The MOPITT and IASI missions have now accumulated 15 and 7 years respectively of near-continuous global data for tropospheric CO. For this comparison we used the retrieval algorithm versions that were running in 2013 (MOPITT v5T and IASI FORLI v20100815) and the retrieved CO profile products, from which integrated total columns are derived, along with their associated averaging kernel matrices (for profiles) or vector (for columns). Only the data from the IASI/MetOp-A mission are analyzed here.

Table 1 provides a detailed description of the retrieved products, the a priori information, and the auxiliary data (temperature, emissivity, cloud content) for each mission. Note that the number of retrieved layers exceeds the number of independent pieces of information available vertically and hence is not representative of the vertical resolution of the observation.

Previous validation studies using ground-based, aircraft and satellite data have shown that CO total columns from MOPITT and IASI are retrieved with an error generally below 10–15 % at mid and tropical latitudes, but can have larger errors in polar regions (MOPITT: Deeter et al., 2012, 2013; Emmons et al., 2004, 2009; IASI: George et al., 2009; Pommier et al., 2010; De Wachter et al., 2012; Kerzenmacher et al., 2012). The profiles are only weakly resolved, with < 1 to ∼ 2.5 independent pieces of information, depending mostly on the thermal state of the atmosphere. A DOFS of less than 1 indicates that the a priori information dominates the calculated total column, whereas a DOFS of 2 or more means that at least two independent partial columns can be retrieved. The highest sensitivity is achieved in the inter-tropical region or at mid-latitudes during daytime and over land: for instance, there is a gain of 0.5 DOFS above the northern mid-latitude continental surfaces between the morning and evening orbits (Hurtmans et al., 2012).

A major difference between MOPITT and IASI retrievals resides in the choice of the a priori, which is fixed for IASI, and variable for MOPITT. Having a variable or a static a priori has implications on the retrieved data set, with both choices presenting advantages and disadvantages as discussed hereafter. Figure 2 represents the a priori profile(s) and the variance-covariance matrices (Sa), for MOPITT (in September 2010) and for IASI (invariant). These were built using chemistry-transport model simulations and other available data. For MOPITT v5T the a priori profile varies as a function of location and time of year and it is based on a monthly climatology of the MOZART-4 chemistry transport model. For each retrieval, the climatology is spatially and temporally interpolated to match the date and location of the observation. The fixed Sa matrix allows for a 30 % variability in each retrieved layer. The off-diagonal elements which define the correlations between the different layers are consistent with a short vertical correlation length which limits the spread of information from one layer to another (Deeter et al., 2010). On the contrary, the IASI a priori consists on a single profile, and a fixed Sa matrix, built from a climatology that uses LMDz-INCA model outputs, MOZAIC aircraft data and ACE-FTS satellite profiles (Turquety et al., 2009). The a priori profile is around 90 ± 20 ppbv from the surface to the middle troposphere, and then smoothly decreases to 40 ppbv from 7 km up to 18 km. The Sa matrix allows a maximum variability in the first layer (63 %), decreases to 35 % between 5 and 6 km, to 30 % (as MOPITT) between 6 and 10 km, and is increasing again, reaching 45 % between 15 and 16 km (see Fig. 2). Off diagonal elements are calculated from the ensemble profiles, and allow the information to be projected from layers with high sensitivity to layers where the sensitivity is much weaker. The correlation length, therefore variable, is about 5 km.

The comparison analysis is performed over the period extending from January 2008 to December 2013, a period when MOPITT and IASI were both in operation. As the two instruments are not onboard the same platform, neither the measurement time nor the location are exactly the same.

The top and middle panels of Fig. 3 show the monthly average for CO total column distribution (daytime data) for April 2010 along with the monthly average of the DOFS for the profile retrieval, for each instrument. As expected, it can be seen that large concentrations of CO are found near emission sources, and plumes are transported downwind. In the NH elevated levels of CO are found above the west and east coasts of the USA, over Europe, and over East Asia. Due to long range transport, high CO concentrations are also observed over the Northern Pacific and Atlantic oceans. In the tropics, elevated CO concentrations are found over the Guinea gulf countries (fires). Note that reduced CO total columns at the location of mountains in North and South America, as well as in the Himalayas, are due to surface height. Figure 4 provides in addition a time series of zonal mean total column CO over the entire period. NH concentrations peak in April, after accumulating during winter, and drop off gradually until late summer as the increasing solar insolation activates tropospheric chemistry (except over Siberia and Alaska fire regions where CO concentrations increase in summer). In the tropics the CO maximum is mainly associated with fires occurring in the Amazon basin, in central and southern Africa and sometimes over Australia, with maximum in August–November. Major fires occurring in Russia in August 2010 (Yurganov et al., 2011; Krol et al., 2013; R'Honi et al., 2013) and in Siberia in July 2012 (Ponomarev, 2013) are also visible on the zonal mean total column plots. The associated DOFS distributions (right panels of Fig. 3) illustrate the strong latitudinal variations due to temperature changes. The patterns look similar, but MOPITT is showing lower associated DOFS than IASI. Note that as the instruments are intrinsically different we do not expect their DOFS values to be the same, and that both the a priori and the measurement covariance matrix (Se) will impact the DOFS values.

Even if the general horizontal spatial concentration patterns agree well, differences in the CO total columns can be seen when comparing the MOPITT and IASI data for the same areas/periods. In Fig. 5a representing the relative differences between IASI and MOPITT v5T for 1-month (April 2010), more than 70 % of the plotted data do not exceed 10 % (ratio calculated from the original grid), which is the CO accuracy specification for both missions (Pan et al., 1995; IASI Science Plan, 1998). Note that here we discuss the agreement between the two products, not the absolute accuracy which was evaluated in previous validation papers (e.g., see references provided in Sect. 2.2). MOPITT concentrations are generally larger than the IASI concentrations over land, in particular close to the location of strong emission sources (USA's east coast, China). In contrast, IASI concentrations are generally larger over the ocean, between 30∘ S and 45∘ N, and above 75∘ N. Major fire events such as in Russia and Siberia (in 2010 and 2012, respectively) appear to be more marked in the IASI data, and likewise for the fires occurring in Africa and Amazonia (Fig. 4). Note that over Antarctica, MOPITT DOFS are close to zero, indicating that the retrieved profile is close to the a priori profile.

It is the aim of this paper to investigate the possible sources of the differences between IASI and MOPITT data measured at the same location. We expect differences to be associated with (i) the different vertical sensitivity of the two sensors, (ii) with the a priori assumptions, (iii) the auxiliary data (e.g., surface temperature, temperature profiles, emissivity, cloud information, etc.) used in the retrieval process, as well as (iv) due to the different air masses sounded (different sounding angles, and between one and 2-hours time lag for the observation time). Because the two instruments fly on different satellites, and rely on different auxiliary data sets (temperature, clouds, etc.), only the differences associated with the a priori assumptions are studied in this paper.

To study the impact of a change of a priori on the retrieval we made a two-step comparison: first with the native retrieved data, and second with a dedicated retrieval chain set-up at the National Center for Atmospheric Research (NCAR), where the MOPITT data were reprocessed using the IASI a priori profile and Sa matrix (hereafter referred to as MOPITT vX1). It is not possible to exactly convert the IASI Sa matrix (expressed in altitude and partial columns) into a MOPITT-compatible matrix (expressed in pressure levels and log(VMR)) since the IASI and MOPITT retrieval algorithms exploit mathematically inconsistent formats to express the vertical distribution of CO molecules. Schemes for interpolating or extrapolating Sa may also violate basic properties of covariance matrices, such as positive definiteness. Therefore we built a new a priori profile and covariance matrix from the original profiles ensemble used for the Sa matrix generation in FORLI, on a common 35-pressure-layer grid (The MOPITT algorithm uses a priori information on a 35-level pressure grid to produce 10-level a priori profiles used in the actual retrieval algorithm).

The CO total column distribution measured by MOPITT in April 2010 and reprocessed with the IASI a priori constraints (MOPITT vX1) is shown in Fig. 3 (bottom part). Figure 5 provides the relative difference plots between IASI and MOPITT v5T, MOPITT v5T and MOPITT vX1, as well as between IASI and MOPITT vX1. Probability density functions by latitude bands are also represented (see Fig. 5d–m). It can be seen that the larger differences between the MOPITT v5T and vX1 concentrations are observed over the polar regions, where the v5T concentrations are larger than the vX1 ones at the North Pole (15 % on average between 60 and 90∘ N) and smaller over Antarctica (-60 % on average between 60 and 90∘ S). Between 60∘ S and 60∘ N, the differences generally range between -5 and 5 %, with the MOPITT v5T columns being larger than the vX1 ones above some emissions sources (USA's east coast, Mexico and China). This can be explained by the MOPITT v5T climatology-based a priori, which is closer to the real atmospheric state, including higher levels of CO above emissions sources.

From Fig. 5a and c we see that the reprocessing of MOPITT data slightly improves the agreement with IASI over the USA's east coast and China, i.e., for regions where emission sources are usually high. Between 20∘ S and 20∘ N and between 20 and 60∘ N, the statistics are alike: when looking at the histograms (Fig. 5i–m), the probability density functions (100 × (IASI-MOPITT v5T)/IASI and 100 × (IASI-MOPITT vX1)/IASI) look similar. But in the Southern Hemisphere between 20 and 60∘ S the reprocessing of MOPITT does not reconcile the differences with IASI, in fact the difference percentages are larger for the comparison with MOPITT vX1 (the probability density functions peaks at -10 %, and it was -5 % with v5T), and it is the same at high northern latitudes (20 % for vX1 compared to 5 % for v5T). Finally, the differences are about the same amplitude in Antarctica, but with the opposite sign (-30 % for vX1 and +30 % for v5T).

These differences will be discussed in details in the next two sections.

In order to investigate the observed differences, a detailed analysis was performed over the 6-year seasonal record, on 12 selected regions spread over the globe (listed in Table 2, and also identified by green boxes in Fig. 5b). The areas are representative of different ecosystems (water, sand, forest) and of various seasonal CO atmospheric content (cities, fire seasonal activity, background). The size of the grid boxes (5∘ × 5∘ for nine regions and 2∘ × 2∘ for three cities) was chosen so that the number of data is statistically significant for each instrument. For each box, 15-day averages of CO total column values are calculated, provided data from both MOPITT and IASI are available for each day. Typically, each grid box contains about 500 MOPITT and 850 IASI pixels. Table 2 lists the biases and the absolute biases, along with their standard deviation (SD), as well as the correlation coefficients for each region.

Figure 6 illustrates the seasonal patterns as seen by both instruments, for a subset of six regions representative of different regimes: Africa (fires), China (high concentrations and large variability), Pacific (remote sea), Siberia, USA and Europe (NH regions with large discrepancies in boreal winter). The figure provides the average and the standard deviation for IASI (in red) and MOPITT v5T (in blue), twice per month. The maxima and minima are driven by the chemical and photochemical reactions described in Sect. 3.1. It can be seen that the agreement is good in general although MOPITT columns are most of the time slightly larger for all the boxes located over land, as already discussed. The correlation coefficients (r5T and rX1 in Table 2) are good (range between 0.72 and 0.95) and generally improved by the reprocessing. The variability inside the box (standard deviation in Fig. 6) is an indicator of the rapid changes in the CO content occurring over the area. It is very low over the remote sea (see the Pacific box) and very high over the polluted area in China. Figure 7a provides the differences in percent for the same six areas for both the MOPITT v5T and the MOPITT vX1 processing, relatively to IASI. The grey envelopes indicate the IASI standard deviation within the box (in %). By analyzing the time periods when the MOPITT v5T vs. IASI differences exceed this “natural” variability (i.e., when the black dots are outside the grey area in Fig. 7), we find as a consistent pattern that the MOPITT total columns sometimes exceed the IASI total columns by ∼ 30  %. This happens each year during the boreal winter period (December–January) for the boxes “Europe”, “USA”, and “Siberia” (see black rectangles and bold figures in Table 2). In the “Siberia” box, the difference can reach 50 % from October to April. This is closely linked to the seasonal evolution of the information content available in the data (how much it can depart from the a priori) as can been seen from IASI DOFS plotted in Fig. 7b for the “Europe”, “USA” and “Siberia” boxes. The largest biases are indeed observed in boreal winter and are associated with low DOFS at this time of year. Although they do not totally disappear, these biases are significantly reduced when the MOPITT data are reprocessed to derive CO using the IASI a priori: differences are reduced by a factor of 2 to 2.5 (“Europe”: absolute mean bias of 35 % in December–January compared to 18.1 % after the reprocessing; “USA”: 27.2 vs. 11.4 %; “Siberia”: 35.9 vs. 12.9 %) (see Table 2). Surprisingly, the use of the same a priori information slightly increases the biases for some other regions (Pacific, Atlantic and Mexico City), for which an in-depth analysis of averaging kernels would be needed for a complete understanding.

A global map of the differences in a priori for both missions is provided in Fig. 8, which shows the global difference between the IASI and MOPITT a priori CO data, for both January and July, at the lowest vertical level and at 400 hPa (∼ 7 km). The larger differences are found near the surface, close to pollution and fire emission sources, mostly in the Northern Hemisphere, and peak in winter over the selected areas as discussed in Sect. 3.2. An in-depth look at the retrieved profiles will provide more information on how the a priori profiles and associated Sa matrix and actual observations combine.

Even more than for total column values, the shape of the retrieved CO profiles will be determined by the vertical instrumental sensitivity, modulated by the thermal contrast which governs the sensitivity to the lower atmospheric layers, and by the a priori assumptions. If the measurement sensitivity is low and/or the background covariance is small relative to that of the measurement, then the retrieval tends toward the a priori profile value at these altitudes. When the a priori profiles differ significantly for IASI and MOPITT, large differences can appear in the retrieved profile products.

As explained in Sect. 2.2, the IASI a priori profile is always the same, and the Sa matrix allows a large variability, in particular near the surface. On the contrary MOPITT v5T a priori profiles rely on a monthly/latitudinal varying climatology, and the Sa matrix has a moderate and constant vertical variability. Thousands of CO data were analyzed to compare the profiles from both the original IASI and MOPITT v5T products, and the profiles obtained after the reprocessing (MOPITT vX1). Figure 9 illustrates a typical finding. It shows the CO profiles for 1 day of observation (9 April 2010) for the “Europe” box, when high levels of CO were observed. The total column means are similar for each product but the shape of the profiles differs. We see that the IASI-retrieved profiles (Fig. 9a), depart from the a priori at all altitudes but especially near the surface given the high variance of its Sa matrix at this altitude. For MOPITT v5T (Fig. 9b), it can be seen that the retrieved profiles remain quite close to the a priori profiles near the surface and depart at around 400 hPa, where its maximum sensitivity lies. This corresponds also to the altitude where the pressure modulated cell (PMC) channels provide most information. The quasi-diagonal MOPITT Sa matrix limits the “extrapolation” effects to the adjacent levels. Interestingly, for MOPITT vX1 (Fig. 9c), the shape of the profiles differs from the MOPITT v5T profiles and departs more from the a priori. However MOPITT vX1 profiles do not show the large concentrations at the surface that IASI profiles do, despite the fact that the same a priori is used. As illustrated in Fig. 9d the averaging kernels for a representative case (in red in Fig. 9a–c) show a non-zero sensitivity at the surface for MOPITT. Another possible explanation lies in the constraint applied to the measurements (the Se in the OE), which might be looser in FORLI, increasing further the range of variability. Looking at the total errors associated with each retrieved profiles (Fig. 9e), we note that the three profiles are within the errors of each other, which indicates the consistency of the data sets. The errors in % are plotted in Fig. 9f. The MOPITT vX1 total error profile is close to the IASI one because the smoothing error dominates.

In order to go further in the analysis we selected three illustrative cases, representative of different situations, for which aircraft profile data from the MOZAIC-IAGOS program (Nedelec et al., 2003; http://www.iagos.org/) were available within a ±12 h time slot. Figures 10 to 12 show for different locations the IASI, MOPITT v5T and MOPITT vX1 averaged profiles with their corresponding a priori profiles, along with the collocated MOZAIC-IAGOS profile. All data within 0.5∘ of the MOZAIC-IAGOS profile path (which corresponds to 36 to 56 km, depending on latitude) were selected and then averaged. Note that the MOZAIC-IAGOS profiles were not smoothed by the IASI/MOPITT averaging kernels here, as we wanted to represent the actual altitude of the pollution plume if any. Representative averaging kernel functions at different altitudes are also provided for each product, in order to evaluate the altitudes where the retrievals are mostly sensitive.

For the “Nagoya” case plotted in Fig. 10, the MOZAIC-IAGOS profile shows a pollution plume around 600 hPa (∼ 4 km) measured on 25 June 2012. The shape of the collocated satellite retrieved profiles differs, with MOPITT peaking around 300–400 hPa and at the surface, and IASI peaking at lower troposphere and at the surface. The MOPITT averaging kernel functions show that the retrieval is most sensitive just above the plume altitude, where the MOPITT v5T profile peaks. Due to the fact that there is no sensitivity at the surface the retrieved CO sticks to its a priori at this altitude. The IASI averaging kernel functions show a sensitivity of the retrieval slightly lower in altitude, with a maximum around 700 hPa, as well as a slight sensitivity near the surface. The IASI retrieved profile underestimates the amount of CO around 600 hPa and overestimates it at surface level. Due to the loosely constrained covariance matrix near the surface, the CO amount “seen” by IASI is extrapolated toward the surface. The MOPITT vX1 profile lies “in between”, with lower concentrations than the v5T one in the first layers close to the surface, and larger concentrations than the IASI profile above 400 hPa.

The case at the Caracas airport (Fig. 11) shows a typical aircraft profile measured at this location, with CO mixing ratios reaching more than 300 ppbv around 900–800 hPa (1–2 km). The total columns retrieved by both MOPITT v5T and IASI are quite similar, but again the shape of the profiles differs. The IASI retrieval shows some sensitivity close to the surface as the averaging kernel functions associated with the lower altitudes peak between 700 and 900 hPa. The IASI profile somewhat departs from the a priori for the first altitude levels but it does not reach the MOZAIC-IAGOS high values. On the other hand, the altitude of MOPITT retrieval sensitivity maximum is higher, around 300 hPa (∼ 9 km) and its sensitivity is low near the surface. MOPITT does not capture the plume (-40 ppbv compared to IASI near the surface), and the retrieved profiles (v5T and vX1) are close to their a priori profiles (and the climatology is far from the observation in this case).

The “Frankfurt” case (Fig. 12) shows large mixing ratios measured by the MOZAIC-IAGOS aircraft near the surface. Both MOPITT and IASI are sensitive in the mid troposphere (between 500 and 300 hPa) but not at the surface. All the retrieved profiles stick to their a priori profiles, especially at the surface. The MOPITT v5T profile agrees very well with the MOZAIC-IAGOS profile, sticking to the a priori profile which in this case shows large mixing ratios at the surface (reaching more than 250 ppb). For IASI, the plume is missed and for MOPITT vX1, the profile behaves similarly to the IASI profile.

These three cases were selected to illustrate the impacts of choosing a single or a variant a priori profile and a strongly or loosely constrained Sa matrix. In summary, when there is a good sensitivity of the satellite instrument at the altitude of the plume, both instruments manage to detect the CO increase, but MOPITT generally puts it where its maximum sensitivity lies (around 300–400 hPa), whereas IASI tends to project high CO observed in the middle-troposphere towards the surface (because of the 5 km correlation length). For the altitudes where the instrument is not sensitive, in particular at the surface level when the thermal contrast is low, each instrument sticks to its a priori. This leads to a better agreement for the MOPITT-retrieved profile when the measured CO profile at one location is close to the climatology used to build the a priori, which is usually the case for seasonal fires and highly polluted areas (e.g., Frankfurt). On the contrary, for situations where unexpected fires or pollution events occur (e.g., near Caracas) the agreement is better with the IASI derived profile.

In order to confirm the important role of the choice of the a priori assumptions and especially the weight of the Sa matrix, we also performed some tests processing the IASI algorithm with the MOPITT Sa matrix (but with the single IASI a priori profile). As expected, the reprocessed IASI profiles (not shown here) show lower CO concentration than the native IASI profiles near the surface because the allowed variability (used for MOPITT) around the a priori profile is lower.