Uncertainties in biospheric fluxes are thought to be greater than in other CO2 source types e.g.,, and the CO2 signal from biospheric fluxes is often larger than from other source types. Hence, we design a set of top-down experiments to examine whether we can detect variations in biospheric CO2 sources and sinks within different regions of the globe and different months of the year using OCO-2 observations. In the present study, these variations are defined as any spatial or temporal patterns in CO2 fluxes that have been gridded to the resolution of a global atmospheric model: 1∘ latitude by 1.25∘ longitude and a 3-hourly time interval.

Detecting variations in CO2 fluxes is a prerequisite for constraining CO2 budgets or flux totals; we must be able to detect variations in CO2 sources and sinks across a region if we are to constrain budgets across any region of smaller size. We begin with two large hemispheric regions and then decrease the size of those regions to create increasingly challenging tests of the OCO-2 observations (Fig. ); we use four and seven global regions, respectively, in each of these experiments. All of these regions are based on a map of global biomes presented in . The seven-region map contains broad global biomes aggregated from those in , while the two- and four-region maps have been aggregated from to form even larger regions. We use a biome-based map because inverse modeling studies often estimate CO2 flux totals for biome-based regions, and these regions have clear ecological significance.

We construct this set of experiments for each of the last three versions of the OCO-2 observations and examine how the results change with the retrieval version. These experiments are identical except for the retrieval version used. Therefore, this setup provides a means to understand how improvements in the observations are improving the constraint on biospheric CO2 fluxes. We examine these questions for each month within the year 2015 – to understand how these results vary by season and by region or biome.

We design a regression framework to determine whether we can detect variations in CO2 fluxes using OCO-2 observations. This section provides an overview of the approach, but provides full descriptive and mathematical detail. This regression will try to match CO2 observations from OCO-2 using numerous atmospheric model outputs. Each model output estimates the enhancement in total column CO2 (XCO2) from fluxes in a particular region and a particular month. We generate all of these model outputs of CO2 using the Parameterized Chemistry and Transport Model (PCTM) . The model setup used here has a spatial resolution of 1∘ latitude by 1.25∘ longitude, and we incorporate CO2 fluxes at a 3-hourly time resolution. The wind fields used to drive PCTM are from the Modern-Era Retrospective analysis for Research and Applications (MERRA) product . This setup is identical to .

We run many atmospheric model simulations using numerous different biospheric CO2 flux estimates. The regression will try to reproduce OCO-2 observations using a linear combination of these model simulations. For example, in the seven-region experiments, we use seven different geographic regions, seven biospheric CO2 flux estimates, and 16 different months (September 2014–December 2015). We discard results from the first 4 months as model spin-up. These combinations equate to 784 total atmospheric model outputs. We further run atmospheric model simulations using a spatially and temporally constant flux in each region and each month, and we allow the regression to use these model outputs as well. The Supplement and describe the CO2 flux estimates and regression in greater detail.

This approach provides a means to evaluate when and where current satellite observations can constrain variations in CO2 fluxes. At least some of the atmospheric model outputs that are driven by biospheric CO2 flux estimates should help reproduce the OCO-2 observations better than the model outputs that are driven by spatially and temporally constant fluxes. If so, a model with spatially and temporally variable fluxes is better able to reproduce OCO-2 observations than a model with constant fluxes. This result would imply that OCO-2 observations can be used to detect variations in biospheric CO2 sources and sinks within a given region for a given month. By contrast, suppose that the atmospheric model outputs driven by biospheric CO2 flux estimates do not reproduce the OCO-2 observations any better than the model outputs with constant CO2 fluxes. This result would imply one of several conclusions. First, the observations may not be sensitive to fluxes from the region or month in question. This outcome may occur if the magnitude of fluxes is small in a given region or if there are no OCO-2 observations near that region. Second, errors in the atmospheric model or in the OCO-2 observations may obscure variations in XCO2 that are due to CO2 fluxes. Lastly, the biospheric CO2 flux estimates used in the atmospheric model may not be skilled and may not reflect real-world biospheric CO2 fluxes. However, in this study, we offer up seven biospheric CO2 flux estimates for each region and each month, and at least one of these estimates should correlate with real-world CO2 fluxes to a reasonable extent. Hence, it is unlikely that this explanation would drive the results. Rather, it is more likely that the observations are not sensitive to fluxes from a given region or that errors in the model–data system are too large.

Note that anthropogenic, biomass burning, ocean, and biospheric fluxes all contribute to XCO2 observed by OCO-2, and we need to account for non-biospheric CO2 fluxes in order to isolate the signal from biospheric fluxes in the regression. We model atmospheric enhancements of XCO2 from anthropogenic emissions using EDGAR v4.2 FT2010 , climatological ocean fluxes using , and biomass burning fluxes using the Global Fire Emissions Database (GFED), version 4.1 ; each of these model outputs is considered in the regression.

We further implement model selection to evaluate when and where current satellite observations can constrain variations in biospheric CO2 fluxes. Model selection will determine which combination of atmospheric model outputs to include in the regression based upon which best reproduces the OCO-2 observations. If this combination includes at least one biospheric CO2 flux model for a given region and season, we conclude that the observations likely can be used to constrain variations in CO2 fluxes. However, if this combination does not include any biospheric CO2 flux model for a given region and season, we conclude that the observations likely cannot be used to constrain flux variations for that region and season.

We specifically employ a form of model selection known as the Bayesian information criterion (BIC), an approach commonly used in regression modeling e.g.,chap. 12 and more recently in atmospheric inverse modeling e.g.,. To this end, we create different combinations of model outputs and use each combination in the regression. We score each combination based upon how well it reproduces the OCO-2 observations; combinations with a lower weighted sum of squares error receive a better score. Each combination is also scored based upon the total number of model outputs in that combination. Specifically, combinations with a greater number of model outputs receive a larger penalty for complexity, and this penalty prevents combinations that overfit the data from receiving an anomalously good score. The best combination of atmospheric model outputs is the one with the lowest score. We subsequently examine this combination and tally whether at least one atmospheric model output using a biospheric flux estimate was selected for each region and each month of the year. and the Supplement describe this approach in greater detail, including the specific equations for the BIC.

The constraint on CO2 fluxes using recent versions of the OCO-2 observations is a step-change improvement relative to previous versions. Overall, there was only a limited ability to detect variations in monthly CO2 fluxes across individual biomes using version 7 of the retrievals Fig. a–c,. However, these capabilities have changed using versions 8 and 9 of the observations (Fig. d–i). Variations in CO2 fluxes are detectable across tropical biomes throughout much of the year and across temperate biomes in the Northern Hemisphere summer when fluxes from these regions are most variable. These results imply that the updated OCO-2 observations can be used to detect and constrain variations in monthly CO2 fluxes from seven biome-based regions in certain circumstances – in about two-thirds of all months in the tropics and during the Northern Hemisphere summer in the extratropics.

The improvement in the flux constraint is particularly evident in the four- and seven-region experiments (Fig. b–c and e–f). In the four-region model selection experiments, the OCO-2 observations can be used to detect variations in tropical fluxes for most months of the year (Fig. e). In other words, at least one biosphere flux model is found to explain a sufficiently large fraction of the observed variability in XCO2 as to be selected via the BIC model selection procedure for the tropical regions for most months. This result indicates that spatiotemporal variability in CO2 fluxes from within each of these regions is preserved in the OCO-2 observations. This represents a marked improvement over results when using observations from version 7 of the OCO-2 retrieval algorithm Fig. b and e,. The results using the newer versions 8 and 9 also show substantial improvements in other regions, including dryland and dry monsoon regions, temperate regions, and high-latitude regions (Fig. e and h).

The seven-region model selection experiments are an even more challenging test of current observations. These experiments examine whether we can detect spatiotemporal variations in biospheric fluxes across seven broad, aggregated global biomes. These experiments produce much better results using versions 8 and 9 of the observations. Specifically, biospheric flux models are selected across tropical and subtropical biomes for at least 1 month of every season. The same is true across all temperate and high-latitude biomes for a minimum of 1 month during the Northern Hemisphere summer.

These improvements appear greatest across tropical biomes. There is a consistent flux signal from many tropical regions throughout the year, and hence we are able to detect variations in fluxes from tropical regions across different seasons using versions 8 and 9 of the observations. By contrast, the atmospheric signal due to biospheric CO2 fluxes in northern mid- and high latitudes has the largest absolute magnitude during the Northern Hemisphere summer. As a result, we see a large improvement in the flux constraint in midlatitudes in the Northern Hemisphere summer but not in other times of year when the absolute magnitude of CO2 fluxes is smaller. Furthermore, there are far fewer land nadir and land glint observations in northern mid- and high latitudes in the Northern Hemisphere winter relative to summer.

One notable feature of all model selection experiments is the result for dryland and dry monsoon regions (Fig. c). At first glance, it may appear surprising that biospheric flux models are selected for so many months in this region, given that some parts of this region are very dry and presumably have small CO2 fluxes. Several semiarid regions within this classification have a very distinct monsoon that can bring over 500 mm of precipitation per month (e.g., northeastern Brazil, western India, and Pakistan). As a result, there is a large spatial contrast in CO2 fluxes across these regions during the Northern Hemisphere spring and summer – large CO2 uptake in places with a spring and summer monsoon and little to no fluxes in places like the Sahara or the Arabian Peninsula.

Note that the results using version 9 of the observations are not very different from those using version 8. The change in the observations between versions 8 and 9 is only incremental (e.g., Fig. b). Version 9 has a lower quality control threshold for surfaces with low albedo, resulting in more observations across tropical rainforests , and this version includes a topography correction that mostly manifests at small spatial scales. The latter change could be very important for studies that estimate point sources or urban emissions using OCO-2. However, these changes are unlikely to make a large difference in this study both given the large size of the regions examined and the 1∘ latitude by 1.25∘ longitude spatial resolution of the atmospheric model simulations. The Supplement includes a detailed discussion of the subtle differences between the model selection results using versions 8 and 9 of the observations.

Numerous factors affect the accuracy of CO2 fluxes estimated from satellite data. These factors include the accuracy and precision of the observations, the atmospheric transport model, and the prior flux estimate used in the inverse model. Improvements in any of these inverse modeling inputs could improve the constraint on biospheric CO2 fluxes. We find that recent improvements to the retrieval are having a particularly large impact on the strength of the CO2 flux constraint. Furthermore, these improvements are not restricted to a single satellite like OCO-2. Rather, the ACOS retrievals and bias correction will be directly applicable to other NASA carbon monitoring missions, including the recently launched OCO-3 mission and the planned GeoCarb mission .

These improvements to the retrieval algorithm have had an effect on both glint and nadir observations from OCO-2 collected in almost every region of the globe. The sheer number of different changes makes it challenging to pinpoint exactly which have had the largest impact on the CO2 flux constraint; there have been numerous updates to the quality control prescreening, the forward spectroscopy model, the retrieval algorithm, and the bias correction. Furthermore, these updates have had multiple effects on the reported CO2 observations, reducing white noise, reducing bias, and changing which observations do or do not pass quality control. detail these changes in much greater detail.

With that said, a few of these improvements appear to have a particularly salient impact on the results of this study. For example, the largest improvements have generally been to the glint-mode observations. A 0.2 to 0.3 ppm bias between land nadir and land glint observations in version 7 has been remedied in version 8, and version 8 glint observations show smaller biases across many ocean regions. Furthermore, version 8 exhibits less random noise in all types of observations, but that noise reduction is largest in glint observations, both over land and over the oceans .

Indeed, we also see the largest improvement in the flux experiments conducted in this study when we include glint mode observations. Figure  displays the results of the model selection experiments when the glint data are excluded. The figure shows results using version 7, 8, and 9 of the observations. The improvement between versions 7 and 8 is much smaller when the glint observations are excluded than when they are included (Fig. ). Even in terrestrial regions, these glint observations may play a key role in the overall flux constraint. For example, the absolute number of nadir and glint observations over land are roughly equal; there are 4.3×106 land nadir observations with a positive quality control flag for 2015 and 4.3×106 land glint observations during the same time period.

Note that this study focuses on detecting variations in CO2 fluxes from terrestrial regions in individual months. To that end, certain types of flux estimation problems are beyond the scope of the current study. For example, there is strong evidence that OCO-2 observations are still biased across northern tropical oceans, and reductions in these biases could improve ocean flux estimates derived from OCO-2 . Furthermore, there is always a possibility that the observations have a bias that is correlated across regions larger than those examined in this study. For example, the observations show a small, time-dependent drift from one year to another . The approach used in this study would be unlikely to detect the impact of those biases.