Assessing Uncertainty Of Different GPP Models In Drought Conditions

Understanding the interactions of Gross primary productivity?GPP?and climate is important to terrestrial carbon cycle and biogeochemical process of terrestrial ecosystems,especially under extreme weather conditions such as droughts.The eco-physiological response of grassland to temporal variations in GPP under drought conditions is still unclear.In this paper,we have adopted high time and high spatial resolution of the VPM model,BESS model and PSN model calculated the seasonal change characteristics of grassland ecosystem GPP.With global FLUXNET2015 flux observation research network to provide global 15 grass site of CO₂ flux observation data as the foundation,Select the observation data from 2000 to 2014 for more than five years grassland site.According to the climate characteristics of the grassland site,it is divided into tropical grassland,temperate grassland and alpine meadow.Firstly,we used VPM model,BESS model and PSN model to estimate the gross primary production of global grassland,then we estimated the gross primary production of different grasses,the temperate meadows,the tropical grasses and the grasses,and analyzed the three models at different ecological biomes.Further,we analyzed the model performance capability of VPM model,BESS model and PSN model in extreme climatic conditions.We compared dynamics of eddy convariance?EC?tower-based GPP(GPP_EC)from FLUXNET2015 dataset with estimated GPP of models?VPM,BESS and PSN?over the 2000-2014 years under different drought conditions?i.e.,drought and non-drought?as well as in different grassland biomes?tropical,temperate,and alpine?.We analyze the uncertainty of VPM model,BESS model and PSN model in drought and non-drought condition.Parameter sensitivity analysis of VPM model,BESS model and PSN model was conducted,and the correlation of vegetation index?NDVI?,enhanced vegetation index?EVI?and land surface water index?LSWI?)between flux tower were used to observe GPP were analyzed,and found out the influence factors of simulated performance differences between models in drought and non-drought conditions.We found that:1)across all grassland ecosystem,GPP_VPM,GPP_BESS and GPP_PSN underestimated GPP with root mean squared errors?RMSE?of 2.88 gCm^-2day^-1?VPM?,3.09 gCm^-2day^-1?BESS?and 3.18 gCm^-2day^-1?PSN?;2)The predicted GPP from these three models had better performance with GPP in drought conditions with RMSE of 2.94 gCm^-2day^-1?VPM?,3.18 gCm^-2day^-1?BESS?and 3.20 gCm^-2day^-1?PSN?than in non-drought conditions with RMSE of 2.82 gCm^-2day^-1?VPM?,3.03 gCm^-2day^-1?BESS?and 3.13 gCm^-2day^-1?PSN?,even the later have a smaller RMSE;3)in each of the three grassland biomes in drought conditions,the VPM model performed best,followed by the PSN model and BESS model,with the coefficient of determination?R²?of 0.59-0.74,0.36-0.52and 0.01-0.65 respectively;at the tropical grassland,the VPM model underestimated GPP by 2%,the BESS model and PSN model underestimated GPP by 21%and 51%respectively.This difference in model performance in drought conditions is attributed to the fact that the VPM model algorithm uses Land Surface Water Index?LSWI?to calculate the effect of water stress on LUE and GPP and the LSWI is more sensitive to capture the drought,while the other two models do not have such a sensitive mechanism.And the underestimation of GPP models?VPM and PSN model?could be partly attributed to the fact that these products did not consider the variation in maximum light use efficiency(^?0)values of different grassland biomes.This study suggests that it is essential for GPP models to incorporate biome-specific^?0and the effect of water stress into the GPP algorithm,in addition to using EVI to estimate FPAR,if the models are used to estimate seasonal variation of GPP in drought conditions.