The Atmospheric Infrared Sounder (AIRS) instrument is carried on board the Aqua satellite. Aqua was launched in May 2002 and flies in a polar, sun synchronous orbit. For most of its mission, it had a local ascending equator crossing time of ∼ 13:35 LT. Starting in 2022, it began to drift to a later crossing time; as of early 2025, it has an equator crossing time of ∼ 14:30 LT (https://aqua.nasa.gov/, last access: 12 March 2025).

AIRS is a grating spectrometer, covering three spectral bands (approximately 650 to 1140 cm-1, 1220 to 1610 cm-1, and 2170 to 2670 cm-1) with 17 detector arrays and a nominal spectral resolution of λ/Δλ = 1200 ranging between 1086 and 1570,. The original calibration is described by and an update is given in . AIRS's radiometric calibration has been very stable over its lifetime, with < 2 mKyr-1 drift between 2017 and between -3 and +6 mK since . Table  compares some of the relevant instrument characteristics of AIRS and CrIS. For a more in-depth description of the differences between the AIRS and CrIS instruments, please see Table S1 of and Table 1 of . In this work, we use AIRS level 1B radiances from version 5 of the AIRIBRAD product .

At time of writing, there are three operational Cross-track Infrared Sounder (CrIS) instruments. The first is on board the Suomi-NPP satellite, launched in October 2011, followed by copies on the JPSS-1/NOAA-20 and JPSS-2/NOAA-21 satellites, launched in November 2017 and November 2022, respectively. All three are in sun synchronous orbits with ascending local equator crossing times around 13:30 LT. Unlike AIRS, CrIS is a Fourier transform spectrometer that observes nine fields of view in a 3 × 3 array simultaneously. It performs an across-track scan of 30 view positions. The fields of view are ∼ 15 km in diameter .

used radiances from the CrIS instrument on board Suomi-NPP (S-NPP) specifically the NASA version 2 level 1B radiances, to retrieve PANs. CrIS measures in three spectral bands: long-wave IR (650 to 1095 cm-1), midwave IR (1210 to 1750 cm-1), and short-wave IR (2155 to 2550 cm-1) . At launch, the CrIS S-NPP instrument was operated in “normal spectral resolution” mode, with the bands measuring at 0.625, 1.25, and 2.5 cm-1, respectively. In December 2014, it was switched to “full spectral resolution” (FSR) mode, with 0.625 cm-1 resolution in all bands . The NASA FSR level 1B product begins a year later (December 2015) after additional upgrades to CrIS calibration in November 2015. Our work uses the MUSES algorithm (Sect. ), which uses radiances from multiple CrIS bands to retrieve atmospheric trace gases and temperature. Currently, this requires the full spectral resolution product, thus we limit ourselves to CrIS data from December 2015 on.

validated the CrIS PANs retrievals against PAN measurements taken during the ATom campaign. The measured profiles had GEOS-Chem profiles appended to the top. From the standard deviation of the differences between CrIS and aircraft free tropospheric PAN column averages, derived a single sounding uncertainty of 0.08 ppb for the CrIS PANs retrieval. This was larger than the uncertainty calculated by the MUSES optimal estimation (OE) algorithm, but attribute the discrepancy to pseudo-random error contributions from the retrieval of interfering species or the temperature profile. Such interferent-driven error was not included in the uncertainty calculated by the MUSES algorithm, as for PAN retrievals, the algorithm calculates uncertainty from noise only.

Further, the comparison with ATom found a negative bias (CrIS lower than aircraft) that correlated with the total column amount of water vapor. The relationship between water vapor and the CrIS PAN bias was further corroborated by examination of the pre-PAN retrieval spectral residuals, which found a positive residual correlated with water vapor column amounts. From the ATom comparisons, derived a bias correction for the CrIS PAN product, c=0.05+0.035×10-23×X, where X is the column density of water vapor in molec.cm-2 and c is the correction in ppb.

The MUSES retrieval is an optimal estimation retrieval with heritage tracing back to the TES retrieval . It is instrument-agnostic, able to solve for the optimal state vector given radiances from a variety of instruments (e.g., AIRS, CrIS, the Ozone Monitoring Instrument (OMI), and the TROPOspheric Monitoring Instrument (TROPOMI)), or multiple instruments (e.g., AIRS + OMI, CrIS + TROPOMI).

The MUSES algorithm allows retrievals to be broken down into smaller steps, each of which define the spectral windows for which to minimize the radiance residuals, the atmospheric parameters to solve for, which of those parameters to update for the next step, along with a number of more technical options. The steps are defined in a “strategy table” which can be quickly edited to test different retrieval approaches. This step-wise design provides flexibility to fix some elements of the state vector while updating others in certain steps, which is particularly useful when retrieving state vector elements with large differences in the magnitude of their Jacobian matrices (e.g., atmospheric temperature vs. PAN) or which interfere with each other (e.g., O3 vs. PAN). These steps are run sequentially; the final state of one step becomes the initial state for the next, save for any state vector elements which the strategy table indicates should not be updated.

Within each step, MUSES uses an iterative solver that applies the trust-region Levenberg–Marquardt scheme to mimimize a cost function 1J(x)=[y-F(x,b)]TSϵ-1[y-F(x,b)]+(x-xa)TSa-1(x-xa) where

x is the retrieved state vector,

The Levenberg–Marquardt solver will iteratively update the state vector, x along a direction in state space expected to mimimize Eq. (). It will continue until the convergence criteria are satisfied or the maximum number of iterations is reached.

An important distinction within MUSES is the difference between the a priori (or constraint) state vector and the initial state vector. The former is xa in Eq. () and is a mathematical constraint on the optimal state vector, the latter is the starting point of x before the first iteration of the Levenberg–Marquardt solver. This distinction is important within MUSES because it is a multi-step retrieval. The strategy table, mentioned above, defines which elements of the state vector will be retrieved in each step and whether or not the retrieved state for step i becomes the initial state for step i+1. For example, the retrieval may begin with an H2O profile taken from a meteorological reanalysis as both the initial guess and the a priori constraint. An early step in the retrieval can then retrieve a new H2O profile which is more consistent with the observed radiances. This new H2O profile can then be used as an initial state for later steps (whether or not those steps retrieve H2O). This can be important for weak absorbers, such as PAN, which need the profiles of strong thermal IR absorbers to be accurate for the scene in question so that the relatively small absorption feature of the weak absorber can be identified. We note that, for a given step, the initial state and a priori constraint can be the same but do not need to be. For later steps of the retrieval, the initial state may have been set by earlier retrieval steps (as in the example given with H2O) but the a priori constraint will remain the same for all steps. Or, the a priori constraint may be chosen to be a relatively simple profile to avoid imposing undue assumptions, while the initial state may be chosen to reflect a better estimate of the atmospheric state in that location to attempt to minimize the number of steps needed by the solver.

MUSES can use different radiative transfer models for F in Eq. (). For this work, we use version 1.2 of the Optimal Spectral Sampling (OSS) model . OSS is designed to use an optimal set of absorption coefficients (per absorbing species and vertical layers) and weights that can be used to compute the radiance for each channel of a spectrometer very efficiently, given the amounts of each absorbing species. These weights are computed by training OSS against a reference line-by-line spectroscopic model. Determining those optimal absorption coefficients and weights requires it to be trained for a given instrument. In version 1.2, the absorption coefficients are calculated from the Line By Line Radiative Transfer Model (LBLRTM) version 12.4 . This allows OSS to efficiently simulate the radiances a specific instrument would observe by reducing the number of monochromatic wavelengths that must be modeled for a given instrument channel, but means that OSS must be trained for each instrument used in a retrieval separately. For details on the approach, readers are encouraged to review and .

The TROPESS project focuses on applying the MUSES algorithm to retrieve a range of atmospheric trace gases from a variety of space-based instruments, including AIRS, OMI, CrIS, and TROPOMI to date. Operational processing for TROPESS is set up to accommodate two distinct goals. The first is to provide a global record of ozone and related trace gases for the first ∼ 20 years of the 21st century. The second is to support rapid iteration on and improvement of the underlying level 2 algorithms while processing more recent data. Due to the computational cost of these retrievals, meeting both goals requires two separate data streams.

The first is a “retrospective” or “reanalysis” stream that retrieves trace gas amounts from ∼ 2002 through ∼ 2021. This stream is processed with a version of the MUSES algorithm frozen at the time the retrospective processing began. The second is a “forward” stream that processes new radiances as they become available with the latest version of the MUSES algorithm, including updates to the algorithm made after the retrospective processing began. The forward stream serves the dual purpose of monitoring significant events affecting air quality and serving as a test bed for improvements to the MUSES algorithm. Due to the difference in the algorithm versions, users must take care not to misinterpret changes in trends between the two streams.

Both streams use a “global survey” sampling approach to process a subset of all available soundings yet provide global coverage, which allows a balance between computational cost and spatial coverage. The default survey strategy processes one sounding in each x° × x° box over land and one out of every four such boxes over ocean. For the current products, x is either 0.7 or 0.8°. In addition, TROPESS produces special collections with full data density for high interest events (e.g., the 2019–2020 Australian Bush Fires and 2020 US West Coast Fires) and a set of megacities around the world.

The CrIS PAN product described in and Sect. , with mostly minor updates, is now routinely produced as part of both the reanalysis and forward TROPESS streams, as well as special products. The reanalysis and forward streams provide twice daily (day and night) global coverage of PAN, using the global survey strategy described in the previous paragraph. Other species retrieved within the TROPESS project include methane, carbon monoxide, deuterated water (HDO), ammonia, and ozone.

Retrievals using AIRS radiances have previously been implemented within the TROPESS MUSES algorithm (§ ), thus the AIRS PAN retrieval can use the existing readers and MUSES OE framework. The components that must be added are (1) the desired windows, (2) the strategy table that instructs the MUSES algorithm to retrieve PAN, and (3) details about PAN retrievals copied from the CrIS PAN retrievals, such as the prior vector values.

For the existing CrIS PAN retrieval, chose two windows on the low frequency side of the PAN spectral feature (Fig. ). However, parts of these windows fall in the AIRS “spectral gap,” where no radiance channels are available. Thus, we had to compromise between windows that will see sufficient signal for PAN absorption and windows that avoid signal from interfering species. Figure  shows the selected windows overlaid on simulated absorption features for the relevant species in this spectral range. The two windows on the left of the PAN feature (below 785 cm-1) only see a weak part of the signal from PAN, but are outside of the CCl4 absorption. The center window at 795 cm-1 is able to capture the core PAN absorption, but has interference from both water and CCl4. The two rightmost microwindows (above 800 cm-1) are able to avoid interference from water, but have minor to moderate interference from CCl4. CCl4 is not retrieved (Table ) but is simulated in the radiative transfer as an interferent, using climatological profiles scaled by yearly scale factors derived from ground based observations. The base climatological profiles vary with latitude and longitude in 30 and 60° bins, respectively, and were developed from MOZART model output .

Early tests with the three windows above 790 cm-1 showed that omitting the 795 cm-1 window gave erroneously high PAN column average values across much of the western United States during a period when the Pole Creek Fire was emitting PAN among other species,. Similar tests also showed no benefit to adding additional microwindows above 805 cm-1. The two microwindows below 785 cm-1 were added later to provide the retrieval with some radiance information with PAN but not CCl4 absorption. The specific frequency ranges for each microwindow are given in Table .

Development of the strategy table was straightforward, requiring only the addition of a PAN retrieval step to the standard AIRS strategy table in use by TROPESS to generate AIRS products. Table  enumerates the steps included in this table; the PAN step added is step number 6. The choice to place the PAN retrieval immediately following the “strong features” step (no. 4) follows . Step number 5 was also added to enable saving of spectral residuals in a wider range of frequencies centered on the PAN feature. Such a diagnostic is helpful to understand what factors might be affecting a given retrieval and proved valuable for filtering (Sect. ).

The a priori constraints used in the AIRS PANs retrieval are mostly the same as the CrIS PANs retrieval, with the exception of surface emissivity. As in , the PAN profile used as the a priori constraint for each sounding is selected from a set of 6 climatological profiles for each month (Fig. ) and the initial PAN profile used as the starting point for the nonlinear optimization is a flat 0.3 ppb in the troposphere. Likewise, the a priori covariance for the PAN VMRs is the same as in . These constraints derive from those used in the retrieval of PAN from the Tropospheric Emissions Spectrometer . For surface emissivity, we used the Combined ASTER MODIS Emissivity over Land (CAMEL) database for our inital and a priori constraint on surface emissivity. used the University of Wisconsin Cooperative Institute for Meteorological and Satellite Studies High Spectral Resolution database . Note that all TROPESS products starting from v1.16 now use the CAMEL database; this included the CrIS PAN retrievals we use for comparison in Sect. .

During development, we found that low, warm clouds over ocean would be misinterpreted by our AIRS retrieval as PAN. For the cases tested, we were able to filter out such soundings by decomposing the AIRS radiances into empirical orthogonal functions (EOFs) and filtering soundings for which the second principle component (PC) was below a threshold. This section describes that approach.

This issue of certain clouds appearing as PAN in the retrieval can be seen in Fig. , which shows free tropospheric column averages of PAN (which we will refer to as XPAN) from 11 September 2020. This was a period with major wildfires throughout the west coast of the United States. In Fig. a, XPAN retrieved from CrIS shows a reasonable plume structure, with clear advection of PAN from the fires on the west coast. We also see some of this in the AIRS retrievals – specifically, the enhanced PAN in the southern half of California, most of Arizona, and the northwest corner of Mexico, as well as over the northern Pacific Ocean (Fig. b).

However, in the black box (20 to 30° N, 142 to 122° W), CrIS shows mostly background column whereas the AIRS retrievals show an enhancement with an unusual structure (not a shape representative of transport from the fires). When we check RGB imagery from the GOES-West Advanced Baseline Imager (https://noaa-goes17.s3.amazonaws.com/index.html#ABI-L2-MCMIPC/2020/255/22/, last access: 8 December 2022), we clearly see that this “plume” seen by the AIRS PAN retrieval matches the shape of the clouds in that area (Fig. a). Further, cloud properties from the MODIS-Aqua MYD06 product plotted in Fig. b–d show that this is a low, warm cloud. This clear spatial correlation between the cloud extent and the spurious PAN plume leads us to conclude that such low, warm clouds cause difficulties for our retrieval with the chosen spectral windows (Table ). Similarly, in the plume around 50°N, AIRS sees enhanced XPAN further west than CrIS (around 150° W) and more to the northwest of the state of Washington (near 50°N, 125°W). From the cloud properties shown in Fig. , these are also potential cases of erroneous impact from clouds.

The AIRS data shown in Fig.  are those soundings which pass prototype quality flags chosen based on quality flags for other thermal retrievals, including sufficiently small radiance residual, surface temperature > 265 K, cloud top pressure (as retrieved in our algorithm) below the tropopause, and the quality of the H2O retrieval in step 4 of Table . (Note that these quality flags were for prototyping purposes only, and are not those used in the final product.)

Since these criteria were insufficient to remove the spurious plume, we investigated an approach inspired by . As they used an empirical orthogonal function (EOF) decomposition to study dominant patterns of variability in the AIRS data, we tested whether an EOF decomposition could identify the low, warm clouds causing the spurious PAN signal in our AIRS PAN retrieval. We do note that a cloud-clearing approach, like that used in CLIMCAPS , could be one approach to address this issue. Such an approach combines radiances from multiple soundings to yield radiances unimpacted by clouds see Sect. 3 and Fig. 1 offor a description of this approach . However, the MUSES algorithm is designed to operate on individual soundings. Therefore, we focused our efforts on the EOF decomposition as a way to screen out these cloud-affected soundings.

Figure  shows the first three EOFs resulting from a decomposition of the AIRS observed radiances (as stored by the MUSES algorithm in its output radiance files) within the domain covering 20 to 60° N and 150 to 110° W. Keeping in mind that the sign of an EOF is arbitrary, as it can be flipped by changing the sign of the principal component (PC) by which it is multiplied, the first two EOFs contain many features which match up closely in shape to the H2O spectral features shown in the top panel. The third EOF appears to relate to CO2, as the dominant feature appears at approximately the same frequency as the CO2 feature shown in Fig. . These EOFs were computed from the window used in step 5 of the strategy table (Table ), which spans 760 to 860 cm-1.

For each AIRS sounding, the observed radiances can be represented as the linear combination of the EOFs with the PCs as the coefficients. Figure  shows the values of the PCs for the first two EOFs needed to reconstruct the AIRS radiances for all the soundings in this domain. The second PC (Fig. , right panel) has a spatial pattern of negative values strikingly similar to the clouds seen in Fig. . In Fig. c, we show the AIRS XPAN with a filter based on the values of PC 2 applied. Filtering out soundings with PC 2 < 0 removes the spurious XPAN.

As a next step, we applied this PC-based filtering to different region. We chose a PAN plume from the Australian Bush Fires in late 2019/early 2020. For this Australian fire case, we also see a collection of soundings with large XPAN values in the AIRS data but not the CrIS data, marked by the black box in Fig. a and b. The MODIS cloud properties (Fig. d–f) confirm that this is again a low, warm cloud. When we apply the PC-based filter, it correctly marks these soundings as bad quality and removes them (Fig. c).

Fortunately, it appears that soundings over land are not as susceptible to this issue with low, warm clouds. Figure  shows XPAN retrieved from both CrIS and AIRS again along with MODIS-Aqua cloud properties, this time over the Amazon. While the AIRS XPAN (Fig. b) shows sporadic high values compared to CrIS (Fig. a), these erroneous high values appear to be random, rather than systematically located where the low, warm clouds are. In particular, the western swath shows mostly low XPAN values despite the presence of low, warm clouds. Therefore, we apply the PC-based filter only to ocean soundings.

Our hypothesis is that the reason the AIRS retrieval is affected by the low, warm clouds and CrIS is not is due either the difference in spectral windows used between the retrievals (Fig. ), the difference in radiance noise between the instruments, or a combination of the two. Further, our hypothesis for why land soundings are much less impacted than ocean soundings is that it is more difficult to distinguish a low, warm cloud from an underlying ocean surface than a land surface.

While the PC-based filter was successful in these cases, we note that it may need additional adjustment in the future. During testing, we found that filtering out soundings with PC 2 < -10 was sufficient for the US West Coast Fires case (Fig. ), but not the Australian Bush Fires case (Fig. ), whereas requiring PC 2 < 0 worked for both. Future work will examine whether the criterion of PC 2 < 0 is sufficient globally, or if further refinement is necessary.

We also note that it was necessary to use the 760 to 860 cm-1 window, rather than the narrower windows used in the PAN retrieval step (Table ). When we tested the latter, this PC-based filter was not effective in the Australian fires case. Therefore, we conclude that information available in the wider window provides the necessary data for the EOFs to correctly fit clouds.

While the PC-based filter addresses the issue of interference from low, warm clouds (Sect. ), it is not sufficient by itself as a quality filter. Ideally, quality filters would be derived by comparing the satellite product to in situ data and checking that the filters ensure good agreement between the satellite and in situ data. To this end, used aircraft profiles from the ATom campaign to validate the CrIS PAN product. This was ideal for CrIS, as the ATom flights provided profiles of PAN over the majority of the troposphere. However, the majority of the ATom profiles are over ocean see Fig. 1 of, and this is also true for HIPPO, a similar campaign that occurred before the start of the CrIS FSR product and so not be used by. As we will lose any profile comparisons that occur over ocean clouds, this would likely limit the number of comparisons we can draw from HIPPO and ATom. To add to the challenge, from Fig. , we can see that outside of strong PAN plumes from, e.g., fires, the single sounding retrievals over land from AIRS have significant sounding-to-sounding noise, indicating that bulk statistics will be necessary for a meaningful comparison.

Thus, instead of relying on aircraft data directly, we decided to use the existing CrIS PAN product as a transfer standard by designing a quality filter that predicts whether the AIRS XPAN value will be within a given threshold of the nearest CrIS XPAN value. This provides the large number of soundings needed for bulk statistics and implicitly makes the AIRS PAN product consistent with the CrIS PAN product, which can allow users to combine the two.

We chose to implement this quality filter using decision trees, with the Scikit Learn package . Using simple decision trees allowed us to investigate what variables were used to classify a sounding as good or bad quality during development. Using decision trees rather than hand-tuned quality filter parameters allowed faster iteration and should, in principle, be more reproducible. Because we saw in Sect.  that ocean soundings required different filtering for clouds than land soundings, we also tested whether using a single decision tree for all soundings or separate decision trees for land and ocean soundings gave better results. We found that separate decision trees for land and ocean soundings retained more soundings with significant XPAN, and that there was little difference in the correlation between AIRS and CrIS XPAN using separate land/ocean trees or a single tree. Therefore, we chose to use separate decision trees. Appendix  shows a subset of results using a single decision tree, and describes the trade offs between using a single decision tree and separate decision trees.

The decision trees were trained on AIRS and CrIS retrievals for one day from each of the 2019/2020 Australian Bush Fires and 2020 US West Coast Fires. Ocean soundings that failed the PC-based filter (Sect. ) were excluded from training. Two different days from these fires, plus retrievals over the Amazon and Africa were used for testing (Table ). The data was divided into training and testing by days and regions rather than a random 70/30 or similar stochastic split to ensure that the training data included at least some soundings with significant XPAN. Since plumes with significant XPAN are outnumbered by background soundings, we were concerned that a fully random split would miss the plume soundings.

As inputs, the decision trees received 16 values commonly used by existing MUSES retrievals as quality metrics, listed in Table . It was trained to predict a binary flag indicating whether the AIRS XPAN was within 0.2 ppb or 50 % of the CrIS XPAN from the CrIS sounding closest to it (by great circle distance). The CrIS soundings are restricted to those that pass basic quality flagging for modeled vs. observed radiance and a check for certain surface features that can cause erroneous retrievals. The CrIS XPAN value compared against includes an averaging kernel adjustment to accommodate different vertical sensitivity between CrIS and AIRS. (see Fig.  for a summary of typical CrIS and AIRS XPAN column averaging kernels.) Specifically, following Eq. (25) of , 2c^CrIS,comp=ca,AIRS+aT(x^CrIS-xa,AIRS) where

ca,AIRS is the a priori XPAN from AIRS,

Note that the CrIS XPAN is not an input to the decision trees; it is used only in training. This permits the decision trees to be applied to AIRS soundings without a coincidence CrIS sounding.

Typically, it is important to “prune” decision trees by limiting the number of decision nodes it can include in order to prevent overfitting to the training data. We tested pruning by limiting both the maximum depth (i.e., the number of nodes along any one path) and maximum number of leaf nodes (i.e., the number of end points for the model). However, we found that either method of pruning the decision trees caused the filter to screen out soundings with enhanced XPAN. Our hypothesis is that, because these soundings are still in the minority of all soundings in the training data, limiting the decision tree's size gave it too little flexibility to account for these somewhat uncommon cases. That is, because soundings with enhanced XPAN are in the minority, a model limited in size lacked the flexibility to develop useful rules for these soundings, and instead was able to achieve better accuracy by simply classifying all such soundings as bad quality. Therefore, we proceed without limiting the model size.

Additionally, we include an explicit check that the retrieved surface emissivity at 1025 cm-1 is > 0.94. This filter is similar to one used in to remove soundings impacted by a silicate feature that produces a surface emissivity with a similar spectral shape to PAN. The same silicate feature also shows up as a low emissivity near 1025 cm-1 (see Appendix ). Although the decision trees are trained on this value as an input, it still retains some soundings clearly affected by the silicate feature. Figure  shows CrIS XPAN in Fig. a, AIRS XPAN in Fig. c and d, and the emissivity value in Fig. b. The red or black box identifies a region with low 1025 cm-1 emissivity values that has very high XPAN values in the AIRS retrieval in Fig. c. When we add an explicit filter on the 1025 cm-1 emissivity, those few remaining soundings are removed.

The final quality filter will be a combination of the PC-based filter from Sect. , the emissivity-based filter, and the decision tree-based filter. Figure  shows how each of these filters affects the soundings passed as good quality for two days with clear PAN plumes. As discussed in Sect. , the PC-based filter is applied only to ocean soundings, where clouds cause a high bias in XPAN. For these two scenes, the emissivity filter has a modest impact, removing some soundings in southern California, northeastern Arizona, and southeastern Utah (Fig. g). In both scenes, the decision tree-based filter does remove a number of the soundings with large XPAN values (Fig. d,h). Therefore, in the public files, we will provide the information for users to adjust the quality flagging to suit their application; specifically the PC value used for flagging, the emissivity value used for flagging, and the binary flag produced by the decision trees.

For the rest of this section, we will focus on the performance of the combined filter. Appendix  contains a brief exploration of the relationship between the input variables and predicted quality flag.

First, we examine the spatial distribution of PAN plumes in our filtered AIRS product versus CrIS. Figure  shows our filtered AIRS PAN data alongside the PAN retrieved from CrIS. The data shown here are from the four testing data region/day pairs in Table ; thus, these are data that the decision trees were not trained on. The first two rows show the Australian 2019/2020 Bush Fires and the 2020 US West Coast Fires, respectively. In both cases, we can see that the AIRS PAN product matches the location of enhanced PAN plumes seen in the CrIS data very well. In the US West Coast Fires case, the large XPAN values in Arizona, central/southern California, and northwestern Mexico are all in the same region where CrIS sees high XPAN values. Likewise, in the Australian fires case, AIRS captures the PAN plume approaching New Zealand's northern island, though compared to CrIS, more of the plume is removed by our filtering criteria.

The last two rows of Fig.  show a day over the Amazon and central/southern Africa, respectively. These are regions not included in the training data for the decision trees (Table ), so these are a good test of whether the filter can generalize to new regions. Neither region has significant PAN plumes in the CrIS data. However, there are small enhancements to ∼ 0.5 ppb in both cases. In the Amazon, there are also a few soundings with ∼ 1 ppb XPAN near 17.5° S, 55° W. AIRS does see this 1 ppb hotspot, though it also retrieves several soundings with > 1 ppb further north, where CrIS does not. The Amazon hotspot in western Brazil cannot be seen in AIRS due to the swath gap. The PAN hotspot seen by CrIS in the African test over Angola, Zambia, and the Democratic Republic of the Congo is not as apparent in the AIRS PAN; however, AIRS does appear to capture some enhancement in that area, particularly compared to further north, near the equator.

Helpfully, in most of these cases, when there is a strong PAN enhancement in CrIS, AIRS also sees an enhancement in CO. For example, in the Amazon test case, only the soundings with enhanced PAN at 17.5° S, 55° W also have a strong CO enhancement; while the false enhancements further north in the AIRS PAN do not. This implies that users looking for PAN plumes in the AIRS data can check for enhanced CO to distinguish whether a small PAN plume is likely real. This is not a entirely self-sufficient condition, as it is possible to have a PAN plume without enhanced CO, but the presence of enhanced CO can give more confidence in an observed PAN plume. (See Sect.  for a summary of recommendations for use.)

We also tested the correlation between AIRS and CrIS XPAN with different amounts of spatial averaging. Figure  shows the results for four different spatial averaging box sizes. While the data from our test cases does have some fire-influenced observations, many of the observations vary primarily from large-scale seasonal or latitudinal variations. At 1° × 1°, the correlation is somewhat weak. The correlation is more significant at 2° × 2°, 5° × 5°, and 10° × 10°. However, averaging to 5° × 5° or 10° × 10° is needed for the root mean squared error (RMSE) between AIRS and CrIS XPAN values to drop below 0.1 ppb and for the visual correlation (especially for high values) to be apparent. This is a fair amount of averaging, but is not surprising, given the sounding-to-sounding variation seen in Fig. . Given the amount of observations, this will still provide useful PAN coverage. We discuss recommendations for use based on this result in Sect. .

In Fig. d, we see that the AIRS XPAN value is biased low compared to CrIS XPAN. It is not clear if this bias in the AIRS data is best parameterized as a function of H2O, as was the case for CrIS, or if another parameter is a better predictor. were able to derive the CrIS bias correction through comparison between CrIS and in situ background XPAN values. In this work, the need to average a significant number of AIRS soundings to reduce the random sounding-to-sounding noise makes it difficult to identify any relationship between AIRS XPAN values and H2O column amounts.

The CrIS radiance noise is lower than the AIRS radiance noise, which is a significant advantage when retrieving species, such as PANs, with only weak absorption features. Figure  shows per-channel median and 25th to 75th percentile noise equivalent spectral radiance (NESR) values. Although we use different frequencies in the AIRS and CrIS retrievals, the AIRS NESR values are systematically greater than the CrIS values. Taking all of our test cases for comparing AIRS and CrIS (Table ), we find that the median ratio of AIRS to CrIS NESR across all channels is ∼ 5.9. Assuming that single sounding uncertainty scales linearly with radiance noise, that suggests that the AIRS single sounding uncertainty in XPAN should be approximately 0.5 ppb, that is, approximately six times the 0.08 ppb value calculated for CrIS. This aligns with the correlation between AIRS and CrIS XPAN shown in Fig. , which shows that AIRS values below 0.5 ppb are dominated by random uncertainty without significant averaging. We also checked the correlation between individual AIRS and CrIS XPAN values in Fig. , and similarly see that the values have a spread of ∼ 0.5 ppb. While we expect the error of individual soundings to vary depending on the specific atmospheric and surface conditions for each sounding, we believe 0.5 ppb to be a reasonable estimate of the typical uncertainty in the AIRS XPAN data.

Figure  compares the pressure-weighted column averaging kernels and the sum across the rows of the averaging kernel for AIRS and CrIS for good quality land soundings within the US West Coast Fires domain on 11 September 2020. The averaging kernels shown are the medians of averaging kernels for soundings binned by surface temperature. For both instruments, maximum sensitivity shifts to lower pressure with decreasing surface temperature. However, compared to CrIS, AIRS maximum sensitivity decreases more quickly as surface temperature decreases. We suspect this is due to the greater noise present in the AIRS radiances, with AIRS sensitivity decreasing more with reduced thermal contrast due to the greater noise. However, we have not confirmed this hypothesis. Note that, for both instruments, the averaging kernels shown in the left panels incorporate the pressure weighting function, which is why the values are well below 1.

Figure  shows the overall degrees of freedom (DOF) of signal for both the AIRS and CrIS products in the 11 September 2020 US West Coast Fire scene. From Fig. a and b, we can see that the DOFs for the CrIS PAN product are grouped around 1, indicating that there is essentially always enough information to retrieval a single piece of vertical information in the form of a column average. In contrast, Fig. c and d show that the AIRS DOFs are lower (centered around ∼ 0.5) with a wider distribution. Greater AIRS XPAN values do tend to be associated with greater DOFs. This implies that the AIRS product will retain influence from the prior, particularly in background conditions, but can detect sufficiently large PAN enhancements.