Journal of Geophysical Research: Atmospheres

Estimation of regional emissions of nitrous oxide from 1997 to 2005 using multinetwork measurements, a chemical transport model, and an inverse method

Authors


Abstract

[1] Nitrous oxide (N2O) is an important ozone-depleting gas and greenhouse gas with multiple uncertain emission processes. Global nitrous oxide observations, the Model of Atmospheric Transport and Chemistry (MATCH) and an inverse method were used to optimally estimate N2O emissions from twelve source regions around the globe. MATCH was used with forecast center reanalysis winds at T62 resolution (192 longitude by 94 latitude surface grid, and 28 vertical levels) from 1 July 1996 to 30 June 2006. The average concentrations of N2O in the lowest four layers of the model were then compared with the monthly mean observations from four national/international networks measuring at 65 surface sites. A 12-month-running-mean smoother was applied to both the model results and the observations, due to the fact that the model was not able to reproduce the very small observed seasonal cycles. The inverse method was then used to solve for the time-averaged regional emissions of N2O for two time periods (1 January 1997 to 31 December 2001 and 1 January 2002 to 31 December 2005). The best estimate inversions assume that the model stratospheric destruction rates, which lead to a global N2O lifetime of 125 years, are correct. It also assumes normalized emission spatial distributions within each region from Bouwman et al. (1995). We conclude that global N2O emissions with 66% probability errors are 16.3−1.2+1.5 and 15.4−1.3+1.7 TgN (N2O) a−1, for 1997–2001 and 2001–2005 respectively. Emissions from the equator to 30°N increased significantly from the initial Bouwman et al. (1995) estimates while emissions from southern oceans (30°S–90°S) decreased significantly. The quoted uncertainties include both the measurement errors and modeling uncertainties estimated using a separate flexible 12-box model. We also found that 23 ± 4% of the N2O global total emissions come from the ocean, which is slightly smaller than the Bouwman et al. (1995) estimate. For the estimation of emissions from the twelve model regions, we conclude that, relative to Bouwman et al. (1995), land emissions from South America, Africa, and China/Japan/South East Asia are larger, while land emissions from Australia/New Zealand are smaller. Our study also shows a shift of the oceanic sources from the extratropical to the tropical oceans relative to Bouwman et al. (1995). Between the periods 1997–2001 and 2002–2005, emissions increased in China/Japan/South East Asia, 0°–30°N oceans, and North West Asia and decreased in Australia/New Zealand, 30°S–90°S oceans, 30°N–90°N oceans, and Africa. The lower tropical ocean emissions in 1997–2001 relative to 2002–2005 could result from the effects of the 1997–1998 El Nino in the earlier period.

1. Introduction

[2] Nitrous oxide (N2O) has the fourth largest contribution to radiative forcing (RF) over the past 250 years among the long-lived greenhouse gases behind CF2Cl2, CH4 and CO2 according to the Intergovernmental Panel on Climate Change (IPCC [Forster et al., 2007]). Nitrous oxide is also the major source of ozone-depleting NO and NO2 in the stratosphere. The recent Ozone Assessment [Daniel et al., 2007] recommended a lifetime of 120 years (114 years including its indirect effects on ozone) for N2O. The uncertainty in this lifetime remains large; specifically Volk et al. [1997] calculate a value of 122 ± 24 years (not including indirect effects) based on stratospheric measurements. There are also large uncertainties in estimates of emissions from the major soil, agricultural, combustion and oceanic sources of N2O. Given these emission uncertainties, Prather et al. [2001] in the earlier IPCC report concluded that its observed relative rate of increase of 0.2 to 0.3% yr−1 was consistent with its better-quantified major sinks (principally stratospheric destruction) and the primary driver for the industrial-era increase of N2O was concluded to be enhanced microbial production in expanding, fertilized agricultural lands.

[3] Since IPCC (2001), understanding of regional N2O fluxes has evolved. The results of various studies that quantified global N2O emissions from coastal upwelling areas, continental shelves, estuaries, and rivers suggest that these coastal areas contribute 0.3–6.6 TgN (N2O) a−1 or 7–61% of the total oceanic emissions [Bange et al., 1996; Nevison et al., 2004b; Kroeze et al., 2005]. Using inverse methods and the AGAGE measurements at Mace Head, Ireland, Manning et al. [2003] estimated European Union emissions of 0.9 ± 0.2 TgN (N2O) a−1, which agrees well with the United Nations FCCC inventory (0.8 ± 0.3 TgN (N2O) a−1, see www.unfccc.int). Melillo et al. [2001] provided evidence from Brazilian land-use sequences that the conversion of tropical forest to pasture leads to an initial increase but a later decline in emissions of N2O relative to the original forest. They also deduced that Brazilian forest soils alone contribute about 10% of total global N2O emission. Estimates of N2O sources and sinks using observations and inverse methods had earlier implied that a large fraction of global N2O emissions in 1978–1988 were tropical: specifically 20–29% in 0°–30°S and 32–39% in 0°–30°N compared to 11–15% in 30°S–90°S and 22–34% in 30°N–90°N [Prinn et al., 1990]. These estimates were uncertain due to their significant sensitivity to assumed troposphere-stratosphere exchange rates that strongly influence interhemispheric gradients. A more recent inverse study by Hirsch et al. [2006] estimated significantly lower emissions in 30°S–90°S (0–4%) and higher emissions in 0°–30°N (50–64%) than Prinn et al. [1990] during 1998–2001, with 26–36% from the oceans suggesting possible long-term variations in emissions.

[4] The stratosphere is also proposed to play an important role in the seasonal cycles of N2O [Nevison et al., 2004a]. For example, its well-defined seasonal cycle in the Southern Hemisphere has been interpreted as resulting from the net effect of seasonal oceanic outgassing of microbially produced N2O, stratospheric intrusion of low-N2O air, and the N2O thermally driven flux which is associated with its changing solubility in warm and cool water [Nevison et al., 2005]. These authors also estimated a southern ocean (30°–90°S) source of 0.9 TgN (N2O) a–1 or about 5% of the global total. The complex seasonal cycle in the Northern Hemisphere is more difficult to reconcile with seasonal variations in the northern latitude soil sources and stratospheric intrusions [Prinn et al., 2000; Liao et al., 2004]. The destruction of N2O in the stratosphere causes enrichment of its heavier isotopomers and isotopologues providing a potential method to differentiate stratospheric and surface flux influences on tropospheric N2O [Morgan et al., 2004].

[5] There are many surface measurements of N2O available both in time and space, from several networks but previous work has used only subsets of these data. In this paper, we utilize all the available intercalibrated N2O data from these networks around the world, together with a three dimensional global transport model, and an optimal inverse method to address the following scientific questions: (1) Are the current N2O ground measurement networks sufficient to resolve N2O sources at regional scale (e.g., in eight land and four ocean regions)? (2) Has the geographic pattern of N2O emissions changed between the 1978–1988 period examined by Prinn et al. [1990] and the 1998–2001 period examined by Hirsch et al. [2006]? (3) What fraction of global N2O emissions is attributable to the oceanic source?

2. Observations and Calibration

[6] In this study, we use measurements of N2O from four research groups (ALE/GAGE/AGAGE, NOAA/GMD/CCGG, NOAA/GMD/HATS and CSIRO), which include both in situ and flask data. In the ALE/GAGE/AGAGE global network program, continuous high frequency gas chromatographic (with electron capture detection, ECD) measurements of nitrous oxide are carried out at five globally distributed sites from 1978 to the present [Prinn et al., 2000; data archive at http://cdiac.ornl.gov/ndps/alegage.html]. AGAGE uses the SIO-98 absolute calibration scale. The current station locations are Cape Grim, Tasmania (CGO), Cape Matatula, American Samoa (SMO), Ragged Point, Barbados (RPB), Mace Head, Ireland (MHD), and Trinidad Head, California (THD). Stations also previously existed at Cape Meares, Oregon (1979–1989), and Adrigole, Ireland (1978–1983). The current Mace Head station replaced the Adrigole station and the station at Trinidad Head replaced the Cape Meares station.

[7] The NOAA Earth System Research Laboratory, Global Monitoring Division (GMD) Carbon Cycle and Greenhouse Gases (CCGG) group measures N2O in discrete flask samples collected in pairs approximately weekly at 48 sites in its global cooperative air sampling network (see Dlugokencky et al. [1994] for sampling details and http://www.esrl.noaa.gov/gmd/dv/site/site_table.html for information on the network). Samples are collected in 2.5 L borosilicate glass flasks fitted with two Teflon O-ring sealed stopcocks by first flushing the flasks and then pressurizing them to ∼0.25 atm overpressure with a portable sampler. Flasks are returned to GMD in Boulder, Colorado, where they are analyzed for N2O by gas chromatography with electron capture detection (ECD). Repeatability of the analytical system is 0.2 ppb (1σ), but the average agreement between members of pairs collected simultaneously is 0.4 ppb (1σ). Measurements are made relative to the NOAA2000 N2O standard scale. On the basis of an intercomparison of measurements from joint AGAGE/CCGG sites, the average difference, AGAGE–CCGG, is 0.15 ± 0.09 ppb (1σ).

[8] The GMD/HATS (Halocarbons and other Atmospheric Trace Species Group) program has both flask and in situ measurements of N2O over the globe (ftp://ftp.cmdl.noaa.gov/hats/n2o). The flask program involves the analysis of air samples collected at weekly to monthly intervals from Point Barrow, Alaska, USA (BRW); Alert, Northwest Territories, Canada (ALT); Mauna Loa, Hawaii, USA (MLO); American Samoa (SMO); South Pole (SPO); Cape Kumukahi, Hawaii, USA (KUM); and Cape Grim, Tasmania, Australia (CGO). The GMD/RITS (Radiatively Important Trace Species) program began measuring N2O in the atmosphere using in situ gas chromatographs at all four GMD baseline observatories (BRW, MLO, SPO and SMO) and at Niwot Ridge, Colorado, USA (NWR) in collaboration with the University of Colorado from late 1980s to present (ftp://ftp.cmdl.noaa.gov/hats/n2o/insituGCs/RITS/). Beginning in 2000, RITS was replaced by CATS (Chromatograph for Atmospheric Trace Species), which uses a four channel gas chromatograph (ftp://ftp.cmdl.noaa.gov/hats/n2o/insituGCs/CATS/).

[9] In addition to ALE/GAGE/AGAGE and GMD data, we also use flask measurements from CSIRO (the Commonwealth Scientific and Industrial Research Organization) which has five collection sites in the Southern Hemisphere (CFA, Cape Ferguson, Australia; MAA, Mawson, Antarctica; CGA, Cape Grim, Australia; SPU, South Pole, Antarctica; and MQA, Macquarie Island, Australia), and five northern hemisphere sites including one at Cape Rama, India which provides valuable information for estimating sources from South Asia (see Table 4).

[10] The measurements from each group are based on different absolute calibration scales. To make the comparison between observations and model calculations consistent, we have used the AGAGE calibration (SIO-98 scale) as a reference and rescaled the measurements from the other groups to this standard. This is a matter of convenience and is not intended to imply that other network standards are not equally valid. Ideally, this rescaling requires both direct intercomparison of the standards between each group, and comparisons of atmospheric N2O data at those sites common to more than one group. The strategy to link the GMD data with the AGAGE data: (1) uses those measurements from the GMD sites which are in common with AGAGE sites, (2) calculates the average (over space) ratio of the measurements as a linear function of time, t, between the two networks, i.e., a linear fit (a + bt) to the ratios of CCGG data to AGAGE at the common sites, and then, (3) applies this time varying ratio to the other GMD sites. Table 1 shows the coefficients (a, b) for the two time periods. The differences between AGAGE and GMD networks are decreasing in time but the exact reasons are not known.

Table 1. Coefficients of the Linear Time Variations in the Scaling Factors Linking AGAGE and the Other Three Networksa
 1997–20012002–2005
abab
  • a

    The HATS refers to the RITS and CATS in situ data but not the flask data.

CCGG/AGAGE0.38313.0807E–040.84117.9063E–05
HATS/AGAGE0.30603.4678E–040.68941.5511E–04
CSIRO/AGAGE1.001701.00170

[11] For the intercomparison between GMD/HATS flask data and AGAGE measurements, the ratio obtained is from the two Southern Hemispheric sites, i.e., SMO (SAM) and CGO (TAS). The averaged scaling factors range from 0.9987 to 1.0001, which is about 0.02–0.40 ppb at the relevant atmospheric N2O mole fractions. When it is applied to the NOAA-HATS Northern Hemispheric stations, i.e., ALT, BRW and KUM, the rescaled HATS flask data are almost 1 ppb lower than the NOAA-CCGG data at the same location. Considering the interhemispheric gradient of N2O is only about 0.7 ppb and the very long lifetime of this species, the 1 ppb difference in mole fraction could be significant in terms of emission estimates. Therefore we decided not to use the HATS flask measurements in this study. We do use the HATS-RITS and HATS-CATS in situ data which do not have this calibration issue.

[12] The well determined relationship between the AGAGE N2O calibration and the N2O calibration used at CSIRO is based upon both the intercomparisons of both standards and the atmospheric N2O data at Cape Grim. A constant ratio of 1.0017 between CSIRO and AGAGE data is used in this study. The analytical method for N2O measurements at CSIRO, the calibration scale, and an assessment of the N2O measurement uncertainties is provided by Francey et al. [2003].

[13] In order to show the interhemispheric gradients of N2O as well as the differences among different networks, we plot the time averaged mixing ratios (after calibration scaling) as a function of latitude for 1997–2001 and 2002–2005 in Figure 1. As we can see, the interhemispheric gradient of N2O is rather small.

Figure 1.

Time averaged latitudinal gradients of the observed N2O mixing ratios (after scaling) from AGAGE (open circle), GMD/CCGG (plus), GMD/HATS (asterisk) and CSIRO (diamond) for 1997–2001 and 2002–2005, respectively.

3. Model Description

[14] The Model of Atmospheric Transport and Chemistry (MATCH) [Mahowald et al., 1997a, 1997b; Rasch et al., 1997; Lawrence et al., 1999] is an offline transport model that uses meteorological fields derived from forecast center analyses. MATCH has been successfully driven by data from the National Centers for Environmental Prediction (NCEP), the European Center for Medium-Range Weather Forecasts (ECMWF) and the GSFC/NASA Data Assimilation Office (DAO) analysis [Rasch et al., 1997]. MATCHcan be configured with either the semi-Lagrangian [Rasch and Williamson, 1990] or the flux-corrected advection scheme (SPITFIRE) [Rasch and Lawrence, 1998]. Subgrid mixing processes, which include dry convective mixing, moist convective mixing and large-scale precipitation processes, are computed in the model. MATCH can be used at a horizontal resolution as fine as T62 (1.8° × 1.8°), with either 42 or 28 levels in the vertical. MATCH has been used for studies of transport in the stratosphere [Rasch et al., 1994; Waugh et al., 1997] and troposphere [Rasch et al., 1997; Mahowald et al., 1997b], stratospheric chemistry [Rasch et al., 1995], source inversions for CFC-11 [Mahowald et al., 1997a], CO2 [Gurney et al., 2002, 2003, 2005; Law et al., 2003; Chen, 2004; Dargaville et al., 2002], tropospheric chemistry with nonmethane hydrocarbons [Lawrence et al., 1999; von Kuhlmann et al., 2003], constraining OH levels using 14CO [Jockel et al., 2002], modeling dimethylsulfide [Lucas and Prinn, 2003, 2005], simulating aerosols [Mahowald et al., 2002; Collins et al., 2001], and most recently in estimating monthly methane (CH4) emissions from around the world [Chen and Prinn, 2005, 2006]. The ability of MATCH to accurately simulate the effects of transport on long-lived trace gases is well illustrated by the latter methane simulations.

[15] In this paper we use MATCH version 4.2 with T62 (192 longitudes by 94 latitudes surface grid and 28 vertical levels up to 2.7 hpa) resolution. This version of MATCH was used with the National Center for Environmental Prediction (NCEP) assimilated winds for the 10 year period from 1 July 1996 until 30 June 2006. We then compare model results to actual measurements of N2O taken at 65 sites around the surface of the Earth (Table 4). In addition to the inversion of data to define emissions from the twelve basis regions, we also estimate the global and semihemispheric fluxes in two separate inversions, in order to compare with Prinn et al. [1990] and Hirsch et al. [2006]. Specifically, a single factor is included in the state vector for the time averaged global inversion and 4 factors are included in the state vector for the time averaged semihemispheric inversion. In both cases, the normalized GEIA emission distributions in the relevant subdivisions are assumed to be correct.

[16] The estimations of N2O emissions were broken into two time periods: from 1997 to 2001 and from 2002 to 2005. The reasons to do that are two fold: first, we wanted one time period to include the large 1997–1998 El Nino episode and the other to be free of large El Nino events, in order to see if El Nino affects the estimations of N2O flux in the tropical oceans; second, we wanted to compare our results with the previously published research by Hirsch et al. [2006] for the period from 1998 to 2001.

[17] In this paper, we do not attempt to estimate the seasonal N2O flux. The amplitude of the N2O seasonal cycle is about 0.4 and 0.3 ppb in the Northern and Southern hemispheres, respectively [Prinn et al., 2000; Nevison et al., 2004a, 2004b]. Compared to its 1996–2006 average background concentration of about 317 ppb, this amplitude is very small. The N2O seasonal cycle is due to the seasonality in both its surface sources and its transport between the Northern and Southern hemispheres and between the troposphere and stratosphere. The latter transport brings down N2O-poor air to the troposphere from the stratosphere at high latitudes. Therefore to simulate the observed seasonal cycle of N2O well, one needs in particular to have a very accurate simulation of the month-by-month exchange between troposphere and stratosphere in the transport model. However, according to our evaluation, MATCH still needs improvement regarding this issue. For example, the model is out of phase with the observed mole fractions at most Southern midlatitude sites which we diagnose as due primarily to troposphere-stratosphere transport inaccuracies. Thus we decided to use the 12-month running means of both the observed and simulated mole fractions in the annual emission estimation. That is also why we do not attempt to estimate seasonal source variations. We also do not attempt to estimate the year by year N2O emissions, again because of lack of knowledge of the interannual changes in troposphere-stratosphere exchange.

3.1. Initialization

[18] For the first time period (1997–2001), the model runs were initialized on 30 June 1996, using the N2O surface latitudinal distribution as given by the 5 AGAGE sites. For initializing the vertical N2O profile we used results from runs of a previous 3D low-resolution spectral model [Golombek and Prinn, 1986; Prinn et al., 1990, 1999]. This model also supplied the necessary stratospheric loss frequencies photodissociation plus reaction with O(1D)) for computing the seasonally and latitudinally dependent N2O stratospheric sink. These loss frequencies were adjusted so that MATCH had an atmospheric annual average lifetime for N2O of 125 years from the above model. The effects of the significant uncertainty in this lifetime are specifically addressed in our error analysis (section 4.3). Stratospheric destruction is the only known atmospheric sink for N2O. MATCH was subsequently run for four years with constant emissions (see next section) to ensure independence from its above initial condition as far as spatial patterns are concerned. The resulting N2O global average mole fraction was then scaled back to its 1 July 1996 values as measured by the five AGAGE stations to yield the initial state used.

[19] This initialization was done in order for the model N2O mole fraction fields to stabilize from their original latitudinal distribution (where all mole fractions along each longitude circle are equal) to a full latitude – longitude detailed distribution as calculated by the MATCH model, using the assimilated NCEP winds. For the second time period (2002 to 2005), the model was initialized from its N2O mole fraction fields as of the end of 30 June 2001 and no adjustments were necessary, only the total source strength at N2O was reduced by ∼6% as compared to the first period (1 July 1996 to 30 June 2001). This is discussed later in detail in section 4.3.1.

3.2. Reference Runs

[20] With this new initial state the MATCH model was run for six years, from 1 July 1996 to 30 June 2002 to produce the “reference run” for subsequent inverse modeling for the first time period. The N2O source distributions for this reference run were taken from the Global Emissions Inventory Activity (GEIA) emission inventory [Bouwman et al., 1995] for 1990. For the second time period inversion, the optimally estimated N2O emissions from the first time period inversion were used as the reference emissions. However, within each source region (see next section), the relative distributions are kept the same as in the GEIA emissions.

[21] The GEIA inventory contains emissions from nine processes: (1) soils under natural vegetation and fertilized agricultural fields, (2) animal excreta, (3) postforest clearing enhanced emissions, (4) ocean, (5) industrial, (6) fossil fuel burning, (7) biofuel burning, (8) agricultural waste burning, and (9) biomass wasting. The GEIA files include global maps for each of these nine processes at 1° × 1° grid resolution, as described by Bouwman et al. [1995]. Because we compare 12-month running mean mole fractions from both MATCH and observations (see section 4.1), these N2O source distributions have no prescribed seasonal dependence. The total oceanic source is about 28% of the global total. The total source strength for N2O used in the reference run was 16.4 TgN (N2O) a−1 (i.e., the GEIA sources were scaled up uniformly from 13.5 TgN a−1) which is what is required for MATCH to yield for N2O the observed global surface average rate of increase of about 0.25% per year, and interhemispheric gradient of about 1 ppb across the equator (higher in the Northern hemisphere). The model reference run successfully balances the above total surface emissions with the stratospheric photochemical sink plus the observed atmospheric accumulation. We specifically compare MATCH model calculations at each of the 65 measuring sites, averaged over the 4 lowest model layers, i.e., an average over a tropospheric layer about 715 m thick from the surface upwards. This is because of prior evidence that the lowest model layers in MATCH are not sufficiently well mixed with the layers above [Chen and Prinn, 2005].

3.3. Source Regions and Sensitivity Runs

[22] The surface of the Earth was divided into eight land regions and four ocean regions for the purpose of running the sensitivity tests for the inversion (Figure 2). From the GEIA emissions, the eight land regions chosen were: North America (9.0% of global source), South America (18.8%), Europe (4.2%), North and West Asia (2.4%), China and Japan and Southeast Asia (14.5%), South Asia (6.2%), Africa (14.8%), and Australia and New-Zealand (2.0%). The four ocean regions were: Oceans 90°S–30°S (12.4%), Oceans 30°S–0° (3.1%), Oceans 0°–30°N (7.0%), and Oceans 30°N–90°N (5.7%). The numbers in parentheses are percentages of the global total in each region according to Bouwman et al. [1995] and represent our assumed a priori estimates (with generous ± 100% assumed for a priori errors) to be improved in the subsequent inversions. Tropical soils are probably the single most important source of N2O to the atmosphere [Prather et al., 2001], and temperate soils add another important portion to the atmospheric N2O budget. Therefore it is reasonable that South America, Africa, and China/Japan/Southeast Asia are assumed to be the top three N2O-emitting regions over the globe. Oceans south of 30°S are the most important oceanic N2O source due to the large ocean area.

Figure 2.

Eight land regions and four semihemispheric ocean regions whose emissions are estimated in this study. The measurement sites listed in Table 4 are also plotted here.

[23] In each of the 12 sensitivity runs, the overall source strength in one of the 12 regions was doubled and the changes in N2O mole fractions at the 65 sites were computed. The sensitivity runs are used to compute the 65 × 12 time varying H matrix in section 4.1, and the elements of H show how sensitive each site is to emission changes in each of the 12 regions. The relative surface source distributions within each region were not changed in these sensitivity runs from those in the reference run. Only the total source strength (not the spatial distribution) in each region was changed. We show the sensitivities of 12 of the 65 sites to changes in emissions from the twelve source regions in Figure 3 for the 1997 to 2001 time period. These sites are chosen because they show the strongest sensitivity among all the stations to the nearby relevant source region. Note that the source region that a site is the most sensitive to does not necessarily determine the accumulated change in mole fraction at that station, especially if the strength of that source region is relatively small.

Figure 3.

Partial derivatives of changes in site mole fractions with respect to changes in regional emissions at 12 illustrative sites (expressed as ppb changes per fractional change in regional emissions and contained in the Kalman filter H matrix). In each subplot, the colored lines represent the sensitivities of that site to emissions from the 12 source regions. The site name is given at the top left corner, and the dominant source region affecting that site is given in color at the bottom right corner (thus also defining the color code for the lines in all subplots).

4. Inversion Method and Results

4.1. MATCH Inversion Methodology and Errors

[24] In this study, a recursive weighted least squares filter is used together with MATCH to optimally estimate the emissions from the 12 basis regions as defined above. The recursive weighted least squares inversion is the simplest form of the Kalman filter, and is effectively the same as a Bayesian bulk (all data) inversion but with the advantage of showing the value of each piece of added data.

[25] Estimates of the unknowns, i.e., the twelve time-averaged dimensionless emission factors which multiply the a priori percentage emissions prescribed in the GEIA emission map, are contained in the state vector x, which has a dimension of 12 by 1. As described by Kaminski et al. [2001], prescribing the intraregional spatial patterns in the 12 large regions in the inversions may lead to possible biases (so-called aggregation error), if the patterns are wrong. Our results are therefore conditional on the GEIA patterns being reasonable. The squares of the estimation errors associated with the state vector comprise the diagonal elements of the state estimation error covariance matrix P (12 by 12 for the regional emission inversion). These arrays are updated with each new month of data using

equation image
equation image

Here, y0 and y are vectors (65 by 1, where 65 is the number of measurement sites) containing the mole fractions observed at each site and calculated by MATCH, respectively. The postscripts (−) and (+) denote values of P and x before and after use of each month's data. The matrix H (65 by 12) (and its transpose HT) contains the model-calculated partial derivatives of the elements of y with respect to the elements of x (see section 3.3). The off-diagonal elements in matrix P represent the correlations between the errors in the state variables contained in vector x. These correlations will be examined in detail in section 4.3.3 when the regional N2O emission estimations are discussed. In equations (1) and (2), K (12 by 65) is the Kalman gain matrix given by

equation image

More details about the recursive weighted least squares and Kalman filter techniques used are given by Prinn [2000].

[26] The time varying matrix R (65 by 65) is assumed to be diagonal with its diagonal elements being the variances (σk2) associated with the observed monthly mean χ values at station k (contained in vector y0). In this study, we include three types of observational errors associated with: (1) instrumental precision errors; (2) the frequency of sampling used to define the monthly mean; and (3), the mismatch error between observations and model [Chen, 2004; Chen and Prinn, 2005, 2006]. The precision of individual nitrous oxide measurements is about the same among the different observing groups and for in situ and flask measurements. We take a uniform instrumental precision error of 0.2 ppb at all sites, except for HATS-RITS data which are reported to have a larger precision error of 0.5–0.8 ppb. The effect of assuming a 0.4 ppb instead of a 0.2 ppb for the NOAA-CCGG flask data (see section 2) were also investigated (see section 4.3.3). The sampling frequency error accounts for how well the monthly means are defined given a small number of measurements within a month. Following Chen [2004]σsamplingfrequency is defined as equation image, where σmon2 is the monthly mean variance and n is the number of measurements during a month. Here the monthly mean variance is calculated from the MATCH reference run, as high frequency measurements are not available at every site. This is justified since the high frequency variability for N2O in the MATCH simulations is very similar to that observed at most high frequency measurement sites. Flask measurements are usually taken once each week and at most 4 times per week, while the number of in situ measurements is from 400 to 1000 each month. Therefore the sampling frequency errors for flask measurements are about 5–16 times larger than for high frequency measurements. Thus the monthly averages based on the in situ measurements will be given significantly more weights than the flask measurements in the inversion. We also account for the mismatch errors in the matrix R. Mismatch errors occur because we are comparing the point measurements with volume-averaged mole fractions at the model grids [Prinn, 2000]. Following Chen and Prinn [2005], we define the mismatch errors as the average differences between the values at a particular grid point and its four surrounding grid points (at the same vertical level) in the MATCH reference run. In general, the mismatch errors are larger over the strongly emitting continental areas and smaller over the remote marine areas.

4.2. Other Modeling Uncertainties

[27] In the inversion process discussed above, we assume a perfect model, i.e., the model errors other than the mismatch errors are not included in the inversion. With 60 months of data in the first time period and 48 months in the second time period, one should expect the posterior errors associated with the multiyear average estimated N2O fluxes to become successively smaller as data is added during the inversion [Prinn, 2000]. This is true for the 12 regional inversions discussed in section 4.3.3 and even more so for the global and semihemispheric inversions discussed in section 4.3.1 and section 4.3.2.

[28] However, all existing transport models including MATCH are far from perfect. The small posterior errors obtained from the inversions do not represent the true errors of the estimated N2O flux. Thus it is particularly important to assess and add in the uncertainties due to these model errors. The most significant source of model error comes from the stratosphere-troposphere exchange (STE) rate calculated in the model [Prinn et al., 1990]. Above 50 hPa, N2O is destroyed photochemically, and then N2O-poor air is brought back to the troposphere through large scale atmosphere circulation. This effect on the tropospheric N2O concentration is almost as important as its surface flux [Prinn et al., 1990]. However, it is not only technically very difficult but also very time consuming to address the influence of the STE uncertainties on estimated fluxes using multiple inverse runs (with differing STE rates) of a three dimensional global transport model, such as MATCH, at a high resolution (T62 in this paper). Therefore we use a separate 12-box model [Prinn et al., 2000] with flexible STE rates (as well as other key parameters) to estimate the uncertainty in fluxes coming from STE uncertainty.

[29] The 12 box model divides the globe into 90°S–30°S, 0°–30°S, 0°–30°N, and 30°N–90°N regions horizontally, and 1000–500 hPa, 500–200 hPa, and 200–0 hPa regions vertically. The STE are explicitly specified in the model, and the ratio of the STE rates in the Northern and Southern Hemispheres is about 1.8 based on previous studies [Prinn et al., 1990]. To make this simple model more representative of MATCH, we first estimate the surface emissions of the four semihemispheres using MATCH and the aforementioned inverse method. The total N2O sources used in the 12 box model is 16.3 TgN/a (1997–2001) and 15.4 TgN/a (2002–2005), which are the same as those estimated in the inversions using the MATCH model. With these best estimates of surface fluxes from MATCH, we then derive the STE rates in both hemispheres inversely using the 12-box model with the observations mentioned in section 2. We found that these newly estimated STE rates give a NH/SH ratio of 2.1 (1997–2001) and 1.7 (2002–2005) instead of the value of 1.8 from Prinn et al. [1990]. This is largely because our estimated N2O surface emissions in the semihemispheres for year 1997–2001 and 2002–2005 are somewhat different from those estimated in Prinn et al. [1990], and we discuss this in more detail in the next section.

[30] Using this reference version of the 12-box model, we then use a Monte Carlo method to estimate the uncertainties in the hemispheric N2O surface emissions due to uncertainties in the following model parameters: exchange time between troposphere and stratosphere (±50%), N2O stratospheric lifetimes (±24 years), and horizontal (±30%) and vertical (±50%) transport rates in the troposphere. The numbers in the parentheses define the 1σ uncertainty range for each parameter. We assume that these parameters are normally distributed and take 10,000 randomly chosen equal-probability samples of each of the above uncertain model parameters. We then estimate the hemispheric surface emissions of N2O using the inverse method for each randomly chosen case. From the two 10,000 Monte Carlo runs, we conclude that the estimated 66% uncertainty range in surface emissions is [−22%, +34%] (1997–2001) and [–24%, +36%] (2002–2005) in the Southern hemisphere (SH), and [−16%, +14%] (1997–2001) and [−17%, +15%] (2002–2005) in the Northern hemisphere (NH). We also produced two other Monte Carlo ensembles in which we only treat the STE times as uncertain variables. This study showed that the 66% uncertainty range is [−17%, +27%] (1997–2001) and [−19%, +29%] (2002–2005) in SH emissions and [−8%, +9%] (1997–2001) and [−9%, +10%] (2002–2005) in NH emissions. This clearly shows again that the STE rate plays a very important role in estimating N2O surface sources from tropospheric observations. The significantly larger error in the Southern Hemisphere occurs because N2O distributions in the Northern Hemisphere are determined more by its surface sources, while the distributions in the Southern Hemisphere are dominated by downward transport from the stratosphere and interhemispheric transport in the troposphere.

4.3. Inversion Results

4.3.1. Global Total Flux

[31] Using the GEIA emission distribution map with a 125 year lifetime, we optimally derive a global total flux of N2O of 16.3 TgN a−1, with an a posteriori 1σ uncertainty of 0.02 TgN a−1 (c.f. the a priori uncertainty of ±16.4 TgN a−1) for 1997 to 2001. The optimally estimated global total flux agrees very well with the one used in the reference case (i.e., 16.4 TgN a−1). As discussed in section 4.2, the small a posteriori uncertainty is due to the assumption of a perfect model in the inversion, as well as to the fact that, when estimating multiyear averages, there are a very large number of N2O measurements available in time and space over the globe. A sensitivity study, which includes the modeling uncertainties discussed in section 4.2, expands the N2O 1σ global total flux range to 16.3−1.2+1.5 TgN a−1 with the dominant contribution coming from the uncertainty in STE. This result also agrees fairly well with Hirsch et al. [2006] for 1998–2001. Their estimated global total source is 15.2–20.4 TgN a−1 assuming a 122 year lifetime which is well within the 125 ± 24 year 1σ lifetime range assumed in our model error analysis.

[32] Following the same procedure for the second time period, a global mean N2O flux of 15.4−1.3+1.7 TgN a−1 is derived. This is about a 5.5% decrease from the 1997–2001 time periods, although the two global mean fluxes are obviously not statistically different from each other.

[33] In addition to total global emissions, we also estimate the ratio of NH to SH emissions by solving for them separately. The optimally estimated ratio from the MATCH inversion for 1997–2001 is 2.1 compared to 1.5 in the reference GEIA emission map. Using the aforementioned 12-box model with the 10,000 member ensemble Monte Carlo uncertainty analysis, the 1σ NH to SH ratio ranges from 1.6 to 2.7, with the lower end agreeing well with the previously reported range of 1.5–2.0 from Prinn et al. [1990] and Butler et al. [1989] for the pre-1990 emissions. The inferred increased Northern Hemispheric N2O emissions in 1997–2001 compared to pre-1990 could be due to increased agricultural fertilizer use and warming soils in the Northern tropics. For the second time period, the increase of Northern Hemispheric emission is getting even larger, as the derived ratio of NH to SH emissions of N2O is now 2.7−0.6+0.8, i.e., 2.1 to 3.5.

4.3.2. Semihemispheric Fluxes

[34] As summarized in Figure 4, we estimated N2O surface emissions in the four semihemispheres for 1997 to 2001 and 2002 to 2005, for comparison to previous studies by Prinn et al. [1990], Bouwman et al. [1995], and Hirsch et al. [2006]. The a priori flux from GEIA [Bouwman et al., 1995] shown in Figure 4 has our arbitrarily assumed 100% initial uncertainty, while our best estimated flux from the MATCH inversion has its total 66% uncertainty estimation from the P matrix plus the 12 box model Monte Carlo uncertainty study results. We also summarize these results in Table 2. We find that our estimated fluxes in the 30°S–90°S region for both time periods are significantly lower than the 1990 GEIA estimate, with the flux for the second time period being smaller than the first time period. On the other hand our estimated fluxes in the 0°–30°N region are about 50% higher than the 1990 GEIA estimate. In the 0°–30°S region, the estimates from this study are similar to the GEIA estimate, while in the 30°N–90°N region, they are about 25% lower than the GEIA estimate. The a posteriori uncertainties from the Kalman P matrix decreased significantly from the ±100% a priori uncertainties. In fact, they are too small to be identified in Figure 4 (0.4%, 1.2%, 1.1% and 0.3% in the four boxes going from south to north). Again, as discussed earlier in the paper, the small a posteriori uncertainties are due to the assumption of a perfect model in the inversion, as well as to the fact that there are a large number of N2O measurements available for multiyear flux estimates for the semihemispheric inversions.

Figure 4.

Results from the four semihemispheric emission estimations expressed as percentages of the global flux. The a priori flux from GEIA [Bouwman et al., 1995] is in blue, with assumed 100% initial uncertainty; our best estimated flux from the MATCH inversions for 1997–2001 and 2002–2005 are in red and pink respectively. The 1σ (66% probability) error bars combine the Kalman filter and modeling errors. The corresponding results from Prinn et al. [1990] are in green and from Hirsch et al. [2006] are in black.

Table 2. Percentages of the Global N2O Surface Flux (TgN (N2O) a−1) in the Four Semihemispheres
StudyPeriod30°S–90°S0–30°S0–30°N30°N–90°N
Bouwman et al. [1995, GEIA]199014263426
This study1997–20010–722–3646–5615–23
This study2002–20050–322–4140–5617–25
Prinn et al. [1990]1978–198811–1520–2932–3922–34
Hirsch et al. [2006]1998–20010–416–3250–6416–23

[35] In comparison with Prinn et al. [1990], our study shows smaller surface emissions in the 30°S–90°S and 30°N–90°N regions, significantly higher fluxes in the 0°–30°N region and similar fluxes in the 0°–30°S region. These differences may be partly caused by the much fewer measurement stations used in the earlier study. They could also be due to the different structure, resolution, and transport in the 3D MATCH and 2D box inversion models used. In comparison with Hirsch et al. [2006], we see reasonably good agreement between their study and ours in the adjustments needed to the GEIA estimates, i.e., large changes needed in surface fluxes in the 90°S to 30°S and equator to 30°N regions, negligible adjustments in the 0°–30°S region and relatively small decreases in the 90°N–30°N region. Considering the fact that Hirsch et al. [2006] used N2O measurements only from GMD/CCGG for the 1998 to 2001 time period and the inversion techniques and transport models used in the two studies differ as well, these good agreements may indicate that the relative contributions of regional N2O surface fluxes to the global total may indeed have changed significantly from the GEIA 1990 estimates.

4.3.3. Regional Fluxes

[36] Figures 4a and 4b show how the MATCH inversion (1997 to 2001) state vector, which contains the estimated relative contributions to the global total N2O from the eight land and four ocean regions (in fractional terms), evolves as each set of measurements is utilized for the 1997 to 2001 time period. The assumed a priori 1σ uncertainty for each state vector element is usually 100% of the a priori flux. The exceptions are for the 0°–30°N and 0°–30°S Ocean region, where a 200% initial uncertainty is needed to ensure that the final estimated flux falls within the 1σ range of the a priori guess. The largest corrections to the state vector and most of the uncertainty reductions occur, when the first several measurements are used. Most of the state variables converge very well as shown in these two figures. However, we do see some unsettled behavior for the estimated fluxes in the 0°–30°S Ocean and South America regions. These two regions are apparently not sufficiently well enough sampled by the networks to provide accurate estimates for them individually. Hence their flux estimates are significantly anticorrelated to each other (Figures 5a and 5b). In fact, the negative correlation between the errors in the flux estimates of these two source regions, expressed in the appropriate off-diagonal elements of the error covariance matrix P (section 4.1), is the largest among all the error correlations (and it is much larger than the second largest correlation in the P matrix).

Figure 5a.

Convergence of estimated regional fluxes (expressed as fractions of the total global flux) with 1σ error bars (from the inverse method only) for 6 of the 8 land areas.

Figure 5b.

Same as Figure 5a but for the other 2 land areas and the 4 oceanic areas.

[37] To address this issue, the regional N2O emission estimations were repeated, with the 0°–30°S Ocean and South America regions aggregated into one source region. The strong negative correlation in the error covariance matrix is now removed. Also, the estimated error for the combined 0°–30°S Ocean and South America regions is smaller than the errors when they were estimated separately as expected. As noted in section 4.1, our aggregation procedure does assume that the relevant normalized GEIA maps are sufficiently accurate to avoid significant aggregation errors [Kaminski et al., 2001]. In the rest of the paper, the N2O emission estimates are therefore based on an eleven source region inversion for this time period. However, the 0°–30°S Ocean and South American regions are still reported separately, assuming that the ratio of the a priori GEIA emissions of N2O from these two source regions still holds. Please note that the ratio of the 0°–30°S Ocean and South American sources in the GEIA estimates may be in error due to poor knowledge of the Peruvian oceanic upwelling region.

[38] In Table 3a (1997 to 2001) and Table 3b (2002–2005) and in Figure 6, we compare the final results for our regional emission estimates with the GEIA estimates. The results reported in Table 3a are different from those shown in Figures 5a and 5b, because there is no aggregation of regions involved in Figures 5a and 5b. For the uncertainty reduction in the inversions for the 1997–2001 time periods assuming a perfect model, we can calculate from Table 3a the ratio of the a posteriori uncertainty from the inverse method to the initial guess of the state vector uncertainty for each region. Expressed as a percentage, this yields the following results: European Union (5%), China/Japan/Southeast Asia (8%), South Asia (11%), Australia/New Zealand (20%), North America (6%), South America and 0°–30°S Ocean (6%), North and West Asia (17%), Africa (9%), 30°S–90°S Ocean (2%), 0°–30°N Ocean (13%), and 30°N–90°N Ocean (5%). For land regions, Australia and New Zealand has the least uncertainty reduction, followed by North and West Asia. For the ocean regions, the error reduction is similar for each of the semihemispheric regions. However, when South America and 0°–30°S Ocean are not aggregated together, the error for the 0°–30°S Ocean is only 27% of its initial guess in the inversion. This shows again that this southern hemispheric region is relatively poorly sampled by the current N2O measurement networks. On the other hand, the European Union and the 30°S–90°S Ocean region are apparently relatively well sampled; hence their a posteriori uncertainties are reduced the most among all the regions. Overall, the error reduction is sufficient in all cases to conclude that the measurements contribute very significantly to lowering emission uncertainties.

Figure 6.

Results from the twelve regional emission estimations (expressed as percentages of the global flux) and their 1σ (66% probability) error bars. The fluxes from GEIA [Bouwman et al., 1995] are in blue with assumed 100% initial uncertainties. Our best estimated fluxes from the MATCH inversion are in red for 1997–2001 and black for 2002–2005. The thick error bars represent the error estimation from the inverse method only, while the thin error bars include both the inverse method and modeling errors.

Table 3a. Percentages of the Global Total N2O Surface Flux in the 12 Model Regions and Their Errors for 1997–2001a
 EUCN/JP SE AsiaSouth AsiaAUS/NZNorth AmericaSouth AmericaNW AsiaAfrica
Reference4.214.56.22.09.018.82.414.8
Errora priori4.214.56.22.09.018.82.414.8
Estimates3.314.87.00.18.522.52.018.4
ErrorKF0.21.20.70.40.61.10.41.4
Errormodel+0.6+2.9+0.9+0.7+1.3+5.8+0.3+3.4
−0.7−2.5−1.0−0.4−1.4−3.9−1.1−2.7
ErrorTotal+0.6+3.1+1.1+1.5+1.4+5.9+0.5+3.7
−0.7−2.8−1.2−1.4−1.5−4.1−1.1−3.0
 30°S–90°S Ocean0–30°S Ocean0–30°N Ocean30–90°N Ocean    
  • a

    The Kalman Filter (KF) and modeling errors are shown separately and combined. All the quoted errors are 1σ or 66% probability. Reference emissions are from the GEIA estimates.

Reference12.43.17.05.7    
Errora priori12.43.114.05.7    
Estimates1.53.716.12.2    
ErrorKF0.30.20.90.3    
Errormodel+4.2+1.1+1.0+0.8    
−2.7−0.7−1.1−0.9    
ErrorTotal+4.2+1.1+1.3+0.9    
−2.7−0.7−1.4−1.0    

[39] Table 3a also compares the contributions to the total estimation uncertainties from measurements (from Kalman Filter P matrix) and transport model (from multiple 12-box model inversions). In general, the emission uncertainties derived from the uncertainties in the chemical transport model are significantly larger than the errors derived from the measurements. The modeling errors in the Southern Hemisphere are the largest (e.g., see South America and the Southern oceans). For North and West Asia, South Asia and the 0°–30°N ocean area, the uncertainties derived from the uncertainties in the measurements are comparable to the errors due to the transport models. These results show that the current N2O measurement networks are sampling most parts of the globe quite well under the assumption of a perfect model in the inversions. On the other hand, modeling uncertainties, especially those in tropospheric and stratospheric exchange rates, play a very important role in limiting the accuracy of N2O surface emission estimates as discussed in section 4.2.

[40] After inclusion of the significant modeling uncertainties estimated from the 12-box model, our inversion shows that emissions for 1997–2001 from most land regions do not differ statistically from the a priori guess (i.e., GEIA estimates) (Table 3a), despite a 13% increase in South Asia, a 17% decrease in North and West Asia, and a 20% increase in South America, respectively. The three regions where their emissions are statistically different from the GEIA estimates are Africa (24% increase), the European Union (21% decrease) and Australia and New Zealand (95% decrease). In contrast, three of the four oceanic regions are statistically quite different from the GEIA estimates (0°–30°S Ocean is the only exception). The northern tropical ocean shows a significant increase (130%) of N2O flux from the a priori guess, and there are also quite large decreases (61–88%) from the two extratropical ocean regions relative to the GEIA estimates.

[41] Table 3b shows the same regional inversions as Table 3a but for the 2002–2005 time periods. There are some differences in the inversion method used in this second time period compared to the first time period. First, the initial best guesses of the N2O emissions are assumed to be the estimated emissions from the 1997–2001 inversions (but the initial emission errors continue to be ±100%). Second, there are no more measurements available at CRI (Cape Rama, India measured by CSRIO) and NMB (Namibia, Africa measured by GMD/CCGG) stations for 2002–2005 (Table 4). From the inversions for 1997–2001, we learned that CRI and NMB sites turned out to be very important in estimating the emissions in the South Asia and Africa regions, respectively. Specifically, without the CRI and NMB sites, the estimated S. Asia source was about 74% larger and the African sources about 21% smaller than the 1997–2001 estimates obtained including these sites. Therefore these two source regions are aggregated into one source region in the inversions for this second time period. We also assume that the ratio of the a priori emissions of N2O from these two source regions (i.e., the ratio of their estimated emissions from 1997–2001) still holds. In the mean time, the high anticorrelations in the estimated errors of N2O surface fluxes between South America and 0–30°S oceans still exist. Thus these two regions are aggregated into one as was done in the first time period inversions. Third, a constraint of no negative sources in the southern oceans is applied in the inversions for the second time period by setting any negative estimates to zero during inversions. In this second period, even stronger decreases of N2O surface fluxes relative to GEIA are estimated in the 30°S–90°S region than in the first time period. Last, the time period is one year shorter for the 2002–2005 inversions than for the 1997–2001 inversions. Thus the estimated errors from the inverse method are expected to be larger than those from the first time period inversions.

Table 3b. Same as Table 2a but for 2002–2005, and With Reference (A Priori) Emissions Being the Optimally Estimated 1997–2001 Estimates
 EUCN/JP SE AsiaSouth AsiaAUS/NZNorth AmericaSouth AmericaNW AsiaAfrica
GEIA4.214.56.22.09.018.82.414.8
Reference3.314.87.00.18.522.52.018.4
Errora priori3.314.87.00.48.522.52.018.4
Estimates3.416.96.20.058.422.73.516.7
Error KF0.21.30.50.30.61.10.51.2
Error model+0.5+3.2+1.1±0.1+1.3+7.5+0.3+4.6
−0.6−2.8−1.2−1.5−5.2−0.63.7
ErrorTotal+0.5+3.5+1.1±0.3+1.4+7.6+0.5+4.7
−0.6−3.1−1.3−1.6−5.3−0.8−3.9
 30°S–90°S Ocean0–30°S Ocean0–30°N Ocean30–90°N Ocean    
GEIA12.43.17.05.7    
Reference1.53.716.12.2    
Errora priori1.53.716.12.2    
Estimates0.03.818.40.4    
ErrorKF0.30.21.00.4    
Errormodel+0.5+1.4+2.5+0.3    
−0.4−0.9−2.8−0.4    
ErrorTotal+0.6+1.4+2.6±0.5    
−0.5−0.9−2.9    
Table 4. Square Root of the Mean Square Residual (SRMSR) of N2O Mole Fractions (ppb) Using the a Priori (SRMSRref) and the Best Estimated Emissions (SRMSRinv) in MATCH at Each Site for the 1997–2001 and 2002–2005 Time Periodsa
SiteLatitudeLongitude1997–20012002–2005
equation imageSRMSRrefSRMSRinvequation imageSRMSRrefSRMSRinv
  • a

    Sites at high altitudes (whose steep mountain topography is not adequately resolved by MATCH at T62 resolution) and sites with short measurement time periods are not used in this study. Also shown are the averaged standard deviations from the measurements equation image.

  • b

    ALE/GAGE/AGAGE, real time. All others are GMD/CCGG flasks.

  • c

    CSIRO, flask. All others are GMD/CCGG flasks.

  • d

    GMD-HATS/RITS/CATS, real time.

HBA−75.6333.50.320.840.220.331.220.53
SYO−69.039.60.290.620.260.291.010.38
MAAd−67.762.90.320.840.250.280.730.19
PSA−64.9296.00.310.810.220.280.900.22
TDF−54.9291.50.280.760.200.301.040.45
CRZ−46.551.90.280.910.240.310.950.19
CGOb−40.4144.60.210.590.180.370.890.10
CGO−40.4144.60.300.590.200.451.090.23
CGAd−40.4144.60.270.590.220.310.760.23
EIC−27.2250.50.320.250.240.321.000.32
NMB−23.615.10.470.350.41N/AN/AN/A
CFAd−19.3147.10.350.410.150.360.610.22
SMOb−14.2189.40.280.090.130.370.760.19
SMO−14.2189.40.420.120.140.350.810.21
SMOc−14.2189.40.970.210.200.300.920.27
ASC−7.9345.50.330.120.150.340.840.18
SEY−4.755.20.390.280.160.420.680.32
CHR1.7202.80.310.920.450.310.610.17
GMI13.4144.80.300.510.150.330.670.08
RPBb13.2301.00.170.400.130.180.720.12
RPBd13.2301.00.290.430.160.350.680.15
CRId15.173.80.900.440.29N/AN/AN/A
KUM19.5205.20.310.470.180.340.750.18
KEY25.7279.80.430.250.460.581.290.68
MID28.2182.60.290.360.170.330.610.12
WIS31.134.90.450.520.300.390.380.43
BME32.3295.20.320.400.180.330.630.10
BMW32.3295.20.320.430.200.380.630.11
TAP36.7295.20.500.350.190.510.590.11
PTA39.0236.20.480.740.680.510.480.56
AZR38.8332.60.350.310.220.340.700.14
THDb41.1235.90.320.370.320.350.530.27
BSC44.228.70.650.320.350.630.280.43
KZD44.577.60.350.680.460.360.410.36
UUM44.5111.10.390.490.320.410.450.26
HUN47.016.70.620.470.510.610.500.23
LEF45.9269.70.510.420.490.500.990.32
EPCd49.4233.50.440.130.17N/AN/AN/A
SHM52.7174.10.360.200.180.350.730.13
MHDb53.3350.10.230.270.280.240.970.32
MHD53.3350.10.410.280.290.430.790.35
BAL55.417.10.600.260.270.610.670.35
CBA55.2197.30.350.260.230.390.920.17
SISd60.2358.80.460.130.08N/AN/AN/A
ICE63.3339.70.320.170.160.320.680.13
STM66.02.00.340.180.210.370.820.19
PAL68.024.1N/AN/AN/A0.380.630.14
BRW71.3203.40.360.360.410.380.890.29
BRWc71.3203.40.720.330.370.45 1.090.47
ZEP78.911.90.330.230.210.300.680.13
ALT82.5297.50.290.200.190.350.800.27
ALCd82.5297.50.280.150.140.280.690.18
POCN000.0205.00.410.580.320.390.650.11
POCN055.0209.00.300.580.190.350.730.11
POCN1010.0211.00.350.520.200.330.690.14
POCN1515.0215.00.290.320.260.320.750.15
POCN2020.0219.00.280.400.420.320.680.10
POCN2525.0221.00.300.250.130.290.650.08
POCN3030.0225.00.300.190.110.330.580.19
POCS05−5.0201.00.290.360.170.350.720.07
POCS10−10.0199.00.350.340.280.360.570.20
POCS15−15.0189.00.300.210.190.340.660.08
POCS20−20.0186.00.320.140.210.370.720.13
POCS25−25.0189.00.310.200.260.310.870.16
POCS30−30.0184.00.310.380.260.320.900.17

[42] From Table 3b, including both the modeling errors and the errors from the measurements, the emission estimates from most of the land regions do not differ statistically from their priors (i.e., from the 1997–2001 inversions), despite a 10% decrease from the South Asia and Africa regions, a 14% increase in the China/Japan/South East Asia region, and a 50% decrease from the already very minor source in the Australia and New Zealand region. The only statistically significant change occurs in the North and West Asia region, where a 75% increase is derived from the inversion for the second time period.

[43] When we assumed an instrumental precision of 0.4 ppb instead of 0.2 ppb for the NOAA CCGG flask data, the estimated regional fluxes were not statistically different from those reported here for all regions except the NW Asia even when model errors are neglected. Therefore we have expanded the errors on the NW Asia emission to include both estimates.

[44] For the oceanic source estimations, we see statistically significant decreases occurring in the high latitude oceans in both hemispheres (100% decrease in 30°S–90°S region and 82% decrease in 30°N–90°N region from the 1997–2001 inversions, respectively). The results in 30°N–90°N region differ statistically at the 1σ level, but not at the 2σ level. No experimental or process modeling flux studies are available to either verify or refute these estimated changes. For the two tropical oceanic regions, there is a small increase in the 0°–30°S region and a relatively large 13% increase in the 0°–30°N region. The relatively low fluxes in the tropical oceans for 1997–2001 could be due to the effect of the large El Nino episode from December 1997 to June 1998. During an El Nino event, the equatorial and coastal upwelling is reduced and hence the surface water supersaturation of N2O is lowered. This could lower the N2O flux by 80% from the equatorial tropical Pacific area (17°S–20°N) according to Butler et al. [1989].

[45] In addition, we also see there is a shift of N2O surface fluxes from the Southern Hemisphere to the Northern Hemisphere, compared to the estimations for the 1997–2001 time periods.

[46] Finally, for each Monte Carlo inversion we adjusted the 12-region inversion results from MATCH to account for the different NH and SH emission estimates from each Monte Carlo run. Then we aggregated the eight land regions together and the four ocean regions together to compute the land and ocean emission uncertainties for both time periods, to see if the N2O sources from the total ocean and total land areas are different from the 1990 GEIA emissions. We retain the MATCH inversion results but adopt these uncertainties from this procedure to conclude that 23 ± 4% of the global N2O flux is from the ocean and 77−12+16% is from the land, for both the 1997–2001 and 2002–2005 time periods. Note that the total emissions in each Monte Carlo run are different so the sum of these ocean and land percentages need not always be 100%. The a priori guess (i.e., GEIA emissions) had 28% from the ocean and 72% from the land sources. Our inversion errors come from both the measurement and the modeling errors as noted above. Hence we see a small shift from the oceanic source to the land source in our estimations relative to GEIA. Our oceanic source is also a little lower than that estimated by Hirsch et al. [2006]. They report 26% to 36% of global emissions coming from the oceanic source. However, including the modeling errors, the two results do not differ statistically.

4.3.4. Model-Measurement Comparisons

[47] An important indication of whether the inversion works correctly is to check if the differences between modeled concentrations and measurements are substantially decreased after the inversion (see Table 4). Globally, the square root of the mean square residual (SRMSR) of N2O decreases from 0.49 ppb to 0.26 ppb for 1997–2001 and from 0.78 ppb to 0.26 ppb for 2001–2005. In addition to SRMSR, the χ2 value (a goodness of fit statistics expressed as equation image) is another way of showing the improvement in the model simulations of the observations. The χ2 value before and after inversion for 1997–2001 is 1.8ppt vs. 0.7ppt and for 2002–2005 is 5.2ppt vs. 0.6ppt, respectively.

[48] While the global SRMSR decrease is significant, the decreases are even more significant if we pay attention to individual sites (Table 4). The relatively larger SRMSR of the reference case for the second time period is due to the fact that the assumed reference (initial) global mean N2O emissions were too large for the observed N2O mole fractions in this time period. In Table 4 we see that, for the 1997–2001 time period, there are 58 out of 64 sites with SRMSR values smaller than the averaged standard deviations from the measurements equation image using the best estimated emissions, compared with only 33 out of 64 sites using the a priori GEIA emissions. For the 2002–2005 time period, 53 out of 61 have SRMSR values smaller than equation image using the best estimated emissions versus 4 out of 61 sites using the estimated emissions from the first time period (that served as the a priori emissions for the second time period).

[49] In Figure 7a (1997–2001) and Figure 7b (2002–2005), we compare the actual residuals between the optimized model mole fractions and observations at twelve sites. The sites are chosen to show a representative subsample of the mole fraction changes that result from the optimized emissions. For 1997–2001, Figure 7a shows that agreement between measurements and model is significantly improved by the inversion at Tae-ahn, South Korea (TAP), Cape Ramba, India (CRI), Cape Furguson, Australia (CFA), Bermuda (BMU), Pacific Ocean between 0° and 5°, Palmer Station, Antarctica (PSA), and Guam (GUM). Some improvements occur at Hungary (HUN), Ulaan Uul, Mongolia (UUM), Ascension Island (ASC) and Azores (AZR); in fact, the reference run is already in good agreement with the observations at these four sites. Sary Taukum, Kazakhstan (KZD) is the site chosen to represent the 6 out of 64 sites with SRMSR values still larger than the mean standard deviations from the measurements after the inversions. For 2002–2005, Figure 7b shows that significant improvements occur at every chosen site after the inversions. This is due to the fact that the (initial reference) emissions used in this time period are too high as mentioned above. However, The SRMSR at Syowa station is still larger than the observed standard deviations after the inversions.

Figure 7a.

Residuals (ppb) between the MATCH model mole fractions (calculated using either the a priori (GEIA) or the optimally estimated regional emissions) and the measured mole fractions at 12 illustrative sites for 1997–2001. The black error bars are the 1σ errors for the measurements, the green lines are the residuals using the a priori GEIA emissions, and the red lines are the residuals using the optimally estimated emissions.

Figure 7b.

Same as Figure 7a except the period is 2002–2005, and the green lines are the residuals using the estimated emissions from the 1997–2001 inversions (that served as a priori emissions for the 2002–2005 inversions).

5. Concluding Remarks

[50] We conclude that when we combine data from all the current N2O measurement networks (i.e., AGAGE, GMD/CCGG, GMD/HATS (RITS, CATS), and CSIRO) we are able to resolve N2O emissions with reasonable accuracy for the four semihemispheric regions as well as for the eleven land and ocean regions for 1997–2001 and 10 regions for 2002–2005. Assuming a perfect model, we see significant reductions in the a priori uncertainty of regional emission estimates after the inversion. However, inclusion of model errors, computed using a Monte Carlo techniques, increases very significantly the estimation errors calculated assuming a perfect model. We also conclude that we need more measurement coverage in the southern tropics to be able to resolve the southern tropical ocean from the southern tropical land sources. Considering the long lifetime and small latitudinal gradient of N2O, we also conclude that all measurements must be put on the same N2O standard scale before they are used in the inversion. Even a small difference in standard scales between networks, e.g., 0.1 ppb, could impose a significant false influence on the regional emission estimates.

[51] From the global emission inversion, we found that the global mean N2O flux remains almost the same for 1997–2001 as the GEIA estimate for 1990, but decreases by 5.5% relative to GEIA for 2002–2005 assuming the same 125 year lifetime for N2O. From the semihemispheric emission inversion, we conclude that the N2O flux from equator to 30°N region has increased significantly from that estimated by GEIA and Prinn et al. [1990], while the emissions from the southern oceans (30°S–90°S) have dropped significantly relative to those earlier estimates. This finding agrees well with Hirsch et al. [2006] who used a different three dimensional global transport model, a different inversion technique, and GMD CCGG measurements alone for 1998–2001. The differences between the two recent and two earlier studies may be either due to real long-term variations or to different modeling and methodological approaches. In addition, we found that 23 ± 4% of the global total N2O emissions come from the oceans, which is smaller than the bottom up GEIA estimates (28%).

[52] From the regional emission inversion, we conclude that, on average from 1997 to 2005, sources from South America, Africa, China/Japan/South East Asia and North and West Asia are larger by about 20% relative to GEIA; but they are still not statistically different from the GEIA estimates when both modeling and measurement errors are included. On the other hand, the sources from Australia/New Zealand and Europe show a statistically significant decrease from the GEIA estimates. For the oceanic source, our study shows a shift from extratropical areas to the tropical areas relative to earlier studies, and this shift gets larger in 2002–2005 than in the 1997–2001 time periods. In addition, the Northern tropical ocean becomes a much more important contributor to the global N2O flux than that estimated in the GEIA emissions (the earlier Hirsch et al. [2006] and Prinn et al. [1990] studies did not estimate land and ocean sources separately in the 0°–30°N region).

[53] Quantifying modeling uncertainties is an important part of our optimal estimation procedure. We conclude that uncertainties in the exchange between troposphere and stratosphere (STE) play a very important role in N2O regional emission estimates. This is the most significant source of uncertainty among all types of modeling uncertainties. If there were more N2O (or other long-lived tracer) measurements available in the stratosphere and upper troposphere, they could be used to improve the accuracy of the simulations of the seasonal and interannual changes of STE in the global chemical transport models. This would enable more accurate estimation of the seasonal and interannual variations in the N2O surface fluxes.

Acknowledgments

[54] AGAGE support comes from the National Aeronautics and Space Administration (NASA) with important contributions also from the Department of Environment, Food and Rural Affairs (DEFRA, United Kingdom), Commonwealth Scientific and Industrial Research Organization (CSIRO, Australia), Bureau of Meteorology (Australia), and the National Oceanic and Atmospheric Administration (NOAA). The NOAA Earth Systems Research Laboratory supports the two GMD networks whose data are used here. The Australian flask network is supported by CSIRO.

Ancillary