Gross primary productivity (GPP) is an important indicator quantitatively describing the carbon sequestration capacity of ecosystems. To analyze the influences of climatic properties on GPP prediction, the models were built by the 10-year average climatic dataset (MC), and the neighboring climatic dataset (FM) and current year climatic dataset (CC), by Random Forest model for each year, respectively. In the climatic properties, CC, MC, FM represent climatic fluctuation, the time-lag effect and spatial heterogeneity, respectively. The prediction period split into two stage, those was the historical GPP prediction from 2000 to 2014 and climate changed GPP prediction from 2015 to 2021 under the SSP126, SSP245, SSP370 and SSP585 emission scenarios. The model had highest explanatory rate indicated that the parameter in the model most accurately predicted the GPP than other climatic properties, and had high explanatory rate in cross-year predictions (eg. the model build on 2010 was used to predict the other years with the corresponding year parameters) was most stable than others. In addition, a climatic property had most close the input MODIS GPP, which dominate and determine the GPP. The results on above analysis showed that: (1) Slope was the most important variables among all variables, and the precipitation is the most important factors among climatic variables. In all GPP predictions, the GPP was more correlated with topographic variables than climatic variables in China; (2) The explanatory rate (R²) of each model was consistently above 0.8 in all predictions, including cross-year predictions among historical and climatic changed models. This indicated that all models were stable during these periods; (3) The GPP prediction for the current year using the current climate dataset had a higher explanatory rate compared to the other predictions. It also shaped more accurately the relationship between current climate and GPP better than the others; (4) However, the explanatory rates of the cross prediction among historical models and climate change models under four future emission scenarios resulted that MC had the highest explanatory rates than the others, and the difference from the others were statistically significant (P<0.05). The models built based on MC was most stable than the others. This clearly indicated that vegetation GPP has a time-lag effect in response to climate, and its impact is significantly greater than the spatial heterogeneity and fluctuation of climate, in other words, the climate in the first 9 years continues to affect vegetation GPP; (5) According to the statistics of the occurrence probability of the three climatic properties which had the minimum variation from MODIS GPP in the historical period, area with high occurrence probability of the MC was larger than that of the others. This same conclusion also appeared in the three landscape types of forest, shrub and grassland. Moreover, the prediction ability of MC for mid-high altitude and steep slope were more accurate than the other areas because of mountainous regions exhibit a time-lag effect by reallocating climatic factors (eg. solar radiation) more noticeably than flat regions, which are dominated by the current climate and spatial heterogeneity of the current climate. Therefore, the time-lag effect of climate explains the relation between climate and GPP than the others. However, the time-lag in climate was found to be 10 years in this paper. It is possible that the optimal time-lag is actually higher or lower than 10 years, and this aspect needs to be further explored. The conclusions of this paper outline the optimal spatial properties of climate for predicting GPP in the context of climate change.