Improving Estimation Of MODIS GPP Product Based On The Global FLUXNET Sites

Terrestrial gross primary productivity(GPP),the amount of carbon and energy absorbed by terrestrial plants by photosynthesis,is the largest carbon flux between the terrestrial biosphere and the atmosphere.It is significant for GPP to the global climate system.GPP also drives the terrestrial food chain and is the foundation for agricultural and wood production.GPP is the beginning of the terrestrial carbon cycle and the estimation of the GPP is directly related to the accuracy of the carbon budget and assessment,which also become the hotspot of the global carbon cycle research.Therefore,accurately quantifying GPP is essential for estimating ecosystem carbon dynamics,climate feedbacks,agricultural productivity,human welfare and global and regional carbon cycle estimation.Light use efficiency(LUE)model is perhaps the most classical model and is widely used in estimating regional and global GPP for few parameters in the model which can be obtained by remote sensing data.The MODIS PSN algorithm from the MODIS GPP product is based on the light use efficiency model.National Aeronautics and Space Administration(NASA)launched the Moderate Resolution Imaging Spectroradiometer(MODIS)and provided global high-frequency monitoring of terrestrial GPP product,MOD17A2H.MODIS-PSN algorithm is at the base of the LUE model.Currently,the MOD17A2H 006 product is the latest global terrestrial ecosystem GPP products,with spatial resolution of 500m and temporal resolution of 8-day.An application of the MODIS GPP product demonstrated that an overestimation of GPP at low productivity sites due to overestimation of FPAR and an underestimation at high productivity sites due to the underestimation of LUE in vegetation lead to the uncertainty of the LUE model.A key factor in the application of MODIS GPP products is the estimation of the parameters and uncertainties of the model.To solve this problem is to optimize the parameters of the model.In recent years,the release of the FLUXNET2015 dataset have been provided the opportunity for parameterization of the MODIS-PSN model at the site-level across the globe.The latest version 6 of the MODIS GPP products were validated at the 111 sites,including 10 biomes across the global FLUXNET2015 and the accuracy of the MODIS GPP varied among the biomes.There are 8 Vegetation Indices(VIs):Normalized Difference Vegetation Index(NDVI),Enhanced Vegetation Index(EVI),2-band Enhanced Vegetation Index(EVI2)and Near Infrared Reflectance of terrestrial Vegetation(NIRV)and above four VIs with Bidirectional Reflectance Distribution Function(BRDF),NDVI_BRDF,EVI_BRDF,EVI2_BRDF and NIR_V,BRDF.The relationships between 8 VIs and tower GPP were comprehensive analyzed at the monthly scale and the annual scale and across the biomes.Meanwhile,the Markov Chain Monte Carlo(MCMC)method is used to optimize the parameters of the MODIS-PSN model,including lower limit of the daily minimum air temperature(Tmin_min),the upper limit of the daily minimum air temperature(Tmin_max),the lower limit value of daytime mean vapor pressure deficit(VPD_min),the upper limit value of daytime mean vapor pressure deficit(VPD_max)and maximum LUE(?_max).We analyzed the constraints of different parameters on MODIS-PSN model across the biomes and evaluated the performance of the MODIS-PSN model after parameterization.Through research and analysis,we get the following conclusions:The accuracy of MODIS GPP varies with the change of vegetation type and site on the global site scale.The relationships of 8 VIs and tower GPP were moderate and strong at the monthly scale but low at the annual scale.The performance of the MODIS-PSN model has been greatly improved after parameterization by the MCMC method.Besides,the accuracy of the MODIS GPP products was further improved due to NDVI_BRDF as a new input taking place of the fraction of the Absorbed Photosynthetic Active Radiation(FPAR).(1)The accuracy of MODIS GPP products varies at sites and cross the biomes at the global scale.The accuracy of MODIS GPP products is highest at forest sites except for evergreen broad-leaved forest(EBF)with R² above 0.68 across biomes.The accuracy of MODIS GPP products is medium(R²>0.55)in wetland,grassland and shrub ecosystem.However,the accuracy of MODIS GPP is poor in savannas and crops(R²<0.5)with high RMSE.In general,MODIS GPP product had a good performance in forest ecosystem.Even so,the accuracy of MODIS GPP varied in 33 forest sites with best performance in US-NR1 site(R²=0.91,RMSE=0.79 RE=0.03 g C/m2/day)but poorest performance in US-Prr site(R²=0.26,RMSE=3.35 RE=1.02 g C/m2/day).Although MODIS GPP product had the worst performance in crop ecosystem,there is some individual site with better performance such as FL-Jok site(R²=0.65,RMSE=2.05RE=0.68 g C/m2/day),indicating that the performance of the MODIS GPP varied at site scale.Compared with the EC tower GPP,most of the MODIS GPP product are systematically underestimated with only a few sites where MODIS GPP can capture the beginning and the peak of the growing season.There are two reasons for underestimation of the MODIS GPP.First,the underestimation of the LUE is the most reason for underestimation of MODIS GPP.The parameters in MODIS GPP algorithm is obtained from the Biome-specified Parameters Look Up Table in different vegetation type.The contains five parameters calibrated by the American Flux Network(Ameri Flux)based on the 11 vegetation types,which cannot represent all vegetation types at the global scale.Second,the FPAR in the MODIS GPP algorithm relied on the product of the MOD15 FPAR,which was easy affected by the noise such as long cloud,foggy and rainy weather,leading the accuracy of the MODIS GPP product.(2)We made the comprehensive global analysis of how the traditional vegetation indices(NDVI,EVI,EVI2,NIR_V),BRDF-corrected vegetation indices(NDVI_BRDF,EVI_BRDF,EVI2_BRDF,NIR_V,BRDF)a total of 111 sites encompassing ten biomes across the globe and at both monthly and annual scales.At the monthly scale,all the seven VIs were moderately or strongly correlated with tower GPP.Compared with the traditional VIs(NDVI,EVI,and EVI2),the BRDF-corrected VIs(NDVI_BRDF,EVI_BRDF,and EVI2_BRDF)generally had stronger correlation with tower GPP,indicating that the VIs based on BRDF-correction had an advantage over the traditional VIs.EVI2_BRDFand NIR_V,BRDFhad the strongest correlation with tower GPP(averaged R²=0.70,p<0.0001),and their performance was comparable to the MODIS GPP product(averaged R²=0.70,p<0.0001)based on a light use model.The VIs explained lower variance in tower GPP at the annual scale than at the monthly scale.BRDF-corrected VIs had no advantage over the traditional counterparts at the annual scale.The capability of the VIs in capturing the interannual variability in GPP was also similar to that of the MODIS GPP product.At the site level,VIs generally exhibited moderate to strong correlation with tower GPP at the annual scale for all biomes except savannas.Similar to the MODIS GPP product,all VIs were strong predictors of GPP at the biome level.Our findings have implications for improving our understanding of the relationships between satellite-derived VIs and tower GPP,the influences of the BRDF effect on the VI-GPP relationship,and the variations of the VI-GPP relationship among indices,biomes and timescales.(3)We use the MCMC sampling technique based on the Bayesian approach to optimize all the five parameters(Tmin_min?Tmin_max?VPD_min?VPD_max and?_max)of the MODIS PSN model at 111 FLUXNET across ten biomes.The parameters of the original MODIS-PSN model are renewed to produce a more comprehensive and representative MODIS-PSN model.Meanwhile,the new parameter lookup table subdivided the grassland into two kinds of parameters,including high latitude cold area and arid desert area and crops are also divided into C₃ and C₄ two types.In the same plant functional types(PFTs),these optimized parameters also vary among different sites.The variation coefficient of Tmin_min is 16.57.The other four parameters,Tmin_max,VPD_min,VPD_max and?_max,are moderately different in different sites,and the variation coefficient is between 0.17 and 0.39.After optimization of the parameters with MCMC technique and using NDVI_BRDF as input factor for FPAR,the accuracy of the MODIS GPP products has been greatly improved.The relationships between the MODIS GPP and tower GPP is much stronger after optimization of parameters at forest sites with R²from 0.68 to 0.78 except for evergreen broad-leaves forest(EBF);at wetlands,grasslands and Shrublands site with R² from 0.55 to 0.6;at corn sites with R² from 0.4to 0.78 but at wood savannas with R² from o.27 to 0.37.Compared with the US-NR1site with the highest accuracy of the MODIS GPP product without parameter optimization,the site US-NR1(R2=0.88,RMSE=8.06,RE=6.15%)has a slightly"over-parameterized"appearance after parameter optimization from the perspective of the site level.However,the overall accuracy is still very high at US-NR1 site.MODIS GPP products at site of US-Prr(R2=0.50,RMSE=17.25,RE=66.79%)with the worst accuracy has greatly improved(R2=0.72,RMSE=11.40,RE=30.42%).We used 111 global FLUXNET sites to fully validate and assess MODIS GPP.We also examined the relationships between VIs and GPP at monthly and annual scale across the biomes,indicating that VIs with BRDF correction can be proxy for GPP at monthly scale.Besides,five parameters(Tmin_max?VPD_min?VPD_max and?_max)were optimized in the MODIS PSN model at the site level and then these parameters were averaged by vegetation types to update BPTL of MODIS-PSN algorithm.With this new BPTL and NDVI_BRDF taking place of FPAR,MODIS PSN has greatly improved the accuracy of the MODIS GPP.Our results will help us better understand the global assessment of MODIS GPP products;the relationship between VIs and flux GPP,the influence of BRDF effect on VI-GPP;the constraints of vegetation functional parameters on energy utilization and environmental factors(temperature and saturated vapor pressure difference)on energy utilization in different vegetation types.