MOPITT is a nadir-sounding satellite instrument flying on the NASA Terra satellite. It uses a gas filter correlation radiometer and measures radiance at both the TIR band near 4.7 µm and the NIR band near 2.3 µm. These observations have a spatial resolution of about 22km×22km with satellite overpass time at approximately 10:30 and 22:30 (local time). To determine a unique CO concentration profile from the MOPITT measured radiances, an optimal estimation-based retrieval algorithm and a fast radiative transfer model are used (Deeter et al., 2003; Edwards et al., 1999). The retrieved state vector (xrtv) for optimal estimation-based retrievals can be expressed as 1xrtv=xa+Axtrue-xa+ϵ, where xa and xtrue are the a priori state vector and the true state vector, respectively. A (which has a size of 10×10) is the retrieval averaging kernel matrix (AK) that represents the sensitivity of retrieved profiles to actual profiles and ϵ is the random error vector. Note that CO quantities in the state vector are retrieved as log10(VMR).

We focus on validating the recently released Version 8 of the MOPITT TIR, NIR, and multispectral TIR–NIR products. We also include comparisons with the MOPITT Version 7 TIR, NIR, and multispectral TIR–NIR products in the Sect. 3.1 for reference. These two versions of MOPITT products were introduced in detail in Deeter et al. (2017, 2019).

Aircraft-sampled profiles of CO concentrations during the DISCOVER-AQ, SEAC4RS, ARIAs, A-FORCE, and KORUS-AQ campaigns are used for comparisons with MOPITT-retrieved profiles. DISCOVER-AQ and SEAC4RS were conducted over the US, while ARIAs, A-FORCE, and KORUS-AQ were conducted over East Asia. Locations of the aircraft profiles from these campaigns are compared with the MODIS (Moderate Resolution Imaging Spectroradiometer) Terra and Aqua Land Cover Type Climate Modeling Grid Yearly Level 3 Version 6 0.05∘×0.05∘ Global product (MCD12C1 v006) (Friedl and Sulla-Menashe, 2015) to determine if a profile was sampled over an urban or non-urban region. Specifically, for each aircraft profile, a 0.5∘×0.5∘ box centered over the location of the aircraft profile (determined by averaged latitude and longitude of aircraft observations in the profile) is selected. If the urban and built-up fraction in the box is larger than 10 %, the profile is considered to be an urban profile. Overall, for each campaign, the averaged aircraft profile over urban regions has higher CO concentrations compared to that over non-urban regions, especially near the surface (see Fig. S1 in the Supplement). Profiles during ARIAs, which are sampled over Hebei province in China, are exceptional, as the averaged profile over non-urban regions has higher CO concentrations especially near the surface, indicating high CO levels in the entire study region. We note that Hebei is one of the most heavily industrialized and polluted regions, and the difference in CO profiles is driven less by urban versus rural than by synoptic and mesoscale meteorology. In addition, Hebei is an arid region and subject to strong nocturnal inversions, so the surface CO can be very high. For aircraft profiles sampled during KORUS-AQ, the CO profiles over urban and non-urban regions are similar, even though the averaged profile over urban regions has slightly higher CO concentration near the surface. This is largely due to the fact that many of the non-urban aircraft profiles are sampled over the Taehwa forest site, which is impacted by CO transported from the nearby Seoul urban region. The urban regions often have different surface parameters (e.g., surface temperature and emissivity) and usually but not always have higher CO concentrations than non-urban regions. However, the surface parameters are unlikely to impact the ultimate quality of MOPITT retrieval products (Pan et al., 1998; Ho et al., 2005). The goal of this study is to understand if MOPITT retrievals are able to represent conditions over urban regions given sampling and cloud cover. In addition, the relatively large spatial and temporal variability of CO concentrations over urban regions makes the validation even more complex. Because of the complexity of urban regions and their connection with non-urban regions nearby, we also provide analysis at high CO concentrations regardless of land cover type. Note that the comparisons include the 600 hPa layer (usually in the free troposphere). It is possible that CO concentrations at this layer are transported from other regions that are not representative of urban regions. Even so, MOPITT retrievals at the 600 hPa layer are still impacted by the CO concentrations at other layers including the surface layer (Eq. 1). Therefore, the comparisons at 600 hPa is necessary.

The campaigns and profiles are summarized in the Table 1 and Fig. 1. During DISCOVER-AQ, SEAC4RS, and KORUS-AQ, CO concentrations were measured by the NASA Differential Absorption Carbon monOxide Measurement (DACOM), whereas during ARIAs and A-FORCE, CO concentrations were measured by Picarro G2401-m and Aero-Laser GmbH AL5002, respectively. Note that the primary goal of DISCOVER-AQ was to provide aircraft observation methodologies for satellite validation (e.g., Lamsal et al., 2014). There are 121 profiles over four urban regions from DISCOVER-AQ, making it particularly useful for the goal of this study. Because of this, our results are heavily driven by aircraft profiles from DISCOVER-AQ. Even though there are only two profiles sampled over urban regions, the A-FORCE campaign obtained 45 profiles in total sampled over East Asia during spring 2009, winter 2013, and summer 2013. The seasonal and spatial coverage of the dataset makes it representative of the region. The ARIAs campaign provides 19 profiles and three of these were sampled over Chinese urban regions. Few previous studies have validated MOPITT products over China (e.g., Hedelius et al., 2019), so aircraft profiles from ARIAs have also been included in this study.

We generally follow the method that has been used in previous MOPITT evaluation and validation studies (Deeter et al., 2010, 2012, 2013, 2014, 2016, 2017, 2019; Emmons et al., 2004, 2007, 2009). There are four main steps in aircraft versus MOPITT comparisons.

Because of aircraft altitude limitations, in situ data from field campaigns do not typically reach the highest altitudes at which MOPITT radiances are sensitive. Therefore, to obtain a complete vertical profile as required for comparison with MOPITT retrievals, each in situ profile is extended vertically using the following steps: (i) the aircraft measurements are interpolated to the 35-level vertical grid used in MOPITT forward model calculations (0.2–1060 hPa); (ii) the levels from the surface to the lowest-altitude aircraft measurement are filled with the value of the in situ measurement at the lowest-altitude aircraft measurement; (iii) for levels above a certain pressure level Pinterp (higher altitude), model or reanalysis data are used directly; (iv) for levels between the highest-altitude aircraft measurement and the altitude of Pinterp, values are linearly interpolated. Unlike the previous MOPITT evaluation studies that used monthly model results from MOZART (Model for OZone And Related chemical Tracers) (Emmons et al., 2010) or CAM-chem (Community Atmosphere Model with chemistry) (Lamarque et al., 2012), here we use 3-hourly Copernicus Atmosphere Monitoring Service (CAMS) reanalysis of CO produced by the European Centre for Medium-Range Weather Forecasts (ECMWF). CAMS CO reanalysis has a horizontal resolution of 80km×80km and 60 vertical grids (from surface to 0.1 hPa). Satellite retrievals of atmospheric composition including MOPITT TIR Version 6 total column CO retrievals are assimilated in the CAMS reanalysis (Inness et al., 2019; https://confluence.ecmwf.int/pages/viewpage.action?pageId=83396018, last access: 18 March 2020). We note that as we do not compare with these higher levels later, the use of CAMS reanalysis is expected to have a minimal impact on the lower levels we use in the comparison (e.g., the surface layer, the 800 hPa layer, and the 600 hPa layer). The final CO profile at the 35-level vertical grid is then regridded onto a coarser 10-level grid (for consistency with the actual MOPITT retrieval grid) by unweighted averaging the fine-grid VMR values in the layers immediately above the corresponding levels in the retrieval grid. We investigate the sensitivity of the results to Pinterp in Sect. 4.1.

The overall comparison results are presented in Table 2. Following Deeter et al. (2017), retrieval biases and standard deviation (SD) are calculated based on mean xrtv and xtransformed for each in situ profile and converted from log10(VMR) to percent. The correlation coefficient (r) is quantified based on xrtv-xa and the corresponding xtransformed-xa to avoid correlations which mainly result from the variability of the a priori. xrtv, xtransformed, and xa are in log10(VMR) space in order to apply the AKs, which are computed for xrtv in log10(VMR). These comparisons for MOPITT Version 8 TIR-only (V8T) and Version 8 TIR–NIR (V8J) are shown in Figs. 4 (for all profiles) and 5 (for urban profiles). Overall biases for V8J products (averaged over all campaigns in Table 1) vary from -0.7 % to 0.0 %, which are lower than biases for V8T (from 2.0 % to 3.5 %). Overall biases for V8J products are also smaller than biases for V7J (from -0.5 % to -5.4 %). For V8J and V7J, biases over urban regions vary from -0.8 % to -2 % and from -1.4 % to -8.9 %, respectively, which are generally larger than biases over non-urban regions (-0.3 %–1.1 % and -3.3 %–0.1 %). Correlation coefficients over non-urban regions are higher than those over urban regions for all six products (V7T, V8T, V7N, V8N, V7J, V8J) at all three levels in Table 2 (the surface layer, the 800 hPa layer, and the 600 hPa layer). We also notice that for TIR–NIR and TIR-only products, V8 have higher correlation coefficients with in situ measurements than V7 over non-urban regions, whereas over urban regions, V8 products have lower correlation coefficients than V7 (except for the 600 hPa layer). Overall, MOPITT products (especially V8J) perform reasonably well over both urban and non-urban regions. Performance over non-urban regions is better than that over urban regions in terms of higher correlation coefficients and smaller biases for V8J and V7J.

We also evaluate MOPITT V8J retrievals during individual field campaigns with results in Fig. 6. The corresponding results for MOPITT V8T are summarized in Fig. S2. The patterns of biases are very similar for MOPITT V8J and V8T. Thus, in this subsection, we focus on V8J unless stated otherwise. Overall, except for comparisons with A-FORCE and ARIAs, biases over urban regions and non-urban regions do not have a significant difference. Neither do biases determined for campaigns over the US and East Asia differ significantly, either.

When compared to DISCOVER-AQ CA (California), MOPITT CO values are generally higher than in situ profiles at the 600 hPa layer (i.e., the 100 hPa uniform layer immediately above 600 hPa) but not at the surface layer (i.e., the uniform layer immediately above the surface). This is likely related to the fact that the DISCOVER-AQ CA aircraft profiles are mostly below 600 hPa, and hence CO values of these in situ profiles at 600 hPa and above are filled with CAMS reanalysis data. In addition, DISCOVER-AQ CA was conducted in the winter when boundary layer height is at lower altitudes, which could also explain the difference, in particular since most of the other campaigns are during times with greater vertical mixing. The lack of aircraft observations at 600 hPa and above also has a smaller impact on the biases at the 800 hPa layer through applying AK (see Fig. 3).

During the A-FORCE campaign, only 2 in situ profiles out of 45 were sampled over urban regions. The locations of the two profiles are close to each other and they are both sampled on or near the coast of South Korea (Fig. 1). MOPITT has large negative biases (-30 % to -40 %) when compared to these two profiles. The averaged xin situ, xa, xtransformed, and xrtv over non-urban regions during A-FORCE and the xin situ,xa, xtransformed, and xrtv of the two profiles over urban regions are shown in Fig. S3. Compared to the averaged xin situ over non-urban regions, the xin situ for the two profiles over the urban regions have large enhancements near the surface and between 600 and 800 hPa. Even though the xa and xrtv for the two profiles have higher CO concentrations (∼400 ppb at the surface layer) than the averaged xa and xrtv (∼200 ppb at the surface layer), they are still lower than the xtransformed.

As for KORUS-AQ, MOPITT also has a negative bias (though smaller) when compared to the profiles over urban regions. Most of these KORUS-AQ profiles were located near the two profiles from A-FORCE but farther from the coast. The negative bias is not seen over non-urban regions during KORUS-AQ at the surface layer.

When compared to the in situ profiles from ARIAs, MOPITT has a large positive bias, especially over urban regions (20 %–30 %). During ARIAs, in situ profiles over urban regions have lower CO values (∼200 ppb at the surface layer) than those in situ profiles over non-urban regions (∼400 ppb at the surface layer; Fig. S4). We note there are only a small number of in situ profiles over urban regions in East Asia used in this study, compared to what is provided by DISCOVER-AQ in the US. The large negative biases against A-FORCE and large positive biases against ARIAs point to the need for more in situ observations over East Asia.

Urban regions are often associated with high CO concentrations. But this is not always the case (e.g., Fig. S4). Here we separate the in situ profiles at the surface layer, the 800 hPa layer, and the 600 hPa layer into lower 50 % CO values and higher 50 % CO values based on CO values at each level to demonstrate the impact of CO concentrations on the MOPITT product validation (Fig. 7). For V8J, MOPITT has smaller biases at higher 50 % CO concentrations for all three levels, whereas for V8T, MOPITT has larger biases at the surface layer and the 600 hPa layer at higher 50 % CO concentrations. For the higher 50 % of measured mixing ratios both V8J and V8T have larger SDs and lower correlation coefficients at the surface layer, the 800 hPa layer, and the 600 hPa layer, suggesting that the agreement between MOPITT and the in situ profiles at higher CO concentrations is not as good as that at lower CO concentrations. In contrast, Deeter et al. (2016) found that the retrieval biases do not visibly increase in the upper range of CO concentrations when compared to aircraft measurements over the Amazon Basin. The vertical error bars in Fig. 7 (caused by the multiple co-located MOPITT profiles with one in situ profile) represent the variability (standard deviation) of the MOPITT data used to calculate each of the plotted mean values. For an in situ profile, the variability of the MOPITT data located within its radius of 100 km and within 12 h is larger when the in situ profile has higher CO values, indicated by larger error bars at higher 50 % CO concentrations. At higher 50 % CO concentrations, the averaged retrieval uncertainties for the 600 hPa, 800 hPa, and surface layers are 28 %, 28 %, and 29 %, respectively. This is smaller than the averaged retrieval uncertainties at lower 50 % CO concentrations (28 %, 29 %, and 30 % for the 600 hPa, 800 hPa, and surface layers, respectively). We therefore conclude that the larger apparent biases at high CO concentrations are related to greater CO variability and representativeness error of the in situ profile within the co-location radius used for analyzing the MOPITT data rather than indicating larger retrieval uncertainties. Theoretically, MOPITT retrievals perform better with higher CO concentrations. The larger biases at high CO concentrations in Fig. 7 imply that the relatively greater CO variability may overcome the impact of high CO concentrations. Addressing representativeness error and spatial variability in the comparisons between satellite and in situ profiles is challenging and will be discussed further in Sect. 5.

We will discuss the sensitivity of radius and time difference for the selection of co-located data in Sect. 4. The difference in the variability at different CO concentrations was not found in Deeter et al. (2016). It could be partially due to the fact that the aircraft profiles over the Amazon Basin used in Deeter et al. (2016) were sampled under more geographically homogeneous conditions, whereas the profiles used in this study are from different campaigns, and high CO concentrations over and near urban regions might be associated with more complex and inhomogeneous conditions.

As discussed in Sect. 2.3, the in situ profiles must be vertically extrapolated or extended to compare with MOPITT products due to aircraft altitude limits. Thus, model or reanalysis data must be merged with the in situ data to generate a complete CO profile for comparisons with MOPITT satellite retrievals. The use of model or reanalysis data may introduce uncertainties in the comparison results as they are not measured directly. The parameter Pinterp controls the impact of the model-based profile extension on the shape and value of in situ profiles (see Fig. S5). Here we test the sensitivity of validation results to various Pinterp values (100, 200, 300, 400, 500 hPa) to demonstrate the potential impact of the profile extension. Note that the model-based profile extension and the value of Pinterp impacts the validation results through changing the augmented observational profile, which is different from the other sensitivity tests in this study that change the selection of MOPITT data. The agreements between the values of MOPITT and in situ profiles at the surface layer are insensitive to the selection of Pinterp (Fig. 8). The overall agreements between the values of MOPITT and in situ profiles at the 800 hPa layer are also not sensitive to Pinterp, except for the results against DISCOVER-AQ CA, which have slightly larger biases when Pinterp is 200 hPa or 100 hPa since the DISCOVER-AQ CA aircraft profiles at 600 hPa and above are mostly extended using reanalysis data. Therefore, the comparisons with DISCOVER-AQ CA are more likely to be affected by Pinterp compared to other campaigns which typically obtained higher maximum aircraft altitudes. At the 600 hPa layer, the agreements between the values of MOPITT and in situ profiles are affected more by Pinterp compared to the those at the surface layer and the 800 hPa layer for comparisons with all the campaigns. The overall validation results using 100 hPa as Pinterp have larger biases than using other values of Pinterp. At 400 hPa layer and 200 hPa layer, the comparisons are even more sensitive to Pinterp for all the campaigns (Fig. S6). The CAMS 3-hourly reanalysis data are constrained by observations, but their use may still introduce the uncertainties in the validation results especially at upper pressure levels (e.g., 200 and 400 hPa). Previous MOPITT evaluation results may be subject to larger uncertainties by using CAM-chem monthly CO fields that are not constrained by observations (e.g., Deeter et al., 2012, 2016).

The criteria for co-location in this study (within a radius of 100 km and within 12 h of the acquisition of the aircraft profile) generally follow previous MOPITT validation studies (e.g., Deeter et al., 2016, 2019) and are chosen empirically. They are selected based on a trade-off between uncertainties generated from CO spatial and/or temporal variability and the number of included MOPITT retrievals that impacts the statistical robustness. Here we test the sensitivity of the results to the two criteria for co-location. The box plot of biases calculated with different radii (200, 100, 50, and 25 km) at the surface layer, the 800 hPa layer, and the 600 hPa layer are shown in Fig. 9. Overall, the biases calculated with a radius of 200, 100, and 50 km are similar, whereas the biases calculated with the radius of 25 km are different from others. The comparisons of MOPITT to in situ profile results using the radius of 25 km generally have larger biases and SD, due to including fewer MOPITT retrievals. In some cases, there are no matched MOPITT retrievals within the radius of 25 km of the aircraft profile (e.g., DISCOVER-AQ CA and ARIAs). In addition, representativeness errors would be expected to go up if there are only a few retrievals over a more polluted and perhaps heterogeneous area. We note that the use of the largest radius (200 km) in this paper does not appear to degrade the overall results, even though representativeness errors generated from CO spatial and/or temporal variability are expected to increase. However, the use of the smallest radius (25 km) degrades the overall results by reducing the number of included MOPITT retrievals.

The box plot of biases calculated with four sets of allowed maximum time difference (12, 6, 3, and 1 h) are shown in Fig. 10. The overall results are not sensitive to the selection of allowed maximum time difference. One exception is the comparisons to the SEAC4RS campaign at the 600 hPa layer, due to a smaller number of MOPITT retrievals in the shorter time window. We note that when comparing to the ARIAs campaign, using 1 h as the allowed maximum time difference decreases the biases at the surface layer, the 800 hPa layer, and the 600 hPa layer, compared to the cases using a longer allowed maximum time difference (i.e., 3, 6, and 12 h). This implies that the temporal variability is relatively large in the region. And the improvement observed for ARIAs for the shortest time also points to the possibility that short-term emission sources might be responsible for the large biases there. On the other hand, when the allowed maximum time difference equals 1 h, there are only six aircraft profiles that match MOPITT retrievals.

Previous MOPITT validation studies have only included MOPITT daytime observations. Over land, MOPITT retrievals for daytime and nighttime overpasses are characterized by significantly different averaging kernels (Fig. 3) and may be subject to different types of retrieval error (Deeter et al., 2007). CO has a long enough lifetime (approximately 1 month; Gamnitzer et al., 2006) in the free troposphere that nighttime observations could be potentially comparable, in general, to the daytime flights for remote sites. However, for urban regions where the spatiotemporal variability of the emissions and evolution of the planetary boundary layer drives large changes in the measured CO, comparisons of MOPITT nighttime observations to aircraft profiles sampled during the daytime may introduce representative uncertainties, especially for areas that are subject to strong nocturnal inversions, and the surface CO can be enhanced. It is difficult to disentangle the effects of the MOPITT daytime or nighttime performance and the uncertainty from the temporal representativeness, based on the comparison of the MOPITT daytime or nighttime retrievals with daytime aircraft profiles. Therefore, we only include the results in Fig. S7 and briefly describe the results here without drawing any further conclusions. Overall, MOPITT nighttime retrievals have larger biases than daytime retrievals, which could be expected since most of the aircraft profiles are sampled during the daytime. Flight campaigns with nighttime observations are needed to validate MOPITT nighttime retrievals.

The MOPITT Level 3 data are generated from Level 2 data, and are available as gridded (1∘×1∘) daily mean and monthly mean files. Pixel filtering and SNR thresholds for Channel 5 and 6 average radiances are used when averaging Level 2 data into Level 3 data, and this increases overall mean DFS values (details can be found in the MOPITT Version 8 Product User's Guide, 2018). Taking the MOPITT V8J daytime product as an example, the Level 3 data product excludes all observations from Pixel 3 (one of the four elements of MOPITT's linear detector array that has highly variable Channel 7 SNR values) or observations where both the Channel 5 average radiance SNR <1000 and the Channel 6 average radiance SNR <400. In Fig. 11, we test the impact of applying the aforementioned SNR filters to the agreement between MOPITT and in situ profiles. Note that we are not suggesting the comparisons between MOPITT Level 3 product and aircraft measurements. Because the MOPITT Level 3 product is gridded data, and it represents the average value in a 1∘×1∘ grid. Comparing the grid average value to an aircraft profile within it may be subject to large representativeness errors. Here we only show the sensitivity of agreement between MOPITT Level 2 data and aircraft profiles to the application of SNR filters. We find that applying the SNR filters does not significantly change the overall agreement between MOPITT retrievals and the in situ profiles used in this study. This is mostly because applying the SNR filters reduces the number of MOPITT retrievals included in the comparisons. This effect is particularly important if there are not many MOPITT retrievals to begin with (such as our comparisons with in situ profiles in this study). Even though applying SNR filter when generating Level 3 data does not significantly change the agreement with the in situ profiles used in this study, excluding low-SNR observations from the Level 3 cell-averaged values raises overall mean DFS values (MOPITT Algorithm Development Team, 2018). In addition, the Level 3 product typically is less affected by random retrieval errors (e.g., due to instrument noise or geophysical noise).