating downward trend in the interannual variation. (4) From 2000 to 2014, the change of NPP in northwest China had spatial heterogeneity, which was mainly reduced. Only 10.79% of the areas passed the significance test. NPP change was weakly persistent. The future change trend of NPP is mainly uncertain, supplemented by improved and declined areas, with the improved area being larger than the declined area. (5) Responses of vegetation NPP to temperature and precipitation varied spatially, which were generally more closely related to precipitation.

Key words: IPAT models, human driving force, environmental pressure, ecological elasticity.