1. Introduction
All-sky assimilation of cross-track scanning humidity sounder radiances has been operational in the Integrated Forecasting System (IFS) since 2015 (Geer et al. 2014). This covers observations from the Microwave Humidity Sounder (MHS) on several NOAA and MetOp platforms and the Microwave Humidity Sounder-2 (MWHS-2) on the operational Fengyun-3 (FY-3) platforms. These sensors hold multiple channels sampling the strong water vapor absorption feature at 183.31 GHz, providing profile information on the distribution of humidity in the troposphere. The MWHS-2 sensors also provide a mixture of temperature and water vapor information via their 118-GHz channels. Channel details are found in Table 1. As of 2024, there are three MHS and three MWHS-2 sensors actively assimilated in the IFS, with FY-3E MWHS-2 added in February 2023. Note that MW humidity-sounding information is also assimilated in the IFS from the Advanced Technology Microwave Sounder (ATMS) instrument, but through a clear-sky approach, and assimilation of this instrument is hence not changed in the present work. Furthermore, the ECMWF system also uses MW humidity-sounding observations from conical scanners like SSMIS and GMI, but given their different observing geometry, they are not considered here.
Channel numbers, central frequencies, and peak of the weighting functions at nadir for MHS and MWHS-2. Weighting function peaks are quite approximate, as these depend on the atmospheric profile of humidity and also vary with the view angle.
The addition of microwave sounders to the assimilation system has shown consistent benefit for forecast skill in the short to medium range (Duncan et al. 2021) when considering temperature and humidity sounders added in tandem. Continued benefit is also seen when adding further humidity sounders (hereafter H-sounders) on their own (Geer et al. 2017b; Lean et al. 2022). This has recently been exemplified by the additions of MWHS-2 on FY-3C and FY-3D at ECMWF (Lawrence et al. 2018; Bormann et al. 2021; Duncan and Bormann 2020), as well as all-sky use of H-sounders in other numerical weather prediction (NWP) systems (e.g., Candy and Migliorini 2021; Carminati and Migliorini 2021). Much of the benefit is realized through the 4D-Var tracer effect, through which displacement errors in humidity, clouds, and precipitation are corrected via increments to the wind field during the assimilation window (Peubey and McNally 2009; Geer et al. 2014). In this paper, which summarizes and expands on work described in an earlier technical report (Duncan et al. 2023), we consider the addition of H-sounder data by increasing the total amount of information assimilated from existing sensors and platforms. This is accomplished purely by changes to the thinning and averaging of currently available observations.
Data for the study are described in the next section (section 2), followed by the description of the study’s method (section 3). Results are presented first at the lower model resolution, covering thinning changes alone (section 4a), then superobbing alone (section 4b), and then thinning in combination with superobbing (section 4c). Results at the current operational resolution of the IFS are presented last (section 4d), followed by conclusions (section 5).
2. Data
a. Observations
Radiances measured by H-sounders have relatively small field of view (FOV) sizes compared to other passive microwave sensors, a feature possible because of the shorter wavelengths measured. For example, the 183-GHz channels of MHS have an effective FOV of 16 km × 16 km at nadir compared to 48 km × 48 km for 50-GHz channels on the temperature sounder AMSU-A (Bennartz 2000). Cross-track sounders have larger FOVs at higher zenith angles (i.e., near scan edge). For example, MHS at scan edge has an FOV of approximately 51 km × 27 km compared to 16 km × 16 km at nadir (Robel and Graumann 2014). The sampling on the ground is much denser near nadir than at higher zenith angles (see Fig. 1 for an example of FOV sizes across the scan for MWHS-2). A slight difference between MHS and MWHS-2 lies in the wider swath of MWHS-2, enabled by observing out to a greater scan angle (see Table 2). This provides greater spatial coverage for MWHS-2 but does mean that observations at swath edge exhibit an even larger FOV. All MHS and MWHS-2 sensors are on low-Earth orbit (LEO), sun-synchronous satellite platforms.
FOVs for nine consecutive scans (black ovals) and all thinned observation centers at N128 thinning scale (red crosses; thinning described later in section 2b) for an overpass of MWHS-2 over South India and Sri Lanka.
Citation: Monthly Weather Review 152, 8; 10.1175/MWR-D-24-0020.1
Scan characteristics for MHS and MWHS-2. FOV sizes are for 183-GHz channels only—118-GHz FOVs on MWHS-2 are larger (32 km × 32 km at nadir) due to a beamwidth of 2.0° versus 1.1° (He et al. 2015).
Sensor noise for assimilated 183-GHz channels varies substantially but is typically in the range of 0.3 to 1.0 K, given by noise equivalent differential temperature (NEDT). Unlike temperature-sounding channels, where NWP impact is often limited by sensor noise due to background errors in temperature on the order of 0.1 K, the utility of humidity-sounding observations for NWP is not typically noise limited (Bormann and Bauer 2010). Background errors at 183 GHz are on the order of 1 K in clear scenes and >2 K with any cloud present. The larger background errors at 183 GHz are in part because of humidity’s higher small-scale variability and more significant forward modeling errors because of greater scattering from ice hydrometeors. In contrast, the 118-GHz sounding channels on MWHS-2 exist on a weaker oxygen absorption line than those near 50 GHz, and with relatively narrow bandwidths, they thus exhibit higher NEDT. As a consequence, the impact of 118-GHz channels on temperature has thus far been marginal (Lawrence et al. 2018; Bormann et al. 2021) but may improve with better radiometric performance (Maddy et al. 2022), though the larger sensitivity to ice clouds will remain a limiting factor. The other sensitivities of 118-GHz channels—water vapor continuum absorption, cloud water, cloud ice, and precipitation—remain an asset to all-sky assimilation despite the higher NEDT levels.
All 183-GHz channels on MHS and MWHS-2 are actively assimilated in the IFS. The outer scan positions for specific channels on MWHS-2 are excluded from assimilation due to strong bias characteristics (Lawrence et al. 2018; Duncan and Bormann 2020), but otherwise the full swath of both sensors is used. On MWHS-2, the 118-GHz channels 2–7 are assimilated. The details of which surfaces these channels are assimilated over can be found in Table 3 of Geer et al. (2022), describing the treatment in the operational system since the June 2023 upgrade (known as Cycle 48r1). It is sufficient to say that higher-peaking channels are used everywhere, whereas lower-peaking channels like MHS channel 5 or MWHS-2 channels 6 and 7 are used more cautiously over nonocean surfaces.
b. Thinning and averaging
Satellite radiances are traditionally thinned prior to assimilation. This serves dual purposes: It limits the computational costs involved in processing tens or hundreds of millions of observations and removes observations with potential spatially correlated observation errors. Thus, since the implementation of all-sky H-sounders in the IFS, the first step of observation processing has been to thin the data using a N128 reduced Gaussian grid (78 km spacing), selecting the observation nearest to each N128 grid point (shown as crosses in Fig. 1) and discarding all others. A reduced Gaussian grid is a latitude/longitude grid with the grid name (N number) specifying the latitude lines between a pole and equator; these grids have fewer points per latitude band as latitude increases, keeping the east–west grid length approximately constant (Hortal and Simmons 1991). Figure 2 shows a swath from MHS, contrasting unthinned observations with three thinning scales. The primary thinning to the N128 grid is visualized by comparing the left two panels of the figure. Note that for this thinning step the observation has to lie within 25 km of the gridbox center, so data coverage is sparser near scan edge. A second step of thinning is then performed, selecting every other observation to be thinned out in a diamond pattern, leaving observations at roughly 110-km spacing for the assimilation (cf. middle two panels in Fig. 2). The “diamond” pattern thinning discards every other radiance along the row of the grid and then staggers this pattern in the following row; for example, positions 1, 3, and 5 are discarded in row 1, with positions 2, 4, and 6 discarded in row 2, and so on. These observations are at native resolution and not averaged in any way. This type of thinning is applied at other NWP centers, with 100- to 150-km thinning scales typical (Candy and Migliorini 2021; Lee et al. 2020; Shahabadi and Buehner 2024).
MHS observation locations from an overpass of northwestern Europe on 3 Jun 2021. (left to right) Observation locations with no thinning applied, N128 thinning (80 km), N128 diamond thinning (110 km), and N200 diamond thinning (70 km). Gray lines show 2.5° spacing in latitude and longitude.
Citation: Monthly Weather Review 152, 8; 10.1175/MWR-D-24-0020.1
In contrast, microwave imager radiances are averaged by a process known as “superobbing,” a simple averaging of observations that has long been used in direct all-sky radiance assimilation (superobservations; Geer and Bauer 2010). The transformation of several observations into a single superob serves to homogenize observations from sensors and channels with differing FOVs, while making observations more representative of the model grid box they are compared against (i.e., treating representation error). The downside of this approach is that finer-scale information is smeared out due to the averaging. In the current version of the IFS (Cycle 48r1), microwave imager radiances are superobbed at N128 resolution and then thinned in a diamond pattern as described above. There is no a priori reason why sounder radiances should not be averaged, with Geer et al. (2014) stating that “following the logic of the microwave imagers, the MHS observations would ideally have been superobbed, but this has been left for future work.”
Comparing the left and center columns of Fig. 3, we can see the difference between observations thinned and averaged at the same spatial resolution. Each superob is required to contain at least two observations in the grid box, but we remove the 25-km requirement of distance from the gridbox center as used in the thinning-only panels, permitting greater data coverage from the superobs at high zenith angles. The superob latitude/longitude locations are taken from the center of the grid box rather than the average of the latitude/longitude values from all points that were used in the superob. The change in noise characteristics for the noise-dominated MWHS-2 channel 3 signal is quite stark, revealing striping patterns otherwise not apparent. This effect to reveal striping patterns is most clear for the noise-dominated 118-GHz channels but also seen (no figure shown) for MetOp-C MHS, which displays so-called 1/f noise (e.g., Yang et al. 2022), often called pink noise. The effect of superobbing also makes certain wave features clearer, such as ripples in observation minus background equivalent (O − B) in the South Caribbean that appear to be inertia–gravity waves (e.g., O’Sullivan and Dunkerton 1995). The right column of Fig. 3 shows a finer resolution of superob at N200, about 50 km in the right column versus 80 km in the center column. Comparing the two superob resolutions, smaller convective features such as those in the ITCZ are better resolved by the finer superob of the right panels. The smaller N200 superobs may prove useful for diagnostic purposes such as identifying cloud features and misplaced convection at finer scales.
O − B over Central America from four channels on FY-3D MWHS-2: (top) 3, (second row) 7, (third row) 11, and (bottom) 15. (left column) Radiances at native resolution thinned at N128, (middle column) superobbed at N128, and (right column) superobbed at N200. Data are from the 0000 UTC long window on 5 Jun 2021.
Citation: Monthly Weather Review 152, 8; 10.1175/MWR-D-24-0020.1
c. Total information content
The goal of data assimilation (DA) could be considered the full, optimal utilization of all available observational information content to constrain and guide an Earth system model. In the context of all-sky microwave assimilation, this means exploiting the full spectral and spatial information content of radiances over all surfaces and for all atmospheric conditions. Moving from clear-sky to all-sky assimilation methodologies is a key element of realizing this goal, as is the gradual addition of more spectral information by adding channels that were previously not assimilated. There is also ongoing development regarding mixed scenes and surface emissivity, part of a larger strategy to assimilate microwave radiances more widely over all surfaces (Geer et al. 2022).
To illustrate the challenge of using all available information content, we can revisit the example of Fig. 1. If we view all observations from a humidity sounder and also consider spatial variations across the scan (visualizing O − B not as dots but as ovals matching the sensor’s FOV), we can see how complex the information really is. Figure 4 shows an upper-tropospheric sounding channel. Small gravity wave ripples not captured by the model are visible on the north–south axis near 72°E, as is convection east of Sri Lanka with possible cold pools to the east and west, likely misplaced convection in the model background near 87°E, and so on. These features span scales up to roughly 300 km. In the figure, crosses mark observations currently ingested into the IFS prior to secondary thinning (see section 2b). Along with Fig. 2, this gives a sense of how much information is discarded by thinning. Quantitatively for the N128 (80 km) thinning, as few as 4% of observations are retained near nadir where radiances have 16-km spacing.
O − B near India and Sri Lanka from MWHS-2 channel 12. Each is roughly the shape of the sensor’s FOV. Crosses mark the observations ingested into the IFS with current thinning procedures, i.e., passing the N128 thinning as in the 48r1 control experiment.
Citation: Monthly Weather Review 152, 8; 10.1175/MWR-D-24-0020.1
In contrast to all-sky usage and better spectral coverage, the spatial aspect of the overall information content from microwave radiances has received less attention. It is common in DA to select a thinning scale that avoids most spatial error correlations. This makes it safer to treat all observation errors as being wholly uncorrelated, guided by studies such as Bormann and Bauer (2010) which found roughly 80–120-km error correlation scales for MHS radiances in clear sky. However, error correlations depend on the synoptic conditions, and thus, selecting specific situations to use denser observations should hold some benefit (Dando et al. 2007). Furthermore, all-sky correlation lengths have greater sensitivity to mesoscale meteorology because of the sharper spatial inhomogeneities of clouds and precipitation. As seen in Bonavita et al. (2020) (their Fig. 4), upper-tropospheric humidity channels exhibit roughly 100-km error correlation lengths with a maximum of about 200 km in the extratropics at 300 hPa, though shorter scales would be expected if considering tropical convection. Smaller-scale information may be extracted from observations with correlated errors (Bédard and Buehner 2020), but there is a long-held view that assimilation of observations finer than the error correlation length scale holds little benefit even if correct covariances are assigned (Bergman and Bonner 1976). In practical applications where full errors are not prescribed, Hoffman (2018) concludes that “the best strategy is superobservation with a spacing approximately equal to the observation error correlation length scale.”
A perfect DA system for radiances would ingest all observations and extract all relevant information content with comprehensive consideration of the interchannel, spatial, and even time-dependent error correlations. To be clear, none of these are treated explicitly in this paper. Rather, this paper aims to explore assimilating more radiances in the IFS without resorting to a more complex treatment of such errors. The exploration of finer thinning scales assumes that current thinning may remove more data than is called for by error correlation lengths typical for all-sky humidity radiances. Coupling this with superobbing follows Hoffman (2018), in that the errors assigned are imperfect and do not account for spatial correlations.
3. Method
a. Experimental setup
The experiments discussed all use IFS Cycle 48r1, and most cover two seasons: June–August 2021 and December 2021–February 2022. This was a time period in which the following H-sounders were active: MHS on MetOp-A, MetOp-B, MetOp-C, and NOAA-19; and MWHS-2 on FY-3C and FY-3D. FY-3D 118-GHz channels (2–7) experienced an anomaly that strongly affected their mean bias in January 2022 and were thus removed from active assimilation. All experiments use a full observing system that follows that used in ECMWF operations. Final model resolution (i.e., outer loop) is at TCo399 (29 km) with 137 vertical levels. The data assimilation uses incremental 4D-Var performed at a final inner-loop resolution of TL255 (80 km) with 12-h delayed cutoff assimilation cycles. These outer- and inner-loop resolutions are coarser than in ECMWF operations but constitute the standard testing framework used for the IFS. Background errors come from the operational ensemble of data assimilations (EDAs) due to availability (i.e., 47r3, rather than a bespoke 48r1 EDA), but this is not expected to significantly alter the results’ interpretation (Duncan et al. 2021). IFS Cycle 48r1 has been operational at ECMWF since June 2023, and full documentation is available (ECMWF 2023).
The experiments (named in Table 3) are analyzed in three sets to gauge the impact of the following on H-sounder all-sky assimilation:
Experiment names and characteristics. Diamond refers to thinning out every other spot in a diamond pattern. Change in assimilated observations (Δ obs) is relative to control and is an approximate mean across all H-sounder channels given as a percent difference. N-T stands for no thin; i.e., no secondary thinning is applied. Approximate spacing is the distance between Gaussian grid points for assimilated radiances within the sensor’s swath.
-
Finer thinning (N128 N-T)—section 4a.
-
Superobbing (N128 Superob)—section 4b.
-
Superobbing and finer thinning (N128 Superob N-T; N200 Superob)—section 4c.
Additionally, to examine the question of optimal superob resolution, a separate set of experiments was run that covered a shorter time period from late February to late April 2022. These experiments are listed in Table 4. In these experiments, thinning was tuned to largely conserve the total number of observations assimilated so that the main signal from changing the superob resolution was the resolution change itself rather than the volume of data. Results for this are found in section 4b(2). It is worth noting that to conserve observation numbers in these particular experiments it is not possible to retain evenly spaced radiances as provided by Gaussian grid thinning and subsequent diamond thinning—the N160 Superob and N256 Superob experiments have unevenly spaced radiances after thinning as a result.
As in Table 3, but changes in assimilated observations (Δ obs) are given relative to N128 Superob. For second thinning, lon = 3 keeps every third longitude point on the grid, lat = 2 keeps every second latitude point, and so on (diamond thinning is equivalent to lon = 2).
A last set of experiments was run at higher resolution. These use the final model resolution of the current operational forecast, 9 km. The DA was also performed at a finer resolution, with the final inner-loop resolution now 40 km as of Cycle 48r1 (ECMWF 2023). Thus, the outer- and inner-loop resolutions of these high-resolution (hereafter HRES) experiments are significantly finer than the others tested, i.e., 9- versus 29-km model and 40- versus 80-km DA. As both the spacing and the superob size could interact with the model and DA resolutions, comparing these experiments with the ones above should illuminate such resolution dependence while also showing the impact in an operation-like setup. Only the final thinning and superobbing configuration (N200 Superob) was tested at HRES due to computational considerations. One additional change to the previous setup is that 118-GHz radiances (MWHS-2 channels 2–7) have an extra thinning applied; this brings 118-GHz radiances to 100-km spacing for assimilation, aligning their treatment more closely with AMSU-A and intending to avoid the spatially correlated errors due to striping noise evident in Fig. 3. The HRES experiments were run for two seasons: June–August 2020 and December 2020–February 2021.
b. Observation error modeling
Changes to the thinning alone do not have a large impact on std(O − B) for sounder channels, and thus, in these cases, there is no need to consider retuning the observation error model. However, superobbing H-sounder radiances does prompt consideration of whether the extant error model is sufficient. For expedience and easier interpretation of comparisons between experiments, the experiments discussed here do not modify the observation error models. However, there are significant changes in std(O − B) for noise-dominated channels as a result of superobbing. The higher-peaking 118-GHz channels on MWHS-2 (2–5) witness a large decrease in std(O − B) from superobbing that can be 50% or greater, seen vividly in Fig. 3 for channel 3. Higher-peaking 183-GHz channels also exhibit a decrease in std(O − B). Notably, this change in std(O − B) behavior is not even across the scan, as more observations are averaged together near nadir and fewer near scan edge, and thus, the decrease in noise is most pronounced at low zenith angles; this could perhaps justify a zenith-dependent adjustment to the error model that mirrors what is done for AMSU-A channel 5 (Duncan et al. 2022b; their Fig. 3).
Figure 5 shows the interaction between observation error modeling and the choice to superob. Comparison of the dotted and solid lines shows the pure impact of superobbing, as these two experiments used the same observation error models. Most noticeable in the plots is the change in magnitude of std(O − B) for MWHS-2 channels 2–6, but the 183-GHz channels also show a shift that is most pronounced for low zenith angles, as much as 12% change for the highest peaking one, channel 11. This indicates that the impact of superobbing on normalized departures differs significantly between channels, and for some channels, it is also a function of zenith angle. To fully exploit this increase in information content due to superobbing, the assigned observation errors should ideally be revisited, but this has been left for future work.
Standard deviation of O − B normalized by the assigned observation errors, given as a function of the zenith angle for assimilated radiances at native resolution (dotted) and N128 superob (solid). Each panel shows different channel groups: (left) 2–5, (center) 6–7, and (right) 11–15. Statistics comprise used data from FY-3D MWHS-2 over sea, 1–10 Aug 2021.
Citation: Monthly Weather Review 152, 8; 10.1175/MWR-D-24-0020.1
4. Results
a. Finer thinning
As a first test, we remove the second thinning (see Table 3). This causes a doubling of H-sounder radiances assimilated, with the net spacing changing from roughly 110 to 80 km. The effect of doubling the radiances assimilated has a noticeable impact and can be seen in background departure statistics for various independent observations as shown in the black lines of Fig. 6. Short-range forecasts of humidity and winds are improved, evidenced in tighter O − B fits of independent observations like geostationary (GEO) infrared radiances and 500-hPa winds. The roughly 0.5%–1% improvement for several infrared and microwave humidity-sounding channels is a remarkable signal. To put this in context, the activation of FY-3D MWHS-2 in the IFS had a smaller impact on these metrics (Duncan and Bormann 2020). Although ATMS humidity channels are not entirely independent in this context, the improvement for GEO and LEO infrared channels is of a similar magnitude and this gives confidence in the signal. Winds are slightly improved at upper-tropospheric levels as well, presumably through the 4D-Var tracing effect from better representation of humidity structures. The only slight degradation seen here is for upper-troposphere and lower-stratosphere (UTLS) temperature channels such as 8–10 on ATMS, which is a minor change of about 0.1%. This small degradation is not seen in equivalent IASI or Cross-track Infrared Sounder (CrIS) channels but may be a consequence of a slightly noisier analysis near the tropopause or weighting the 118-GHz channels too heavily in the analysis.
Change in std(O − B) for independent observations caused by the removal of secondary thinning (N128 N-T; black line) and superobbing (N128 Superob; gray line). Both are compared to control, the 100% line. These panels show global data from (a) ATMS, (b) IASI, (c) nonsatellite winds, (d) GEO infrared, (e) radiosonde humidity, and (f) radiosonde temperature. For GEO infrared satellites, M is for Meteosat, H is for Himawari, and G is for GOES, with bands given in micrometers. Horizontal bars denote 95% confidence intervals. Nonsatellite winds include observations from radiosondes, aircraft, and wind profilers.
Citation: Monthly Weather Review 152, 8; 10.1175/MWR-D-24-0020.1
The changes in medium-range forecast scores for winds and geopotential height are shown in Fig. 7. For most of the scores and time ranges shown, the change in forecast scores does not exceed the 95% confidence limits (following Geer 2016); however, the Southern Hemisphere does appear slightly improved for low-level winds and Z500 RMSE in the medium range. In contrast, the impact in the Northern Hemisphere and tropics looks quite neutral. Results for lead times < 2 days are not shown to avoid own-analysis artifacts.
Forecast score changes for either VWs or geopotential height Z RMSE at (top) 200, (middle) 500, and (bottom) 850 hPa. The reference (0% line) is control (48r1), with N128 N-T given by the black line. Verification is against own analysis with confidence intervals of 95% following Geer (2016) in the vertical lines.
Citation: Monthly Weather Review 152, 8; 10.1175/MWR-D-24-0020.1
b. Superobbing
1) To superob or not
As seen in Fig. 3, superobbing H-sounder observations cleans up the image in some ways that are potentially helpful for diagnostic purposes while allowing more total information to enter the assimilation. However, will the shift to superobbed radiances have any positive effects on assimilation? To examine this, first we directly compare histograms of observed radiances at different superob resolutions with those from the model background.
The PDF of observed radiances over sea is relatively unaffected by the choice to superob or indeed the superob resolution (Fig. 8). For the most part, it is only at the very cold end of the Tb distribution (below about 220 K for MHS channels) that the superob resolution has a noticeable effect, but in these areas coarser averaging helps to bring the observations closer to the model PDF. Interestingly, there are quite rare channel 3 observations > 280 K that are almost never seen in the background or superob PDFs; these appear to be observations of small-scale, thick liquid clouds with a very dry free atmosphere above them. The main mismatch between observations and background does not appear to be due to averaging of radiances, but rather systematic underrepresentation of highly scattering scenes in the model background and too many scenes with moderate levels of scattering, i.e., those about 230–240 K for the channels shown. This recalls Geer et al. (2017a, their Fig. 16) and Geer and Bauer (2010, their Fig. 14), who showed that the correspondence between PDFs of observed and modeled Tb was relatively poor for a microwave imager at 10 or 19 GHz, regardless of how the observations were averaged together; they pointed to the significance of model bias in simulating clouds, which is a factor that overwhelms the small effect to the PDF caused by superobbing. Geer and Bauer (2010) reasoned that theoretical considerations should then guide the choice of superob size, with N128 being a compromise spatial resolution that was adequate for the scale of clouds in the operational model at that time while not overly increasing computational cost.
PDFs of MHS observations from channels (top) 3 and (bottom) 5 at native resolution, two superob resolutions, and modeled radiances from the background at two model resolutions. Data are limited to ice-free ocean scenes from MetOp-B, MetOp-C, and NOAA-19, spanning 3 weeks in December 2021.
Citation: Monthly Weather Review 152, 8; 10.1175/MWR-D-24-0020.1
To view the short-range forecast impact from superobbing H-sounder observations in isolation, we can again look at the fits to observations in Fig. 6, now focusing on the gray line. This reveals a similar impact to that seen from doubling the radiances assimilated, but of a smaller magnitude. Both changes improve the short-range forecast’s fit to humidity-sensitive observations of the middle and upper troposphere such as other 183-GHz channels and infrared channels on GEO and LEO platforms (such as IASI wavenumbers 1334 and above); fits to wind observations and radiosondes are neutral to slightly positive for superobbing but more positive for the thinning change. The magnitude of the improvement for humidity-sensitive channels is about half that of the data doubling, but this is still noteworthy as the only change here is averaging the radiances. It is logical to wonder what impact these two changes might have when combined, and that is addressed in the next section. Medium-range forecast impacts from superobbing alone were entirely neutral, so no further figures are shown here.
A potential, intriguing side effect of superobbing H-sounder radiances may be increased impact from the 118-GHz channels on MWHS-2. These channels are limited in their influence due to high radiometric noise, and thus, a significant decrease in the effective noise caused by superobbing (seen in Figs. 3 and 5) could permit a greater impact on the analysis from these channels. However, we do not see direct evidence of any greater impact from the 118-GHz channels due to superobbing. To analyze this aspect specifically would require dedicated experimentation and possibly also retuning of observation errors for these channels, and so this can be considered an avenue for future investigation.
2) Optimal superob resolution
Following the clear benefits of superobbing H-sounder radiances seen in the previous section, it is natural to consider whether some spatial resolution of averaging might be optimal. It is worth considering that the native resolution of H-sounders near their scan edge is considerably larger than at nadir, as large as 51 km × 27 km for MHS at the maximum scan angle (Robel and Graumann 2014) and larger still for the 118-GHz channels on MWHS-2. Thus, superob resolutions that are finer than FOV size (at least for some scan positions) could be considered problematic. As discussed by Geer et al. (2017a), the ideal situation would be observations that match the “effective model resolution” so that spatial representation error is minimized. Effective model resolution is understood to be roughly three to four times the final model resolution when judged by kinetic energy spectra (Ricard et al. 2013; Klaver et al. 2020), in that modeled clouds and atmospheric waves remain larger than the grid spacing due to spectral filtering of small waves.
Figure 8 shows the impact of model resolution on the PDF of background brightness temperatures if we compare the two blue lines. Specifically, this compares the current operational forecast resolution of the IFS, HRES at 9-km model resolution, against that used in the majority of the experimentation presented here, 29 km. As we might expect, the higher-resolution model produces a slightly wider PDF of simulated Tbs, in particular a larger proportion of strongly scattering scenes. This resolution dependence of how well the observations can match the model background is something to bear in mind when discussing lower-resolution experiments.
A series of experiments was run (Table 4) in which the thinning was varied in tandem with the superob resolution to largely conserve the volume of data assimilated, with superobs ranging from about 80 to 40 km. This is intended to isolate the signal from the change in superob resolution. Hence, in Fig. 9, the point of comparison, i.e., 100% line, is the N128 Superob experiment; in other words, we are testing against the result given in the previous section. Figure 9 shows that the choice of superob resolution has surprisingly little impact on the fits to independent observations, indicating that the superob resolution by itself is not a major factor in the quality of the short-range forecast, at least within the range of superobbing scales considered. The only panel in Fig. 9 that appears noteworthy is a small degradation of fits to GEO infrared radiances (Fig. 9e), primarily those from Himawari-8. It is not clear why GEO infrared fits degrade slightly with finer H-sounder superobs. There is no systematic signal in the observational fits, whereas we might have expected some observation types to favor finer or coarser superobs. No forecast scores are shown here, as there were no significant impacts seen over 2 months of experimentation.
As in Fig. 6, but with different superob resolutions compared against N128 Superob; i.e., 100% represents N128 Superob.
Citation: Monthly Weather Review 152, 8; 10.1175/MWR-D-24-0020.1
It may be significant for the interpretation of these results that all experimentation discussed is at TCo399 final model resolution. Would the conclusions change if these comparisons were done at significantly higher or lower model resolution? This is considered in a limited way in section 4d. If the abovementioned theory of matching the model effective resolution is roughly correct, then the baseline N128 superob resolution of 78 km (Table 4) is already most consistent with the effective resolution of the TCo399 model of roughly 90–120 km, i.e., three–four times the model grid spacing, but a finer superob resolution would likely be preferable for higher-resolution model grids.
c. Superobbing and finer thinning
With the prior results showing the benefit of superobbing H-sounder radiances and increasing the volume of data assimilated, now we will assess the impact of these changes combined. As the previous sensitivity analysis demonstrated, the results are almost entirely agnostic to the superob size, so here we use two different superob sizes more as a way to influence the eventual spacing. Comparing to the control, here we roughly double the data in the N128 Superob N-T experiment and add even more in the N200 Superob experiment, 137% more than in the control. This translates to the spacing of assimilated radiances going from roughly 110 km to about 80 and 70 km, respectively.
Figure 10 shows the change in background fits to other observations globally, again relative to the control. The signals here are familiar from earlier, with much-improved fits to humidity-sensitive observations especially. Again, the impact on humidity is quite strong, with about a 1% decrease in std(O − B) for ATMS channels 19–22 and several IASI channels that are similar between the two experiments. Radiosonde observations indicate a small improvement in humidity at midlevels and short-range temperature forecasts that are largely neutral (Figs. 10e,f). The one observation type that shows a distinct difference between N128 and N200 is GEO infrared, which favors the larger N128 superobs; this is consistent with Fig. 9 where it was the only observation with sensitivity to superob size. Also, as noted before, ATMS UTLS channels (8–10) show a minor degradation, but this is not corroborated by infrared or radiosonde measurements (or radio occultation, not shown). There is an improvement in fits to wind observations (Fig. 10c), corroborated by atmospheric motion vector winds (VWs) and Aeolus observations (not shown). There is little impact on temperature short-range forecasts, as seen in the IASI observations (tropospheric temperature sensitive wavenumbers 705–760). Overall, the results here appear as a combination of the results from previous sections.
As in Fig. 6, panels give the change in std(O − B) for independent observations relative to 48r1 control, shown for N128 Superob N-T (black line) and N200 Superob (gray line).
Citation: Monthly Weather Review 152, 8; 10.1175/MWR-D-24-0020.1
To look at the impact on medium-range forecast scores, Fig. 11 provides the change in forecast RMSE for three pressure levels. This shows that despite the 6-month period of experimentation and the significant improvements seen in the short range via fits to observations, medium-range forecast skill was not greatly affected. Mid- and lower-tropospheric forecasts appear improved in the Southern Hemisphere, with day-4 Southern Hemisphere winds on the verge of 95% confidence intervals at 850 hPa, but a longer period would be required to establish significant forecast impacts beyond doubt. There are no significant differences seen when comparing N128 Superob N-T and N200 Superob for medium-range forecast scores.
As in Fig. 7, but with N128 Superob N-T given by the black line and N200 Superob in gray, with score changes in reference to control (48r1) and verification against own analysis.
Citation: Monthly Weather Review 152, 8; 10.1175/MWR-D-24-0020.1
d. Operational model resolution
Last, we examine the finer combination of superobbing and thinning, N200 Superob, with the model and DA in a configuration that mimics current ECMWF operations. N200 Superob was selected due to nearly identical performance as N128 Superob N-T in the lower-resolution testing while providing finer-resolution departure maps for diagnostic purposes. So again we compare against the control, but now with higher inner- and outer-loop resolutions in both the experiment and the control. As before, all else is equal—only the preprocessing treatment of H-sounder radiances is altered in the experiment.
For the HRES experiment N200 Superob, the impact on background fits to independent observations is seen in Fig. 12. These panels can be compared to Fig. 10 to contrast the impact in experiments of differing model resolutions. As in the lower-resolution experiments, fits to humidity- and wind-sensitive observations are significantly improved in the middle to upper troposphere, seen across different observation types. While most signals are common between the higher- and lower-resolution experiments, the most striking feature here is that the HRES experiment shows slightly larger magnitude improvements. For example, fits to several humidity channels on ATMS and IASI are improved by almost 2% in the HRES experiments, and fits to GEO humidity channels are also improved further. Nonsatellite wind observation fits are now improved from 700 hPa through the tropopause for this global sample.
As in Fig. 10, but for HRES experiment N200 Superob against HRES control.
Citation: Monthly Weather Review 152, 8; 10.1175/MWR-D-24-0020.1
There is a small but consistent signal of improved short-range forecasts of tropospheric temperatures, seen in ATMS channels 6 and 7, tropospheric IASI wavenumbers 705–760, and even radiosondes around 500 hPa; it is not clear why the temperature signal is stronger in the HRES experiments, but it may be notable that the observing system had much less aircraft data in this time period—summer 2020 especially—due to the pandemic (Ingleby et al. 2021). It is also notable that the slightly degraded fit to ATMS UTLS channels has disappeared in the HRES experimentation, possibly due to the further thinning of 118-GHz channels. In fact, fits to ATMS stratospheric channels (11–15) are very slightly but significantly improved, possibly showing a beneficial impact of higher-peaking 118-GHz channels after superobbing and with more appropriate thinning applied.
The difference in medium-range forecast scores is seen in Fig. 13 for the HRES experiments. Impacts are qualitatively similar to the lower-resolution experimentation seen in Fig. 11, albeit the experiments span different seasons. The HRES N200 Superob shows a small but statistically significant improvement in Southern Hemisphere 850-hPa wind RMSE at days 2 and 3, as well as Z500 at day 3 in the Northern Hemisphere. This appears to show that some of the short-range impact seen in the fits to observations does translate to improved forecasts up to day 3. Last, Fig. 14 gives the zonal average change in RMSE for relative humidity from 12 to 72 h. Here, the verification is against the operational analysis to avoid own-analysis artifacts prevalent in short-range scores; from 48 h onward, the results are very similar regardless of verification reference. This view of zonally averaged changes in forecast skill reinforces the earlier results, with a strong benefit seen at short ranges throughout the midtroposphere but limited areas of significant improvement at day 2 or day 3. Similar but slightly weaker signals are also seen for temperature and winds.
As in Fig. 11, but for the HRES experiments with verification against own analysis. The blue line signifies the change in forecast scores relative to the 48r1 HRES control.
Citation: Monthly Weather Review 152, 8; 10.1175/MWR-D-24-0020.1
Change in RMSE of relative humidity for HRES N200 Superob vs HRES control at forecast ranges from 12 to 72 h. Verification is against the operational ECMWF analysis. Hatching indicates statistical significance at the 95% confidence level following Geer (2016).
Citation: Monthly Weather Review 152, 8; 10.1175/MWR-D-24-0020.1
Due to the positive results given here, humidity sounders will move to 50-km superobs (N200) and 70-km spacing (diamond thinning) in the next operational version of the IFS, Cycle 49r1. The change is illustrated in Fig. 15 for MetOp-B MHS channel 4, contrasting the control (i.e., 48r1) and the N200 Superob experiments from the HRES experimentation as the left and right columns, respectively. The more finely spaced radiances provide a wealth of extra information for diagnostic analysis, shown in the top row of the plot, as well as over double the radiances assimilated, as seen in the bottom row. In addition to the diamond thinning, the assimilated data points in the bottom row account for quality control procedures in the data assimilation, including first guess check, variational quality control, and emissivity checks over land, which act to remove outlier observations.
O − B over Central America from MetOp-B MHS channel 4. (top) All radiances, with (bottom) used (i.e., assimilated) radiances. The columns contrast the treatment of radiances in Cycle 48r1 (native resolution thinned to 110 km spacing) and Cycle 49r1 (50 km superobs at 70-km spacing). Data are from the 0000 UTC long window on 3 Jun 2020 with the 9-km model. The pixel size changes in each panel to minimize white space.
Citation: Monthly Weather Review 152, 8; 10.1175/MWR-D-24-0020.1
5. Conclusions
This study has examined the thinning and superobbing aspects of the all-sky assimilation of MW humidity sounder radiances in the IFS. In line with recent studies that have found benefits from additional microwave sounder data in assimilation (Duncan et al. 2021; Lean et al. 2022), the additional H-sounder data are shown to improve short-range forecasts of humidity and winds considerably. “Additional data” refers to less thinning of sounder radiances and also the change to superobbing of radiances, which effectively adds information by averaging together many radiances that were previously discarded. In combination, superobbing and finer-spaced H-sounder radiances appear as largely additive benefits for short-range forecasts, better constraining mid- to upper-tropospheric moisture and winds as seen by independent observations (cf. black lines from Figs. 6 and 10). There are also some small benefits to forecast quality in the medium range visible in the Southern Hemisphere. Such impacts are of a similar or larger magnitude than typically seen for a new microwave sounder activation in the IFS, underscoring the importance of how observations are processed prior to assimilation.
The shift to using spatially averaged H-sounder radiances is one that was foreseen by Geer et al. (2014), and this change by itself has a positive effect on assimilation. More than a mere technical change, superobbing leads to some clear improvements by itself (Fig. 6). The interpretation of this outcome is that superobbing acts to beat down the noise slightly and decrease representation error, while providing a smoother field of increments. It can also be considered a fuller use of each sensor’s total information content. The previous operational thinning left about 4% of the total observations available for assimilation. The new treatment that combines superobbing and secondary thinning retains nearly 50% of total observations, as almost all contribute to a superob prior to the diamond thinning step that excludes every second superob. Finer spacing also means that about 2.4 times more superobservations are now assimilated than unaveraged observations were previously.
Comparison of the thinning and superobbing results shows that the reduced thinning dominates. Aspects such as averaging play a role, but doubling of the radiances assimilated was the largest signal. This adds to the evidence that adding observations, where feasible, is one of the most important things that can be done to enhance NWP impact (e.g., Harnisch et al. 2013; Geer et al. 2017b). It echoes findings from Duncan et al. (2021, 2022a) where the addition of data is the largest signal and elements such as the background errors or choice of observation error model are of secondary importance, albeit also crucial for an optimal assimilation. However, there are of course cases in which the addition of observations degrades the analysis, such as with no thinning of radiances whatsoever. This type of analysis—to explore optimal thinning distances and pinpoint where spatial error correlations overwhelm the benefit of additional data in the manner of Liu and Rabier (2002)—is the subject of a separate study (Steele et al. 2023). The interplay between model resolution and optimal superob size was not investigated here comprehensively but is also ripe for future analysis, as is possible interaction with the assimilation resolution itself (in the context of the IFS, the inner-loop resolution).
Considering that most of the experimentation in this study was at 29-km resolution and NWP models continue moving toward higher resolutions, the finest superob resolution tested here (50 km) was tested in the operational HRES configuration with 9-km model and a much finer inner-loop resolution as well. The HRES testing proved that the 50-km superob is indeed suitable for operational use, with positive impacts seen on medium-range forecast scores as well as slightly stronger positive impacts seen in fits to observations than in the lower-resolution testing. The latter suggests that giving attention to thinning and resolution of observations becomes even more important at high model and data assimilation resolutions, although with the caveat that the experimentation was in different seasons and the observing system is always changing. The HRES experimentation also implicitly shows that a superob near the model effective resolution is a reasonable choice. This hence suggests that the optimal superobbing scale is dependent on the model resolution. As model resolutions increase, the optimal combination of thinning and superobbing scale should be revisited accordingly.
An open question concerning the results presented is why the large improvements in fits to independent observations do not necessarily translate to improvements in medium-range forecast scores. A point of comparison is the assimilation of lidar wind retrievals from Aeolus (Rennie et al. 2021), which had a comparable impact on fits to observations in the IFS, but saw a significant medium-range impact in UTLS temperatures and midtropospheric tropical winds especially. Two potential reasons are mentioned here. First, although fits to humidity-sensitive observations were significantly improved here, temperature-sensitive observations showed a more neutral impact, and if atmospheric mass is the primary conduit for transmitting longer-lived atmospheric patterns, this could explain the limited impact later in the forecast. Second, it is possible that heavier use of H-sounders adjusts the analysis to better fit inertia–gravity wave activity in the upper troposphere, but that this does not benefit forecasts because the IFS cannot handle such waves sufficiently. It may be necessary to examine future results against a depleted observing system to shed more light on this connection between short-range humidity forecasts and medium-range humidity forecasts.
As a result of this study, the treatment of radiances from all-sky humidity sounders will change in IFS Cycle 49r1. This leads to a significant increase in assimilated all-sky radiances in the IFS—140% more from cross-track 183-GHz channels and 20% more from 118-GHz channels. Furthermore, the finer spacing of 50-km superobs permits 2.3 times more data through the screening trajectory, which should enable finer diagnostic analysis of mesoscale features.
Acknowledgments.
The EUMETSAT Fellowship Programme supported David and has been instrumental in supporting the development of all-sky assimilation. Thanks to Peter Lean for technical assistance and Cristina Lupu for discussions. Thanks to Tony McNally for encouragement and for reviewing the manuscript. This paper has been adapted and expanded from the ECMWF Technical Memorandum in Duncan et al. (2023).
Data availability statement.
It is not currently possible to permanently archive or curate the large volume of data produced by experimental runs of an NWP system. Monitoring web pages for observations assimilated by the ECMWF operational assimilation system are publicly available (https://charts.ecmwf.int/catalogue/packages/obstat/). The version of the ECMWF model used here is Cycle 48r1 (ECMWF 2023), and a version of the ECMWF model is available for researchers (https://confluence.ecmwf.int/display/OIFS).
REFERENCES
Bédard, J., and M. Buehner, 2020: A practical assimilation approach to extract smaller-scale information from observations with spatially correlated errors: An idealized study. Quart. J. Roy. Meteor. Soc., 146, 468–482, https://doi.org/10.1002/qj.3687.
Bennartz, R., 2000: Optimal convolution of AMSU-B to AMSU-A. J. Atmos. Oceanic Technol., 17, 1215–1225, https://doi.org/10.1175/1520-0426(2000)017<1215:OCOABT>2.0.CO;2.
Bergman, K. H., and W. D. Bonner, 1976: Analysis error as a function of observation density for satellite temperature soundings with spatially correlated errors. Mon. Wea. Rev., 104, 1308–1316, https://doi.org/10.1175/1520-0493(1976)104<1308:AEAAFO>2.0.CO;2.
Bonavita, M., A. J. Geer, and M. Hamrud, 2020: All-sky microwave radiances assimilated with an ensemble Kalman filter. Mon. Wea. Rev., 148, 2737–2760, https://doi.org/10.1175/MWR-D-19-0413.1.
Bormann, N., and P. Bauer, 2010: Estimates of spatial and interchannel observation-error characteristics for current sounder radiances for numerical weather prediction. I: Methods and application to ATOVS data. Quart. J. Roy. Meteor. Soc., 136, 1036–1050, https://doi.org/10.1002/qj.616.
Bormann, N., D. Duncan, S. English, S. Healy, K. Lonitz, K. Chen, H. Lawrence, and Q. Lu, 2021: Growing operational use of FY-3 data in the ECMWF system. Adv. Atmos. Sci., 38, 1285–1298, https://doi.org/10.1007/s00376-020-0207-3.
Candy, B., and S. Migliorini, 2021: The assimilation of microwave humidity sounder observations in all-sky conditions. Quart. J. Roy. Meteor. Soc., 147, 3049–3066, https://doi.org/10.1002/qj.4115.
Carminati, F., and S. Migliorini, 2021: All-sky data assimilation of MWTS-2 and MWHS-2 in the Met Office global NWP system. Adv. Atmos. Sci., 38, 1682–1694, https://doi.org/10.1007/s00376-021-1071-5.
Dando, M. L., A. J. Thorpe, and J. R. Eyre, 2007: The optimal density of atmospheric sounder observations in the Met Office NWP system. Quart. J. Roy. Meteor. Soc., 133, 1933–1943, https://doi.org/10.1002/qj.175.
Duncan, D., and N. Bormann, 2020: On the addition of microwave sounders and NWP skill, including assessment of FY-3D sounders. EUMETSAT/ECMWF Fellowship Programme Research Rep. 55, 23 pp., https://doi.org/10.21957/iimhohmqt.
Duncan, D., N. Bormann, and E. Hólm, 2021: On the addition of microwave sounders and numerical weather prediction skill. Quart. J. Roy. Meteor. Soc., 147, 3703–3718, https://doi.org/10.1002/qj.4149.
Duncan, D., N. Bormann, and A. J. Geer, 2022a: All-sky assimilation of AMSU-A window channels. EUMETSAT/ECMWF Fellowship Programme Research Rep. 59, 36 pp., https://doi.org/10.21957/daefm16p8.
Duncan, D., N. Bormann, A. J. Geer, and P. Weston, 2022b: Assimilation of AMSU-A in all-sky conditions. Mon. Wea. Rev., 150, 1023–1041, https://doi.org/10.1175/MWR-D-21-0273.1.
Duncan, D., N. Bormann, A. J. Geer, and P. Weston, 2023: Superobbing and finer thinning for all-sky humidity sounder assimilation. ECMWF Tech. Memo. 905, 31 pp., https://doi.org/10.21957/5c3b9c8d9f.
ECMWF, 2023: IFS Documentation CY48R1—Part II: Data assimilation, ECMWF Doc. 2, 108 pp., https://doi.org/10.21957/a744f32e74.
Geer, A. J., 2016: Significance of changes in medium-range forecast scores. Tellus, 68A, 30229, https://doi.org/10.3402/tellusa.v68.30229.
Geer, A. J., and P. Bauer, 2010: Enhanced use of all-sky microwave observations sensitive to water vapour, cloud and precipitation., EUMETSAT/ECMWF Fellowship Programme Research Rep. 20, 43 pp., https://www.ecmwf.int/node/9506.
Geer, A. J., F. Baordo, N. Bormann, and S. English, 2014: All-sky assimilation of microwave humidity sounders. ECMWF Tech. Memo. 741, 59 pp., https://doi.org/10.21957/obosmx154.
Geer, A. J., and Coauthors, 2017a: Assimilating observations sensitive to cloud and precipitation. ECMWF Tech. Memo. 815, 62 pp., https://doi.org/10.21957/sz7cr1dym.
Geer, A. J., and Coauthors, 2017b: The growing impact of satellite observations sensitive to humidity, cloud and precipitation. Quart. J. Roy. Meteor. Soc., 143, 3189–3206, https://doi.org/10.1002/qj.3172.
Geer, A. J., K. Lonitz, D. I. Duncan, and N. Bormann, 2022: Improved surface treatment for all-sky microwave observations. ECMWF Tech. Memo. 894, 59 pp., https://doi.org/10.21957/zi7q6hau.
Harnisch, F., S. B. Healy, P. Bauer, and S. J. English, 2013: Scaling of GNSS radio occultation impact with observation number using an ensemble of data assimilations. Mon. Wea. Rev., 141, 4395–4413, https://doi.org/10.1175/MWR-D-13-00098.1.
He, J., S. Zhang, and Z. Wang, 2015: Advanced Microwave Atmospheric Sounder (AMAS) channel specifications and T/V calibration results on FY-3C satellite. IEEE Trans. Geosci. Remote Sens., 53, 481–493, https://doi.org/10.1109/TGRS.2014.2324173.
Hoffman, R. N., 2018: The effect of thinning and superobservations in a simple one-dimensional data analysis with mischaracterized error. Mon. Wea. Rev., 146, 1181–1195, https://doi.org/10.1175/MWR-D-17-0363.1.
Hortal, M., and A. J. Simmons, 1991: Use of reduced Gaussian grids in spectral models. Mon. Wea. Rev., 119, 1057–1074, https://doi.org/10.1175/1520-0493(1991)119<1057:UORGGI>2.0.CO;2.
Ingleby, B., and Coauthors, 2021: The impact of COVID-19 on weather forecasts: A balanced view. Geophys. Res. Lett., 48, e2020GL090699, https://doi.org/10.1029/2020GL090699.
Klaver, R., R. Haarsma, P. L. Vidale, and W. Hazeleger, 2020: Effective resolution in high resolution global atmospheric models for climate studies. Atmos. Sci. Lett., 21, e952, https://doi.org/10.1002/asl.952.
Lawrence, H., N. Bormann, A. J. Geer, Q. Lu, and S. J. English, 2018: Evaluation and assimilation of the microwave sounder MWHS-2 onboard FY-3C in the ECMWF numerical weather prediction system. IEEE Trans. Geosci. Remote Sens., 56, 3333–3349, https://doi.org/10.1109/TGRS.2018.2798292.
Lean, K., N. Bormann, and S. Healy, 2022: WP-2000 calibration of EDA spread and adaptation of the observation error model. ESA Contract 4000130590/20/NL/IA, 25 pp., https://doi.org/10.21957/1auh0nztg.
Lee, S., H.-J. Song, H.-W. Chun, I.-H. Kwon, J.-H. Kang, and S. Lim, 2020: All-sky microwave humidity sounder assimilation in the Korean Integrated Model forecast system. Quart. J. Roy. Meteor. Soc., 146, 3570–3586, https://doi.org/10.1002/qj.3862.
Liu, Z.-Q., and F. Rabier, 2002: The interaction between model resolution, observation resolution and observation density in data assimilation: A one-dimensional study. Quart. J. Roy. Meteor. Soc., 128, 1367–1386, https://doi.org/10.1256/003590002320373337.
Maddy, E. S., S. Bunin, V. J. Mikles, N. Shahroudi, B. Shiotani, and S.-A. Boukabara, 2022: Temperature and moisture sounding performance of current and prospective microwave instruments under all-sky conditions. Remote Sens., 14, 1624, https://doi.org/10.3390/rs14071624.
O’Sullivan, D., and T. J. Dunkerton, 1995: Generation of inertia–gravity waves in a simulated life cycle of baroclinic instability. J. Atmos. Sci., 52, 3695–3716, https://doi.org/10.1175/1520-0469(1995)052<3695:GOIWIA>2.0.CO;2.
Peubey, C., and A. McNally, 2009: Characterization of the impact of geostationary clear-sky radiances on wind analyses in a 4D-Var context. Quart. J. Roy. Meteor. Soc., 135, 1863–1876, https://doi.org/10.1002/qj.500.
Rennie, M. P., L. Isaksen, F. Weiler, J. de Kloe, T. Kanitz, and O. Reitebuch, 2021: The impact of Aeolus wind retrievals on ECMWF global weather forecasts. Quart. J. Roy. Meteor. Soc., 147, 3555–3586, https://doi.org/10.1002/qj.4142.
Ricard, D., C. Lac, S. Riette, R. Legrand, and A. Mary, 2013: Kinetic energy spectra characteristics of two convection-permitting limited-area models AROME and Meso-NH. Quart. J. Roy. Meteor. Soc., 139, 1327–1341, https://doi.org/10.1002/qj.2025.
Robel, J., and A. Graumann, 2014: NOAA KLM user’s guide with NOAA-N, N Prime, and MetOp supplements. NOAA NESDIS Tech. Rep., 2530 pp., https://lccn.loc.gov/2009578433.
Shahabadi, M. B., and M. Buehner, 2024: Implementation of all-sky assimilation of microwave humidity sounding channels in Environment Canada’s global deterministic weather prediction system. Mon. Wea. Rev., 152, 1027–1038, https://doi.org/10.1175/MWR-D-23-0227.1.
Steele, L., N. Bormann, and D. I. Duncan, 2023: Assimilating FY-3E MWHS-2 observations, and assessing all-sky humidity sounder thinning scales. EUMETSAT/ECMWF Fellowship Programme Research Rep. 62, 33 pp., https://doi.org/10.21957/f42a9d9542.
Yang, J. X., Y. You, W. Blackwell, S. Misra, and R. A. Kroodsma, 2022: Quantifying and characterizing striping of microwave humidity sounder with observation and simulation. IEEE Trans. Geosci. Remote Sens., 60, 5302413, https://doi.org/10.1109/TGRS.2021.3132560.
