remotesensing-11-00517-v2.pdf

remote sensing ArticleA Global,0.05-Degree Product of Solar-InducedChlorophyll Fluorescence Derived from OCO-2,MODIS,and Reanalysis DataXing Li and Jingfeng Xiao*Earth Systems Research Center,Institute for the Study of Earth,Oceans,and Space,University of NewHampshire,Durham,NH 03824,USA;*Correspondence:j.xiaounh.edu;Tel.:[phone]Received:20 January 2019;Accepted:27 February 2019;Published:4 March 2019?Abstract:Solar-induced chlorophyll fluorescence(SIF)brings major advancements in measuringterrestrial photosynthesis.Several recent studies have evaluated the potential of SIF retrievalsfrom the Orbiting Carbon Observatory-2(OCO-2)in estimating gross primary productivity(GPP)based on GPP data from eddy covariance(EC)flux towers.However,the spatially and temporallysparse nature of OCO-2 data makes it challenging to use these data for many applications fromthe ecosystem to the global scale.Here,we developed a new global OCO-2 SIF data set(GOSIF)with high spatial and temporal resolutions(i.e.,0.05,8-day)over the period 20002017 based ona data-driven approach.The predictive SIF model was developed based on discrete OCO-2 SIFsoundings,remote sensing data from the Moderate Resolution Imaging Spectroradiometer(MODIS),and meteorological reanalysis data.Our model performed well in estimating SIF(R2=0.79,rootmean squared error(RMSE)=0.07 W m2m1sr1).The model was then used to estimate SIF foreach 0.050.05grid cell and each 8-day interval for the study period.The resulting GOSIF producthas reasonable seasonal cycles,and captures the similar seasonality as both the coarse-resolutionOCO-2 SIF(1),directly aggregated from the discrete OCO-2 soundings,and tower-based GPP.OurSIF estimates are highly correlated with GPP from 91 EC flux sites(R2=0.73,p 0.5 W m2m1sr1)than those based on smaller sample sizes.Since the majority of theRemote Sens.2019,11,5177 of 24test set had relatively lower SIF values(Figure 2b:89.1%was 0.3 W m2m1sr1and 78.8%was0.2 W m2m1sr1),the models based on more samples had lower accumulated errors(Figure 2c).The model based on half of the training samples had nearly the same performance as that based on allthe training samples(two lines are almost completely coincident in Figure 2c),and therefore we usedhalf of the training samples to develop the model for both low computation demand and high accuracy.Table1.The statistical measures for model development and validation.All the models were developedwith training samples in 2015 and 2016 and validated with testing data in 2017.MAE,RE,and Rrepresent mean absolute error,relative error,and correlation coefficient,respectively.R2and RMSErepresent the coefficient of determination and root mean squared error,respectively.MAE,RE,and Rare used for fitting,while R2and RMSE are used for validation.With Land Cover TypeFittingValidationSizeMAERERR2RMSE10,0000.090.340.930.770.08100,0000.080.390.910.800.07250,0000.080.420.890.810.07500,0000.070.420.890.800.07Half0.050.450.890.800.07All0.050.440.890.800.07Without Land Cover TypeFittingValidationSizeMAERERR2RMSE10,0000.090.350.920.760.09100,0000.080.400.900.790.07250,0000.080.420.890.790.07500,0000.070.420.890.790.07Half0.050.450.890.790.07All0.050.450.890.790.07Remote Sens.2019,11,x FOR PEER REVIEW 7 of 23 Table 1.The statistical measures for model development and validation.All the models were developed with training samples in 2015 and 2016 and validated with testing data in 2017.MAE,RE,and R represent mean absolute error,relative error,and correlation coefficient,respectively.R2 and RMSE represent the coefficient of determination and root mean squared error,respectively.MAE,RE,and R are used for fitting,while R2 and RMSE are used for validation.With Land Cover Type Fitting Validation Size MAE RE R R2 RMSE 10,000 0.09 0.34 0.93 0.77 0.08 100,000 0.08 0.39 0.91 0.80 0.07 250,000 0.08 0.42 0.89 0.81 0.07 500,000 0.07 0.42 0.89 0.80 0.07 Half 0.05 0.45 0.89 0.80 0.07 All 0.05 0.44 0.89 0.80 0.07 Without Land Cover Type Fitting Validation Size MAE RE R R2 RMSE 10,000 0.09 0.35 0.92 0.76 0.09 100,000 0.08 0.40 0.90 0.79 0.07 250,000 0.08 0.42 0.89 0.79 0.07 500,000 0.07 0.42 0.89 0.79 0.07 Half 0.05 0.45 0.89 0.79 0.07 All 0.05 0.45 0.89 0.79 0.07 Figure 2.The RMSE between observed SIF and predicted SIF in each 0.05 SIF interval by different models(a).The percentage of each 0.05 SIF interval relative to all the observed SIF in the test set(b).(c)showed that the accumulated errors of predicted SIF by different models decreased with increasing sample size.Table 1 also shows that the models without land cover type had almost identical statistical measures and thus comparable performance to those with land cover type.This difference was better visualized in Figure 3(only the results for 250,000 training samples and half of training samples are provided here).The models with land cover type had slightly lower accumulated errors than those without land cover type when observed SIF was over 0.3 W m2 m1 sr1.This difference was much smaller than that resulting from different training sample sizes(also see Table 1),suggesting that the land cover type only played a very minor role in predicting the SIF,with other explanatory variables already included in the model.In addition,global land cover maps typically have significant classification uncertainty.Therefore,we selected the model without land cover type consisting of only four explanatory variablesEVI,PAR,VPD,and air temperatureto predict the SIF.Figure 2.The RMSE between observed SIF and predicted SIF in each 0.05SIF interval by differentmodels(a).The percentage of each 0.05SIF interval relative to all the observed SIF in the test set(b).(c)showed that the accumulated errors of predicted SIF by different models decreased with increasingsample size.Remote Sens.2019,11,5178 of 24Table 1 also shows that the models without land cover type had almost identical statisticalmeasures and thus comparable performance to those with land cover type.This difference was bettervisualized in Figure 3(only the results for 250,000 training samples and half of training samples areprovided here).The models with land cover type had slightly lower accumulated errors than thosewithout land cover type when observed SIF was over 0.3 W m2m1sr1.This difference wasmuch smaller than that resulting from different training sample sizes(also see Table 1),suggestingthat the land cover type only played a very minor role in predicting the SIF,with other explanatoryvariables already included in the model.In addition,global land cover maps typically have significantclassification uncertainty.Therefore,we selected the model without land cover type consisting of onlyfour explanatory variablesEVI,PAR,VPD,and air temperatureto predict the SIF.Remote Sens.2019,11,x FOR PEER REVIEW 8 of 23 Figure 3.The accumulated errors of predicted SIF by different models:(a)Using 250,000 training samples without land cover type,(b)using 250,000 training samples with land cover type,(c)using half of the training samples without land cover type,and(d)using half of the training samples with land cover type.3.2.Model Validation The scatterplot between observed OCO-2 SIF and predicted SIF by the selected model(i.e.,the model based on half of the training samples without land cover type)was shown in Figure 4.The plot density figure showed that our model estimated SIF fairly well(R2=0.79,RMSE=0.07 W m2 m1 sr1).The model slightly under-and overestimated values greater than 0.6 W m2 m1 sr1 and lower than 0.3 W m2 m1 sr1,respectively.We further validated the model for each biome,and the results of the model with land cover type were also provided for comparison purposes(Table 2).Our model had a high predictive performance for deciduous needleleaf forests,deciduous broadleaf forests,mixed forests,woody savannas,savannas,and croplands(all with R2 greater than 0.75),and a moderate performance for evergreen broadleaf forests(R2=0.43,RMSE=0.08 W m2 m1 sr1)and open shrublands(R2=0.46,RMSE=0.06 W m2 m1 sr1).The model with land cover type had almost identical performance with our selected model in each biome(Table 2).Figure 3.The accumulated errors of predicted SIF by different models:(a)Using 250,000 trainingsamples without land cover type,(b)using 250,000 training samples with land cover type,(c)usinghalf of the training samples without land cover type,and(d)using half of the training samples withland cover type.3.2.Model ValidationThe scatterplot between observed OCO-2 SIF and predicted SIF by the selected model(i.e.,themodel based on half of the training samples without land cover type)was shown in Figure 4.The plotdensity figure showed that our model estimated SIF fairly well(R2=0.79,RMSE=0.07 W m2m1sr1).The model slightly under-and overestimated values greater than 0.6 W m2m1sr1andlower than 0.3 W m2m1sr1,respectively.We further validated the model for each biome,andthe results of the model with land cover type were also provided for comparison purposes(Table 2).Our model had a high predictive performance for deciduous needleleaf forests,deciduous broadleafforests,mixed forests,woody savannas,savannas,and croplands(all with R2greater than 0.75),and amoderate performance for evergreen broadleaf forests(R2=0.43,RMSE=0.08 W m2m1sr1)andopen shrublands(R2=0.46,RMSE=0.06 W m2m1sr1).The model with land cover type hadalmost identical performance with our selected model in each biome(Table 2).Remote Sens.2019,11,5179 of 24Remote Sens.2019,11,x FOR PEER REVIEW 9 of 23 Figure 4.The validation of the predictive SIF model:Scatterplot of observed OCO-2 SIF versus predicted SIF in 2017.The dashed line is the 1:1 line,and the solid line is the regression line.Table 2.Validation of the predictive SIF model for each biome:Evergreen needleleaf forests(ENF),evergreen broadleaf forests(EBF),deciduous needleleaf forests(DNF),deciduous broadleaf forests(DBF),mixed forests(MF),closed shrublands(CSH),open shrublands(OSH),woody savannas(WSA),savannas(SAV),grasslands(GRA),croplands(CRO),and wetlands(WET).The units of the RMSE are W m2 m1 sr1.Biome Without Land Cover Type With Land Cover Type R2 RMSE R2 RMSE ENF 0.66 0.07 0.67 0.07 EBF 0.43 0.08 0.45 0.08 DNF 0.82 0.07 0.83 0.07 DBF 0.87 0.08 0.88 0.08 MF 0.81 0.08 0.82 0.08 CSH 0.62 0.05 0.63 0.05 OSH 0.46 0.06 0.46 0.06 WSA 0.79 0.07 0.79 0.07 SAV 0.75 0.07 0.75 0.07 GRA 0.69 0.06 0.69 0.06 WET 0.54 0.07 0.56 0.07 CRO 0.83 0.08 0.84 0.07 All 0.79 0.07 0.80 0.07 3.3.Global SIF Product:Spatial Patterns and Seasonal Cycles of SIF Figure 5 shows that our predictive model predicted SIF well,with the absolute estimation error within 0.05 W m2 m1 sr1 for the majority of the grid cells.The distribution of the estimation errors was close to the normal distribution.Figure 4.The validation of the predictive SIF model:Scatterplot of observed OCO-2 SIF versuspredicted SIF in 2017.The dashed line is the 1:1 line,and the solid line is the regression line.Table 2.Validation of the predictive SIF model for each biome:Evergreen needleleaf forests(ENF),evergreen broadleaf forests(EBF),deciduous needleleaf forests(DNF),deciduous broadleaf forests(DBF),mixed forests(MF),closed shrublands(CSH),open shrublands(OSH),woody savannas(WSA),savannas(SAV),grasslands(GRA),croplands(CRO),and wetlands(WET).The units of the RMSE areW m2m1sr1.BiomeWithout Land Cover TypeWith Land Cover TypeR2RMSER2RMSEENF0.660.070.670.07EBF0.430.080.450.08DNF0.820.070.830.07DBF0.870.080.880.08MF0.810.080.820.08CSH0.620.050.630.05OSH0.460.060.460.06WSA0.790.070.790.07SAV0.750.070.750.07GRA0.690.060.690.06WET0.540.070.560.07CRO0.830.080.840.07All0.790.070.800.073.3.Global SIF Product:Spatial Patterns and Seasonal Cycles of SIFFigure 5 shows that our predictive model predicted SIF well,with the absolute estimation errorwithin0.05 W m2m1sr1for the majority of the grid cells.The distribution of the estimationerrors was close to the normal distribution.Remote Sens.2019,11,51710 of 24Remote Sens.2019,11,x FOR PEER REVIEW 10 of 23 Figure 5.The probability distribution of errors(estimated SIF minus the observed SIF)in 2016.Compared with the coarse-resolution SIF maps directly aggregated from discrete OCO-2 SIF soundings(Figure 1),GOSIF provides spatially continuous SIF estimates with much finer spatial resolution and 8-day intervals for a much longer period(20002017)(Figure 6).GOSIF exhibited expected spatial and temporal variations in SIF across the globe(Figure 6).In the northern hemisphere,SIF had relatively strong seasonal cycles.In the spring(March,April,and May),SIF exhibited small values at the start of the growing season but then gradually increased over time.The summer(June,July,and August)exhibited the highest values,owing to favorable temperature and soil moisture conditions and long day length.In the autumn(September,October,and November),SIF showed intermediate values first and then approached zero at the end of the season because of the gradual senescence of vegetation.The majority of the areas in mid-and high latitudes showed the lowest values in the winter(December,January,and February)as the canopies of most ecosystems were dormant.The prominently high SIF values were found in the U.S.Corn Belt region in July and August,with SIF ranging from 0.6 to 1 W m2 m1 sr1.High SIF values were also observed in other agricultural regions,such as northeastern China and central Europe.In the southern hemisphere,SIF exhibited smaller seasonal fluctuations.The pan-tropics areas showed high SIF values for most of the year and had slightly lower values in the dry seasons(austral winter:June,July,and August).Low SIF values were found in a large part of Australia throughout the year.Figure 5.The probability distribution of errors(estimated SIF minus the observed SIF)in 2016.Compared with the coarse-resolution SIF maps directly aggregated from discrete OCO-2 SIFsoundings(Figure 1),GOSIF provides spatially continuous SIF estimates with much finer spatialresolution and 8-day intervals for a much longer period(20002017)(Figure 6).GOSIF exhibitedexpected spatial and temporal variations in SIF across the globe(Figure 6).In the northern hemisphere,SIF had relatively strong seasonal cycles.In the spring(March,April,and May),SIF exhibited smallvalues at the start of the growing season but then gradually increased over time.The summer(June,July,and August)exhibited the highest values,owing to favorable temperature and soil moistureconditions and long day length.In the autumn(September,October,and November),SIF showedintermediate values first and then approached zero at the end of the season because of the gradualsenescence of vegetation.The majority of the areas in mid-and high latitudes showed the lowest valuesin the winter(December,January,and February)as the canopies of most ecosystems were dormant.The prominently high SIF values were found in the U.S.Corn Belt region in July and August,withSIF ranging from 0.6 to 1 W m2m1sr1.High SIF values were also observed in other agriculturalregions,such