Reanalysis is the process whereby an unchanging data assimilation system is used to provide a consistent reprocessing of meteorological observations, typically spanning an extended segment of the historical data record. The process relies on an underlying forecast model to combine disparate observations in a physically consistent manner, enabling production of gridded datasets for a broad range of variables, including ones that are sparsely or not directly observed. As such, and with appropriate consideration of the inherent uncertainties, reanalysis products have not only become a staple of the atmospheric research community, but are used increasingly for climate monitoring as well as for business applications in, for example, energy and agriculture. Recent reanalyses from the National Oceanic and Atmospheric Administration (NOAA)/National Centers for Environmental Prediction (NCEP), the European Centre for Medium-Range Weather Forecasts (ECMWF), the National Aeronautics and Space Administration (NASA) Global Modeling and Assimilation Office (GMAO), and the Japan Meteorological Agency (JMA) provide a rich ensemble of climate data products beginning more or less with the period of regular conventional and satellite observations in the mid to late twentieth century (Saha et al. 2010; Dee et al. 2011; Rienecker et al. 2011; Kobayashi et al. 2015). However, there have also been successful efforts to extend atmospheric reanalyses back to the late nineteenth and early twentieth centuries using only surface pressure observations (Compo et al. 2011) or surface and mean sea level pressure observations plus surface marine winds (Poli et al. 2013). As noted by Dee et al. (2011), these century-long reanalyses have also sparked remarkable data recovery and digitization efforts by various groups around the world.
The GMAO’s reanalysis development effort began (under its predecessor organization, the Data Assimilation Office) with the production of the Goddard Earth Observing System, version 1 (GEOS-1), reanalysis (Schubert et al. 1993) but advanced significantly with the more recent production of the Modern-Era Retrospective Analysis for Research and Applications (MERRA; Rienecker et al. 2011). MERRA encompassed the period 1979–2016 and was undertaken with two primary objectives: to place NASA’s Earth Observing System (EOS) satellite observations in a climate context and to improve the representation of the atmospheric branch of the hydrological cycle compared with previous reanalyses. MERRA succeeded in meeting these objectives overall and was found to be of comparable quality to contemporaneous reanalyses produced by NCEP and ECMWF (e.g., Decker et al. 2011). However, it also suffered from a number of known, but not necessarily unique, deficiencies. These include unphysical jumps and trends in precipitation in response to changes in the observing system, biases and imbalances in certain atmospheric and land surface hydrological quantities, and a poor representation of the upper stratosphere (e.g., Bosilovich et al. 2011; Robertson et al. 2011; Reichle et al. 2011; Rienecker et al. 2011). In addition, the long-term viability of MERRA was limited by system constraints that precluded the incorporation of new satellite data sources beyond NOAA-18, which launched in 2005. At the time of its termination in March 2016, MERRA was at risk of suffering a significant degradation in quality were certain observing platforms to fail, including, for example, EOS Aqua, which was already well beyond its designed lifetime and provided MERRA with its only sources of hyperspectral infrared and afternoon-orbit microwave radiances.
The Modern-Era Retrospective Analysis for Research and Applications, version 2 (MERRA-2), was undertaken to provide a timely replacement for MERRA and to sustain GMAO’s commitment to having an ongoing near-real-time climate analysis. MERRA-2 is intended as an intermediate reanalysis, one that leverages recent developments at GMAO in modeling and data assimilation to address some of the known limitations of MERRA but also provides a stepping stone to GMAO’s longer-term goal of developing an integrated Earth system analysis (IESA) capability that couples assimilation systems for the atmosphere, ocean, land, and chemistry. Toward the latter goal, MERRA-2 includes aerosol data assimilation, thereby providing a multidecadal reanalysis in which aerosol and meteorological observations are jointly assimilated within a global data assimilation system. Other new developments in MERRA-2 relevant to IESA focus on aspects of the cryosphere and stratosphere, including the representation of ozone, and on the use of precipitation observations to force the land surface. At the same time, basic aspects of the MERRA-2 system, such as the variational analysis algorithm and observation handling, are largely unchanged since MERRA. Also unchanged is the preparation of most conventional data sources used originally in MERRA.
This paper presents an overview of MERRA-2, including a description of the data assimilation system and various measures of performance. Some of these measures focus on difficulties encountered in MERRA, while others highlight new capabilities, such as the assimilation of aerosol observations. This paper also serves as an introduction to a series of companion papers that provide more detailed analyses of the topics covered in this overview as well as others. For example, a detailed description of the MERRA-2 aerosol analysis system and its validation are presented in Randles et al. (2017) and Buchard et al. (2017, manuscript submitted to J. Climate, hereinafter B17). Reichle et al. (2017a,b) assess the land surface precipitation and land surface hydrology, while Draper et al. (2017, manuscript submitted to J. Climate) examine the land surface energy budget. Bosilovich et al. (2017) evaluate the global water balance and water cycle variability in MERRA-2. Collow et al. (2016) examine MERRA-2’s representation of U.S. summertime extreme precipitation events, and Lim et al. (2017) investigate aspects of major El Niño events. Collow and Miller (2016) examine the radiation budget and cloud radiative effect over the Amazon. Several papers investigate aspects of the stratosphere in MERRA-2: Wargan et al. (2017) examine the representation of lower-stratospheric ozone and the effect of assimilating ozone observations from NASA’s Aura satellite; Coy et al. (2016) examine the representation of the quasi-biennial oscillation in MERRA-2; and Wargan and Coy (2016) present a case study of the 2009 sudden stratospheric warming.
Section 2 provides an overview of the MERRA-2 data assimilation system, focusing primarily on developments since MERRA, including new observation sources. Basic metrics of the assimilation system performance are presented in section 3. The MERRA-2 aerosol analysis is described in section 4, along with sample results and validation statistics. Section 5 examines global and regional aspects of the representation of precipitation in MERRA-2, focusing on areas of difficulty in MERRA. Stratospheric processes and the representation of ozone are discussed in section 6. Section 7 addresses the representation of the cryosphere in MERRA-2, with focus on glaciated land surface processes. Section 8 provides information about MERRA-2 products and how they can be accessed. It is noted here that each MERRA-2 data collection has its own digital object identifier (DOI) number, so data used in scientific publications can be cited exactly. Most of the results shown for MERRA-2 in this paper are derived from these collections, which are individually cited in the corresponding figure captions. Finally, a brief summary and perspective on future work are presented in section 9. A list of acronyms is given in the appendix.
2. MERRA-2 system description
MERRA-2 is produced with version 5.12.4 of the GEOS atmospheric data assimilation system. The key components of the system are the GEOS atmospheric model (Rienecker et al. 2008; Molod et al. 2015) and the GSI analysis scheme (Wu et al. 2002; Kleist et al. 2009b). The model includes the finite-volume dynamical core of Putman and Lin (2007), which uses a cubed-sphere horizontal discretization at an approximate resolution of 0.5° × 0.625° and 72 hybrid-eta levels from the surface to 0.01 hPa. The analysis is computed on a latitude–longitude grid at the same spatial resolution as the atmospheric model using a 3DVAR algorithm based on the GSI with a 6-h update cycle and the so-called FGAT procedure for computing temporally accurate observation-minus-background departures. The analysis is applied as a correction to the background state using an IAU procedure (Bloom et al. 1996).
The MERRA-2 system has many of the same basic features as the MERRA system described in Rienecker et al. (2011) but includes a number of important updates. An overview of these updates is provided here, with additional details provided in companion publications as cited. Unless otherwise stated, other aspects of the system configuration and preparation of the input data are as described in Rienecker et al. (2011). The updates discussed here include changes to the forecast model (section 2a), the analysis algorithm (section 2b), the observing system (section 2c), the radiance assimilation (section 2d), the bias correction of aircraft observations (section 2e), the mass conservation and water balance (section 2f), the precipitation used to force the land surface and drive wet aerosol deposition (section 2g), the boundary conditions for sea surface temperature and sea ice concentration (section 2h), and reanalysis production (section 2i).
a. Forecast model
Since MERRA, the GEOS model has undergone changes to both its dynamical core and its physical parameterizations. Whereas in MERRA the horizontal discretization of the model was computed on a latitude–longitude grid, MERRA-2 uses a cubed-sphere grid. This allows relatively uniform grid spacing at all latitudes and mitigates the more severe grid spacing singularities that occur on a latitude–longitude grid. Upgrades to the physical parameterization schemes include increased reevaporation of frozen precipitation and cloud condensate, changes to the background gravity wave drag, and an improved relationship between the ocean surface roughness and ocean surface stress (Molod et al. 2015). The MERRA-2 model also includes a Tokioka-type trigger on deep convection as part of the RAS convective parameterization scheme (Moorthi and Suárez 1992), which governs the lower limit on the allowable entrainment plumes (Bacmeister and Stephens 2011). A new glaciated land representation and seasonally varying sea ice albedo have been implemented, leading to improved air temperatures and reduced biases in the net energy flux over these surfaces (Cullather et al. 2014).
b. Analysis algorithm
The control variable for moisture used in recent versions of GSI and MERRA-2 differs from the one used in MERRA. Whereas MERRA used the so-called pseudorelative humidity (Dee and da Silva 2003) defined by the water vapor mixing ratio scaled by its saturation value, MERRA-2 uses the normalized pseudorelative humidity (Holm 2003) defined by the pseudorelative humidity scaled by its background error standard deviation. The latter has a near Gaussian error distribution, making it more suitable for the minimization procedure employed in the assimilation scheme. Also within the GSI, a tangent linear normal mode constraint (Kleist et al. 2009a) is applied during the minimization procedure to control noise and improve the overall use of observations. The background error statistics used in the GSI have been updated as well in MERRA-2. As in MERRA, the statistics are estimated using the NMC method (Parrish and Derber 1992) by calculating variances and covariances from the differences between 24- and 48-h forecasts, but from a more recent version of GEOS. Compared with the MERRA system, the background error statistics for the MERRA-2 system exhibit generally smaller standard deviations for most variables, but both larger and smaller correlation length scales depending on the variable, latitude, and vertical level.
c. Observing system
MERRA included no new satellite observation sources after the introduction of NOAA-18 in 2005. MERRA-2, in contrast, includes numerous additional satellite observations both before and after this time. The complete set of input observations assimilated in MERRA-2 is summarized in Table 1, while a detailed description of their use is provided in McCarty et al. (2016). Additions to the MERRA-2 observing system compared with MERRA include the following:
atmospheric motion vectors from AVHRR;
surface wind speeds from SSMIS;
surface wind vectors from the Meteorological Operational Satellite-A (MetOp-A) ASCAT and WindSat;
temperature and ozone profiles from EOS Aura MLS;
total column ozone from EOS Aura OMI;
bending angle from GPSRO;
microwave and infrared sounding radiances from ATOVS on NOAA-19, MetOp-A, and MetOp-B;
microwave sounding radiances from ATMS on SNPP;
hyperspectral infrared radiances from IASI on MetOp-A and MetOp-B and from CrIS on SNPP; and
geostationary radiances from MSG SEVIRI and Geostationary Operational Environmental Satellites (GOES-11, GOES-13, and GOES-15).
Observation types assimilated in MERRA-2, including their usage dates and sources. Boldface fonts indicate observation types not assimilated in MERRA. Acronyms are defined in the appendix.
Time series of the various types of observations assimilated in MERRA and MERRA-2 are shown in Fig. 1. The number of assimilated observations in MERRA-2 grows from approximately 2 million per 6-h cycle in 2002 to almost 5 million in 2015, while MERRA assimilates approximately 1.5 million observations per 6-h cycle from 2002 onward. The GSI in MERRA-2 is also capable of assimilating microwave and hyperspectral infrared radiances from planned future satellites, including MetOp-C and JPSS. The temporary spike in the number of QuikSCAT data assimilated in MERRA-2 in late 2000 is due to an error in preprocessing, which led to observations beyond the midswath “sweet spot” being used in the analysis. This has no discernible impact on the quality of the analyzed fields or on the use of other observations in the assimilation system.
MERRA-2 also assimilates reprocessed versions of some of the same satellite observation types used in MERRA. In MERRA-2, Remote Sensing Systems version 7 (RSS v7) recalibrated radiances and retrieved surface wind speeds from SSM/I are used, whereas MERRA used RSS v6. The use of retrieved ozone from the SBUV also differs, with MERRA-2 assimilating version 8.6 on 21 layers from 1980 through 2004 before switching to OMI and MLS in October 2004. In contrast, MERRA used SBUV version 8 throughout, in a form degraded from its original 21 layers to 12.
d. Radiance assimilation
Radiative transfer calculations necessary for the assimilation of satellite radiances in MERRA-2 are performed using the CRTM (Han et al. 2006; Chen et al. 2008). MERRA-2 uses version 2.1.3 of the CRTM for assimilation of all satellite radiances, whereas MERRA used a prototype version of the CRTM for all radiances except those from SSU, for which the GLATOVS (Susskind et al. 1983) was used. Differences between the prototype and version 2.1.3 of the CRTM are too numerous to mention here, but a detailed description of the latter can be found in Liu and Boukabara (2014).
The actively assimilated channels for each satellite sensor type in MERRA-2 are summarized in Table 2. Microwave temperature sounding channels with strong surface sensitivity—so-called window channels—are not assimilated in MERRA-2, in part because of the strong sensitivity of global precipitation and humidity to these data found in MERRA (Robertson et al. 2011). These include channels 1–3 and 15 on AMSU-A, channels 1–4 and 16 on ATMS, and channel 1 on MSU. For microwave humidity sounders including AMSU-B and MHS, window channels are actively assimilated along with the sounding channels. For heritage infrared sounders, channels 13–15 on HIRS were assimilated in MERRA but are excluded in MERRA-2. The channel selections for hyperspectral infrared sounders and performance assessments for selected instruments are provided in McCarty et al. (2016).
Nominal channel selections for satellite radiances assimilated in MERRA-2. Usage can vary for individual satellite platforms as a result of sensor failure or quality control decisions.
Like MERRA, MERRA-2 uses an automated bias correction scheme for the assimilation of most satellite radiance observations. Bias estimates for individual sensor channels are represented by a small number of predictors, which can depend on the atmospheric state, the radiative transfer model, and the sensor characteristics. Airmass- and viewing-angle-dependent biases are estimated using a variational scheme in which the predictor coefficients are updated as part of the control vector used to minimize the analysis cost function (Derber and Wu 1998). Satellite scan-position-dependent bias is estimated directly as an exponential moving average filter of the observation-minus-background departures for brightness temperature. For both the variational and scan-position predictors, initial values of the coefficients for MERRA-2 were derived from GEOS operations and other long production runs using system versions similar to that used for MERRA-2. In the few cases where no recent coefficient information was available, initial values were derived from MERRA. Note that no bias correction is applied to a small number of sensor channels that peak in the upper stratosphere, including channel 14 on AMSU-A, channel 15 on ATMS, and channel 3 on SSU. This is done to prevent the variational bias correction scheme—which is formulated to remove systematic discrepancies between the observations and the background state irrespective of the source—from making erroneous adjustments to the observations at levels where model biases are known to be large.
e. Bias correction of aircraft temperature observations
A bias correction scheme for aircraft temperature observations has been implemented in MERRA-2, motivated by the known warm bias of these measurements compared with other data sources (Cardinali et al. 2003; Ballish and Kumar 2008; Rienecker et al. 2011). The scheme uses the mean observed-minus-background departures to estimate the bias for temperature reports from individual aircraft, identified by their tail number. The bias estimates are updated after each analysis. The scheme is used to correct AMDAR and ACARS reports only, since other sources of aircraft observations in MERRA-2 do not have unique identifiers by which they can be tracked. As of 2015, bias corrections for approximately 3700 separate aircraft are tracked in MERRA-2.
The performance of the scheme is discussed in McCarty et al. (2016). As expected, the scheme is shown to reduce the bias between the corrected aircraft observations and the background forecast, as well as to reduce the variance of the corrected background departures, allowing more aircraft observations to be used in the analysis. Unfortunately, the MERRA-2 background state was found to have a larger than expected positive bias in the mid to upper troposphere, which feeds back to the bias estimates. The result is that the bias correction actually increases the aircraft temperatures in some cases, and the fit to other unbiased observation types, such as radiosondes, is degraded. This is discussed further in section 3.
f. Mass conservation and water balance
Studies have documented the difficulty of maintaining realistic balances between variations in total mass and total water content in previous reanalyses (e.g., Trenberth and Smith 2005; Bosilovich et al. 2011; Berrisford et al. 2011). These studies show that analysis adjustments to moisture are often large (when, ideally, they should be small), highly sensitive to changes in the observing system, and mostly balanced by unphysical changes in precipitation. Takacs et al. (2016) argue that, in attempting to analyze the total mass of the atmosphere from surface pressure observations, reanalyses may violate the simple physical constraint that, to an excellent approximation, the total dry mass of the atmosphere is invariant, and so changes in total mass must be essentially equivalent to changes in total water mass. At the same time, Berrisford et al. (2011) argue that, while the observing system may not provide the data to determine exactly the total mass of the atmosphere, the degree to which dry mass is preserved in a reanalysis provides a useful diagnostic of reanalysis quality.
Reconsideration of these issues during the development of MERRA-2 prompted modifications to GEOS to conserve atmospheric dry mass and to guarantee that the net source of water from precipitation and surface evaporation equals the change in total atmospheric water. As described by Takacs et al. (2016), this has been achieved by making the following changes to the forecast model and assimilation procedure:
Sources and sinks of atmospheric water have been added to the model continuity equation so that changes in total mass are driven purely by changes in total water.
A constraint that penalizes analysis increments of dry air has been added to the GSI.
Tendencies in the IAU are rescaled so that the global mean is removed from the analysis increment of water.
The global impact of these modifications is illustrated in Figs. 2 and 3, which compare different components of atmospheric mass in MERRA and MERRA-2. Figure 2 shows monthly mean anomalies from the mean seasonal cycle for total mass, total water, and dry-air mass in the two reanalyses. In MERRA, there is an increase in total water over the period, with significant interannual variations, but these features do not necessarily match the changes in total mass. There also are spurious anomalies in dry-air mass throughout, some of which track closely with the changes in total mass. In MERRA-2, changes in total mass and total water track each other almost perfectly, by design, and the dry-air mass remains a constant whose value must be specified. For the latter, the value 983.24 hPa is chosen based on MERRA. This value falls within 0.1% of the values derived from other recent reanalyses (Takacs et al. 2016).
Figure 3 shows monthly mean values of evaporation minus precipitation (E − P, or water source term), the vertically integrated analysis increment of water, and the atmospheric water storage. Note that the atmospheric water storage has similar magnitude in both reanalyses and is dominated by the seasonal cycle. In MERRA, however, the storage is determined by a near balance between the large and highly variable contributions from the analysis increment on the one hand, and unphysical variations in E − P of the opposite sign on the other hand. This includes an abrupt change in the sign of these quantities after the introduction of AMSU-A in 1998 (Robertson et al. 2011). In MERRA-2, the globally integrated analysis increment is zero, by design, and the water storage is determined as in nature by small seasonal differences in E and P. It should be noted that removing the global mean analysis increments of total mass and water mass does not imply that the analysis increments of water vapor or surface pressure vanish locally, as shown in section 3 of this paper and discussed in further detail by Bosilovich et al. (2017).
g. Observation-corrected precipitation forcing
The precipitation generated by the atmospheric model during the IAU segment of the assimilation procedure is subject to considerable errors that can propagate into land surface hydrological fields and beyond (Reichle et al. 2011). To mitigate these effects in MERRA-2, the model-generated precipitation is corrected with observations before being used to force the land surface or affect the wet deposition of aerosols over land and ocean. Both the model-generated precipitation and the precipitation seen by the land surface and the aerosols are available in the MERRA-2 output. MERRA-2 is one of several recent applications of GEOS that uses observation-corrected precipitation estimates. Others include the GMAO seasonal forecasting system (Ham et al. 2014), the MERRA-Land data product (Reichle et al. 2011), and the aerosol reanalysis MERRAero (Buchard et al. 2015). Precipitation observations have also been used in reanalyses produced by NOAA, including the North American Regional Reanalyis (Mesinger et al. 2006) and in the CFSR (Saha et al. 2010; Meng et al. 2012), although in both cases the approaches differ from that used in MERRA-2. Some discussion of the differences between the approaches used in MERRA-2 and CFSR can be found in Reichle et al. (2017a).
The corrected precipitation in MERRA-2 is derived from publicly available, observationally based global precipitation products disaggregated from daily or pentad totals to hourly accumulations using precipitation estimates from MERRA (Reichle and Liu 2014; Reichle et al. 2017a). The land surface in MERRA-2 sees a combination of corrected and model-generated precipitation depending on latitude, with the land surface forced primarily by the corrected estimates at low to midlatitudes, by the MERRA-2 model-generated precipitation at high latitudes, and by a weighted mixture in between to prevent spatial discontinuities in climatological means. This is illustrated in Fig. 4, which shows the annual average adjustment made to the model-generated precipitation in MERRA-2 for the period 1980–2015 using this technique. The greatest adjustments are made in the tropics, where precipitation is greatest and the corrected estimates are given most weight, while no adjustments are made poleward of 62.5° latitude in either hemisphere.
Based on the evaluation of several metrics, Reichle et al. (2017a) found the observation-corrected precipitation to be more realistic overall than that generated by the model within the cycling MERRA-2 system, or that of the MERRA and MERRA-Land data products. Exceptions include discontinuities in the MERRA-2 corrected precipitation that result from errors in the underlying gauge products, for example, in Myanmar and South America. Another issue is the high bias in MERRA-2 summer precipitation in the high latitudes (where precipitation observations are not used). Moreover, the diurnal cycle of the MERRA-2 corrected precipitation has reasonable amplitudes compared to independent observations, but the time of day of maximum precipitation is inherited from MERRA and is unrealistic.
The improvements in the precipitation forcing are also reflected in the MERRA-2 land surface estimates. Reichle et al. (2017b) show that soil moisture, snow, terrestrial water storage, and runoff in MERRA-2 agree better with independent observations than estimates from MERRA. Draper et al. (2017, manuscript submitted to J. Climate) further demonstrate that the temporal behavior and long-term mean values of the land–atmosphere turbulent fluxes in MERRA-2 are improved. Moreover, by applying the precipitation corrections within the coupled atmosphere–land modeling system, MERRA-2 can provide more self-consistent surface meteorological data than were used for MERRA-Land (Reichle et al. 2017a). This self-consistency is important for applications such as forcing land-only model simulations.
Finally, it should be noted that the atmospheric water and energy prognostic variables associated with the creation of precipitation in MERRA-2 are not directly modified by the corrected estimates, although they can be indirectly modified through subsequent feedback with the land surface.
h. Sea surface temperature and sea ice concentration
The boundary conditions for SST and SIC in MERRA were based on the 1° resolution weekly (or monthly) product of Reynolds et al. (2002). In MERRA-2, SST and SIC boundary conditions are instead based on currently available high-resolution (finer than 1° resolution) daily products. However, as there exists no continuous source of daily global high-resolution SST and SIC for the entire period of MERRA-2—and no source of daily data whatsoever prior to 1982—the following products were used in combination (Table 3): monthly 1° resolution data from CMIP as in Taylor et al. (2000) for the period prior to 1982; daily ¼° resolution data from NOAA OISST as in Reynolds et al. (2007) from 1982 through March 2006; and daily 1/20° resolution data from OSTIA as in Donlon et al. (2012) from April 2006 onward. Note that different versions of the NOAA OISST product are used prior to and after January 2003, the latter including satellite data from both AVHRR and AMSR-E on NASA’s Aqua satellite, and the former including satellite data from AVHRR only. The processing of these products into a unified gridded set of daily SST and SIC boundary conditions for MERRA-2 is described in Bosilovich et al. (2015). Care was taken to use both SST and SIC from the same data source to avoid potential inconsistencies, especially in marginal ice zones.
Sea surface temperature and sea ice concentration data products used in MERRA-2.
Figure 5 shows 12-month running averaged values of SST between 60°N and 60°S for MERRA-2 and several other reanalyses, including MERRA. In all cases, there is a positive trend in SST throughout the period. The running means for all the reanalyses are within 1 K for the 30 years spanning 1980–2010, and the anomalies (not shown) are separated by less than 0.2 K. At the same time, there are clear systematic differences between reanalyses, with the MERRA-2 SST’s on the one hand being cooler than those used in the other reanalyses shown except CFSR (which used similar input datasets), especially before the transition to OSTIA in 2006. The values for JRA-55, on the other hand, are on the order of 0.1 K higher than other reanalyses throughout the 35-yr period. It can also be seen that the MERRA-2 SSTs increase slightly with the change in NOAA OISST versions after 2003. The reader is referred to Bosilovich et al. (2015) for a more detailed list of known issues with the SST and SIC boundary conditions for MERRA-2.
MERRA-2 was produced in four separate streams, each of which was spun up for a year at full resolution beginning on 1 January 1979 (stream 1), 1 January 1991 (stream 2), 1 January 2000 (stream 3), and 1 January 2010 (stream 4). The land surface restart files for each MERRA-2 stream were themselves spun up for at least 20 years using the offline MERRA-2 land model forced by MERRA surface meteorological fields, and with the precipitation replaced by the observation-corrected estimates described in section 2g. The final MERRA-2 product distribution is from stream 1 for 1 January 1980–31 December 1991, followed by stream 2 for 1 January 1992–31 December 2000, then stream 3 for 1 January 2001–31 December 2010, and finally stream 4 for 1 January 2011–present. With streams 1–3 complete, MERRA-2 production continues as a near-real-time climate analysis from stream 4 alone. The decision to begin stream 1 in January 1979 and distribute products beginning in January 1980—a year later than the schedule followed in MERRA—was based on the fact that the products used to create the observation-corrected precipitation estimates for MERRA-2 only start on 1 January 1979, leaving no viable way to initialize the land surface properly before this time (which requires several months of spinup after initialization from climatological conditions).
The overlap periods between successive streams were examined to determine the adequacy of the spinup procedure and to quantify the uncertainty in individual fields. Differences between overlapping MERRA-2 streams were found to be minimal for most fields after one year, with the exception of certain land surface variables including the deep-level soil temperature and land surface soil moisture storage at high latitudes. The spinup of the land surface is addressed separately in Reichle et al. (2017a); section 3d and Fig. 13 of that paper discuss specific examples of the aforementioned discontinuities across consecutive MERRA-2 streams. Users should be aware of these discontinuities when the data are used for specific applications.
3. Data assimilation diagnostics
By-products of the data assimilation procedure in the form of differences between forecasts and observations, analysis increments, and estimates of bias can be used effectively to monitor the quality of both the input and output of the assimilation. In this section, examples of such diagnostics are presented for MERRA-2, focused mainly on feedbacks with respect to in situ conventional observations and on the net correction, or increment, brought by the entirety of the assimilated observations. The reader is referred to McCarty et al. (2016) for examples of feedbacks related to the treatment of satellite radiance observations.
a. Background departure statistics
Differences between the assimilated observations and the background forecast, referred to as innovations or background departures, provide important information about the quality of the assimilation. In particular, it is important that the assimilation system be able to predict high-quality observations, especially for conventional data types that provide direct measurements of the analyzed variables. In addition to affecting the analysis directly, many conventional data play an important role in anchoring the variational bias estimates used in the assimilation of satellite radiances. Generally speaking, smaller background departures indicate a higher-quality assimilation. The results shown here are selected to highlight both strengths and weaknesses of MERRA-2 in this regard. As in MERRA, for convenience, gridded versions of the observations and corresponding departures used in MERRA-2 will be made available to users.
Figure 6 shows time series of monthly mean and RMS background departure statistics for all assimilated surface pressure observations in MERRA and MERRA-2 for both the Northern and Southern Hemisphere. Also shown are the monthly mean numbers of surface pressure observations assimilated in each 6-h assimilation cycle in MERRA-2. The RMS values decrease with time in both reanalyses, especially in the Southern Hemisphere after the early to mid-1990s, when the number of observations begins to increase significantly. The RMS values in the Northern Hemisphere (Fig. 6a) are smaller than in the Southern Hemisphere initially and decrease more slowly with time, reflecting the greater number of conventional observations available over land throughout the period. This decrease is slightly more pronounced in MERRA-2 after the mid-1990s when the number of surfaces pressure observations from land stations increases significantly. In the Southern Hemisphere (Fig. 6b), the RMS values are larger in MERRA-2 than in MERRA before the mid-1990s but smaller by the end of the period. The larger values early on are due to the use of larger observation errors for surface ship observations (and some other conventional data types) in MERRA-2, allowing more “outliers” with larger departure values to pass the quality control procedure in the analysis.1 The impact diminishes by the mid-1990s as other observation types, including from satellites, become more abundant. There is no similar effect in the Northern Hemisphere, where surface pressure observations from land stations are dominant early in the period; the observation errors specified for these data are the same in MERRA and MERRA-2. Finally, the jump in RMS values in the Southern Hemisphere evident in both reanalyses at the beginning of 1985 coincides with the introduction of regularly spaced synthetic surface pressure observations over Southern Ocean areas.
The mean background departures for surface pressure in the Northern Hemisphere are consistently less biased in MERRA-2 than in MERRA, especially after the mid-1990s. In the Southern Hemisphere, however, the departures for MERRA-2 show a negative bias throughout the period; this is discussed further in section 3b. The mean departures in MERRA-2 also show a more pronounced annual cycle in this hemisphere. As a point of reference, the background departure statistics for other reanalyses, including, for example, ERA-Interim (Dee et al. 2011), exhibit a clear annual cycle, but with somewhat smaller amplitude than in MERRA-2.
Figure 7 shows global background departure statistics for radiosonde temperatures for MERRA and MERRA-2 at selected pressure levels in the troposphere (300 and 700 hPa) and stratosphere (10 and 50 hPa). Also shown for each level are the monthly mean numbers of radiosonde temperature observations assimilated in each 6-h assimilation cycle in MERRA-2. In the troposphere (Figs. 7c,d), the performance of MERRA-2 is degraded compared to that of MERRA, especially at 300 hPa. The RMS values for MERRA-2 decrease with time but remain 10%–20% larger than those for MERRA during much of the period. Again, this is due at least partially to the use of larger observation errors for radiosonde temperatures and other conventional data types in MERRA-2. Noticeable improvements occur first in the mid-1990s, when satellite observations become more abundant, and again in 2006, when the number of GPSRO observations increases significantly.
The mean departure values at 300 hPa for both MERRA and MERRA-2 exhibit a clear negative bias. The bias is generally larger in MERRA-2, reaching a maximum amplitude of greater than 0.5 K during the early 2000s. This is due to a warm model bias in the upper troposphere, which worsened during the course of development between MERRA and MERRA-2 (see also Fig. 10). However, aspects of the assimilation process may exacerbate the problem. It can be seen, for example, that the bias in the background departures at 300 hPa increases noticeably after the mid-1990s, especially in MERRA-2, when the numbers of both aircraft temperature observations and satellite radiances begin to increase significantly (Fig. 1). The design of the bias correction procedures for both observation types is such that they result in an adjustment of the observations regardless of the source of the bias. In the presence of a strong model bias this can reinforce the actual observational bias and cause the assimilation system to drift further toward the model state, as noted in the case of the aircraft bias corrections described in section 2e. A similar, though less direct, effect may occur through the observational bias corrections used to assimilate satellite radiances, although other aspects of the variational scheme used to adjust these data act to reduce this risk (Dee and Uppala 2009). At 700 hPa, the mean departures for both reanalyses are generally more comparable and considerably less biased.
In the stratosphere (Figs. 7a,b), there are fewer significant differences between the results for MERRA and MERRA-2, although the departures at 10 hPa for MERRA-2 show a larger negative bias from −0.2 to −0.3 K prior to the early 2000s. After 2002, when assimilation of AIRS radiances begins, the biases at 10 hPa in both reanalyses exhibit an upward trend and eventually become positive, first in MERRA around 2003 and then in MERRA-2 in 2005. There is a discernible jump in the mean departures at this level for MERRA-2 in 2005. This is around the time when assimilation of both MLS temperature retrievals (above 5 hPa) and GPSRO bending angle observations (up to approximately 10 hPa) begins in MERRA-2, but this does not appear to improve the fit to radiosondes at 10 hPa compared with MERRA. After 2006, the biases in both reanalyses have average values of 0.2–0.3 K. Finally, at 50 hPa, the departure values for both reanalyses are very similar and exhibit only a small positive bias throughout.
Figure 8 shows statistics for radiosonde specific humidity background departures at 500 and 850 hPa in the tropics. The performance of MERRA-2 is slightly worse than that of MERRA in the midtroposphere in terms of both RMS and bias, but similar or slightly better in the lower troposphere. Again, the mean departure values are consistent with known biases in the GEOS model.
b. Analysis increments
The analysis increments represent the net adjustment to the background state by the assimilation scheme in response to all the observations. As this adjustment depends in a complex way on assumed or crudely estimated errors in the observations and background state, and on the forward operator that transforms the model variables to observation space, the increments do not necessarily represent errors in the background state. Nonetheless, their spatial and temporal variations provide an important diagnostic of system performance, including how changes in the observing system may affect the consistency of the analysis. Systematic increments often indicate the presence of biases in the model or observations that may complicate the use of reanalyses for estimating budgets and identifying trends (Dee et al. 2011).
As described in section 2, the GEOS assimilation system uses an IAU procedure, which, instead of correcting the initial condition, applies the analysis increment to the model as a constant tendency term during the 6-h assimilation window. It is this contribution to the time tendency from the analysis that is provided as a standard output quantity in MERRA-2, examples of which are presented here. For convenience, these are referred to as simply the analysis increments in the discussion that follows.
Figure 9 shows the mean and standard deviation in time of the monthly mean analysis increment of surface pressure in MERRA-2 for the period January 1980 through December 2015. The monthly means themselves have been computed from subdaily data, eight times per day. The pattern of the mean increments indicates that the analysis tends to move mass from the oceans to the continents, as noted also by Takacs et al. (2016), although this pattern is arguably most robust in the Southern Hemisphere. (The mostly negative surface pressure increments over Canada provide an obvious counterexample.) These results are consistent with those in Fig. 6 showing a negative bias in the Southern Hemisphere background departures in MERRA-2. The standard deviation of the increments shows that the largest variations in surface pressure occur in the middle and high latitudes, and especially over coastal Antarctica and the mountainous regions of southern and eastern Asia, as well as southern Alaska.
Time series of the global monthly mean and standard deviation of the analysis increments of temperature from the surface to 70 hPa in MERRA-2 are shown in Fig. 10. The most striking feature in the mean increments is the persistent cooling by the analysis in the layer between 250 and 400 hPa. This is consistent with the negative bias in the background departures at 300 hPa shown in Fig. 7 and provides further evidence of the warm model bias at these levels. Except for seasonal variations, the magnitude of the cooling remains relatively constant throughout much of the period, although noticeable changes occur, for example, beginning in the mid-to-late 1990s as the number of aircraft and satellite observations increase, and again in 2006, possibly in response to the introduction of data from IASI and GPSRO. Warming by the analysis is evident above 200 hPa and below 700 hPa. In this global view, the mean increments close to the surface exhibit a negative trend, with strong warming before the early 1990s turning to slight cooling after 2010, but this is in fact the net effect of distinct regional differences in the increments (not shown). In particular, near-surface warming by the analysis in response to a cold model bias over northern midlatitude landmasses is offset by cooling over southern oceans that generally increases with time, beginning with the assimilation of data from the first microwave humidity sensors in the late 1980s. These differences also contribute to the large variability of the increments below 700 hPa (Fig. 10b). The variability in the midtroposphere is noticeable but small compared with that at low levels, again highlighting the consistency of the cooling by the observations between 250 and 400 hPa.
The increments of specific humidity in the tropics are shown in Fig. 11 for levels between the surface and 250 hPa (the values become exceedingly small above this level). The mean increments indicate distinct biases in the middle and lower troposphere, with systematic drying between 600 and 300 hPa and mostly moistening below 700 hPa. The corrections are generally larger during the second half of the period and especially after the late 1990s as more satellite observations of humidity become available. There is an abrupt increase in the variability of the increments corresponding to the introduction of the first SSM/I instrument in mid-1987, with additional increases corresponding to the use of a second and third SSM/I instrument in late 1990 and mid-1995, respectively. The use of multiple SSM/I instruments from the early 1990s to late 2000s also corresponds to a strong drying and a marked increase in variability at levels very close to the surface. The introduction of AMSU-B data in 1998 corresponds to marked increases in the mean and variability of the increments, the latter being most pronounced in the layer between 800 and 900 hPa. The sensitivity of the precipitation to these observing system changes is discussed in section 5.
4. Aerosol data assimilation
In addition to a standard meteorological analysis, MERRA-2 includes an aerosol analysis as described in Randles et al. (2016, 2017) and B17. The multidecadal coverage and the coupling between aerosols and the circulation is a step forward compared to previous EOS-era reanalyses, such as MERRAero, the NAAPS reanalysis (Lynch et al. 2016), the MACC reanalysis (Inness et al. 2013), and the more recent CAMS reanalysis (Flemming et al. 2017). The MERRA-2 system produces 3-hourly analyses and gridded output of both observable parameters and aerosol diagnostics not easily observed, especially on a global scale, with potential applications ranging from air quality forecasting to studies of aerosol–climate and aerosol–weather interactions (e.g., Bocquet et al. 2015).
An analysis splitting technique (Randles et al. 2017) is used to assimilate AOD at 550 nm, in which a two-dimensional analysis is performed first using error covariances derived from innovation data, and then the horizontal increments are projected vertically and across species using an ensemble method. AOD observations are derived from several sources, including
reflectances from AVHRR [1979–2002, ocean-only (Heidinger et al. 2002)],
reflectances from MODIS on Terra (2000–present) and Aqua (2002–present) (Remer et al. 2005; Levy et al. 2007),
AOD retrievals from MISR [2000–14, bright, desert regions only (Kahn et al. 2005)], and
direct AOD measurements from the ground-based AERONET [1999–2014 (Holben et al. 1998)].
MODIS provides the vast majority of AOD observations assimilated in MERRA-2, especially after 2002, when data from both the Terra and Aqua satellites become available. Prior to 2000, only AVHRR reflectances over ocean are used in MERRA-2. AOD for both MODIS and AVHRR are derived from cloud-cleared reflectances using a neural net procedure trained on AERONET measurements (Randles et al. 2017). By construction, these AOD retrievals are unbiased with respect to AERONET observations. AOD from MISR and AERONET observations are used without bias correction. Additional details about the aerosol observing system in MERRA-2 can be found in Randles et al. (2016, 2017).
The GOCART model (Chin et al. 2002; Colarco et al. 2010) is coupled with the GEOS atmospheric model to simulate the life cycles of five externally mixed aerosol species, including dust, sea salt, black carbon, organic carbon, and sulfate. The model carries three-dimensional mass mixing ratios of these five aerosol species as prognostic aerosol tracers. The AOD at 550 nm is a column- and species-integrated optical quantity, which is calculated as the summed product of each species mass and its extinction coefficient based on aerosol optical properties derived largely from the OPAC dataset [see Randles et al. (2017) and references therein]. Emissions of both dust and sea salt are wind-driven for each of five size bins, parameterized following Marticorena and Bergametti (1995) and Gong (2003), respectively. Sulfate and carbonaceous aerosol emissions derive from both natural and anthropogenic sources, as described in Randles et al. (2017). In particular, MERRA-2 includes volcanic sources (Diehl et al. 2012) and biomass burning emissions that utilize satellite observations and are based on the Reanalysis of the Tropospheric Chemical Composition, version 2 (RETRO-2; Schultz et al. 2008), the Global Fire Emissions Database, version 3.1 (GFED-3.1; van der Werf et al. 2006), and the Quick Fire Emission Dataset, version 2.4r6 (QFED-2.4.r6; Darmenov and da Silva 2015).
It should be noted that AOD observations can only directly constrain the total, species-integrated and vertically integrated aerosol extinction—a quantity that can be related to column aerosol mass by assuming a set of optical properties. Nonanalyzed aerosol properties, such as the vertical distribution, aerosol speciation, and absorption, are not fully constrained by the observations and are chiefly determined by the underlying model physics and error covariance assumptions. Despite this fact, B17 show that the MERRA-2 aerosol reanalysis has considerable skill in simulating numerous observable aerosol properties. Randles et al. (2017) show that the AOD fields in MERRA-2 generally have both high correlation and low bias relative to independent (nonassimilated) sun photometer and aircraft observations.
As in the case of the meteorological analysis discussed in section 3, statistics of background and analysis departures provide a basic metric of the quality of the aerosol assimilation. Figure 12 shows probability distribution functions of collocated observation-minus-forecast and observation-minus-analysis departures from MERRA-2 for each sensor in the aerosol observing system. Statistics are shown in terms of the log-transform AOD analysis variable [i.e., ln(AOD + 0.01)] which is approximately normally distributed (Randles et al. 2017). Note that AOD is a dimensionless quantity and log-transformed AOD is typically in the range (−4, 2). As expected, compared to the forecast departures, the analysis departures show reduced bias with respect to the observations. Note also that the innovation variances are much larger over land than ocean, a direct consequence of the signal-to-noise limitation of aerosol retrievals over land.
Regional aspects of the global distribution of aerosols are illustrated in Fig. 13, which shows time series of analyzed AOD from MERRA-2 area averaged over several major aerosol source regions. The contribution of each aerosol species to the total AOD is indicated by the colored shading. The seasonal cycles of dust and biomass burning (carbonaceous) AOD are apparent in all regions. Large increases in sulfate aerosol occur in all regions after the El Chichón (1982) and Mount Pinatubo (1991) volcanic eruptions. Over the Asian region (Fig. 13a), the analysis captures high carbonaceous aerosol associated with the 2003 Siberian fires and the increasing trend in AOD between the late 1990s and present [commensurate with increasing anthropogenic aerosol emissions reported by Diehl et al. (2012)]. The AOD over northern Africa (Fig. 13b) is dominated by dust, and major dust transport events such as in 2010 are captured (see B17 for details). Carbonaceous aerosol from biomass burning in major source regions such as the Amazon basin are also well captured (Fig. 13c), especially after 2000, when emissions inventories derive from MODIS observations (Darmenov and da Silva 2015).
Figure 14 compares values of AOD from several recent aerosol reanalyses for the period 2003–10. Where such information is available, the results are partitioned by species and identified as either fine or coarse mode (see caption for details). Also shown are multimodel average results from phase I of AeroCom (Kinne et al. 2006), as well as both model (Yu_Model) and observational (Yu_Obs) estimates from Yu et al. (2006). The latter study includes an attempt to account for satellite clear-sky biases by combining MODIS and MISR observations with the GOCART model. Compared to MERRAero, for example, MERRA-2 has slightly higher global average AOD because of increased contributions from dust (related to the assimilation of MISR AOD over bright surfaces) and sea salt (related to changes in model physics). MERRA-2 and NAAPS show similar global average AOD, both for fine- and coarse-mode aerosol. Models without assimilation (AeroCom and Yu_Model) underestimate global average AOD compared to both observational estimates (Yu_Obs) and the aerosol reanalyses. The MACC aerosol reanalysis has the highest global mean AOD (Bellouin et al. 2013), which is close to the MODIS-only value of 0.188 for the period 2003–10 (Yu et al. 2006). MACC also has more dust and sea salt aerosol compared to the other reanalyses, particularly over the ocean (not shown).
The direct aerosol impact on the radiative energy balance of Earth is dependent on the vertical distribution of aerosol scattering and absorption, which is not fully constrained by the vertically integrated AOD measurements that MERRA-2 assimilates. An assessment of the aerosol vertical structure and absorption is presented in a companion paper (B17). Long-term aerosol reanalyses can potentially reduce uncertainty in how aerosol direct effects have changed over time, particularly once better observational constraints on aerosol absorption become available. The DRE of all aerosols is defined as the flux difference (in W m−2) between clear-sky and clear clean-sky conditions (no aerosols or clouds). In the absence of clouds, this quantity is less sensitive to the vertical distribution of aerosol absorption, although it remains sensitive to absorbing aerosols over surfaces with high albedo (Chýlek and Coakley 1974).
Table 4 compares the DRE from MERRA-2, MERRAero, MACC, model intercomparisons, and the observationally constrained estimate of Yu et al. (2006). Listed are the TOA, surface (SFC), and atmospheric (ATM) estimates of DRE for the period 2003–10, averaged over land and ocean separately. Note that the atmospheric contribution to the DRE is defined as the difference between top-of-the-atmosphere and surface values, ATM = TOA − SFC. Over land, the DRE estimate from MACC best agrees with the observationally constrained estimate. TOA and SFC forcing in MERRA-2 and MERRAero are lower than in MACC because of their lower AOD, although the atmospheric forcing is similar. Over ocean, the DRE estimates from MERRA-2 and MACC are lower and higher, respectively, than the observational estimate, and both reanalyses have lower estimates of atmospheric absorption. Much of the uncertainty in the DRE reported by the IPCC arises from differences between estimates from global models and satellite-based estimates (Myhre 2009). However, as aerosol reanalyses like MERRA-2 continue to mature and incorporate additional observations (e.g., from lidars and multispectral sensors), we expect a narrowing of the gap between simulated and satellite-based estimates of the DRE.
Clear-sky DRE from reanalyses and observations. Observed median and std dev from satellite-derived estimates in Yu et al. (2006), median and std dev from four global models in Yu et al. (2006), MERRA-2 and MERRAero climatological global area-weighted average (plus or minus monthly std dev) for 2003–10, and, for MACC, the 2003–10 global mean and uncertainty is given following Bellouin et al. (2013).
The representation of precipitation in a reanalysis is key to applications in weather and climate, as it ties together aspects of both the water and energy cycles. It also presents a significant challenge, however, as estimates of precipitation are only indirectly constrained by observations and are strongly dependent on model physics whose parameterizations have known errors and can be highly sensitive to even small changes in large-scale temperature and humidity fields. The observations themselves can sometimes introduce additional uncertainty in these estimates as a result of heterogeneous sampling, changes in instrumentation, and time-varying calibration (Bosilovich et al. 2017).
While improved representation of the hydrological cycle was a primary focus of MERRA, the character of its global precipitation in particular was found to be highly sensitive to the assimilated observations and thus to changes in the observing system (e.g., Robertson et al. 2011). Among the development aspects of MERRA-2 intended to address this issue are modifications to GEOS to conserve atmospheric dry mass and ensure that changes in global atmospheric total mass are equivalent to changes in total water (section 2f), exclusion from the analysis of microwave temperature sounding channels with strong surface sensitivity (section 2d), and, less directly, forcing of the land surface by observation-corrected precipitation estimates (section 2g).
a. Global aspects
Bosilovich et al. (2015, 2017) have investigated the global water cycle variability in MERRA-2 using comparisons with observational datasets and other recent reanalyses. Those studies present a broad range of metrics on this topic, a small subset of which are summarized here. Figure 15 shows time series of global mean precipitation for several recent reanalyses and the observation-based estimates from GPCP (Adler et al. 2003). MERRA-2 exhibits larger temporal variability than GPCP but similar temporal variability to other recent reanalyses and noticeably less spurious temporal variability than MERRA. The largest improvements compared with MERRA in this regard relate to the decreased sensitivity of MERRA-2 to the introduction of AMSU-A radiances on NOAA-15 and NOAA-16 in the late 1990s, and to the loss of SSM/I radiances in the late 2000s. There is still an obvious sensitivity in MERRA-2 to the introduction of SSM/I in 1987, but the response to these data is comparable in magnitude to those of the other reanalyses shown. The response in MERRA-2 appears accentuated as a result of the decrease and subsequent recovery of precipitation through the mid-1980s. This behavior is not reflected in the GPCP time series but is evident to lesser degrees in CFSR and ERA-Interim, especially after 1983. For MERRA-2 and CFSR, this may be related to the fact that the SST boundary conditions used in these reanalyses reach their global minimum value for the entire reanalysis period after 1985 (Fig. 5), but further investigation is required to confirm this. The increasing trend in global precipitation in MERRA-2 from approximately 2.9 mm day−1 in 1988 to approximately 3.0 mm day−1 in 1998 is likely due to increasing evaporation over oceans driven by the assimilation of additional SSM/I wind speed observations and the tight coupling of evaporation and precipitation in MERRA-2 through the global mass constraint (Bosilovich et al. 2017). Overall, the global mean precipitation values are higher than those of GPCP but well within the envelope of other recent reanalyses.
Spatial comparisons provide additional insight into the strengths and weaknesses of the representation of precipitation globally in MERRA-2. Figure 16 shows maps of time-averaged differences in precipitation during boreal summer for MERRA and MERRA-2 compared with GPCP. MERRA-2 shows general improvement compared to MERRA over oceanic regions in both the tropics and extratropics, but an increase in positive bias over northern high latitudes. A notable deficiency in MERRA-2 is the excessive precipitation in the vicinity of high topography in the tropics, especially along the Andes and over the Maritime Continent. This is related to the partitioning between resolved (large scale) and parameterized (convective) precipitation in the MERRA-2 model, which, being more heavily skewed toward the former, results in large-scale precipitation over high topography that is difficult to control. In comparing these features with available gauge data, Bosilovich et al. (2015) point out that the maximum precipitation values in MERRA-2 do not always coincide with the maximum terrain height, so other effects also may play a role locally. Despite this deficiency over tropical land areas, the positive bias over the warm pool present in MERRA is slightly improved in MERRA-2. Additionally, the high precipitation bias over the Central American Seas in MERRA has been reduced significantly in MERRA-2, and precipitation over the Bay of Bengal and Arabian Sea is slightly improved. Results for other seasons (not shown) are qualitatively similar to those in Fig. 16.
b. U.S. summertime precipitation variability
Deficiencies in the ability of MERRA to reproduce certain aspects of the summertime seasonal precipitation over the United States have been well documented (Bosilovich 2013). In particular, MERRA was unable to produce seasonal highs and lows in regional precipitation that were similar to observations. For example, droughts and floods were only weakly reproduced.
Figure 17 shows the time series of summertime seasonal precipitation anomalies over the U.S. Midwest as derived from the NOAA CPC gauge observations and from MERRA and MERRA-2 model-generated precipitation. (The correlation values between various reanalyses and the gauge data for this and other regions of the United States are shown in Fig. 18.) The limitations of MERRA are apparent, especially when comparing values for 1988 (regional drought) and 1993 (large-scale flooding) with the observed values. In contrast, MERRA-2 is able to reproduce the 1988 and 1993 anomalies and is generally much better at tracking the overall variability of the observed anomalies. The poor performance of MERRA-2 in 1980 is a notable exception. A significant drought occurred in the southern Great Plains that year, but its location in MERRA-2 extended too far northeastward into the Midwest.
Figure 18 presents regional summary statistics for U.S. summer seasonal precipitation anomalies for selected reanalyses. The regions are defined as in Bosilovich (2013). For each region, the temporal mean, standard deviation, and anomaly correlation with respect to the CPC data are derived from time series like those shown in Fig. 17. In general, precipitation mean values across the United States are improved in MERRA-2 compared with MERRA (Fig. 18a), and in many regions the values for MERRA-2 improve over those of other reanalyses as well. There is also a marked increase in the standard deviation of the MERRA-2 time series relative to MERRA (Fig. 18b). As discussed above, for example, MERRA-2 more realistically reproduces the seasonal extremes in Midwest precipitation. Note, however, that MERRA-2 overestimates the standard deviation with respect to the CPC estimates in some regions. Ancillary results indicate that this is due to an excess in the number of days with rain in MERRA-2. Improvements in MERRA-2 are most evident in the anomaly correlation of the seasonal time series (Fig. 18c). In this measure, the two most recent reanalyses, JRA-55 and MERRA-2, generally outperform the others. MERRA-2 produces the highest correlations of the reanalyses shown in most regions, with substantially higher correlations in a few of these regions.
The detection and analysis of extreme weather, including extreme precipitation events, is a topic of societal interest and another potential application of reanalyses. At least some of this interest is related to assessing changes in the risk of such events in the context of climate change. For example, observation-based studies cite strong evidence of an upward trend in the frequency and intensity of extreme precipitation events averaged over the United States during the last 50 years (Kunkel et al. 2013), although the causes of the observed trends are less certain. Figure 19 shows the accumulated precipitation amounts for the largest precipitation events (at the 99th percentile) as derived from gauge observations, MERRA, and MERRA-2. Compared with the observations, MERRA shows very low values and very little structure across the continental United States. MERRA-2, on the other hand, exhibits a spatial pattern more similar to the observations, and the magnitude of the extreme rainfall is also more similar to the observations. MERRA-2 does, however, overestimate the precipitation values over the Midwest. While the results in Fig. 19 provide an indicator of how the representation of extreme events has improved in MERRA-2 compared with MERRA, the relatively coarse resolution of both reanalyses limits their utility for studying such events in detail. Presumably, the trend toward increasing resolution, among other improvements, will reduce these limitations in future global reanalyses.
6. The stratosphere
In MERRA-2 the stratospheric meteorology and ozone have benefited from improvements to the GEOS atmospheric model and GSI analysis scheme, as well as from the addition of observations that were not incorporated into MERRA. The model changes most relevant to the stratosphere are the use of the cubed-sphere grid and the retuning of the GWD parameterization. The amplitude of the nonorographic GWD was increased in the tropics, enabling a model-generated QBO that was not found in the model version used for MERRA (Molod et al. 2015). Having a model-generated QBO, in turn, results in smaller lower-stratospheric analysis wind increments in MERRA-2 than in MERRA (Coy et al. 2016). The strength of the orographic GWD was also increased in the Southern Hemisphere to better model the strong, late-winter westerlies found there (Molod et al. 2015).
The main GSI change relevant to the stratosphere is the use in MERRA-2 of the CRTM for the assimilation of SSU radiances, while in MERRA the SSU assimilation was based on GLATOVS (section 2d). These SSU radiance channels are a major source of stratospheric information during the 1980s and 1990s, although the SSU instruments during these decades span several satellite platforms, each with different bias characteristics (Kobayashi et al. 2009). The CRTM has been enhanced for SSU data assimilation since MERRA and now accounts for these biasing factors.
The main additional observations relevant to the stratosphere for MERRA-2 are GPSRO bending angle observations from the suite of platforms beginning in July 2004 and temperature and ozone measurements of the middle atmosphere from MLS and OMI on the EOS Aura satellite beginning later the same year (Froidevaux et al. 2006; Schwartz et al. 2008; McPeters et al. 2008). MERRA-2 assimilates GPSRO bending angle observations up to 30 km. Details of the GPSRO platforms assimilated by MERRA-2 can be found in McCarty et al. (2016). The GPSRO observations aid lower-stratospheric bias correction by providing a stable source of temperature and moisture measurements. The MLS-retrieved temperature profiles are assimilated in MERRA-2 at altitudes above 5 hPa, providing a strong constraint on the dynamics of the stratopause and lower mesosphere. As shown below in section 6a, this improves the quality of the synoptic meteorological fields at these altitudes but may complicate the study of trends. The MLS and OMI contributions to ozone assimilation are discussed in section 6b.
The cubed-sphere discretization of the MERRA-2 model eliminates computational instabilities near the poles, a characteristic of latitude–longitude grids. This is especially important for stratospheric analysis where strong cross-polar flow events occur frequently, especially during major sudden warming events, as planetary-scale Rossby waves disturb the polar vortex. Ertel’s potential vorticity (EPV), a scalar based on the horizontal vorticity, is often used to characterize the stratospheric circulation (Andrews et al. 1987), where stronger EPV gradients imply stronger flow. Figure 20 illustrates a case where the analyzed wind speeds in MERRA-2 reached nearly 170 m s−1 close to the polar stratopause at 1200 UTC 2 January 1995. On a global scale (Figs. 20a,b), the MERRA and MERRA-2 EPV fields appear similar, with the polar vortex (indicated by green and orange colors) displaced well off the pole. In both cases, strong winds cross the North Pole as they circle around the region of high EPV. However, a closer look reveals that the EPV in MERRA (Fig. 20c) has anomalous radial perturbations near the pole, while the EPV in MERRA-2 (Fig. 20d) shows a smooth and strong EPV gradient in this region. Note also that, while the largest discontinuities in the MERRA EPV field occur close to the pole itself, their effects can extend well beyond this location.
Figure 21 provides an example of how the assimilation of MLS temperature measurements in MERRA-2 improves the representation of the dynamics near the stratopause. The figure shows the time–height evolution of polar temperatures during the 2005/06 Northern Hemisphere winter in which a major stratospheric sudden warming occurred. In a comprehensive study of this winter based on MLS observations, Manney et al. (2008) documented the disappearance of the warm polar stratopause during the warming and its later high-altitude reformation and subsequent descent. This breakdown and high-altitude reformation in early February 2006 is now well captured in MERRA-2 (Fig. 21b), in contrast to MERRA (Fig. 21a).
The characteristics of the assimilation on longer time scales is illustrated in Fig. 22, which shows the time–height evolution of global monthly averaged temperature anomalies in MERRA-2. The 35-yr mean and annual cycle for the period 1980–2015 have been subtracted from each pressure level. The global temperatures in the lower stratosphere (100–10 hPa) show no obvious discontinuities as different instruments become available. There is a slight cooling with time over the 35 years, which is generally consistent with recent analyses of the satellite-based stratospheric climate data record [see Seidel et al. (2016) and references therein]. There are also episodic temperature increases associated with the two large volcanic eruptions, El Chichón in 1982 and Mount Pinatubo in 1991. In the upper stratos