Toward Aerosol-Aware Thermal Infrared Radiance Data Assimilation

Abstract

Aerosols considerably reduce the upwelling radiance in the thermal infrared (IR) window; thus, it is worthwhile to understand the effects and challenges of assimilating aerosol-affected (i.e., hazy-sky) IR observations for all-sky data assimilation (DA). This study introduces an aerosol-aware DA framework for the Infrared Atmospheric Sounder Interferometer (IASI) to exploit hazy-sky IR observations and investigate the impact of assimilating hazy-sky IR observations on analyses and subsequent forecasts. The DA framework consists of the detection of hazy-sky pixels and an observation error model as the function of the aerosol effect. Compared to the baseline experiment, the experiment utilized an aerosol-aware framework that reduces biases in the sea surface temperature in the tropical region, particularly over the areas affected by heavy dust plumes. There are no significant differences in the evaluation of the analyses and the 7-day forecasts between the experiments. To further improve the aerosol-aware framework, the enhancements in quality control (e.g., aerosol detection) and bias correction need to be addressed in the future.

1. Introduction

Aerosols affect the radiation budget of the Earth through their transmittance effects (absorption and scattering). Many studies have demonstrated that aerosols can considerably reduce the simulated radiance at the top of the atmosphere, as a form of brightness temperature (BT) [1,2,3,4,5]. Field campaigns, such as the Saharan Dust Experiment (SHADE) [6] and the Saharan Mineral Dust Experiment (SAMUM-2) [7] also reported the BT reduction over the thermal infrared (IR) window region (800–1200 cm⁻¹) due to mineral dust aerosols.

The direct assimilation of the clear-sky satellite BT has been operationally implemented in numerical weather prediction (NWP) centers since the early 1990s [8,9,10]. It utilizes a radiative transfer model (RTM) as the radiance observation operator to simulate the BT at the top of the atmosphere in the model. The RTM also provides the linearized operator (i.e., Jacobians) through the tangent-linear and adjoint methods, which are needed for the conversion between the model and the observation space [10,11]. In the last decade, the clear-sky radiance assimilation has been extended to cloudy-sky scenarios for microwave [12,13,14] and thermal infrared sensors [15,16,17].

Several studies have investigated aerosol effects in analysis systems. Weaver et al. [18] assimilated aerosol-affected temperature retrieval profiles and found warmer temperatures at the sea surface and in the lower atmosphere. Kim et al. [19] and Wei et al. [20] directly assimilated IR measurements by incorporating modeled aerosol information into the simulation of radiance. These studies demonstrated similar warming features in the lower atmosphere due to the cooler BTs of the model compared to the experiments without any aerosol information (i.e., aerosol-blind experiments). In addition the warming feature, they also reported substantial changes in the quality control and bias correction. Similarly, Sun et al. [21] and Peng et al. [22] demonstrated that considering the reanalysis of the aerosol mixing ratio from the Copernicus Atmosphere Monitoring Service (CAMS) and the Modern-Era Retrospective Analysis for Research and Applications, Version 2 (MERRA-2) datasets can substantially reduce the biases of BT simulations for IR instruments.

Wei et al. [5] investigated the aerosol effects on the RTM simulation of BTs and Jacobians. They further assessed the impacts of aerosol-affected Jacobians on the temperature and water vapor analyses via a single-observation test. It shows that simulated BTs are most sensitive to dust aerosols, the total aerosol loading, and the peak altitude of aerosol plumes. The analysis of the aerosol-affected Jacobians for temperature and water vapor indicates that BTs become more sensitive to the temperature of the aerosol layer and less sensitive to the water vapor profile. Their single-observation test indicated that the multivariate relationship between the temperature and the water vapor would dominate the impacts of aerosol-affected Jacobians on the analysis. These modifications due to aerosols in Jacobians also played a role in the quality control and bias correction shown in [19,20]. Given the effects of aerosols on the BT simulations in prior studies, the aerosols’ effect on the IR measurements in the data assimilation (DA) system is not negligible. To constrain the aerosol effects, we need to identify the IR measurements affected by aerosols (i.e., hazy-sky conditions). A common method is using BT differences (BTD) between multiple thermal infrared channels to identify the location of hazy-sky measurements, especially the data with mineral dust [3,23,24,25].

Prior studies from the European Centre for Medium-range Weather Forecasts (ECMWF) discussed the application of the aerosol detection method in their DA system. Letertre-Danczak [26] applied the aerosol detection method based on BTDs to Infrared Atmospheric Sounder Interferometer (IASI) observations and reported that rejecting those hazy-sky data can improve the analyses in the lower atmosphere. The same aerosol detection method, with different criteria for BTDs, was applied to the assimilation of Cross-track Infrared Sounder (CrIS) radiance [27]. As shown in Figure 2d in [15], aerosol-contaminated data occupied more than 60% of the aerosol-laden region, which was identified by using the aerosol detection method (ADM) [26,28] for those high-spectral-resolution IR sounder measurements.

The underutilization of a considerable volume of aerosol contamination data over aerosol-laden regions motivates the central question of this study: “Can we improve analyses and forecasts by including those hazy-sky IR observations in a DA system?”. To achieve this, we introduce an aerosol-aware framework (described below) to exploit hazy-sky IR observations. The all-sky DA algorithms aim to improve the analyzed water vapor fields by assimilating cloudy-sky radiance data. In contrast to the all-sky DA studies above, our aerosol-aware framework assimilates hazy-sky data, but it focuses on the analysis of meteorological fields, rather than the analysis of aerosol mixing ratios.

This manuscript documents an aerosol-aware framework and highlights the enhancement and potential challenges. The system and the dataset are described in Section 2. The framework is described in Section 3. The experimental design is in Section 4. The results and the evaluation are reported in Section 5. The conclusion and the discussion are in Section 6.

2. Methods

2.1. Global Data Assimilation System (GDAS)

In this study, we developed the aerosol-aware framework based on version 16 of the National Centers for Environmental Prediction (NCEP) Global Forecast System (GFS v16; implemented in 2021). The forecast model is the Finite-Volume Cubed Sphere (FV3) dynamical core with the GFS physics suite (~13 km). The analysis system, the Global Data Assimilation System (GDAS), is a hybrid, four-dimensional ensemble–variational (4DEnVar) analysis system with 80 ensemble members (~25 km) based on Gridpoint Statistical Interpolation (GSI) [29,30]. Readers can find more detailed information about the NCEP GFS and the GDAS and the presentation of v16 upgrades at https://www.emc.ncep.noaa.gov/emc/pages/numerical_forecast_systems/gfs.php (accessed on 24 April 2025). The conventional and satellite observation data within the 6 h window (±3 h) are assimilated in the NCEP GDAS to generate a meteorological analysis every 6 h (i.e., 00Z, 06Z, 12Z, and 18Z). Readers can find the detailed lists of conventional and satellite dataset ingested in NCEP GDAS at https://www.emc.ncep.noaa.gov/emc/pages/infrastructure/obs-data/prepbufr.doc/table_2.php (accessed on 24 April 2025) and https://www.emc.ncep.noaa.gov/emc/pages/infrastructure/obs-data/prepbufr.doc/table_18.php (accessed on 24 April 2025), respectively.

In GSI, the Community Radiative Transfer Model (CRTM) is used to provide the BT simulation and the linear operator (i.e., Jacobian) of satellite measurements ranging from ultraviolet to microwave spectrum. CRTM is a single column RTM, which was developed at the Joint Center for Satellite Data Assimilation (JCSDA) with contributions from scientists at JCSDA partner institutes [31,32]. It considers the absorption for the gaseous constituents, absorption and scattering for clouds and aerosols, surface emission, and the surface interaction with downwelling atmospheric radiation. In this study, we use CRTM v2.4.0 to simulate the radiance observations based on model states.

For aerosols, CRTM considers the transmittance effects from the ultraviolet to the IR region. The default specification of aerosol optical properties in CRTM is based on the Goddard Chemistry Aerosol Radiation and Transport model (GOCART) [33,34]. Briefly, the implemented five GOCART aerosol species include 5 bins of dust, 4 bins of sea salt, hydrophobic and hydrophilic black and organic carbon, and sulfate. The refractive indices are adopted from the Optical Properties of Aerosols and Clouds (OPACs) [35]. The spherical particles and lognormal size distribution are assumed. For dust aerosols, the effective radius for each bin is 0.55, 1.4, 2.4, 4.5, and 8.0 µm with radii ranges of 0.1–1.0, 1.0–1.8, 1.8–3, 3–6, and 6–10 µm, respectively. For other species, the effective radius is determined based on the ambient relative humidity. Note that look-up tables for aerosol optical properties have been updated based on the latest version of GOCART in CRTM v2.4.0 [36]. More detailed descriptions for aerosol optical properties in the CRTM are documented in [37,38].

The GFSv16 global workflow maintained by the NCEP Environmental Modeling Center (EMC) (https://github.com/NOAA-EMC/global-workflow, accessed on 24 April 2025) is utilized to conduct the experiments. The verification package, EMC_verif-global, (https://github.com/NOAA-EMC/EMC_verif-global, accessed on 24 April 2025), is used to evaluate the forecast in Section 5.4.

2.2. Datasets

In this study, several datasets including models and observations are used. The 4-dimensional aerosol mixing ratios from the MERRA-2 reanalysis [39,40,41] are ingested to GSI as part of model states (i.e., aerosols are considered in the BT simulations of CRTM). MERRA-2 is the reanalysis dataset produced by the NASA Global Modeling and Assimilation Office (GMAO) based on the Goddard Earth-Observing System, version 5 (GEOS-5). As described in [41], it assimilates the bias-corrected aerosol optical depth (AOD) from the ground-based Aerosol Robotic Network (AERONET), Multiangle Imaging Spectro Radiometer (MISR), Moderate Resolution Imaging Spectroradiometer (MODIS), and Advanced Very-High-Resolution Radiometer (AVHRR) instruments. MERRA-2 is publicly available at NASA’s Goddard Earth Sciences Data and Information Services Center (https://disc.gsfc.nasa.gov/, accessed on 24 April 2025).

Two observations are used to evaluate the aerosol detection method (described in Section 3.1). First, the IASI level 1c data (i.e., apodised) is downloaded through the user portal of the European Organisation for the Exploitation of Meteorological Satellites (EUMETSAT; https://user.eumetsat.int/, accessed on 24 April 2025). It is the same product used in GDAS but with full spectrum coverage (645 to 2760 cm⁻¹). Second, the AOD at 10 µm of the retrieval product of IASI from Laboratoire de Météorologie Dynamique (LMD; LMDAERO [42]) is utilized as a benchmark to evaluate the detection method. The LMDAERO data can be requested through the IASI portal (https://iasi.aeris-data.fr/, accessed on 24 April 2025). It only retrieves the AOD at cloud-free pixels by using multiple sensors information and BT differences from several window channels. Additionally, the level 3 product of the daily VIIRS AOD at a 1° resolution is used to compare the MERRA-2 simulation qualitatively. The VIIRS data can be accessed at level 1 and via the Atmosphere Archive & Distribution System Distributed Active Archive Center (LAADS DAAC) website (https://ladsweb.modaps.eosdis.nasa.gov/, accessed on 24 April 2025).

The ECMWF Reanalysis version 5 (ERA5) reanalysis produced by ECMWF [43] is used to evaluate the meteorological analyses of temperatures, humidity, and horizontal winds from GDAS experiments. The ERA5 reanalysis data can be retrieved from the Copernicus Climate Change Service Climate Data Store (https://cds.climate.copernicus.eu/, accessed on 24 April 2025).

3. Aerosol-Aware Framework

To exploit the information from the hazy-sky IR radiance observations, the aerosol-aware framework proposed in this study consists of the following components:

• The identification of hazy-sky (i.e., hazy-sky) observations: In addition to the existing quality control (QC) for clear-sky radiances from IR sounders in the GSI system, we introduce the aerosol detection method (described in Section 3.1) as the first check of the quality control. The aerosol detection method is based on BTDs to identify hazy-sky observations. As a result, only the observations not identified as not being hazy-sky are passed to the clear-sky QC checks.
• The incorporation of aerosol mass mixing ratios: For those hazy-sky observations, we incorporate the 4-dimensional information of the aerosol mass mixing ratios from the MERRA-2 reanalysis as a part of the first-guess fields into the radiance observation operator (i.e., CRTM). Because only meteorological fields are adjusted through the analysis processes, this configuration can confine the impacts of aerosols on the identified hazy-sky observations.
• The quantification of the aerosol effects: An aerosol effect parameter (A_e) is introduced as the proxy to symmetrically quantify the aerosol impacts on each observation (described in Section 3.2) that is contributed by both the model and observations.
• The inflation of the observation errors: Define the observation error as a function of A_e. The observation error increases with A_e linearly from the value used for clear-sky DA.

In addition to incorporating the modeling aerosol mixing ratios (i.e., item 2) into the DA system, as performed in previous studies [19,20,21,22], the proposed framework can constrain the uncertainties induced by MERRA-2 aerosols and inflate the observation error when aerosols interfere with the measurement. The following sub-sections describe the overview of items 1, 3, and 4. More detailed information can be found in Wei [44].

3.1. Aerosol Detection Method

The ADM within the Clouds and Aerosols Detection Software v3.0 (CADS [28]) from the Satellite Application Facility on numerical weather prediction (NWP SAF) was implemented to the GSI system. In this study, we focus on the level 1c data of the IASI sensor [45], which is a hyperspectral sounder onboard MetOp series satellites. It covers the thermal infrared spectrum from 645 to 2760 cm⁻¹ with an apodised spectral resolution of 0.5 cm⁻¹. Given the high-spectral resolution and wide coverage, it can resolve the spectral signature of aerosols [46], especially the “V-shape” spectral signature of mineral dust around the IR window region [4,47]. In the ADM for the IASI sensor, the BTDs calculated from the 11-channel average around 4 key channels (C1: 980 cm⁻¹; C2: 1232 cm⁻¹; C3: 1090.5 cm⁻¹; and C4: 1234 cm⁻¹), which are used to identify hazy-sky observations of thermal infrared sounders through the spectral signature of the BT. A pixel is identified as hazy-sky data when its BTDs fulfill the following thresholds:

$$\mathrm{C}1-\mathrm{C}2:\overline{{\mathrm{B}\mathrm{T}}_{980}}-\overline{{\mathrm{B}\mathrm{T}}_{1232}}<0.2\mathrm{K}$$

(1)

$$\mathrm{C}3-\mathrm{C}4:\overline{{\mathrm{B}\mathrm{T}}_{1090.5}}-\overline{{\mathrm{B}\mathrm{T}}_{1234}}<-1.55\mathrm{K}$$

(2)

The IASI data used in the NCEP GDAS system does not retain the complete set of 11 channels for the 4 key channels (Table S1). This may lead to misidentifications due to noise from the limited number of channels. To investigate the impact on the identification, we compare the ADM identification using the full spectrum channels and NCEP GDAS channels (NCEPIASI) based on a 6 h window covering the “Godzilla” dust plume in 2020 June [48]. The results reveal that the ADM with the NCEPIASI can capture the major events and include more pixels with light aerosol loading (Figures S1 and S2).

Table 1. The comparison of identified hazy-sky observations from two sets of channels: NCEPIASI and full spectrum.

LMD AOD at 10 µm >0 Counts = 48,615	NCEPIASI	Full Spectrum
Matched	12,779 (26.29%)	6478 (13.33%)
Missed	35,836 (73.71%)	42,137 (86.67%)
Zero AOD or not retrieved	55,676 (81.33%) 1	8897 (57.8%) 1

¹ The fraction of identified hazy-sky data pixels from each dataset.

provides descriptive statistics of the performance of the ADM with respect to the LMDAERO 10 µm AOD retrieval. The evaluation in [49] demonstrated that the ADM can identify about 20% hazy-sky observations against the retrieval product of the IASI. Among 48,615 pixels from LMDAERO (AOD at 10 µm > 0), twice the amount of hazy-sky data (i.e., matched) and less missing data (73% vs. 86%) are identified using NCEPIASI. However, the third row of Table 1 caveats that more than 80% of the identified hazy-sky data in the NCEPIASI were not retrieved due to cloud contamination, or the corresponding AODs at 10 µm were zero in the LMDAERO product. It implies that additional QC checks are needed to exclude those misidentified hazy-sky data. Therefore, we keep the cloud QC in GSI and introduce the “bust” QC [16] to constrain the misidentified hazy-sky observations.

3.2. Aerosol Effect Parameters

To quantify the aerosol effects on each hazy-sky observation, we introduced a proxy, an aerosol effect parameter (A_e), which is adapted from the cloud effect parameter used in all-sky radiance DA [16,17,50]. The aerosol effect parameter is calculated as follows:

$${\mathrm{A}}_{\mathrm{e}}={\displaystyle \frac{\left|{\mathrm{B}}_{\mathrm{a}\mathrm{e}\mathrm{r}}-{\mathrm{B}}_{\mathrm{c}\mathrm{l}\mathrm{r}}\right|-\left|\mathrm{O}-{\mathrm{B}}_{\mathrm{c}\mathrm{l}\mathrm{r}}\right|}{2}}$$

(3)

where B_aer and B_clr are the hazy-sky and clear-sky BT simulations, respectively; O is the BT observation. The A_e is conducted in two parts, using the BT differences. The first term, B_aer − B_clr, represents the aerosol effect from the model’s aerosol information (i.e., the model term). The second term, O − B_clr, shows the aerosol impacts on the observation side (i.e., the observation term). The combination of these two terms symmetrically quantifies the aerosol impacts on the BT differences from the model and observations.

Figure 1 displays the A_e and the corresponding contribution from two terms based on clear-sky and hazy-sky data for the cycle of 12Z 22 June 2020. The A_e shows a strong negative signal being contributed by both terms in the data over the trans-Atlantic region and West Africa. This indicates that MERRA-2 captures the dust plumes in the real world. It can also lead to better agreement between the simulated and the observed BT over these regions by considering MERRA-2 aerosols compared to no aerosols. Some data over the west Pacific and Southern Ocean also show a large A_e, which is mainly due to the extreme negative values (over 10 K) of the observation term (Figure 1c). It implies those observations could be contaminated by clouds and misidentified by the ADM because of the limited information in NCEP dataset and the indiscernible BT differences from the model term.

As discussed in Section 3.1, it is desired to introduce an additional QC to constrain the misidentified hazy-sky observations from the ADM. To do this, the “bust” QC [16] is adapted in the proposed aerosol-aware framework. The “bust” QC rejects observations where the absolute first-guess departure is greater than 1.8 A_e and the A_e is greater than 1.5 K. This “bust” QC mainly rejects the hazy-sky data identified by the ADM under the two circumstances below:

• B_aer ~ B_clr and |O − B_clr| > 3 K: The A_e is dominated by the observation term when MERRA-2 only provides a low aerosol loading. It implies the aerosol loading is too low or the observation is contaminated by clouds but misidentified as hazy-sky data.
• O ~ B_clr and |B_aer − B_clr| > 3 K: This indicates that MERRA-2 provides a considerable aerosol loading at this observation that cools the B_aer down, so the A_e is mainly contributed to by model term. It implies that the aerosol loading is over-estimated, or the ADM identifies clear-sky data as hazy-sky data.

The “bust” QC, however, cannot capture cases where the ADM misidentifies cloud-contaminated data as hazy-sky data. It happens when the observation is identified as hazy-sky data but is affected by clouds, and the aerosols from MERRA-2 create insufficient cooling effects on the BT. In this case, the first-guess departure is likely smaller than 1.8 A_e, and this information must subsequently be assimilated into the system. Therefore, the cloud QC in the GSI [51] used for the clear-sky DA is applied to the hazy-sky data to reduce the impacts of misidentification by the ADM on the results. The detail of the proposed aerosol-aware QC procedures is provided in the Supplementary Material (Figure S3).

3.3. Observation Error Model

To adequately describe the observation error in the hazy-sky data of each IASI channel, we conducted an “observer” experiment for two months on episodic dust outbreaks (10 June to 10 July) and wildfire events (22 August to 21 September) in 2020 using the proposed QC procedures. The “observer” experiment means that the GSI system is utilized as a simulator to generate the first-guess departures (observed BTs minus simulated BTs), which are then used to calculate the standard deviation (SD) for the observation error model. The information of the aerosol mixing ratios from MERRA-2 is provided in the “observer” experiment. Given the good agreement between the MERRA-2 reanalysis and the VIIRS AOD product (Figure S4), the observation error model contains the signal from the low aerosol loading condition to the dust and biomass-burning aerosols’ dominant scenarios.

The results of the “observer” experiment provides insights into the proposed QC procedures. Figure 2 displays the 2-dimensional histogram of the first-guess departures as a function of the A_e from the “observer” run. First, the negative tail indicates the poor performance of the atmosphere and aerosol information. Despite a reasonable mass loading of aerosols, MERRA-2 is considered during the BT simulation, and the measured BTs are still much cooler than the simulated BTs. Second, the negative tail could be attributed to the potential cloud-contaminated points that were misidentified by the ADM. Hence, it suggests the necessity of the cloud QC for the aerosol-aware framework in this study. As Figure 2b shows, the negative tail is removed because the cloud QC rejects a considerable amount of hazy-sky data.

For the hazy-sky pixels retained through the proposed QC procedures, we adapt the piecewise observation error model [16,50] below to define the A_e-dependent observation error for the IASI’s 174 assimilated channels:

$$\mathrm{S}\mathrm{D}=\left\{\begin{array}{l}{\mathrm{S}\mathrm{D}}_{\mathrm{m}\mathrm{i}\mathrm{n}} \mathrm{for} {\mathrm{A}}_{\mathrm{e}}\le {\mathrm{A}}_{\mathrm{e}1}\\ {\mathrm{S}\mathrm{D}}_{\mathrm{m}\mathrm{i}\mathrm{n}}+{\displaystyle \frac{{\mathrm{S}\mathrm{D}}_{\mathrm{m}\mathrm{a}\mathrm{x}}-{\mathrm{S}\mathrm{D}}_{\mathrm{m}\mathrm{i}\mathrm{n}}}{{\mathrm{A}}_{\mathrm{e}2}-{\mathrm{A}}_{\mathrm{e}1}}}\left({\mathrm{A}}_{\mathrm{e}}-{\mathrm{A}}_{\mathrm{e}1}\right) \mathrm{for} {\mathrm{A}}_{\mathrm{e}1}<{\mathrm{A}}_{\mathrm{e}}<{\mathrm{A}}_{\mathrm{e}2}\\ {\mathrm{S}\mathrm{D}}_{\mathrm{m}\mathrm{a}\mathrm{x}} \mathrm{for} {\mathrm{A}}_{\mathrm{e}}\ge {\mathrm{A}}_{\mathrm{e}2}\end{array}\right.$$

(4)

A_e1 is the A_e where the SD of the first-guess departures is equal to the original prescribed SD (SD_min); A_e2 is the A_e with the maximum SD (SD_max) of the first-guess departures. Based on A_e1 and A_e2, the observation error is determined with three segments. When the A_e is smaller than A_e1, the observation is considered as clear-sky and the observation error is assigned with the SD_min, which was used in the original clear-sky framework. When the A_e is larger than A_e2, the observation is considered as fully hazy-sky and its observation error is assigned with the SD_max. For an observation with an A_e of between A_e1 and A_e2, its observation error is linearly interpolated with the SD_min and the SD_max.

Among the 174 assimilated channels of the IASI, 130 channels are considered as aerosol-sensitive channels because the SD_max exceeds the SD_min. Figure 3 shows the histogram and the piecewise observation error model as the function of the A_e for 962.5 and 1096 cm⁻¹. The histogram of both channels depicts substantial rejection by cloud QC (light vs. deep grey bars). After applying the whole QC procedure in Figure S3, most observations with an A_e of larger than 5 K were rejected. As a consequence, the strong negative-mean first-guess departures are considerably reduced (thin vs. thick blue lines), and the standard deviation of the first-guess departures (grey vs. red line) becomes smaller. Based on the standard deviation of the post-QC dataset (red line), A_e1 is determined by the A_e bin, where its SD equals the prescribed SD; A_e2 is the bin with the maximum SD. The look-up-table for other aerosol-sensitive assimilated channels is then generated (Table S2).

4. Experimental Design

To evaluate the aerosol-aware framework, we conducted two fully cycled GDAS experiments producing a meteorological analysis based on GFS v16 with C192/C96 (~50 km/~100 km) with 60 ensemble members. Because of the limited computational resources, the experiments are in a lower resolution and of a smaller ensemble size compared to the operational configuration. The definition of the hybrid sigma levels uses 64 layers, which is used in a prior GFS version. The study period was 10 June to 10 July 2020. We only focused on the dust period because of (i) the stronger sensitivity to dust aerosols in BT simulations, (ii) the significant Saharan dust transportation crossing Atlantic Ocean, and (iii) the ADM’s limitation with regard to identifying hazy-sky pixels over non-smoke- (North and South America in Figure S5a) and smoke- (North and South America and tropical Africa in Figure S5b) affected areas [28]. In addition, we produced 7-day forecasts of the 00Z cycle during the period, which were initialized from the analyses in the two experiments.

The first experiment, CTL, was conducted as the baseline with the default aerosol-blind configuration. The second experiment, AER, applied the aerosol-aware framework to the IASI sensors onboard MetOp-A and B. The assimilation of the other sensors stayed the same as the CTL experiment, which is different to impacting all the IR sensors over all the surface types in [21]. Due to the limitation of the ADM over land, only the hazy-sky pixels of the IASI over water and the clear-sky pixels were assimilated in the AER experiment. Note that the hazy-sky data that passed the QC is excluded from the biases estimation from the variational bias correction to prevent an inappropriate correction on the clear-sky data. Hence, the biases are estimated based on the clear-sky data and applied to the whole assimilated dataset.

5. Results

This section presents the comparison of the assimilated dataset between the CTL and AER experiments and the evaluation of both experiments from the aspects of the in situ measurements, the reanalysis, and the 7-day forecast skills. Overall, it shows limited impacts on the analysis and the forecast from this prototype aerosol-aware framework. More detailed evaluation statistics across multiple cases and time periods will be conducted once the planned refinements (discussed in Section 6) are implemented.

5.1. Aerosol Impacts on Assimilated Dataset

Figure 4 shows the differences in the simulated BTs between the scenarios with and without aerosol information in the AER experiment as a function of the assimilated IASI channels. The statistics are averaged globally for all hazy-sky data over the whole period. As prior studies have reported [1,3,4], aerosols mainly impact the thermal infrared radiance between 750 and 1200 cm⁻¹, which is in the window region. Among the assimilated channels in the GDAS, 962.5 and 1096 cm⁻¹ are two channels with relatively larger cooling effects. Note that the smaller magnitude of the cooling effect compared to that reported in [20] is due to the lack of the stratification of the AOD, which includes many points with a light aerosol loading.

Figure 5a displays the horizontal distribution of the simulated BT differences (B_aer − B_clr) for the assimilated data (i.e., after QC process) at 962.5 cm⁻¹ in the AER experiment. As expected, the strong signal of the cooler BTs is over the trans-Atlantic region and the Arabian Sea, where dust aerosols were present. The maximum cooling effect of over 2 K is found near the African coast. For these assimilated observations, Figure 5b shows the inflation of the observation error from the piecewise model (Section 3.3) with respect to the prescribed (i.e., clear-sky) observation error. There is a 10–30% increase over the trans-Atlantic region. Some grid points, however, could have increased by more than 90% compared to the CTL experiment. This could be attributed to the calculation of correlated observation error [52], which induces a smaller observation error compared to the prescribed value (dashed line in Figure 3).

To illustrate how the aerosol-aware QC works in the system, Figure 6 shows the QC categories of 962.5 cm⁻¹ from the IASI on top of the AVHRR natural-color RGB image from the partial swath near the African coast from the analysis cycle of 12Z 18 June 2020. The AVHRR natural-color RGB image is made by combining the 0.63, 0.87, and 1.61 µm channels for the blue, green, and red colors, respectively. Because both the IASI and the AVHRR data used here are from a MetOp-A satellite, the same time period means that the IASI and the AVHRR measured the same atmosphere. In the AVHRR natural-color RGB image, there is a strong dust plume between 10 and 20 °N near the African coast. Most of the pixels over this region were rejected by the cloud QC in the CTL (blue dots) and the AER (purple dots) experiments. This implies that the aerosol plumes can produce similar signals as clouds when they are thick enough. The observations over the same area in the AER experiment were identified as hazy-sky data but rejected by the cloud QC (purple dots). Only some of the hazy-sky observations were retained in the AER (magenta dots) due to the changes in the first-guesses under the fully cycled configuration. This reflects the strict QC strategy with the rejection of the potential cloud contamination in hazy-sky data. This also contributes to the smaller BT cooling shown in Figure 5, when compared the magnitude to the prior studies [5,20].

Figure 7 shows the QC categories fraction results between the CTL and AER experiments on the 962.5 cm⁻¹ channel. In the CTL experiment, it mainly distinguishes the IASI observations into clear-sky and cloudy-sky data. More than 80–90% of the data on this channel is rejected due to cloud contamination over mid-latitude and dust-laden regions (i.e., African coast, Sahara Desert, and Arabic Sea). In contrast, more than 50% of the data is identified as hazy-sky over dust-laden regions in the AER experiment. Among those hazy-sky data, less than 10% are assimilated (Figure 7c) with aerosol effects, while the other data are mainly rejected by the cloud QC (Figure 7f). Note that some points near the middle of the Atlantic that are assimilated as clear-sky in the CTL experiment are assimilated as hazy-sky in the AER experiment.

For other assimilated window channels, the results follow a similar pattern. Figure 8 shows the comparison of the observation usage and the assimilated first-guess departures (i.e., observation minus forecast/first-guess, OMFs) between the experiments as a function of the assimilated channels between 750 and 1200 cm⁻¹ from the IASI MetOp-A. There are no noticeable changes between the two experiments, but the differences show that the size of the AER’s assimilated dataset is about 0.5 to 1.5% smaller than that in the CTL experiment. Within a similar amount of data, Figure 8b indicates that the AER includes about 5% of the points that were identified as hazy-sky and passed the QC over most of the assimilated window channels. Given the similar dataset, Figure 8c shows the AER experiment assimilated slightly warmer first-guess departures with respect to the CTL experiment, which would lead to a warmer analyzed temperature, as reported in prior work [19,20]. However, we can expect that the impacts will be smaller compared to prior studies because of the strict QC strategy and the inflation of observation error. Note that similar datasets and the exclusion of hazy-sky data in the BC lead to the fewest impacts on the probability density function of the pre- and post-BC first-guess departures between the experiments (Figure S6).

Figure 9 displays the temporal evolution for the total bias estimates from the variational bias correction (VarBC) in the assimilated channels between 750 and 1200 cm⁻¹. The biases in the CTL experiment (Figure 9a) show that most of the channels between 1000 and 1200 cm⁻¹ have negative biases, and those channels around 780 to 1000 cm⁻¹ have more positive biases. Overall, the biases continually grow toward the positive side throughout the period. Compared to the CTL experiment, Figure 9b shows the absolute biases of the AER experiment become larger between 1000 and 1200 cm⁻¹ and smaller between 780 and 1000 cm⁻¹. Combining the features of the CTL experiment in both regions, the absolute differences indicate that the bias estimates are more negative in the AER experiment, which could be attributed to the changes in the first-guess states.

5.2. Against Conventional Dataset

While the assimilated dataset is similar among the experiments, the small amount (~5%) of hazy-sky data in the AER experiment still plays a role in the tropical analysis, especially over the trans-Atlantic region covered by a transported heavy dust plume during the period. The first-guess departures can represent the impacts on the DA system through the adjustment in the analyses. This section presents the evaluation of the in situ measurements of the sea surface temperature (SST) and the upper air temperature and specific humidity in the conventional dataset.

Figure 10 shows the first-guess departures of the SST measurements over the tropical region (20 °S to 20 °N) and the spatial differences in the RMS between the experiments averaged over the period. The mean and the RMS first-guess departures illustrate that the AER experiment agrees with the SST measurements better. Particularly in the period around June 20 to 25th, the historical “Godzilla” dust plume, with an AOD over 3 [48], was transported across the Atlantic Ocean, which is the area with the lowest RMS first-guess departures, as shown in Figure 10b. In addition, the correlation coefficient between the observed and the first-guess SSTs indicates a slight improvement in the AER experiment (0.969 vs. 0.966). These features imply that assimilating those hazy-sky data induces warmer SSTs in the analyses and propagates to short-range forecasts (6 h).

For the upper air measurements, Figure 11 shows the mean and the RMS first-guess departures for the temperature and the specific humidity in the tropical region. Compared to the SST, it shows drastically smaller differences between the two experiments. However, the AER experiment shows a warmer and more humid atmosphere at a lower level, which is in better agreement with the measurements given the smaller RMS and the biases. Note that the smaller differences could be attributed to the limited coverage over the oceanic areas, where larger impacts from the hazy-sky IR data are seen.

5.3. Evaluation of Analysis

We also evaluate the analyses of the experiments with the ERA5 reanalysis. Figure 12 shows the evaluation of the geopotential height, temperature, specific humidity, and zonal and meridional winds over the tropical region. Similarly to the neutral results in the first-guess departures of the upper air measurements (Figure 11), there are no significant differences between the experiments against the ERA5 reanalysis. However, the subtraction of the biases between the experiments illustrates the changes induced by assimilating the hazy-sky data. Compared to the CTL experiment, the analyses of the AER experiment have a thicker atmosphere, which may be attributed to the warmer temperatures below 600 hPa, the higher humidity below 850 hPa, and the lower humidity above; there are also westward and northward changes to the horizontal winds around 600 hPa.

In terms of the spatial distribution, Figure 13 shows the differences between the experiments for several fields. As we showed in the evaluation against the SST measurements (Figure 10), the differences indicate a systematic warming feature in the AER experiment over the trans-Atlantic region, where some hazy-sky data were assimilated. Unlike the surface temperature, the temperature at 700 hPa shows a scattered but slightly warming feature globally. Regarding the wind field at 600 hPa, as the profiles show, there are westward changes to the zonal wind (u-wind) near the equator and northward changes to the meridional wind (v-wind) in the tropics. This feature indicates that the low-level jet of the African easterly wave in the AER experiment may show more displacement compared to the CTL experiment. Like the evaluation with the conventional dataset, the horizontal differences show that the SST has a stronger response than the atmosphere. It indicates the SST analysis is more sensitive to the additional aerosol information than the atmosphere analysis.

5.4. Evaluation of Forecast

We also conducted 7-day forecasts at each 00Z between 10 June and 10 July 2020 to investigate the impacts of the aerosol-aware framework on the medium-range weather forecast. The comparison of the anomaly correlation coefficient indicates small differences in the forecast skill between the two experiments. Figure 14 shows the global mean root mean square error (RMSE) against the upper air measurements for the forecasts of the geopotential height, temperature, specific humidity, and winds from the CTL experiment and the differences in the AER experiment (AER minus CTL). The RMSE of each pressure level is shown as a function of the forecast hour. As illustrated in Figure 14, the default clear-sky framework (i.e., CTL) has a larger RMSE around 300 to 200 hPa for the geopotential height, temperature, and winds after 96 h of forecasts. The RMSE for the specific humidity is larger below 600 hPa and shows a similar magnitude after 60 h of forecasts. The differences in each variable show that both experiments are compatible before 120 h of forecasts. Beyond 120 h, there are more discernible changes in the AER experiment. It shows some improvements (blue colors) at the levels where the CTL experiment has a larger RMSE, which are contributed by the improvements over the tropical region. However, those smaller RMSEs in the AER experiment are not statistically significant, except the improvements at the low-level temperatures and the specific humidity.

6. Discussions and Conclusions

In this study, we introduced an aerosol-aware framework to assimilate the thermal infrared satellite observations under the hazy-sky condition. This framework was developed based on the all-sky DA works [16,17,50] and prior studies of aerosol impacts on radiance DA [19,20]. In addition, by incorporating the model’s aerosol information into the radiance observation operator [19,20,21,22], this framework includes the identification by aerosol detection, the proxy for aerosol effects, the screening process with new quality control checks, and the inflation of the observation error by the piecewise observation error model for the assimilated dataset.

The aerosol detection method in the framework, based on the CADS from the NWPSAF, provides the reasonable identification of hazy-sky observations when compared with the fraction of hazy-sky pixels to the VIIRS-gridded AOD retrieval product. Given the incomplete spectral information in the IASI dataset used in the NCEP, it can identify observations affected by dust aerosols over the ocean properly. However, it shows a high rate of misidentification, with many pixels that are not retrieved or show a zero AOD in the LMD product. Moreover, it has difficulty identifying hazy-sky pixels over land and pixels affected by smoke. These limitations lead to a strict QC strategy with the clear-sky cloud QC in this study.

Based on the aerosol-aware framework, the AER experiment was conducted for a period with strong activity of Sahara dust transportation over the Atlantic Ocean for comparison with the experiment of the clear-sky framework (i.e., CTL). For the hazy-sky pixels in the AER, the MERRA-2 reanalysis was used to provide the 3d distribution of the aerosol mixing ratios in the BT simulation of the radiance observation operator.

We evaluated the results of the AER experiment from various aspects, including the quality control and bias correction of the assimilated dataset, analyses against conventional and reanalysis dataset, and the performance of forecasts. Overall, the AER experiment shows little differences compared to the CTL experiment. Given the strict quality control strategy, about 5% of the assimilated IASI data in the AER were identified under hazy-sky and assimilated with the aerosol information. It leads to more positive first-guess departures and more negative estimated biases from the VarBC. The changes in the estimated biases between 780 and 1000 cm⁻¹ indicate the assimilated dataset in the AER experiment needs less bias correction. Assimilating this dataset results in a warmer analyzed sea surface temperature, which shows better agreement with the measurements in the tropical region. It also shows limited changes to the analyses of the atmosphere, while the differences in the zonal wind suggest the displacement of a low-level jet around 600 hPa. More diagnoses are needed to understand the impacts on circulation.

The different responses to changes in the sea surface temperature and the air temperature reveal the limitation of the proposed aerosol-aware framework. From the comparison of the AOD from MERRA-2 and the VIIRS, we learned that MERRA-2 has a reasonable total loading of aerosols. In this case, the properly represented total attenuation can lead to the improvement of the analyzed sea surface temperature in the AER experiment, which is more sensitive to the column’s integral impacts on the radiance. However, it is more challenging to improve the air temperature analysis because of the uncertainties regarding the aerosol’s vertical structure.

Based on the partial swath image of the quality control categories over the top of the AVHRR natural-color image (Figure 6), the application of the quality control for clouds constrains the impacts of the high misidentification rate in the ADM (Table 1) on the system. However, this strict quality control removes many hazy-sky pixels. This indicates that the refinement of the aerosol detection is needed for the framework. Like we showed in Table 1, the full spectrum data of the IASI over the key channels can provide a better identification rate. Therefore, we plan to improve the aerosol detection over different surface types and for different aerosol types by using the full spectrum of the IASI data near the key channels, adapting thresholds regionally, and exploring machine learning methods. Then we will revisit the identification performance by an extensive comparison against independent AOD retrievals from space and ground (e.g., VIIRS, AERONET). By improving the aerosol detection method, we can omit the cloud QC in the hazy-sky dataset and assess the impact of different QC strategies on the analyses.

After including those hazy-sky pixels rejected by the cloud QC, we should revisit the bias correction given the substantial biases in the IR window. We conducted a simple sensitivity study without the cloud QC on hazy-sky pixels and applied the bias correction based on the A_e and the interception and the slope of first-guess departures from the “observer” run. This test shows how relaxing the QC filters affects the sample size and analysis skills and clarifies the trade-off between coverage and purity. By assimilating more hazy-sky pixels and applying the linear bias correction, it introduces an over-correction (Figure S7) and degrades the agreement with the SST measurements in the tropics. It reveals that a proper predictor (e.g., the A_e and the modeled AOD) for the VarBC is needed to properly assimilate those hazy-sky data.

In terms of the operational implementation, an aerosol forecast model with the DA system regularly producing the analysis of the aerosol mixing ratios is preferred to replace the aerosol reanalysis from MERRA-2. In addition, the optimization of the aerosol-aware framework needs to be addressed, because it costs around double the run time (2902.6 vs. 1582.3 s per cycle) for each step of the GSI analysis. This is attributed to the additional I/O of the aerosol mixing ratios from MERRA-2, calculations of the A_e, and the newly introduced QC steps.

In summary, we plan to revisit the following aspects to enhance the aerosol-aware framework in the future:

• Updating the IASI and the other IR data to be full-spectrum around the four key channels used in the ADM;
• Evaluating the ADM’s performance over different surfaces and for different aerosol types by comparing it with independent AOD observations;
• Adding bias correction predictors based on the A_e or the modeled AOD into the VarBC;
• Exploring the use of machine learning methods in aerosol detection and dynamic error estimation;
• Optimizing the framework.

The enhanced aerosol-aware framework will be investigated via a series of sensitivity tests that is similar to that used by Wei et al. [5] to reveal the relationship among the aerosol vertical structure, Ae, Jacobians, Ae-dependent observation error, and the analysis increments. Furthermore, a more detailed evaluation over multiple time periods to demonstrate the impacts on the analysis under different scenarios will be conducted.

Unlike prior studies, this study first constrained the aerosol effects in the thermal infrared radiance of a fully cycled DA system. This unveiled the challenges of a comprehensive all-sky IR data assimilation that includes the data under hazy-sky conditions, rather than tossing them out. When the comprehensive all-sky DA is fully implemented, the step of hazy-sky identification is not needed, as shown in this study, because every available data point should be considered. Furthermore, the aerosol effect parameter can be combined with the cloud effect parameter [17] to provide a proxy for the total column attenuation. In addition, through our work with the all-sky IR DA, this study lays a foundation for coupled DA work between composition and meteorological analyses. In a coupled DA system, the improvements in the 3-dimensional aerosol distribution identified through the aerosol DA can feedback and constrain the meteorological analysis by considering the aerosol information during the simulation of the IR radiance.