Consistency of Tropospheric Water Vapor between Reanalyses and Himawari-8/AHI Measurements over East Asia

Although water vapor (WV) comprises only 1% to 4% (by volume) of the whole atmosphere, it is the strongest greenhouse gas and plays a critical role in the atmospheric energy budget equilibrium (Held and Soden, 2006; Trenberth et al., 2009). The amount of outgoing longwave radiation is largely determined by the WV concentration in the atmosphere (Held and Soden, 2000), since the radiance signals relevant to WV variations are quite strong and occur over most of the infrared (IR) longwave spectrum. WV is also significant in the hydrological cycle through its complicated interactions with clouds and precipitation (Chahine et al., 2006; Oki and Kanae, 2006). An accurate description of the distribution and variability of WV is essential for weather and climate diagnostics.

Global data are widely used for weather and climate diagnostics. Gridded reanalysis datasets combine various observations and numerical weather prediction (NWP) model outputs to provide comparatively reliable estimates of past atmospheric states (Fujiwara et al., 2017) using a consistent data assimilation system. In recent decades, more than 10 different atmospheric reanalysis datasets have been developed, freely released, and widely used in the climate research community to understand atmospheric processes and variability, to validate chemistry–climate models, and to investigate and identify climate change (Randel et al., 2004; Erying et al., 2010; Davis et al., 2017). Unfortunately, conflicting, or incompatible results are commonly found in the same weather or climatic diagnostics using different reanalysis datasets, such as the atmospheric energy budget and hydrological cycle (Trenberth et al., 2011) and temperature trends (Randel et al., 2009; Xu and Powell, 2011); and those differences might be related to inconsistent technical schemes adopted in different reanalysis processing systems (Fujiwara et al., 2012, 2017). Note that WV is one of the most poorly understood atmospheric variables in reanalysis datasets due to insufficient WV measurements in assimilation and imprecise physical parameterization schemes in current numerical model systems (Takahashi et al., 2016; Polavarapu and Pulido, 2017). Therefore, it is meaningful to assess the abilities of reanalysis datasets to reproduce the distribution and variability of WV, identify WV biases or differences among various datasets, and suggest future improvements for reanalysis processing systems.

Numerous objective assessments of upper-tropospheric WV (around 200–400 hPa) in reanalysis datasets have been conducted (Soden, 2000; Tian et al., 2004; Chung et al., 2007, 2009, 2011). In these assessments, radiances at the top of the atmosphere (TOA) in the 6.7-μm WV absorption band measured from geostationary Earth orbit (GEO) or low Earth orbit satellites are commonly used as a reference. The insufficient spatial resolution of in-situ observations limits their application in assessing the distribution and availability of WV on large scales. The basic physical theory in this assessment is that the upward radiance emitted from the lower-tropospheric WV would be appreciably absorbed by upper-tropospheric WV, especially in the 6.7-μm band, implying a dominant role of upper-tropospheric WV content in the radiance observed by the satellite. Therefore, the distribution and variability of upper-tropospheric WV in reanalysis datasets can be evaluated based on long-term satellite WV band observations.

Soden and Bretherton (1996) demonstrated the relationship between the 6.5-µm band radiance [or brightness temperature (BT)] and WV in the upper troposphere, and Xue et al. (2020a, b) used the global homogenized 6.5-μm radiances from international GEO satellites to evaluate monthly and diurnal variations of upper-tropospheric WV in several reanalysis datasets on a near global scale. The emphasis was placed on the close relationship between radiances from the 6.5-μm WV band and the moisture in the upper troposphere between roughly 200 and 600 hPa [wide range; see Fig. 1 in Xue et al. (2020a)]. The unique aspect of their study was the direct comparisons between observations and reanalysis datasets in radiance space on a near global scale. However, due to a lack of homogenized multiple WV bands, their study was limited to one WV band (i.e., based on the GOES-15 imager 6.5-μm WV band), which can only provide single-layer WV information. Therefore, their evaluation was confined to upper-tropospheric moisture.

Figure 1.  Normalized spectral response functions (SRFs) of WV absorption bands for AHI/H8 (colors) and S-VISSR/FY2E (black). The gray line in the panel represents IASI (Infrared Atmospheric Sounding Interferometer) BTs of upwelling radiances at the TOA simulated with the standard atmospheric profile of midlatitude winter.

The advanced imaging systems on the new-generation GEO meteorological satellites, such as Himawari-8/-9 (H8/9) (Bessho et al., 2016), Geostationary Operational Environmental Satellite-R (GOES-R series, GOES-16/-17/-18) (Schmit et al., 2005), FengYun-4A/-4B (FY-4A/-4B) (Yang et al., 2017), and the pending Meteosat Third Generation imaging mission (MTG-I) (Holmlund et al., 2021), carry multiple WV absorption IR bands. The observational data from multiple GEO WV bands provide a great opportunity for the evaluation of multilayer atmospheric WV in reanalysis datasets. It could improve our understanding in various reanalysis datasets of the deficiencies in depicting WV; update the knowledge of WV biases in the upper, middle and lower troposphere; and improve the parameterization of physical processes relevant to regulating the vertical distribution of WV in the respective assimilation–forecast systems.

Our focus in the present paper is on assessing multilayer WV depiction in six representative reanalysis datasets against the measurements from three independent Advanced Himawari Imager (AHI) WV bands (using the AHI WV band measurement as a reference) over East Asia (0°–45°N, 90°–140°E), and in particular focusing on the monthly variability over the Tibetan Plateau (TP). Section 2 describes the AHI observations for the reanalysis assessment, the six reanalysis datasets, and the evaluation methods. The assessment results from multilayer WV over East Asia are elaborated in section 3. Further assessments of WV monthly variability over the TP are presented in section 4. The shifts of the subtropical high found in both reanalysis and observations are presented in section 5. Finally, the major findings are summarized in section 6.

H8, the first new-generation GEO weather satellite operated by the Japan Meteorological Agency (JMA), was successfully launched on 7 October 2014. As the main Earth-viewing sensor on H8, AHI is remarkably improved over previous sensors carried on the first-generation GEO weather satellites, such as the FY-2 series satellites (Min et al., 2017). Specifically, as shown in Fig. 1, one broad WV band around 6.7 μm was used to monitor WV, which is illustrated by the Stretched Visible and Infrared Spin Scan Radiometer (S-VISSR) onboard the FY-2E satellite. The AHI’s ability to resolve the WV structure has been significantly enhanced with the increased spatial resolution and signal-to-noise ratio in the IR WV absorption bands. The H8/AHI has three narrower IR WV bands (centered around 6.21, 6.93, and 7.34 μm) to detect multilayer moisture in the troposphere. In addition, the AHI observes the atmosphere with high spatial (0.5 km for visible bands and 2.0 km for IR bands) and temporal (10 min for full disk) resolutions. Moreover, the measurement precision of AHI WV bands is affirmed by inter-calibration with measurements from several hyperspectral sounders (Hewison et al., 2013); near real-time assessment results for the AHI measurements conducted by the Global Space-based Inter-Calibration System (GSICS) are routinely released (http://www.data.jma.go.jp/mscweb/data/monitoring/gsics/ir/monit_geoleoir.html). The absolute radiometric calibration biases for the AHI’s three WV bands are found to be 0.1–0.2 K according to the official GSICS website; this demonstrates their sufficient accuracy for evaluating the WV found in reanalysis datasets or forecasts from NWP models. In addition, AHI measurements can be used as an independent reference since only a very small portion of the high spatiotemporal resolution [2 km (10 min)⁻¹] AHI data are used in global NWP models after thinning.

WV information content is reduced and the best WV information layer (BWIL) is lifted when the terrain height is elevated (Di et al., 2016). Therefore, it is necessary to separate evaluations over the plains and TP. The same midlatitude summer and winter atmospheric profiles (shown in Figs. 2a and b) but with different terrains (plains and TP, respectively) are used to calculate the Jacobian functions of the AHI’s three WV bands. The slightly negative WV Jacobian value means the measured radiance at the TOA decreases as the atmospheric WV concentration increases. Figures 2c–e show that the heights of the BWILs for the three H8/AHI WV bands increase from winter to summer. In summer, the main WV information layer observed by the 6.21-μm (band 08), 6.93-μm (band 09), and 7.34-μm (band 10) bands of the AHI are around 200–350 hPa, 250–400 hPa, and 300–500 hPa, respectively. In winter, meanwhile, the three WV bands sense comparatively lower layers (about a 50–100 hPa increase). In addition, the seasonal and regional differences in BWILs are also remarkable. Figures 2c–e demonstrate that the topographic elevation tends to lift the height of the BWIL and narrow the full width of the Jacobian function peak by comparing the WV Jacobians over two regions in winter, especially for band 10. It indicates that band 10 (7.34-μm band) could detect WV in the 300–500-hPa layer throughout the year over the TP, while over the plains this band could monitor WV in the lower troposphere when the atmospheric environment is relatively dry. Furthermore, it suggests a possible influence of the surface on band 10, especially over the TP. In terms of annually averaged results, the 6.21-μm band mainly observes upper-tropospheric WV (around 200–350 hPa); the 6.93-μm band monitors WV in the upper and middle troposphere (around 300–450 hPa); and the 7.34-μm band detects WV in the middle troposphere (around 400–600 hPa). Of course, for each band, the BWIL varies depending on the atmospheric WV content.

Figure 2.  Two atmospheric (a) temperature and (b) WV profiles of midlatitude winter and summer, respectively, and (c–e) the corresponding WV Jacobian functions [d(Tb) ∕d(lnq)] of the AHI’s three WV bands (6.21, 6.93, and 7.34 μm; red for summer and blue for winter). Tb stands for brightness temperature and q stands for water vapor mixing ratio (k kg⁻¹). Two types of terrain (solid line for plains with surface pressure of 1000 hPa; dashed lines for the TP with surface pressure of 600 hPa, respectively) are assumed for the WV Jacobian calculations.

In this investigation, six representative reanalysis datasets produced at several major meteorological forecast centers [the European Centre for Medium-Range Weather Forecasts (ECMWF), National Centers for Environmental Prediction (NCEP), National Aeronautics and Space Administration (NASA), JMA, and China Meteorological Administration (CMA)] are evaluated. Some details of these reanalysis datasets including their spatial/temporal/vertical resolutions are summarized in Table 1. Some highlights are given below. References providing deeper reviews of these reanalysis datasets are listed in Table 1 for more detail.

The reanalysis datasets evaluated in this study are modern reanalyses generated from state-of-the-art numerical model and assimilation systems. All of these reanalysis datasets have assimilated surface and upper-air conventional observations as well as satellite observations, and their performances are significantly improved. Nevertheless, the proportion of satellite observations that have been assimilated is limited, especially in the upper troposphere where the observations are sparse and reanalysis data rely heavily on their forecast model. Moreover, the forecast models, assimilation systems, and assimilated observation sources are quite different in the various reanalysis datasets (Fujiwara et al., 2017). Such discrepancies further lead to inter-reanalysis differences (Jiang et al., 2015). Previous studies have demonstrated that deficiencies in convection parameterization schemes and large-scale dynamics could cause errors in WV fields of the reanalysis models. The uncertainties in observations that are currently assimilated in the reanalysis systems could be another error source.

ERA-Interim and ERA5 were produced by ECMWF. The humidity analysis scheme in the ECMWF reanalysis involves a nonlinear transformation of the humidity control variable to make the humidity background errors nearly Gaussian (Hólm, 2002). The model cycle used in ERA-Interim only includes supersaturation in the cloud scheme of the nonlinear forecast model, which will cause an increase in relative humidity in the upper troposphere (Tompkins et al., 2007). When compared with ERA-Interim, the latest ECMWF reanalysis, ERA5, is significantly improved in terms of its resolution, data assimilation system, and model physics. ERA5 provides better performance in the global balance of precipitation and evaporation among other key differences.

The Climate Forecast System Reanalysis (CFSR) is the first global reanalysis generated from the coupled atmosphere–ocean–sea-ice system (Fujiwara et al., 2017). CFSR was migrated to the operational CFSv2 analysis system from 2011. The reanalysis evaluated in this study is CFSv2, which has minor changes to physical parametrizations and can be treated as a continuation of CFSR. CFSR only assimilates radiosonde humidity at pressure levels of 250 hPa and larger.

The MERRA-2 dataset is produced at NASA’s Global Modeling and Assimilation Office using the latest version of the Goddard Earth Observing System (GEOS-5) data assimilation system, which can ingest recently available observations such as modern hyperspectral radiance and microwave observations (Lin, 2004), while most of the reanalysis models use spectral dynamical cores (e.g., Machenhauer, 1979). It should be noted that the rain and snow re-evaporation was intentionally increased in MERRA-2 to make the upper troposphere wet enough to strengthen the stationary wave and the boreal winter circulation pattern (Molod et al., 2015).

The JRA-55 dataset, released in 2013, is the second reanalysis dataset created at JMA using the Global Spectral Model (GSM). Two major biases (a cold bias in the lower stratosphere and a dry bias in the Amazon basin) in JRA-25 were reduced because of the upgraded assimilation system in JRA-55. Observations of humidity at pressures lower than 100 hPa were not assimilated in JRA-55 and the vertical correlations of humidity background errors were set to zero at pressures lower than 5 hPa to prevent the spurious increments at higher levels.

The CRA dataset is the first-generation atmospheric reanalysis dataset developed by the CMA from early 2014. The CRA reanalysis project aims to produce a 40-year (1979–2018) global atmosphere and land reanalysis dataset, and then continue production in near real-time. So far, a 10-year interim product has been generated with a spatial resolution of 0.3125° latitude × 0.3125° longitude. The current CRA system is based on NCEP’s Global Forecast System (GFS) plus a Gridpoint Statistical Interpolation 3DVAR data assimilation system. In comparison to other reanalysis datasets, CRA has assimilated more observations over East Asia.

AHI radiances are not assimilated in these reanalysis datasets, except ERA5. Even when AHI radiances are assimilated, the proportion of AHI radiances used is very limited, considering the thinning strategy in the data assimilation system; therefore, AHI can provide independent information to evaluate WV fields produced by the reanalysis systems.

A point-to-point direct comparison between satellite observations and the reanalysis datasets is impossible because the reanalysis datasets use humidity quantities to depict the WV field while satellite sensors observe TOA radiances that contain the WV information (e.g., mixing ratio). Usually, evaluations are conducted by converting the radiance measurements into humidity quantities for direct comparisons or converting atmospheric profiles from a reanalysis dataset into radiances for indirect comparisons. The former approach needs both a radiative transfer model (RTM) and an accurate inversion algorithm, while the latter approach only needs an RTM. Due to the coarser spectral resolution, the accuracy of the retrieved moisture profile from sensors like the AHI is not as good as that from a hyperspectral IR sounder (Schmit et al., 2009). To avoid retrieval errors, the evaluations are thus based on the profile-to-radiance approach.

In this study, the classic Radiative Transfer for TOVS (RTTOV) model developed by the NWP Satellite Applications Facility from EUMETSAT (Eyre, 1991; Saunders et al., 1999) is used as the forward RTM. According to the assessments shown on the RTTOV official website, the root-mean-square errors (RMSEs) for WV band simulations are within 0.2 K of the most accurate Line-by-Line RTM (Clough et al., 1989, 1981, 1992, 2005; Di et al., 2018). This demonstrates that the RTTOV model is accurate enough to do the model-to-radiance conversion. The RTTOV model input includes atmospheric vertical profiles of temperature, WV concentration, surface properties, and satellite zenith angle for TOA radiance simulations. The surface emissivity is from the global IR land surface emissivity database (Seemann et al., 2008) developed at the Cooperative Institute for Meteorological Satellite Studies at the University of Wisconsin–Madison. Although the simulation accuracy of the RTM in cloudy skies has been greatly improved, its accuracy is still not as high as that in clear skies. In this study, we only carry out assessments under clear-sky conditions and avoid cloud contamination. It should be noted that the RMSE of 0.2 K reflects the RTM model uncertainty. When compared with the observations, the main source of uncertainty in radiative transfer simulation comes from input parameters such as atmospheric temperature and moisture profiles for WV absorption bands. The physical basis for evaluating moisture from reanalysis is through comparison between simulated and observed BTs. The level-2 cloud-mask products specific to the AHI developed at the CMA (Min et al., 2017) are used to determine the clear-sky pixels (only measurements with confirmed clear pixels are used). Intercomparisons among collocated products from FY-4A, H8, and the Moderate-resolution Imaging Spectroradiometer (MODIS) (version 6.0) suggest that the AHI cloud mask has an overall accuracy of better than 90% (Wang et al., 2019), which implies that the instantaneous uncertainty of the clear portion of the BT within the reanalysis sub-grid is less than 0.2 K (Zhang et al., 2006).

Confining our assessment to clear skies could pose a problem when using IR WV measurements or moisture-relevant products (e.g., upper-tropospheric humidity) as the verification resource. An inherent dry bias in IR satellite observations will be introduced by clear-sky sampling; specifically, 9%–30% drier in terms of relative humidity due to clear-sky sampling (e.g., Sohn et al., 2006; John et al., 2011; Shi et al., 2022). While the bias is less noticeable for shorter time scales, it becomes more obvious as the averaging time increases to the climate scale (Lanzante and Gahrs, 2000; John et al., 2011). However, the dry bias will not affect our findings since, because of our procedure, the simulations from the six reanalysis datasets and AHI observations are both confined to clear skies. However, it should be noted that our assessment conclusions only apply to clear-sky situations.

All WV band observations from H8/AHI at 10-min intervals in 2016 over East Asia are included in this study. Given the remarkable differences in spatial and temporal resolutions between the six reanalysis datasets and AHI band observations, the spatial collocation should be conducted first for both observations and simulations after model-to-radiance calculations are done. The simulated BTs from each reanalysis and the observed BTs from the AHI are spatially interpolated into a 0.5° × 0.5° horizontal grid resolution using an inverse distance square weighted method [see Eq. (1)] for every day in 2016 at 0000, 0600, 1200 and 1800 UTC.

where BT_ori and BT_grid are the BT observations (or simulations) at the original location and interpolated to the grid, respectively; R is the target interpolation grid distance (0.5° × 0.5° is used here), and r is the distance from the original location to the grid center. The evaluations of multilayer WV in the reanalysis data are mainly based on BT differences (BTDs; simulation minus observation), along with BT pattern differences. The average fraction of cloud-free grid points over the study region for BT comparisons is 0.48 in spring, 0.33 in summer, 0.36 in fall, and 0.50 in winter.

As explained below, the BTD between simulations and observations is useful for interpreting the WV bias in the reanalysis datasets. Sensitivity tests are conducted to discuss the impacts of temperature and humidity perturbations on BT at the TOA. Figures 3a–c show that BT will increase as the atmospheric temperature increases. This result is reasonable since the atmospheric temperature directly contributes to the radiance through emissions, and thus warmer temperatures lead to positive responses at the TOA. On the other hand, WV absorbs the emissions, and Fig. 2 shows that more WV lifts the Jacobians. Since temperature decreases with altitude in the troposphere, more moisture leads to a negative response at the TOA. Therefore, the TOA is reacting to a combination of atmospheric temperature and moisture changes. Figures 3d–f show that the relationship between WV perturbation (%) uniformly applied throughout the vertical profile and BT change is logarithmic, which is consistent with the earlier findings in Soden and Bretherton (1996). This test indicates that the BT in the WV band is responding more to the humidity concentration than temperature (e.g., 15% moisture change versus 1 K temperature change). Because the atmospheric temperature distributions are generally well depicted in the reanalysis datasets, the atmospheric temperatures from the reanalysis datasets are very close to each other and are all close to radiosonde observations (Trenberth et al., 2001; Santer et al., 2003). For example, intercomparisons between reanalysis datasets and radiosonde observations over China indicate small atmospheric temperature differences (~1 K) (Guo et al., 2016). Figure 4 presents direct intercomparisons of area-weighted and monthly averaged atmospheric temperatures and moisture at 200 hPa, 300 hPa, 500 hPa and 750 hPa over East Asia from the six reanalysis datasets. The inconsistencies among their WV are much larger than for temperature on annual and monthly time scales, especially for the upper and middle troposphere (the annual result is not shown). It is therefore reasonable to attribute the BTDs between simulations and observations to the moisture uncertainty in the reanalysis datasets. Specifically, a negative BTD (simulation minus observation) means that the reanalysis dataset tends to overestimate the WV concentration and shows a wet bias in comparison to the observations. In addition, for the same magnitude of humidity perturbation, the impact on BT caused by the increase in WV is less than that caused by a decrease in WV. This finding implies that, even for the same BT bias, the corresponding WV bias in the reanalysis data is more significant in summer than in winter. Similarly, for the same BT bias, the corresponding WV bias in the reanalysis data is more significant over the TP than over the plains, especially for the middle and lower troposphere.

Figure 3.  BT perturbations caused by (a–c) temperature perturbations and (d–f) humidity perturbations for the three AHI WV bands at 6.21, 6.93 and 7.34 μm, respectively. The test profiles are the abovementioned midlatitude standard atmospheric profiles in winter and summer with a different surface pressure assumption.

Figure 4.  Area-weighted and monthly mean atmospheric temperature (left) and moisture (right) at (a, b) 200 hPa, (c, d) 300 hPa, (e, f) 500 hPa and (g, h) 750 hPa from six reanalysis datasets over East Asia.

The evaluation begins by first comparing the probability distribution functions (PDFs) of all the gridded BT observations and simulations for the three AHI WV bands in 2016. The aim is to assess the overall accuracy of the atmospheric multilayer WV in the six reanalysis datasets. The PDFs of the BT observations and simulations for the three WV bands are presented in Figs. 5a, c and e, respectively. The figures show that, except for the PDFs of the MERRA-2 dataset, which clearly deviates from the observations significantly, the PDFs of the reanalysis datasets are quite close to the observations. For the MERRA-2 dataset, the BT simulations are significantly colder than the observations for all three WV bands, with mean BTD biases of −2.07 K for band 08, −1.6 K for band 09, and −0.6 K for band 10. These negative biases correspond to enhanced wetness by about 40% for WV08 (WV absorption band 08), 20% for WV09, and 10% for WV10 according to Fig. 3, and indicate a relatively large wet bias in the troposphere. The results in Fig. 4 also prove that MERRA-2 shows a much larger wet bias compared with the other reanalysis datasets, especially in the upper troposphere. In addition, the upper-tropospheric temperature in MERRA-2 is slightly higher than in the other reanalysis datasets (Fig. 4). If only the temperature bias in the MERRA-2 dataset are taken into consideration, it will lead to a higher BT than other re-analyses, which is a completely opposite results to Fig 5. . Therefore, this indicates that the excessively large humidity in MERRA-2 dominates the BTD and results in a large negative BTD value. For the other reanalysis datasets, the WV08 BTDs are negative, with absolute values less than 1.0 K, indicating enhanced wetness of less than 20% according to Fig. 3. Similarly, the WV09 BTDs are negative, with absolute values less than 0.5 K, indicating an enhanced wetness of 10%, except in ERA-Interim. Also, the WV10 BTDs are less than 10%, wetter or drier, depending on the reanalysis. The PDF skill scores for the six reanalysis datasets (right-hand panels of Fig. 5) are calculated from the proportion of overlap between the target PDF (simulation) and the reference PDF (observation), and normalized by the integral area of the reference PDF (Mao et al., 2010). This score provides a quantitative assessment of the similarity between two different PDFs. A higher PDF skill score indicates greater accuracy in the WV field depicted in the reanalysis dataset. Among the six reanalysis datasets, the simulations from ERA-5 are closest to the AHI observations, based on the PDF skill score. In addition, for all the reanalysis datasets, the PDF skill scores for WV10 are larger than for WV09 or WV08. Therefore, the overall performance of the six reanalysis datasets is best in depicting WV10.

Figure 5.  PDFs of BT observations and simulations from the six reanalysis datasets for (a) band 08, (c) band 09, and (e) band 10, at 6.21, 6.93 and 7.34 μm, respectively. The right-hand panels show the corresponding PDF skill scores specific to the six reanalysis datasets using the observation as a reference for (b) band 08, (d) band 09, and (f) band 10. (g) Mean BTDs (simulation minus observation) and (h) BTD STDs for the six reanalysis datasets.

The BTDs (simulation minus observation) and standard deviations (STDs) for the six reanalysis datasets are shown in Figs. 5g and h, respectively. For WV09 and WV08, all reanalysis datasets show significant wet biases, with MERRA-2 showing the largest ones. For WV10, ERA-Interim and MERRA-2 exhibit a smaller wet bias, in contrast to ERA5, CFSR, JRA-55 and CRA, which show a dry bias.

The evaluations in Fig. 5 are focused on overall WV PDF biases but ignore the temporal and spatial variations of the WV bias. Next, the ability to capture climate features (e.g., the annual cycle) with regard to WV variations for the six reanalysis datasets is evaluated. The annual cycle patterns of BT observations and simulations for the three AHI WV bands are shown in Figs. 6a–c, respectively. The relatively consistent variation phase of the BT observations and simulations suggests that these six reanalysis datasets can reproduce the seasonal variation tendencies of the WV fields. However, the BT simulations from the six reanalysis datasets for band 08 and band 09 are colder than observed in all months, which results from the consistent wet biases in these reanalysis datasets. The corresponding BTDs for the three AHI WV bands are also shown in the right-hand panels of Fig. 6. For WV08 and WV09, the wet biases of all reanalysis datasets consistently decrease in boreal winter and increase in boreal summer, with a peak approximately in July. However, for WV10, the seasonal variations in WV biases vary among the six reanalysis datasets: MERRA-2 and ERA-Interim show wet biases that are comparatively large in summer; CFSR, JRA-55 and CRA present a dry bias that is lowest in summer; and ERA5 biases are small and stable all year. It should be noted that the influences of the surface on band 10 are non-negligible, and the bias from surface parameters likely contributes to the reanalysis–observation discrepancies for band 10, especially over the TP. This aspect is discussed in section 3.3.

Figure 6.  (a–c) Annual cycles of BT simulations from the six reanalysis datasets and AHI observations for band 08, band 09 and band 10, respectively. (d–f) Annual cycles of mean BTD between simulations and observations for band 08, band 09 and band 10, respectively.

The spatial patterns of annually averaged BT observations and annually averaged BTDs for the six reanalysis datasets over East Asia are shown in Fig. 7. For WV08, all reanalysis datasets tend to overestimate the WV concentration over the entire area. The wet biases are comparatively large in MERRA-2, especially over the ocean (or land–sea contrast). For WV09, the dry biases begin to appear in CFSR, JRA-55 and CRA over the TP, and JRA-55 has the largest spatial coverage of dry bias, covering the TP, East China, Korea, and the West Pacific Ocean. For WV10, the spatial coverage of the dry biases for the six reanalysis datasets increases remarkably compared with WV09 or WV08 and are more predominant in ERA5, CFSR, JRA-55 and CRA. One common phenomenon among the six reanalysis datasets is that the large WV10 biases mainly occur over the TP. It is also interesting to see that in ERA-Interim, ERA5 and MERRA-2, the bias sign (negative or positive) over the TP is opposite to that of the surrounding regions. For example, the dominant wet bias over the TP in ERA5 transitions to a widespread dry bias in the surrounding regions. As suggested by Fig. 2, the bias from surface parameters should contribute to the reanalysis–observation discrepancies for band 10, especially in winter over the TP. The monthly averaged and area-weighted skin temperatures of the reanalysis datasets over the TP are different, especially for CFSR (Fig. 8). Combined with Fig. 7, we can see that, over the TP, cold surface temperatures correspond to a cold bias in ERA5; and similarly, warm surface temperatures correspond to a warm bias in CFSR. However, over the plains, this relationship is not obvious. Therefore, the bias in band 10 is partly attributable to the bias in surface temperatures, not only the WV bias.

Figure 7.  Spatial distributions of annually averaged BT observations and the BTDs for the simulations from ERA-Interim, ERA5, CFSR, MERRA-2, JRA-55 and CRA over East Asia in 2016. From left to right, the three columns depict band 08, band 09 and band 10, respectively.

Figure 8.  Area-weighted and monthly mean (a) surface pressure, (b) skin temperature, (c) 2-m air temperature and (d) 2-m moisture from six reanalysis datasets over the TP.

Various studies have suggested that upper-tropospheric humidity bias is strongly affected by bias in the large-scale circulation (e.g., Schmetz and van de Berg, 1994; Soden and Bretherton, 1994; Allan and Ringer, 2003; Pierce et al., 2006; Shi et al., 2022). For example, Chung et al. (2011) proved that the geographical distribution of humidity bias exhibits a close association with differences in 500-hPa vertical pressure velocity, suggesting that much of the bias in relative humidity in the tropical upper troposphere can be attributed to errors in simulating the intensity of large-scale tropical circulation. The underlying cause of upper tropospheric humidity bias might be partly attributable to problems in the parameterization of deep convection. Combining the information from the multilayer WV helps to indicate the deficiencies in capturing the WV vertical transport or convection events in each reanalysis dataset. It should be noted that we do not try to understand the possible parameterization scheme deficiency here, as this is beyond the scope of the present study. Instead, we try to analyze the underlying question when a reanalysis dataset with such a vertical structure of WV in NWP or nowcasting. For example, it is very common to use such a reanalysis dataset as the truth in machine learning–based approaches for moisture retrieval (e.g. Zhu et al., 2023), so how will it then influence convection nowcasting. Compared with observations, except for ERA-Interim and MERRA-2, the reanalyses show drier air in the middle troposphere and moister air in the upper troposphere over most of East Asia. The origins of the reanalysis model biases are still unclear and their impact on prediction/projection remains unknown, which are aspects that remain to be evaluated (Lin et al., 2017). However, studies indicate that vertical moisture has a significant effect on conditional instability. Theoretically, atmospheric structures with a larger WV vertical gradient (for example, nearly saturated air conditions in the lower troposphere) might have a greater likelihood for conditional instability when the atmospheric temperature structure is fixed. WV measurements from dropsondes abutting and across the ITCZ also support this view (Stevens et al., 2017). From this perspective, the ERA5, CFSR, JRA-55 and CRA datasets may underestimate the conditional instability compared with observations if using them in nowcasting applications.

The WV bias features over the TP are distinct from other regions of East Asia in terms of their magnitude and sign as well as levels of inconsistency across the various reanalysis datasets. Therefore, further study is specifically made into analyzing the WV bias in these reanalysis datasets over the TP. However, because H8/AHI cannot observe the whole TP because of its equatorial location at 140.7°E, the study region is thus limited to (25°–40°N, 90°–110°E) in order to keep the local zenith angle less than 65°.

It is necessary to point out that, unlike the observation–reanalysis discrepancies resulting from obvious viewing angle dependency, the BT observations themselves are highly dependent on the viewing angle of the GEO satellite sensor. Because the absorption path is longer when the satellite observes the atmosphere at a large zenith angle, the radiance received by the sensor will be reduced. In sensitivity tests, radiance measurements at the TOA for the three AHI WV bands over the TP are simulated (given the satellite zenith angle and the atmospheric profile), and the BT differences between the slant path and subsatellite point are shown in Fig. 9. BT differences are much larger when the AHI observes the southeastern and northwestern parts of the TP, even when the two regions have the same atmospheric conditions, especially in summer. However, in the case of the meridional and zonal average (shown in Fig. 10), the meridional or zonal BT differences introduced by the variation in the satellite’s zenith angles are reduced significantly and almost remain unchanged in the two seasons (reflected by the very similar variation in the blue and red lines). The results in Fig. 9 help us to understand the findings in Fig. 10.

Figure 9.  BTDs caused by the zenith angle variation of the AHI (with the atmospheric profiles unchanged) over the TP. Two seasons are considered, along with the corresponding meridional and latitudinal BT differences’ variation patterns for the three bands.

Figure 10.  The meridional distribution (averaged over 90°–110°E) of monthly averaged BT for observations and simulations from ERA-Interim, ERA5, CFSR, MERRA-2, JRA-55, and CRA. From left to right, the three columns di band 08, band 09 and band 10, respectively. The green box shows the pattern with higher BT in the north and lower BT in the south in observations from AHI band 09.

Next, the meridional distribution (averaged over 90°–110°E) and zonal distribution (averaged over 25°–40°N) of monthly BT for both AHI observations and simulations from the six reanalysis datasets over the TP are compared. Figure 10 shows that the BT gradients between the northern and southern regions of the TP show a remarkable seasonal transition. The large BT gradients in winter gradually reduce in spring. The patterns with higher BT in the south and lower BT in the north are weakened or disappear in summer (see the green box in AHI band 09 observations). Instead, both the northern and southern slopes of the TP are dry with observation of high BT. This transition is probably connected to the large-scale atmosphere in the middle troposphere being warmed by the sensible heat over the TP in summer (Yanai et al., 1992; Li and Yanai, 1996). Most of the reanalysis datasets show a similar transition with observations, but to a smaller extent. ERA5, JRA-55 and CRA show the more consistent features with observations, while ERA-Interim, CFSR and MERRA-2 cannot reproduce this feature very well. This suggests that the distribution of WV over the southern part of the TP is consistently overestimated, especially in ERA-Interim, CFSR and MERRA-2.

To further analyze these differences, the meridional distribution (averaged over 90°–110°E) of monthly averaged BTD between observations and simulations from ERA-Interim, ERA5, CFSR, MERRA-2, JRA-55, and CRA are exhibited in Fig. 11. The results show that, from April to October, the WV of the middle and upper troposphere over the TP is generally overestimated. In addition, the patterns of wet bias are more similar for WV08 and WV09 in the six reanalysis datasets. For ERA-interim, CFSR, MERRA-2 and CRA, the WV over the southern slopes of the TP is overestimated more than over the northern slope. For ERA5, in summer, for WV08 and WV09, a larger wet bias appears over the northern TP and moves to the central TP in autumn. For JRA-55, the situation is more complicated. The wet bias peak location varies, with both the northern and southern TP exhibiting it in spring and summer, while it appears in the northern TP in autumn and winter. The bias patterns in band 10 are quite inconsistent, especially in winter. In particular, CFSR shows a distinctive dry bias through winter, with a winter dry bias also evident in JRA-55 and CRA. ERA5 exhibits a larger wet bias in winter. Considering that band 10 might be influenced by the surface as shown in Fig. 2, especially in winter over the TP, surface temperatures in the six reanalysis datasets are further analyzed in Fig. 8. The surface temperature of CFSR is significantly higher than for other reanalysis data in winter. This may explain why CFSR shows a warmer BTD through the winter, which is also the case for ERA5. Therefore, the large variations in band 10 among the different reanalysis datasets shown in Fig. 11 might be due to surface influences.

Figure 11.  The meridional distribution (averaged over 90°–110°E) of monthly averaged BTD between observations and simulations from ERA-Interim, ERA5, CFSR, MERRA-2, JRA-55, and CRA. From left to right, the three columns depict band 08, band 09 and band 10, respectively.

As a further comparison, based on the zonal distribution (averaged over 25°–40°N) of monthly averaged BTDs between observations and simulations from the six reanalysis datasets (not shown), WV08 and WV09 are more likely overestimated in the eastern part of the TP than the western part.

The Asian summer monsoon is well known for its prominent subseasonal variations, the active and break monsoons (e.g., Krishnamurti and Bhalme, 1976), and abrupt changes during its seasonal evolution. Here, we attempt to understand if the reanalysis data can describe these abrupt changes in monsoon systems, such as the northward shift of the subtropical high, using moisture as an indicator. According to the China Climate Bulletin, in the summer of 2016, the area of the western North Pacific subtropical high (WNPSH) was significantly larger, the intensity was stronger, and its position was extended more to the west. Moreover, the subtropical high had an obvious northward shift in the third pentad in June and September, separately. During the summer monsoon, WV is transported via the western margin of the WNPSH (Zhou and Yu, 2005; Chu et al., 2021). The WV field will change significantly during an abrupt change of the subtropical high (as proven by the 850-hPa WV flux results; not shown). We choose this case to assess the performance of the reanalysis datasets.

Figure 12 shows the pentad-averaged BTs (third and fourth pentad in June) and the temporal variation in BTs in the two pentads for the observations from AHI WV 08 (results are similar for WV 09 and 10) and simulations from the six reanalysis datasets over East Asia under a clear sky. It is apparent that, except for MERRA-2, the position and intensity of the WNPSH indicated by the BT in the reanalysis data are closely consistent with observations. However, MERRA-2 obviously overestimates the WV within (15°–30°N, 130°–140°E), and thus may underestimate the strength of the subtropical high. There are two areas with obviously low BTs in this process, located to the southwest and northwest of the subtropical high, respectively, which might reflect an integrated WV region. MERRA-2 and JRA-55 tend to overestimate the WV to the northwest of the WNPSH in the 4th pentad.

Figure 12.  Spatial distributions of two pentad–averaged BTs and the temporal variation of BTs in the two pentads for the observations from AHI band 08 and simulations from ERA-Interim, ERA5, CFSR, MERRA-2, JRA-55, and CRA over East Asia in 2016. The first two columns are for the third and fourth pentad, along with a geopotential height line of 577 gpm in grey and black, respectively. The third column presents the temporal variation of the BTs in the two pentads (4th pentad minus 3rd pentad).

Determining whether the northward shift of the subtropical high can be captured in the reanalyses is also studied. The third row in Fig. 12 displays the variation in BT from the third to fourth pentad in June overlaid with the 577-hPa geopotential height (from ERA5), marked as gray lines for the third pentad and black dotted lines for the fourth pentad. The 577-hPa geopotential height line roughly indicates the position of the subtropical high. The subtropical high has a northward shift from 15°N to 25°N in the third and fourth pentad. During the northerly shift of the subtropical high, the observed BTs in the south of the subtropical high decrease (as WV increases), and the observed BTs in the north of the subtropical high increase (as WV decreases). The simulated BTs in the six reanalysis datasets describe the northward shift of the subtropical high, as well as the accompanying variation in the intensity and location of WV. In addition, the distribution of BTs in the reanalysis datasets caused by the distribution of WV in the subtropical high is also relatively consistent with the observations. Overall, the reanalysis data show small differences with observations in the three bands, indicating good agreement in terms of BT variation in the northerly shift of the subtropical high. Based on RMSE results (not shown), MERRA-2 shows the largest difference among the six reanalysis datasets. JRA-55 also has large BTDs in the high latitudes, while ERA5 has the best accuracy.

This study focused on assessing the multilayer WV in reanalysis datasets over East Asia using measurements from three AHI WV bands. Six typical reanalysis datasets from major meteorological forecast centers were evaluated: ERA-Interim, ERA5, CFSR, MERRA-2, JRA-55, and CRA. The assessment was confined to clear skies and the major findings can be summarized as follows:

(1) In general, overestimations of WV08 and WV09 occur in all six reanalysis datasets, especially in summer. The overall performance in depicting WV10 is consistently better than depicting WV08 and WV09 in all datasets. ERA5 shows the best accuracy in reproducing multilayer WV distributions, in terms of the magnitude, dispersion, and pattern similarity of WV biases.

(2) The six reanalysis datasets can reproduce the seasonal variation tendencies of the WV fields. Except for ERA-Interim and MERRA-2, the reanalyses show drier air in the middle troposphere and moister air in the upper troposphere over the East Plain, except the TP region. From the perspective of further nowcasting applications, ERA5, CFSR, JRA-55 and CRA may underestimate the conditional instability compared with observations.

(3) WV bias features over the TP are distinct from other regions of East Asia in the six reanalysis datasets. First, the biases over the TP are consistently larger. Second, WV10 amounts and biases over the TP show large discontinuities with the surrounding regions. Third, there is disagreement among the six reanalysis datasets as to whether biases over the TP are wet or dry. In summer especially, most of the reanalysis datasets have a larger wet bias for the middle and upper troposphere. For ERA-Interim, CFSR, MERRA-2 and CRA, the WV over the southern slopes of the TP is overestimated compared to the northern slope, while the results from ERA5 are opposite. JRA-55 has wet bias peak location changes between the northern and southern slopes of the TP with the seasons. Most reanalysis datasets have an obvious wet bias in the western rather than the eastern part of the TP. The surface temperatures in reanalysis datasets might be one of the causes for the reanalysis–observation discrepancies for band 10.

(4) Overall, the reanalysis datasets capture the northward shift of the subtropical high in terms of the moisture field. MERRA-2 shows the largest differences from observations, while ERA5 has the best accuracy.

Acknowledgements. This study was partly supported by the National Natural Science Foundation of China (Grant Nos. 41975020 and 41975031) (Jun LI). The surface IR emissivity data are from the UW-Madison Baseline Fit Emissivity database (available at ftp://ftp.ssec.wisc.edu/pub/g_emis/). CMA is thanked for providing the CRA reanalysis dataset. The other reanalysis datasets can be downloaded (https://reanalyses.org/atmosphere/overview-current-atmospheric-reanalyses#). We thank W. P. MENZEL for useful input to the discussion and conclusions related to this work.