Comparison of CO 2 from NOAA Carbon Tracker reanalysis model and satellites over Africa

The scarcity of ground-based observations, poor global coverage and resolution of satellite observations necessitate the use of data generated from models to assess spatio-temporal variations of atmosphericCO2 concentrations in a near continuous manner in a global and regional scale. Africa is one of the most data scarce region as satellite observation at the equator is limited by cloud cover and there are very limited number of ground based measurements. As a result, use of simulations from models are mandatory to fill this data gap. However, the first step in the use of data from models requires assessment of model 5 skill in capturing limited existing observations. Even though, the NOAA Carbon Tracker model is evaluated using TCCON and satellite observations at a global level, its performance should be assessed at a regional scale, specifically in a regions like Africa with a highly varying climatic responses and a growing local source. In this study, NOAA CT2016 CO2 is compared with the ACOS GOSAT observation over Africa using five years datasets covering the period from April 2009 to June 2014. In addition, NOAA CT2016 CO2 is compared with OCO-2 observation over Africa using two years data covering the period 10 from January 2015 to December 2016. The results show that the XCO2 retrieved from GOSAT and OCO-2 are lower than CT2016 model simulation by 0.42 and 0.93 ppm on average respectively, which lie within the range of the errors associated with the GOSAT and OCO-2 XCO2 retrievals. The mean correlations of 0.73 and 0.6, a regional precisions of 3.49 and 3.77 ppm, and the relative accuracies of 1.22 and 1.95 ppm were found between the model and the two data sets implying the performance of the model in Africa’s land regions is reasonably good despite shortage of in-situ observations over the region 15 assimilated in the model. These differences, however, exhibit spatial and seasonal scale variations. Moreover, the model shows some weakness in capturing the whole distribution. For example, the probability of detection ranges from 0.6 to 1 and critical success index ranges from 0.4 to 1 over the continent when the analysis includes data above the 95 percentile and the whole data respectively. This shows the model misses the higher extreme ends of the CO2 distribution. Spatially, GOSAT and OCO-2 XCO2 are lower than that of CT2016 by upto 4 ppm over North Africa (10 0−35 N) whereas it exceeds CT2016 XCO2 by 20 3 ppm over Equatorial Africa (10 0S−10 N). Larger spatial mean biases of 2.11 and 1.8 ppm, 1.25 and 0.73 ppm in CT2016 XCO2 with respect to that of GOSAT and OCO-2 are observed during winter (DJF) and spring (MAM) while small biases of -0.15 and 0.21 ppm, and 0.2 and -1.14 ppm are observed during summer (JJA) and autumn (SON) respectively. The model simulation has the ability to capture seasonal cycles with a small discrepancy over the North Africa and during winter seasons over all regions. In these cases, the model overestimates the local emissions and underestimate CO2 loss. 25 1 Atmos. Meas. Tech. Discuss., https://doi.org/10.5194/amt-2018-84 Manuscript under review for journal Atmos. Meas. Tech. Discussion started: 16 May 2018 c © Author(s) 2018. CC BY 4.0 License.


Introduction
An understanding of the regional contributions and trends of carbon dioxide (CO 2 ) is critical to design mitigation strategies aimed at stabilizing atmospheric greenhouse gases.The present-day concentration of atmospheric CO 2 continues to rise.
Ongoing emissions of CO 2 and other greenhouse gases will influence the global climate system during the next decades and centuries.Approximately one-half of the CO 2 emissions from human activities is accumulating in the Earth's atmosphere, whereas the remaining portion of emitted CO 2 is absorbed by sink on land and in the ocean (Raupach et al., 2007).
Several studies (e.g., Tsutsumi et al., 2009;Stocker, 2014;Allison, 2015) have shown that the column averaged dry air mole fractions of CO 2 (XCO 2 ) has undergone rapid changes from pre-industrial period value of 280 ppm to 396 ppm as recently as 2013.This change has been attributed to anthropogenic factors such as fossil fuel combustion, land use change, biomass burning, emission from industries such as cement.Notably, this value exceeds the highest XCO 2 level retrieved from ice cores representing the past 800,000 years.This increasing trend has been persistent.For example according to the World Meteorological Organization (WMO) 2016 report, the average concentrations of CO 2 hit 403.3 ppm, up from 400 ppm in 2015 surpassing an annual positive increasing rate of 1.99 ± 0.43 ppmyr −1 (http://www.esrl.noaa.gov/gmd/ccgg/trends/global.htmlHoughton, 2007).These positive trends have led to imbalance of 0.58 ± 0.15 W m 2 in energy budget between 2005 and 2010 at the top of the atmosphere (Hansen et al., 2011).To this end, changes in atmospheric temperature, hydrology, sea ice, and sea levels are attributed to climate forcing agents dominated by CO 2 (Santer et al., 2013;Stocker et al., 2013).
However, understanding the climate response to anthropogenic forcing in a more traceable manner is still difficult due to a major uncertainty in carbon-climate feedbacks (Friedlingstein et al., 2006).Part of this uncertainty is due to lack of sufficient data on regional and global carbon cycle.This is compounded with inappropriate modelling practices to capture spatio-temporal variability of carbon cycle.These problems can be solved through strengthening carbon monitoring networks and setting up proper modelling.A transport model, with appropriate physical and mathematical formulations and sufficiently tuned by observations, can be used to understand the spatio-temporal nature of atmospheric CO 2 source and sink as well its associated drivers.
Towards this, a number of national and international efforts have been initiated in the recent past by different government and non-government agencies across the globe.Among these efforts, ground-based observations of greenhouse gas using Total Carbon Column Observing Network (TCCON) is a notable one since it provides accurate and high-frequency measurements of column-integrated CO 2 mixing ratio.For example, it has been established that TCCON has a precision of 0.25% for measurements taken under clear sky conditions (Wunch et al., 2011).However, the number of TCCON sites is limited and can not establish accurate CO 2 amount and flux on subcontinental or regional scale.Moreover, some studies shows that the large uncertainty is amplified due to uneven global distribution of TCCON sites (Gurney et al., 2002;Hungershoefer et al., 2010).
In addition, none of these ground based observation networks are found in Africa.
On the other hand, the CO 2 concentration retrieved from the satellite-based CO 2 absorption spectra have the advantages of unified, long-term, and the global coverage observation as compared to ground-based measurements.It has been established from theoretical studies that accurate and precise satellite derived atmospheric CO 2 can appreciably minimize the uncertainties in estimated CO 2 surface flux (Rayner and O'Brien, 2001;Chevallier, 2007).Other studies have revealed that significant improvement in estimation of weekly and monthly CO 2 fluxes can be achieved subject to CO 2 retrieval error of less than 4 ppm from satellite and modelling scheme whereby CO 2 concentration is an independent parameter of carbon cycle model (Houweling et al., 2004;Hungershoefer et al., 2010).However, XCO 2 shows temporal variability on different time scales: diurnal, synoptic, seasonal, inter-annual, and long term (Olsen and Randerson, 2004;Keppel-Aleks et al., 2011).More recent missions such as the Greenhouse gases Observing SATellite (GOSAT) (Hamazaki et al., 2005), the Orbiting Carbon Observatory-2 (OCO-2) (Boesch et al., 2011) and planned missions such as the Active Sensing of CO 2 Emissions over Nights, Days, and Seasons (ASCENDS) (Dobbs et al., 2008) have been and are being developed specifically to resolve surface sources and sinks of CO 2 and provide information on these different scales of temporal variability.For example, GOSAT observations started in 2009 and provide XCO 2 based on spectra in the Short-Wavelength InfraRed (SWIR) region with a standard deviation of about 2 ppm with respect to ground-based and in-situ air-borne observations (Yokota et al., 2009;NIES GOSAT Project, 2012).
Moreover, atmospheric transport model, such as the NOAA Carbon Tracker (CT) is an integrated modelling system that assimilate CO 2 from other observations in order to compliment satellite observations in understanding CO 2 surface sources and sinks as well as its spatio-temporal variabilities.However, both satellite and model data should be validated against other independent satellite observations and/or in-situ observations before using them to answer scientific questions.As a result, a number of validation and intercomparison have been conducted in previous studies.For example, Kulawik et al. (2016) found root mean square deviation of 1.7, and 0.9 ppm in GOSAT and CT2013b XCO 2 relative to TCCON respectively.Other authors have undertaken validation exercises and found bias of −8.85 ± 4.75 ppm in NIES XCO 2 with respect to TCCON (Morino et al., 2010); root mean square deviation of −1.48 and 2.09 ppm in NIES Level 2 V02.xx XCO 2 (Yoshida et al., 2013); and bias of −0.68 ± 2.56 ppm in NIES level 2 V02.xx XCO 2 with respect to aircraft observations (Inoue et al., 2013).
Moreover, strong consistency between the ACOS and NIES XCO 2 monthly averages time series over different regions was reported.For example, Deng et al. (2016a) found the greatest mean difference (1.43 ± 0.60 ppm) over China and the least over Brazil (−0.03 ± 0.64 ppm) in the two time series of monthly means.Globally, ACOS XCO 2 is higher than NIES by about 1 ppm and has smaller bias than NIES data.Moreover, comparison of NIES Level 2 V02.xx XCO 2 with XCO 2 from Goddard Earth Observing System-Chemistry model(GEOS-5 v09-01-01) simulations revealed that the satellite XCO 2 was lower than the model by 2 ppm on average globally (Lei et al., 2014).Liang et al. (2017) compared XCO 2 from OCO-2 and GOSAT with that of TCCON and found mean measurement accuracy of 0.27 and -0.41 ppm with RMSD of 1.56 and 2.22 ppm respectively.However, they found the measurement accuracy of GOSAT decreased to -0.62 ppm with RMSD of 2.3 ppm for period from 2014 to 2016.Liang et al. (2017) also indicates GOSAT shows a larger seasonal variability in describing amplitudes than OCO-2, with greater amplitude in the northern hemisphere than the southern hemisphere.Lei et al. ( 2014) also showed regional difference of XCO 2 between the ACOS and NIES datasets.For example, a larger regional difference from 0.6 to 5.6 ppm was obtained over China land region, while it is from 1.6 to 3.7 ppm over global land region and from 1.4 to 2.7 ppm over US land region.These findings suggest that it is important to assess how satellite and model XCO 2 compare with each other over other regions.
Therefore, this paper aims to assess the performance of Carbon Tracker model in simulating observed XCO 2 from GOSAT and OCO-2 satellites over Africa using various statistical metrics and identify weakness and strengths of the respective data sets.Moreover, the skill of the model in capturing the amplitudes and phases of observed seasonal cycles over different parts of the continent is evaluated and the consistence of the modelled spatio-temporal variability with the known seasonal climatology of the regions that determines carbon source and sink levels is assessed.

Carbon Tracker Model and Data
Carbon Tracker is an annually updated analysis of atmospheric carbon dioxide distributions and their surface fluxes (Peters et al., 2007).It is a data assimilation system that combines observed carbon dioxide concentrations from 81 sites around the world with model predictions of what concentrations would be based on a preliminary set of assumptions ("the first guess") about sources and sinks for carbon dioxide.Carbon Tracker compares the model predictions with reality and then systematically tweaks and evaluates the preliminary assumptions until it finds the combination that best matches the real world data.It has modules for atmospheric transport of carbon dioxide via weather systems, photosynthesis and respiration, air-sea exchange, fossil fuel combustion and fires.Transport of atmospheric CO 2 is simulated by using the global two-way nested transport model (TM5).TM5 is an off line atmospheric tracer transport model (Krol et al., 2005) driven by meteorology from the European Centre for Medium-Range Weather Forecasts (ECM W F ) operational forecast model and from the ERA Interim reanalysis (Dee et al., 2011) to propagate surface emissions.TM5 is based on a global 3 0 × 2 0 and a 1 0 × 1 0 spatial grids over North America.
The data from CT (version:CT2015) (http://carbontracker.noaa.gov;Peters et al. (2007) is used to extend aircraft profiles from the stratosphere to the top of the atmosphere (Inoue et al., 2013;Frankenberg et al., 2016) and to quantify co-location error (Kulawik et al., 2016).The older data versions have been used and also compared with different data sets over other parts of the globe in previous studies (Peters et al., 2007;Nayak et al., 2014;Kulawik et al., 2016;Krishnapriya et al., 2017).Most of the studies confirm that CT XCO 2 captures observations reasonably well.In this study we use Carbon Tracker release version CT2016 , here after (CT2016) and near real time version (CT-NRT.v2017).Both versions of NOAA CT provides 3 hourly CO 2 mole-fractions data for global atmosphere at 25 pressure levels for a period covering 2000 to 2016.The data can be accessed freely at the public domain (ftp://aftp.cmdl.noaa.gov/products/carbontracker).
GOSAT is the world's first spacecraft dedicated solely to measure the concentrations of carbon dioxide and methane, the two major greenhouse gases, from space.The spacecraft was launched successfully on January 23, 2009, and has been operating since then.GOSAT records reflected sunlight using three near-infrared band sensors.The field of view at nadir allows a circular footprint of about 10.5 km diameter (Kuze et al., 2009;Yokota et al., 2009;Crisp et al., 2012).GOSAT consists of two instruments.The sensors for the two instruments can be broadly labelled as thermal, near infrared and imager.The first two sensors are used as part of Fourier Transform Spectrometer for carbon monitoring which is referred to as TANSO-FTS while the imager for cloud and aerosol observations is referred to as TANSO-CAI.The details on spectral coverage, resolution, field of view, and different products of TANSO-FTS in the three SWIR bands can be found in a number of previous studies (Kuze et al., 2009;Saitoh et al., 2009;Yokota et al., 2009Yokota et al., , 2011;;Crisp et al., 2012;Nayak et al., 2014;Deng et al., 2016a, and references therein).In this study bias corrected ACOS B3.5 Lite XCO 2 from GOSAT Level 2 (L2) retrieval based on the SWIR spectra of FTS observations and made available by Atmospheric CO 2 Observations from Space (ACOS) of NASA is used.ACOS B3.5 Lite XCO 2 has lower bias and better consistency than NIES GOSAT SWIR L2 CO 2 globally (Deng et al., 2016a).Therefore, our choice of the ACOS B3.5 Lite, here after (GOSAT) XCO 2 is motivated by these differences.

OCO-2 measurements
OCO-2, the second world's full-time dedicated CO 2 measurement satellite, was successfully launched by National Aeronautics and Space Administration (NASA) on 2 July 2014.OCO-2 measures atmospheric carbon dioxide with the accuracy, resolution, and coverage required to detect CO 2 source and sink on global and regional scale.OCO-2 has three-band spectrometer, which measures reflected sunlight in three separate bands.The O 2 A-band measures molecular absorption of oxygen from reflected sunlight near 0.76 µm while the CO 2 bands are located near 1.61 µm and 2.06 µm (Liang et al., 2017).In this study, bias corrected OCO-2 XCO 2 V7 lite level 2 covering the period from January 2015 to December 2016, here after referred to as OCO-2 XCO 2 are used.Due to scarcity of data, CT values from the two releases CT2016 for the year 2015 and CT-NRT.v2017for the year 2016, here after (CT16NRT17) are employed in this study.The OCO-2 project team at Jet Propulsion Laboratory, California Institute of Technology, produced the OCO-2 XCO 2 data used in this study.The data can be accessed from NASA Goddard Earth Science Data and Information Service Center.

Methods
The GOSAT and CT model XCO 2 time series used in this investigation spans five years, ranging from May 2009 to April 2014.Atmospheric CO 2 concentrations of NOAA Carbon-Tracker have a global coverage with a 3 0 × 2 0 Longitude/Latitude resolution which covers 428 grid points in our study area.Satellite observations, however, is different from model assimilation, and have gaps because of various reasons (e.g., cloud and the observational mode of satellite).As a result, there is no one to one spatio-temporal match between the two data sets.For example, CO 2 products from the two dataset are not directly comparable since CT is a 3 hourly smooth and regular grid dataset whereas GOSAT XCO 2 is irregularly distributed in space and time.
Thus, the CT CO 2 is extracted on the time and location of GOSAT XCO 2 data.Using the grid point of CT as a reference bin, the corresponding GOSAT XCO 2 found with in a rectangle of 1.5 0 × 1.5 0 with centre at the reference bin and with temporal mismatch of a maximum of 3 hrs is extracted.Moreover, CT has higher vertical resolutions than GOSAT.As a result, the two can not be directly compared.It is customary to smooth the high resolution data (in this case CT) with averaging kernels and a priori profiles of the low resolution satellite measurements (in this case GOSAT).In addition, due to a difference between CT and GOSAT on the number vertical levels, CT CO 2 is interpolated to vertical levels of GOSAT.The CT XCO 2 (XCO model 2 ) used in the comparison is computed from the interpolated CT CO 2 (CO interp 2 ), pressure weighting function (w), XCO 2 a priori (XCO 2a ), column averaging kernel of the satellites retrievals (A) and a priori profile (CO 2a ) of the retrievals as per procedure discussed by Rodgers and Connor (2003); Connor et al. (2008);O'Dell et al. (2012); Chevallier (2015); Jing et al. (2018) and given as: where i is the index of the satellite retrieval vertical level.
Correlation coefficients (R), bias and root mean square deviation (RMSD) are used to assess the level of agreement between the two data sets.The mean bias determines the average deviations in XCO 2 between Carbon Tracker simulation and satellite observations.In this work the bias at the j th grid point is computed as: where S i and O i are CT and GOSAT XCO 2 values over the j th pixel at the i th time respectively.To quantify the extent to which XCO 2 of CT and GOSAT agree, the pattern correlations at the j th grid point are computed as: where S and Ō are the mean values of S i and O i over the j th pixel.The root mean square deviation (RMSD) which shows the standard error of the model with respect the observation at the j th grid point is computed as: This study also applies categorical contingency table for evaluation of performance of CT2016 in capturing the different parts of observed XCO 2 distribution.XCO 2 is a continuous physical quantity for which categorical metrics are not applicable.
In such cases, scatter plots are used as a means of visual inspection of the model skill with no quantitative information in terms of spatial distribution and magnitude of the scatter apart from a single standard deviation.Recently, extended categorical contingency table is proposed to overcome this drawback and assess whether a model simulation/satellite retrieval can capture or fail to capture observed physical quantity exceeding a specified quantile threshold.This procedure effectively reduces the continuous physical quantity with two outcomes i.e., yes (capture) or no (fail).In this study, the skill of CT in correctly simulating whether the observed XCO 2 values are above a selected threshold will be determined using the categorical metrics.
The categorical metrics includes Quantile Bias (QBias), Quantile Probability of Detection (QPOD), Quantile False Alarm Ratio (QFAR), Quantile Critical Success Index (QCSI) also known as the Threat Score and quantile Categorical miss (QMISS).The QBias is defined as the ratio of number of observations (satellite data, OBS) to simulations (Carbon Tracker, SIM) above a certain threshold.Mathematically, QPOD, QFAR, QMISS and QCSI are defined as and where ) is the number of data detected by both simulation (SIM ) and observation (OBS) above a threshold t, N oM iss ) is the number of data detected by observation but missed by simulation and N oF alse ) refers to number of data available only from model above the threshold t and n is the number of exceedances.
QPOD quantifies the fraction of reference observations (in this case GOSAT observations) detected correctly by the simulation above a selected threshold.Meaningful values of QPOD ranges from 0 (zero agreement) to 1 (perfect agreement); QFAR identifies the fraction of events captured by simulation but not available in reference observations above the threshold.Sound values of QFAR is bounded by 0 to 1 with 0 implying perfect score.QCSI combines different aspects of the QPOD and QFAR to characterize the overall performance of the simulation in capturing observation.The QCSI is constrained to have values between 0 (zero agreement) to 1 (perfect agreement) by definition.QMISS quantifies events identified by reference observation but missed by the simulation.Therefore, by definition, quantile categorical miss ranges from 0 (perfect score) to 1 (zero agreement).More details about these categorical statistical metrics can be found in works by other authors (e.g., AghaKouchak et al., 2011;Wilks, 2011;AghaKouchak and Mehran, 2013, and references therein).Using similar coincidence criteria and statistical methods, CT16NRT17 and OCO-2 XCO 2 are also compared.
3 Results and discussions

Comparison of XCO 2 mean climatology from NOAA CT2016 and GOSAT
The mole fraction of CO 2 obtained from the NOAA Carbon Tracker model and GOSAT observation was compared.The results are based on 426 grid pints uniformly distributed to cover the whole Africa's land region.The analysis was based on five years daily data starting from May 2009 to April 2014.The XCO 2 comparison was done only when there are more than ten XCO 2 retrievals that fulfils the spatio-temporal matching criteria defined in Section 2.4.Fig. 1 shows temporal average of CT2016 (Fig. 1a) and GOSAT (Fig. 1b) XCO 2 distribution.The major common spatial feature in the mean map of XCO 2 from GOSAT and CT2016 reanalysis is dipole structure characterized by high XCO 2 northward of equator and low XCO 2 southward of equator with the exception of Congo basin which is characterized by spatially anomalous high XCO 2 .The Southern Africa region is characterized by weak anthropogenic CO 2 emission and high CO 2 uptake by the vegetation.This contributed to the observed dipole distribution.Another important pattern is anomalous peak over annual average location of Inter-tropical convergence zone (ITCZ) (Fig. 1b) which appears to fade over Eastern Africa.This is in agreement with fact that carbon stocks and net primary production per unit land area are higher over Equatorial Africa and decrease towards northward and southward of the equator over arid environments (Williams et al., 2007).However, Fig. 1a shows that CT2016 has some limitations in simulating this spatial pattern in comparison to GOSAT.Fig. 1c shows mean difference (CT2016-GOSAT) XCO 2 which ranges from -4 to 2 ppm.The highest difference between the CT2016 and GOSAT XCO 2 (as high as -4 ppm) is observed over Equatorial Africa, western Ethiopia and South Sudan which are also known for near-year-round rainfall and relatively dense vegetation.The regions are known for their rain forest.The likely explanation could be CO 2 flux from respiration (photosynthesis) of forest in the region which is underestimated (overestimated) in the reanalysis.However, the mean (over five years) may also be slightly positively biased due to fewer observations as shown in Fig. 1d.The strategy and methods for cloud screening in GOSAT retrievals could lead to smaller number of observation in the equatorial region (Crisp et al., 2012;O'Dell et al., 2012;Yoshida et al., 2013;Chevallier, 2015;Deng et al., 2016b).The number of datasets used for comparison range from 14 to 4288 from grid to grid with a spatial mean of 1109 data over the continent.Fig. 1c also shows CT2016 simulations are overall lower than the values of GOSAT observation over most regions with an exception in Gabon, Congo, southern Kenya and southern Tanzania where CT2016 simulations are higher than GOSAT observation by more than 1 ppm.The spatial distribution of global atmospheric CO 2 is not uniform because of the irregularly distributed sources of CO 2 emissions, such as large power plant and forest fire, and biospherical assimilation as clearly noted above.Fig. 2a shows the histograms of differences of CT2016 and GOSAT XCO 2 .The mean difference between CT2016 and GOSAT means is about -0.27 ppm with the standard deviation of 0.98 ppm indicating good regional consistency and low potential outliers.Moreover, a negative mean of the difference implies that XCO 2 simulated from CT2016 is lower than that of GOSAT retrievals over Africa land mass.
Because of selection criteria which permits a difference of 3 degree long and wide, the two datasets are not exactly at the same point.The impact of the relative distance between them should be assessed before performing any statistical comparison.datasets.For these datasets the 50 th percentile has a relative distance of 1.19 0 which means 50% of the data has a relative distance shorter than 1.19 0 .The maximum relative distance between them is 2.12 0 .However, there is no indication that this has been the case since the scatter is not a function of relative distance between the data sets.For example, data points with blue color with lowest location difference is scattered everywhere instead of along the 1:1 line.Furthermore, we found the bias of -0.26 ppm, correlation coefficient of 0.86 and RMSD of 2.19 ppm for datasets which has a relative distance shorter than 1.19 0 .On the other hand, the bias , correlation coefficient and RMSD are -0.33 ppm, 0.86 and 2.22 ppm for those which are longer than 1.19 0 .The above statistics was performed merely to test the influence of location mismatch.Fig. 3 shows a statistical comparison of XCO 2 from the CT2016 and GOSAT over Africa.The number of data used in this comparison are shown in Fig. 1d.As it is depicted in Fig. 3a, the bias ranges from -4 to 2 ppm with a mean bias of -0.28 ppm (see Table 1).A larger negative bias of about -2 ppm was found along the annual mean position of ITCZ.The correlation varies from 0.4 over some isolated pockets in Congo, Tanzania, Mozambique, Uganda and western Ethiopia to 0.9 over northern half of Africa northward of 13 0 N , Eastern Ethiopia and Kalahari Desert.Fig. 3b depicts correlation coefficient between GOSAT and Carbon Tracker XCO 2 .The region with poor correlation also exhibits high RMSD as shown in Fig. 3c.To understand whether this discrepancy originates from model weakness alone, we have looked at the GOSAT posteriori estimate of XCO 2 error, which are high over the same regions with high bias and RMSD between GOSAT and Carbon Tracker XCO 2 (Fig. 3d).GOSAT's posteriori estimate of XCO 2 error is a combination of instrument noise, smoothing error and interference errors (Connor et al., 2008;O'Dell et al., 2012).This posteriori estimate of XCO 2 error does not include forward model error which may lead to underestimation of the true error of satellite XCO 2 by a factor of two (O'Dell et al., 2012).Therefore, part of the discrepancy is clearly linked to satellite own uncertainty, which might have been amplified due to small number of data points used to calculate the mean error of GOSAT XCO 2 measurements (see Fig. 1d).In general, the two data sets are characterized by high spatial mean correlation of 0.83, a global offset of -0.28 ppm, which is the average bias, a regional precision of 2.30 ppm, and a relative accuracy of 1.05 ppm as depicted in Table 1.
Table 1.Summary of statistical relation between CT2016 and GOSAT observation.The statistical tools shown are the mean correlation coefficient (R), the spatial average of bias (Bias), the spatial average root mean square deviation (RMSD), the standard deviation in bias (std of Bias), GOSAT posteriori estimate of XCO2 error (GOSAT err), the standard deviation in CT2016 XCO2 (CT2016 std) and the standard deviation in GOSAT XCO2 (GOSAT std).The number of data used in the statistics is 472,821 over 426 pixels covering the study period; distribution at each grid point is shown in Fig. 1d.Negative bias indicates that CT2016 XCO2 is lower than GOSAT XCO2 values.In this way, XCO 2 distribution is effectively rendered to be dichotomous variable for which we can use the categorical metrics.
Fig. 4 displays values for QBias, QPOD, QCSI, QMISS and QFAR for distribution exceeding 5% (first row), 75% (second row), 90% (third row) and 95% (fourth row) quantiles.We filter out pixels in which the total number of observations are less than 10 to avoid unreliable statistics.The thresholds are set based on the quantiles of the GOSAT observation.QPOD and QCSI decrease at higher quantiles.In contrast, QFAR and QMISS increase at higher quantiles.Specially the decrease in QCSI is significant.It ranges from 0.8 to 1.0 at 5 th percentile ( i.e., QCSI for values exceeding the 5 th percentile) and smaller than 0.4 at 90 th percentile.
On the other hand, the QMISS which is below 0.07 at threshold of 5 th percentile shows a value above 0.58 at threshold value of the 90 th percentile for most of the regions (Fig. 4 ).This indicates that the agreement between CT2016 and GOSAT XCO 2 deteriorates as the comparison data range includes only higher extremes of the XCO 2 distribution.For example, on average over the continent at 95 th percentile the QFAR is lower than 0.24 indicating that 24% (see also simulated by the model to be above 95 th percentile threshold were not actually in the observation.This indicates that a notable difference between CT2016 and GOSAT XCO 2 exists at the tails of the distribution.The information from these statistics are crucial for modelers and/or scientists working on satellite remote sensing to improve the model and/or the retrieval strategy. For example, in satellite retrievals, smoothing constraint or regularization (e.g., Tikhonov first or second order, the so called shape constraint) heavily penalizes the portion of the profile at low and high extremes thereby restricting the possibility of observing profiles unusually different from the a priori profile.It is worth-noting that the a priori profile is usually constructed from climatology of CO 2 profile.In some case, only a handful of a priori profile represents a region (e.g., tropical model atmosphere, mid-latitude model atmosphere, polar model atmosphere) with wide range of variability.This could create huge discrepancy between the model and satellite observations in the tail region of the XCO 2 distribution.The observed difference between CT and GOSAT may well be partly attributed to such factors.However, when the data covers lower extremes, QPOD and QCSI have substantially improved indicating the existence of better agreement between CT2016 and GOSAT at the lower end of the XCO 2 distribution.
Fig. 4a shows the QPOD and QCSI are above 0.93 for lower quantiles which indicates that the model simulates above 93% of the observations (see also Table 2).However, at higher quantiles the change in contingency metrics show a clear deviation between the model simulation and observation.Figs.4b show the scarce datasets in Equatorial Africa which, to certain extent, is the main reason for large biases observed around the Equator in Fig. 3a and also for the corresponding posteriori retrieval error in GOSAT XCO 2 in Fig. 3d.Note that individual posteriori retrieval errors are smoothed out during averaging over large number of coincident observations.Fig. 4 shows that QBias is 1 at lower quantiles and decreases with increasing quantiles implying that the number of CT2016 data that matches the GOSAT observation decreases with increasing threshold.Fig. 4b shows a spatial mean bias of 0.85 (see also Table 2).Fig. 4b also depicts that QCSI is lower than 0.5 and the QMISS is above 0.3 over most regions of Southern Africa.However, QCSI exceeds 0.6 over regions northward of 10 o N .The results indicates that, the agreement between CT2016 and GOSAT XCO 2 shows a regional disparity which is better over Northern Africa than the Southern Africa.In general, the discrepancy between CT2016 and GOSAT XCO 2 is significant over the whole continent towards the extreme high ends of the XCO 2 distribution.
In addition, QBias, QPOD, QCSI, QMISS and QFAR are calculated for all data, i.e., data that includes values higher than 5, 10, 25, 50, 75, 90 and 95 percentiles and averaged over the whole African land mass, as shown in Fig. 5.There is one major conclusion that can be drawn from Fig. 5, i.e., the QFAR and QMISS increase with increase in the quantile thresholds while QBias, QPOD and QCSI decrease.

Comparison of monthly average time series of NOAA CT2016 and GOSAT XCO 2
Africa is one of the largest continents covering both northern and southern hemispheres.As a result, the continent is under the influence of semi-permanent high pressure cells which led to the Sahara Desert in the North and the Kalahari in the South.
The equatorial low pressure cell which allows formation of the seasonally migrating inter-tropical convergence zone is part of the major large scale atmospheric circulation systems.These large scale pressure systems, Oceanic circulations and their interaction with the atmosphere coupled with diverse topographies of the region allow for the formation of different climates (e.g., equatorial, tropical wet, tropical dry, monsoon, semi desert (semi arid), desert (hyper arid), subtropical high climates).
Geographically, the Sahel, a narrow steppe, is located just south of Sahara; the central part of the continent constitutes the largest rainforest next to Amazon whereas most southern areas contain savana plains.The continent get rainfall from migrating ITCZ, west Africa monsoon, intrusion of mid-latitude frontal systems, travelling low pressure systems (Mitchell, 2001, and references therein).Since CO 2 fluxes exhibit seasonal variability and Africa experiences different seasons as noted above, it is important to divide Africa into three major regions, namely North Africa (10 to 35 0 N ), Equatorial Africa (10 0 S to 10 0 N ), and Southern Africa (35 to 10 0 S) and conduct the comparison of the two XCO 2 datasets.overall very good agreement for the monthly averages with respect to amplitudes and phase of XCO 2 .However, XCO 2 from the two data sets slightly disagree in capturing seasonal cycle over Southern Africa.
Fig. 6a   value of -0.04 ppm is found in August (see also Table 4).In addition, both datasets show XCO 2 increases from October to April and decreases from May to September.Moreover, the two dataset shows a monthly mean regional mean bias of -0.36 ppm with a correlation of 1.0 and root mean square deviation of 0.36 ppm (see Table 3).
Fig. 7a shows XCO 2 concentration reaches maximum (392.99 ppm) for CT2016 in March and (393.53 ppm) for GOSAT in January while minimum (389.56 ppm for CT2016 and 389.32 ppm for GOSAT) in October over Equatorial Africa.The largest monthly mean difference of -1.34 ppm and the smallest of -0.05 ppm between the two datasets observed in December in April respectively (Table 4).Moreover, both datasets show that XCO 2 increases from October to March while it decreases from June to October.This similarity in the seasonal variability of the two datasets shows that they are in good agreement in terms of amplitude and phase.In addition, the two data sets show a monthly average regional average bias of -0.17 ppm, correlation of 0.98 and a small root mean square deviation 0.71 ppm over Equatorial Africa (see Table 3).Fig. 8a shows maximum XCO 2 concentration in April (391.04 ppm) for CT2016 and in October (391.28 ppm) for GOSAT, while minimum in May (389.30ppm) for CT2016 and ( 388.46 ppm) for GOSAT over Southern Africa.The largest monthly mean difference of 1.53 ppm and 0.03 ppm between the two datasets is observed in April and in July (Table 4) respectively.Both datasets show concentration of CO 2 increases from May to July while it decreases from October to November.However, the XCO 2 from CT2016 shows a gradual increasing trend from January to April.Conversely, GOSAT XCO 2 shows a decreasing trend.This is most likely CT2016 simulation respond to the growing size of sinks following the rainy season.Moreover, the two data sets shows a monthly mean regional mean bias of 0.07 ppm, correlation of 0.97 and RMSD of 0.87 ppm over southern Africa (see Table 3).
Figs. 6b -8b show regional averaged bias in the monthly mean XCO 2 time series between CT2016 and GOSAT covering the period of May 2009 to April 2014.Fig. 6b shows the presence of seasonally varying negative bias over North Africa.A high (<-0.5 ppm) negative bias in dry seasons (April to June) and low (>=-0.1 ppm) negative bias in wet seasons (August to September) are observed.Moreover, the strength of bias increases from February to June.Conversely, the bias decreases from June to September.Similarly, Figs.7b and 8b show seasonally fluctuating bias over Equatorial and Southern Africa regions.
For example, Fig. 8b shows a positive bias from February to July and negative bias from August to December over Southern Africa.
Figs. 6c -8c show the histogram of difference.The mean difference between CT2016 simulation and GOSAT observation of XCO 2 is -0.36 ppm with a standard deviation of 0.35 ppm over North Africa (see Fig. 6c); Fig. 7c presents a mean difference of -0.17 ppm with standard deviation of 0.71 ppm over Equatorial Africa; and Fig. 8c reveals a mean difference of 0.01 ppm and a standard deviation of 0.85 ppm which indicate that XCO 2 from CT2016 was slightly higher than that of GOSAT over Southern Africa on average.In addition, the low standard deviation of monthly mean difference over North Africa typically indicates good regional consistency between CT2016 and GOSAT.This is mainly because Northern Africa is dominated by the Sahara desert which is known for its weak source/sink of CO 2 .However, the spatial mean of monthly mean bias is slightly higher (-0.36 ppm) over North Africa than over Equatorial Africa (-0.17 ppm ) and Southern Africa (0.01 ppm).This is likely due to the presence of strong local source from emissions and long range transport from the Northern Hemisphere as reported in other studies (Williams et al., 2007;Carré et al., 2010).Loa data sets (Kulawik et al., 2015).The growth rate may not be conclusive due to short length of the data sets used.However, it reflects how the CT and GOSAT observations perform with respect to each other.

Comparison of seasonal climatology
Seasonal cycle has important implications for flux estimates (Keppel-Aleks et al., 2012).It is important to analyse whether there are seasonally dependent biases that are affecting the seasonal cycle, and whether the data sets are capturing the same seasonal cycle.The four seasons considered here are: winter (December, January and February or in short DJF), spring (March, April and May or in short MAM ), summer (June, July and August or in short JJA), and autumn (September, October and November or in short SON).Fig. 9 shows the seasonal distributions of CT2016 (left panels) and GOSAT (middle panels) XCO 2 and their difference (CT2016 -GOSAT, right panels).The distribution clearly shows that XCO 2 concentration is maximum during spring (MAM) and minimum during autumn (SON) over the North Africa.On the other hand maxima is found during autumn (SON) and minima during winter (DJF) over the Southern Africa.These features are in good agreement with the rainfall climatology of northern and southern hemispheres.Moreover, Table 5 shows seasonally varying biases.Seasonal biases affect the seasonal cycle and amplitudes, which are important for biospheric flux attribution (Lindqvist et al., 2015).The right panels in Fig. 9 show that the seasonal mean difference (CT2016 -GOSAT) ranges from -4 to 6 ppm.A maximum difference of 6 ppm over the gulf of Guinea and Congo during JJA.However, such maximum difference was observed over Southern Africa during DJF.A minimum of -4 ppm over annual mean ITCZ region was observed during DJF and MAM.
Moreover, the difference is above 1 ppm over Southern Africa regions during DJF and MAM (wet season of the region).This implies high spatial variability in the seasonal mean difference (see also Table 5).It also suggests that the discrepancy between the CT2016 and GOSAT becomes significant when vegetation cover is weak during DJF and MAM (dry seasons) over North Africa.
During SON the seasonal difference in most Africa's land region ranges from -2 to 1 ppm.The result implies CT2016 simulates lower values of XCO 2 than that of GOSAT observation indicating that there is a better spatial consistency during this season.Furthermore, during these seasons both the Northern and Southern Africa have a moderate vegetation cover following their respective summer seasons.The two datasets show lower regional variation (i.e., only from -2 to 2 ppm) over most of Africa land mass.However, the Equatorial Africa exhibits the mean difference lower than -2 ppm during DJF and MAM.This  2015) have also shown that ERA-Interim precipitable water is higher than measurements from radio-sonde, FTIR and GPS observations.Therefore, such wet bias in the driving ERA-Interim GCM might have forced NOAA CT2016 to generate dense vegetation which serve as CO 2 sink.In other study, Nagarajan and Aiyyer (2004) found ECMWF has a cold bias in the lower atmosphere between 1000 to 750 hPa against independent upper-air sounding data which may affects CO 2 .
Fig. 10 shows mean difference between CT2016 and GOSAT XCO 2 seasonal means which ranges from -0.37 to 0.04 ppm with a standard deviation within a range of 1.00 to 1.91 ppm over the continent.The highest mean difference of XCO 2 (-0.37 ppm) occurs during SON and the lowest (0.04 ppm) occurs during MAM.Table 5 presents the summary of statistical values for spatial mean of each season mean.The comparison between the two data sets also shows there is a strong correlation (>0.5) during each season over the continent.However, there is moderate correlations (0.3 to 0.5) during DJF and MAM over North Africa and during DJF over Southern Africa.The low correlation over Northern Africa may be linked to a weak absorption by vegetation and a strong emission from human activities during winter as reported elsewhere (Liu et al., 2009;Kong et al., 2010).
Moreover, Table 5 shows that the seasonal biases are negative over North Africa while they are mostly positive over Equatorial and Southern Africa.Negative biases are observed during DJF and SON over Equatorial and Southern Africa respectively implying that XCO2 from CT2016 are lower than GOSAT during dry seasons.3.5 Comparison of mean XCO 2 from NOAA CT16NRT17 and OCO-2 The strong El Niño event occurred during 2015-2016 provides an opportunity to compare the performance of CT16NRT17 during strong El Niño events.Because of the decline in terrestrial productivity and enhancement of soil respiration, the concentration of CO 2 increases during El Niño events (Jones et al., 2001).In this section we compare mean XCO 2 of NOAA CT16NRT17 and NASA's OCO-2 covering the period from January 2015 to December 2016.OCO-2 is the most recent full-5 time dedicated CO 2 measuring satellite with greater spatio-temporal resolution.
The comparison was done based on the selection criteria discussed in Section 2.4.Fig. 11 shows mean distribution of XCO 2 from CT16NRT17 (Fig. 11a) and OCO-2 (Fig. 11b) over Africa's land mass.CT16NRT17 shows high ( > 400 ppm) XCO 2 values over North Africa while these high XCO 2 values are observed over Equatorial Africa in the case of OCO-2 concentration reported in past study (Williams et al., 2007) implying that the CT16NRT17 likely underestimates XCO 2 values over Equatorial Africa.It is also possible that the discrepancy is a compounded effect of OCO-2 XCO 2 positive bias over the region (O'Dell et al., 2012;Chevallier, 2015).Fig. 11c shows the mean difference between two years mean of XCO 2 from CT16NRT17 and OCO-2, which is in the range from -2 to 2 ppm.However, high (<-2 ppm) negative mean difference between the two data sets over rain forest regions (Gulf of Guinea and Congo basin) and ITCZ zone over Eastern Africa (South Sudan and southeastern Sudan) is observed implying that CT16NRT17 simulates lower XCO 2 values than that of OCO-2 observation over regions where vegetation uptake is strong.Conversely, high (>1) positive mean difference over the Sahara desert, Somalia and Tanzania implies CT16NRT17 simulates higher XCO 2 values than OCO-2 observation where the vegetation uptake is weak.Moreover, a positive (>2) mean difference over Egypt, Libya, Sudan, Chad, Niger, Mali and Mauritania is likely due to overestimates of XCO 2 emission from local sources by CT16NRT17.Overall, the two datasets show a fairly reasonable agreement with a correlation of 0.60 and offset of 0.36 ppm, a regional precision of 2.51 ppm and a regional accuracy of 1.21 ppm.Because of presence of spatial and temporal mismatch of some level between CT16NRT17 and OCO-2 datasets, it is important to assess the effect of relative distance between the datasets.Fig. 12b shows a color coded distribution of the two datasets.
In the figure color codes indicate the relative distance.The random scatter of blue dots implies that the statistical discrepancies do not arise from the relative distance between the two datasets.More specifically, a statistical comparison of datasets lower  with a mean bias of 0.34 ppm.However higher biases (<-2 ppm) are observed over Equatorial Africa along the annual average location of ITCZ.Fig. 13b shows the correlation map with values from 0.2 to 0.8 over Africa's land mass.A good correlation exceeding 0.6 are seen over many regions of the continent while weak correlation of less than 0.2 and higher root mean square error (> 3 ppm ) are observed over small pockets of Equatorial and Eastern Africa regions (see Fig. 13c).These regions also show high (> 0.65 ppm) error in satellite retrieval (see Fig. 13d).In addition, Fig. 11d shows the number of observations are small (< 1000 ) over the regions.This may contribute to the observed discrepancy over these regions.However, weak correlations are also observed over a wider area in North Africa such as Mauritania, Mali, Algeria and some regions of Niger where satellite errors are low and sufficient data are obtained.This indicates the necessity of incorporating more measurement in Carbon Tracker assimilation over North Africa in order to tune the assimilation model such that it captures the sub-regional carbon cycles.The poor correlation and high RMSD values observed over Ethiopia highland is likely due to the inefficiency of retrieving XCO 2 from satellites over high-latitude lands (Chevallier, 2015).
3.6 Categorical comparison of XCO 2 from NOAA CT16NRT17 and OCO-2 Fig. 14 depicts the QBias, QPOD, QCSI, QMISS and QFAR between CT16NRT17 and OCO-2 at different thresholds.The maps clearly show that QFAR and QMISS increase with increasing threshold.On the other hand, QBias, QPOD and QCSI decrease with increasing thresholds (see also Fig. 15).
In our analysis, threshold is determined based on OCO-2 observation.The value of XCO 2 at 90 th percentile is 404.25 ppm and most CT16NRT17 XCO 2 value are below 404 ppm over North Africa.Therefore, categorical analysis shows white space over these regions for thresholds above 90 th percentile.Fig. 14 also shows that CT16NRT17 simulates the observation at lower quantiles.However, discrepancy between CT16NRT17 and OCO-2 is significant at higher quantiles.At 75 th quantile the QPOD and QCSI values are less than 0.6 over Africa's land mass which implies more than 40% of the OCO-2 observation were not captured in the CT16NRT17 simulation.into the atmosphere (Chatterjee et al., 2017).Fig. 16a also shows that XCO 2 from CT16NRT17 simulation are higher than OCO-2 observation over North Africa.
Fig. 16b shows the monthly mean difference between CT16NRT17 and OCO-2 which ranges from -0.5 to 2 ppm.As a consequence of strong El Niño, the region misses the short rain season during spring (MAM) when vegetation are still under the influence of the dormancy of winter (DJF).However, following the summer vegetation awakens from the long dormancy and becomes a sink of CO 2 .Starting from August 2015, the difference between the two datasets is minimum; this implies that CT simulates low values of XCO 2 when the vegetation uptake is strong.On the other hand, a maximum difference of exceeding 1 ppm was observed during MAM, implying higher XCO 2 values from CT16NRT17 simulation than that of OCO-2 when vegetation uptake is weak following the strong El Niño over North Africa.Moreover, Fig. 16c displays a monthly mean regional mean bias of 0.87 ppm, correlation of 0.95 and a root mean square deviation of 0.72 ppm between CT16NRT17 and OCO-2 XCO 2 .This implies that CT16NRT17 is in a good agreement with OCO-2.However, a small discrepancies arose due to likely a strong anthropogenic emission from Nigeria, Egypt and Algeria together with the establishment of plantation over   North Africa, which recently exceeded deforestation, and resulted in net flux of carbon sink (Canadell et al., 2009).This might have contributed to the observed discrepancy over North Africa.
Figs. 17a -18a show monthly mean time series of XCO 2 from the model and OCO-2 instrument over Equatorial Africa and Southern Africa which are also in good agreement in terms of pattern.However, the figures show that CT16NRT17 simulations are lower than those of OCO-2 during October, November and December whereas it is opposite during April, May and June over Equatorial Africa and Southern Africa.Figs.17b and 18b depict a seasonal bias in the monthly time series over Equatorial Africa and Southern Africa respectively.Positive biases are observed during dry seasons while negative biases are during wet seasons.Moreover, the datasets have monthly averaged regional mean biases of 0.13 and 0.11 ppm, correlation of 0.90 and 0.94, RMSD of 0.84 and 0.73 ppm over Equatorial Africa and Southern Africa respectively.This shows that existence of better agreement between CT16NRT17 and OCO-2 over these regions in terms of monthly average regional mean values.Figs.
16d-18d show both CT16NRT17 and OCO-2 are in good agreement in estimating the annual growth rate.Patra et al. (2017) found a global mean of more than 3 ppm of CO 2 added to the atmosphere due to the strong El Niño event that occurred during 2015-2016.In agreement with this, both CT16NRT17 and OCO-2 shows an annual growth rate that ranges from 3.10 to 3.42 ppm year −1 of XCO 2 over Africa's land mass (see also Table 7).However, over all regions of Africa's land mass CT16NRT17 shows lower XCO 2 annual growth rate than those of OCO-2.(e.g., Mali during JJA) is due to insufficient coincident satellite data according to the selection criteria during these seasons.
XCO 2 increases from winter to spring and then decreases from spring peak to summer minimum over the whole continent.
The decrease from spring maximum to summer continued into autumn over northern half of Africa in contrast to southern half of Africa which exhibits an increase in XCO 2 .The decrease from spring to autumn (northward of equator) and until summer (southward of equator) is likely to be a consequence of the land vegetation awakening from dormancy of winter and 5 partly spring.Conversely, the decomposition of died and decayed vegetation which began in autumn and continued throughout winter adds extra CO 2 leading to a maximum concentration during spring (Idso et al., 1999).In agreement with this, both CT16NRT17 and OCO-2 show maximum XCO 2 during MAM over North Africa and during SON over Southern Africa.
Conversely, minimum XCO 2 are observed during SON over North Africa and during DJF over South Africa.
Fig. 19 (right panels) shows the seasonal mean difference of CT16NRT17 and OCO-2.A higher mean difference of greater 10 than 1 ppm are observed over North Africa during DJF and MAM when the vegetation cover over the region decreases.This indicates that XCO 2 values from CT16NRT17 are higher than that of OCO-2 when vegetation uptake is weak.On the other   events.Comparison of Carbon Tracker with the two satellites reveals biases of -0.28 and 0.34 ppm, correlations of 0.83 and 0.60 and root mean square deviations of 2.30 and 2.57 ppm with respect to GOSAT and OCO-2 respectively.The performance of the model in capturing the whole distribution is also assessed.It is found that Carbon Tracker can capture more than 93% of the observation higher than the lowest (5 th percentile) thresholds.However, the quantile probability of detection and the quantile critical success index decrease with increasing threshold implying the model and satellite show some discrepancies over the extreme values exceeding 90 th percentile.
The monthly average time series of CT2016 over North Africa, Equatorial Africa and Southern Africa are separately compared with XCO 2 from the two satellites.CT2016 agrees well with measurements from the two instruments in terms of pattern and amplitude.However, this agreement deteriorates over Equatorial and Southern Africa in terms of amplitude.It is also found that there is a seasonal dependent bias between them which is negative during dry seasons while it is positive during wet seasons.This indicates results of CT2016 are mostly lower than the GOSAT observation.High spatial mean of seasonal mean RMSD of 1.91 during DJF and 1.75 ppm during MAM and low RMSD of 1.00 and 1.07 ppm during SON in the model XCO 2 with respect to GOSAT and OCO-2 are observed respectively thereby indicating better agreement between CT and the satellites during autumn.CT2016 has the ability to capture monthly time series and seasonal cycles.However, XCO 2 from CT2016 is lower than GOSAT observations over North Africa during all seasons whereas XCO 2 from CT2016 is higher than that of GOSAT over Equatorial and Southern Africa with the exceptions of DJF over Equatorial Africa and SON over Southern Africa.In addition, CT2016 simulates lower XCO 2 than the observations over some regions (e.g., Congo, South Sudan and southwestern Ethiopia) and during summer season over the whole continent following large vegetation uptake.In contrast, XCO 2 from CT16NRT17 is higher than that of OCO-2 over North Africa whereas it is lower than that of OCO-2 during DJF and SON over Equatorial and Southern Africa respectively.
In general, XCO 2 from NOAA CT shows a very small bias with respect to GOSAT and OCO-2 observation over Africa's land mass.Moreover, there is a good agreement between CT simulation and observations in terms spatial distribution, monthly average time series and seasonal climatology.However, there are some discrepancies between the model and the two XCO 2 datasets from GOSAT and OCO-2 implying that the accuracy of the model data needs further improvements for the rain forest regions (e.g., Congo) through assimilation of in-situ observations and tuning of the model through process studies.Further 5 work may also be needed to improve the XCO 2 from satellites and models in the extreme parts of XCO 2 distribution.

Figure 1 .
Figure 1.Distribution of five-years averages of CT2016 (a) and GOSAT (b) XCO2 and their difference (c) gridded in 3 0 × 2 0 bins over Africa's Land mass; and the total number of datasets at each grid (d).

Fig
Fig. 2b depicted color-coded scatter plot of CT2016 model simulation verses GOSAT to determine if the discrepancy between the data sets arise from spatial mismatch.The color code indicates the relative distance between the model and observation

Figure 2 .
Figure 2. Histogram of the difference of CT2016 relative to GOSAT (left panel) and color code scatter diagram of XCO2 concentration as derived from CT2016 and GOSAT (right panel).Color indicates the relative distance as shown in colorbar between datasets.

Figure 3 .
Figure 3. Spatial patterns of bias (a), correlation (b), RMSD (c) of the two data sets, and mean posteriori estimate of XCO2 error from GOSAT (d).

3. 2
Categorical comparison of XCO 2 from NOAA CT and GOSAT Following the methods described in AghaKouchak et al. (2011), QBias, QPOD, QCSI, QMISS and QFAR are determined from the coincident XCO 2 data to assess how Carbon Tracker model and satellite observations perform with respect each other in capturing different parts of XCO 2 distribution.It is worth noting that these categorical metrics are used to evaluate the level of agreement between CT2016 and GOSAT XCO 2 in certain part of XCO 2 distribution (e.g., above a given quantile threshold).

Figs. 6
Figs. 6 -8 show time series of XCO 2 monthly means covering the period from May 2009 to April 2014 for both CT2016 and GOSAT over North Africa, Equatorial Africa and Southern Africa respectively.Figs.6a -8a depict the existence of an

Figure 6 .
Figure 6.The monthly mean time series of CT2016 and GOSAT from May 2009 to April 2014 averaged over North Africa (a), bias associated to the monthly means (b), the histogram of difference (c) and the annual growth rate obtained by subtracting the mean from the mean of the next year (d).The error bars in (a) shows the GOSAT a posteriori XCO2 uncertainty.

Figure 7 .
Figure 7.The same as Fig. 6 but over Equatorial Africa.

Figure 8 .
Figure 8.The same as Fig. 6 but over Southern Africa.

Figure 11 .
Figure 11.Distribution of two years average XCO2 of CT16NRT17 (a) and OCO-2 (b) XCO2 and their difference (c) gridded in 3 0 × 2 0 bins; and (d) the total number of datasets at each grid

Fig. 12a shows
Fig.12ashows the histogram of two years mean difference, which is characterized by a positive mean of 0.34 ppm and a standard deviation of 1.21 ppm.This suggests that CT16NRT17 simulates high XCO 2 as compared to observations from OCO-2 over Africa's land mass.

Figure 12 .
Figure 12.Histogram of the difference of CT16NRT17 relative to OCO-2 (left panel) and color code scatter diagram of XCO2 concentration as derived from CT16NRT17 and OCO-2 (right panel).Color indicates the relative distance as shown in colorbar between datasets.

Fig. 13
Fig. 13 shows the comparison of mean XCO 2 from CT16NRT17 and OCO-2 covering the period from January 2015 to December 2016.The number of data used are displayed in Fig. 11d.Fig. 13a depicts the bias which ranges from -2 to 2 ppm

Figure 16 .
Figure 16.The monthly mean time series of CT16NRT17 and OCO-2 from January 2015 to December 2016 averaged over North Africa (a), bias associated to the monthly means (b), the histogram of difference (c) and the annual growth rate obtained by subtracting the mean from the mean of the next year (d).The error bars in (a) shows the OCO-2 a posteriori XCO2 uncertainty.

Figure 17 .Figure 18 .
Figure 17.The same as in Fig. 16 but over Equatorial Africa.

Fig. 20
Fig.20shows the histogram of seasonal mean difference of CT16NRT17 and OCO-2.The smaller standard deviation of 1.49 and 1.07 are observed during JJA and SON.On the other hand, higher standard deviation of 1.69 and 1.75 ppm are observed during DJF and MAM respectively.The results indicate that CT16NRT17 and OCO-2 show a better consistency during wet seasons and this consistency decreases as the vegetation cover decreases over most regions of Africa land mass during dry seasons.

Table 2 .
Summary of extended contingency metrics for the relation between CT2016 model simulation and GOSAT observation.

Table 3 .
Summary of statistical relation between CT2016 and GOSAT observation.The statistical analysis were made using monthly average of time series of 60 months (i.e., months from May 2009 to April 2014).

Table 4 .
Five years monthly averaged XCO2 concentration in ppm obtained from CT2016 (CT) and GOSAT (GO) and their difference CT − GO (D) in ppm over Africa (A), North Africa (NA), Equatorial Africa(EA) and Southern Africa (SA).

Table 5 .
Summary of statistical relation between CT2016 and GOSAT XCO2: Bias, correlation (R), Root mean square deviation (RMSD), standard deviation of XCO2 from CT2016 simulation (CT2016 std), standard deviation of XCO2 from GOSAT observation (GOSAT std), aggregate number of coincident observations (number of data) and number of grids over the region (grid).Negative bias means CT2016 is lower than GOSAT.The statistics are on the basis of spatial average of seasonal averages of bias, correlation, RMSD and standard deviations.

Table 6 .
Summary of statistical relation between CT16NRT17 and OCO-2 observation.The statistical tools shown are the mean correlation coefficient (R), the average of bias (Bias), the average root mean square deviation (RMSD), the standard deviation in bias (std of Bias), mean posteriori estimate of XCO2 error from OCO-2 (OCO-2 err), the standard deviation in CT16NRT17 XCO2 (CT16NRT17 std) and the standard deviation in OCO-2 XCO2 (OCO-2 std).Positive Bias indicates that CT16NRT17 is higher than OCO-2.The number of data used in the statistics is 1,659,411 over 426 pixels covering the study period.Distribution at each grid point is shown in Fig 11d.
3.7 Comparison of monthly average time series of NOAA CT16NRT17 and OCO-2 XCO 2 Figs.16 -18 show a two year monthly average time series comparison of XCO 2 from CT16NRT17 and OCO-2 over North Africa, Equatorial Africa and Southern Africa respectively.Fig.16ashows the existence of good agreement between the two datasets in describing pattern over North Africa.Moreover, both datasets show a decreasing trend of XCO 2 from May to September while increasing trend from October to April.On the other hand, consistent with the climate condition and associated CO 2 exchange, the monthly mean XCO 2 shows a maximum value of 403.37 ppm for CT16NRT17 and 402.06 ppm for OCO-2 during May.Conversely, a minimum concentration of 398.77 ppm from CT16NRT17 simulation and 398.27 ppm from OCO-2 observation are found in September.In addition, both CT16NRT17 and OCO-2 show maximum XCO 2 values (402.15ppm for CT16NRT17 and 402.03 ppm for OCO-2) in December.These pick values in December are not surprising, because the 2015-2016 El Niño started on March 2015 and reached pick in December 2015 which added extra CO 2

Table 7 .
Annual growth rate (AGR) of XCO2 over Africa land mass from CT16NRT17 and OCO-2.The results are obtained as the mean