ERA-40 (Uppala et al., 2005) is an extended full-input reanalysis covering 45 years from September 1957 through August 2002. ERA-40 was released by ECMWF in 2003 and represents an important improvement relative to the first generation of modern reanalysis systems, including FGGE (Bengtsson et al., 1982) and ERA-15 (Gibson et al., 1997). ERA-40 did not assimilate satellite data prior to January 1973. The ERA-40 reanalysis from September 1957 through December 1972 is therefore a conventional-input reanalysis. ERA-40 products continue to be used in many studies that require long-term atmospheric data.

ERA-Interim (Dee et al., 2011), initially released by ECMWF in 2008, is a full-input reanalysis of the satellite era that includes several corrections and modifications to the system used for ERA-40. In particular, ERA-Interim uses a 4D-Var data assimilation system, which makes more complete use of observations collected between analysis times than the 3D-FGAT (first guess at appropriate time) approach used in ERA-40 (see Sect. 4). Major focus areas during the production of ERA-Interim included achieving more realistic representations of the hydrologic cycle and the stratospheric circulation relative to ERA-40, as well as improving the consistency of the reanalysis products in time.

ERA-20C (Poli et al., 2015, 2016) is a surface-input reanalysis produced by ECMWF and released in 2014. ERA-20C uses a 4D-Var data assimilation system but takes its spatially and temporally varying background errors from a prior ensemble data assimilation (Isaksen et al., 2010; Poli et al., 2013). Because ERA-20C directly assimilates only surface pressure and surface wind observations, it can generate reanalyses of the climate state that extend further back in time (in this case to the beginning of the 20th century). Assimilation of surface data indirectly constrains the upper-atmospheric state, but these constraints are relatively weak on longer-than-synoptic timescales. While data from ERA-20C extend up to 0.01 hPa, these data should be used with caution in the upper troposphere and above. The ERA-20C model also uses sea surface temperature and sea ice concentration analyses as well as radiative forcings prescribed for CMIP5. The companion product ERA-20CM (Hersbach et al., 2015) provides an ensemble of AMIP-style simulations using similar forcings and lower boundary conditions. Ensemble members are spun-up from the same initial state and differ only in the prescribed evolution of sea surface temperature (SST) and sea ice, which are drawn from the HadISST2 ensemble (Titchner and Rayner, 2014; Hersbach et al., 2015).

JRA-25 (Onogi et al., 2007), released in 2006, is a full-input reanalysis of the satellite era and the first reanalysis produced by JMA (in cooperation with CRIEPI). This reanalysis originally covered 25 years from 1979 through 2004, and was extended by an additional 10 years (through the end of January 2014) as JCDAS using an identical fixed model–assimilation system.

JRA-55 (Kobayashi et al., 2015), released in 2013, is an extended full-input reanalysis produced by JMA. JRA-55 is the most recent reanalysis that both assimilates upper-air observations and includes coverage of the pre-TOVS era (i.e. before November 1978), starting from the International Geophysical Year (IGY) in January 1958. To date, JRA-55 is the only reanalysis system to apply a 4D-Var data assimilation scheme to upper-air data during the pre-satellite era (ERA-20C has also applied 4D-Var but only to surface observations). Two companion products are also available: JRA-55C (Kobayashi et al., 2014), a conventional-input reanalysis, and JRA-55AMIP, an AMIP-style forecast model simulation without data assimilation. Both JRA-55C and JRA-55AMIP were released to the public in 2015. JRA-55C is available starting from November 1972, 2 months before JRA-55 began assimilating satellite observations (before this date, JRA-55 only assimilated conventional observations so that JRA-55 and JRA-55C are identical) and extends through December 2012. JRA-55AMIP extends from January 1958 through December 2012. Extensions beyond December 2012 are planned for both JRA-55C and JRA-55AMIP, but details have not been determined as of this writing.

MERRA (Rienecker et al., 2011), released in 2009, is a full-input reanalysis of the satellite era developed by NASA's GMAO using the GEOS-5 data assimilation system. MERRA was conceived with the intention of leveraging the large amounts of data produced by NASA's Earth Observing System (EOS) satellite constellation and improving the representations of the water and energy cycles relative to earlier reanalyses. The top level used in MERRA (0.01 hPa, approximately 80 km) is higher than the top levels used in most other reanalyses, which facilitates studies extending into the mesosphere. An earlier NASA reanalysis (Schubert et al., 1993, 1995) covering 1980–1995 was produced by NASA's DAO (now GMAO) using the GEOS-1 data assimilation system; this reanalysis is no longer publicly available and is not included in the S-RIP intercomparison.

MERRA-2 (Bosilovich et al., 2015), released in 2015, is a full-input reanalysis of the satellite era from NASA's GMAO. As the follow-on to MERRA, the production of MERRA-2 was motivated by the inability of the MERRA system to ingest some recent data types. MERRA-2 includes substantial upgrades to the model (Molod et al., 2015) and changes to the data assimilation system and input data. New constraints are applied to ensure conservation of global dry-air mass and to close the balance between surface water fluxes (precipitation minus evaporation) and changes in total atmospheric water (Takacs et al., 2016). Other new features in MERRA-2 relative to MERRA include a modified gravity wave scheme that substantially improves the model representation of the QBO (Molod et al., 2015; Coy et al., 2016); the assimilation of MLS temperature retrievals at high altitudes (pressures less than or equal to 5 hPa) to better constrain the reanalysis at upper levels; the assimilation of MLS stratospheric ozone profiles and OMI column ozone since the beginning of the Aura mission in late 2004 to improve representation of fine-scale ozone features, especially in the region around the tropopause; and the assimilation of aerosol optical depth (AOD; Randles et al., 2016), with analysed aerosols fed back to the forecast model radiation scheme.

NCEP–NCAR R1 (Kalnay et al., 1996; Kistler et al., 2001), which was first released in 1995, is the first modern reanalysis system with extended temporal coverage (1948–present) and is produced using a modified 1995 version of the NCEP forecast model. NCEP–DOE R2 (Kanamitsu et al., 2002), released in 2000, covers the satellite era (1979–present) using a 1998 version of the same model and includes corrections for some important errors and limitations identified in R1. Both R1 and R2 remain in widespread use; however, these systems have relatively low top levels (3 hPa) and relatively coarse vertical resolutions (28 levels), and they assimilate retrieved temperatures rather than radiances from the operational nadir sounders, rendering them unsuitable for most studies of the middle atmosphere.

CFSR (Saha et al., 2010), released in 2009, is a full-input reanalysis of the satellite era that uses a 2007 version of the NCEP CFS. CFSR contains a number of improvements relative to R1 and R2 in both the forecast model and data assimilation system, including higher horizontal and vertical resolutions, a higher model top, more sophisticated model physics, and the ability to assimilate satellite radiances directly. CFSR is also the first global reanalysis of the coupled atmosphere–ocean–sea-ice system. Official data coverage by CFSR only extends through December 2009, but output from the same analysis system was continued through December 2010 before being migrated to the operational CFSv2 analysis system (Saha et al., 2014) from January 2011. This transition from CFSR to CFSv2 should not be confused with the transfer of CFSv2 production from NCEP EMC to NCEP operations, which occurred at the start of April 2011. CFSv2 has a different horizontal resolution and includes minor changes to physical parametrizations (some of which are described below) but is intended to serve as a continuation of CFSR and can be treated as such for most purposes. To distinguish CFSR and its CFSv2 continuation from other (mainly forecast) applications of the NCEP CFS that do not use the full data assimilation system, CFSR may also be referred to as CDAS-T382 and its continuation as CDAS-T574.

NOAA–CIRES 20CR v2 (Compo et al., 2011), released in 2009, is the first reanalysis to span more than 100 years. Like ERA-20C, 20CR is a surface-input reanalysis. Unlike ERA-20C, which uses a 4D-Var approach to assimilate both surface pressure and surface winds, 20CR uses an ensemble Kalman filter (EnKF) approach (see Sect. 4) and assimilates only surface pressure. The forecast model used in 20CR is similar in many ways to that used in CFSR, but with much coarser vertical and horizontal resolutions. 20CR provides reanalysis fields back to the mid-19th century. With only surface observations assimilated and a modest vertical resolution, 20CR is likely to be of limited utility for most studies above the tropopause and many in the upper troposphere.

Table 3 provides references for some of the physical parametrizations used in the forecast models. Many of the families of models use similar parametrizations across generations, but these are often modified and updated for use in newer systems. For example, both ERA-40 and ERA-Interim use shortwave radiation schemes based on Fouquart and Bonnel (1980) and calculate longwave radiative transfer using the RRTM (Mlawer et al., 1997), but ERA-Interim uses an updated version of the shortwave scheme and makes hourly radiation calculations rather than 3-hourly ones (Dee et al., 2011). ERA-20C replaces the Fouquart and Bonnel (1980) shortwave scheme with a modified version of the RRTM (Morcrette et al., 2008) and is the first ECMWF reanalysis to use the Monte Carlo Independent Column Approximation (McICA) for representing the radiative effects of clouds. Both JRA-25 and JRA-55 use shortwave radiative transfer schemes based on Briegleb (1992), but JRA-55 uses an updated parametrization of shortwave absorption by O2, O3, and CO2 (Freidenreich and Ramaswamy, 1999). JRA-55 also uses an updated longwave radiation model (Chou et al., 2001), which replaces the line absorption model used in JRA-25 (Goody, 1952). R1 and R2 use identical longwave radiation schemes (Fels and Schwartzkopf, 1975; Schwartzkopf and Fels, 1991), but R1 performs radiation calculations 3-hourly on a coarser linear grid, while R2 performs radiation calculations hourly on the full model grid. R2 also uses a different shortwave scheme (Chou and Lee, 1996) from that used in R1 (Lacis and Hansen, 1974). Both the longwave and shortwave schemes have been replaced by a modified version of the RRTMG (Clough et al., 2005) in CFSR/CFSv2 and 20CR. The McICA approach for representing cloud radiative effects has been implemented in CFSv2 but has not been used in CFSR or 20CR. MERRA and MERRA-2 both use the CLIRAD shortwave and longwave radiation schemes (Chou and Suarez, 1999; Chou et al., 2001).

Cloud parametrizations in the ECMWF family of reanalyses follow Tiedtke (1989) for convective clouds and Tiedtke (1993) for non-convective clouds, but with substantial differences among the three reanalyses. For example, modifications to the convective parametrization between ERA-40 and ERA-Interim yielded improvements in the diurnal cycle of convection, increases in precipitation efficiency, and the capability to distinguish shallow, mid-level, and deep convective clouds (Dee et al., 2011). ERA-Interim modified the non-convective cloud parametrization to include supersaturation with respect to ice (Tompkins et al., 2007), resulting in substantial changes in the water budget of the upper troposphere. This scheme has been further modified for ERA-20C to permit separate estimates of liquid and ice water in non-convective clouds, resulting in a more physically realistic representation of mixed-phase clouds. JRA-25 and JRA-55 both use variations of the same prognostic mass-flux type Arakawa–Schubert convection scheme (Arakawa and Schubert, 1974), but JRA-55 implements a new triggering mechanism (Xie and Zhang, 2000). MERRA-2 uses the same cloud schemes as MERRA (Table 3) but has a new set of total-water probability density functions as proposed by Molod (2012). The representation of deep convection in R1, R2, CFSR and 20CR follows Arakawa and Schubert (1974), while the representation of shallow convection follows Tiedtke (1983). The versions of these parametrizations used in CFSR and 20CR have been updated relative to those used in R1 and R2, including the addition of convective momentum mixing to the deep-convection scheme and modifications to the shallow-convection scheme that improve the representation of marine stratocumulus (Moorthi et al., 2010; Saha et al., 2010). R1 and R2 both use simple empirical relationships to diagnose non-convective cloud cover from grid-scale relative humidity, but these relationships are slightly different between the two systems. CFSR and 20CR replace these empirical relationships with a simple cloud physics parametrization with prognostic cloud condensate (Xu and Randall, 1996; Zhao and Carr, 1997).

Gravity wave drag and ozone parametrizations are of particular interest to the SPARC community. For example, the details of the parametrization of gravity wave drag (and particularly non-orographic gravity wave drag) can greatly influence the simulation of the QBO. All of the reanalysis systems include representations of orographic gravity wave sources and drag, but only MERRA, MERRA-2, CFSv2, and ERA-20C include parametrizations of non-orographic gravity wave drag. MERRA-2 uses a modified version of the gravity wave drag schemes used in MERRA (McFarlane, 1987; Garcia and Boville, 1994), with enhanced intermittency and a larger non-orographic gravity wave background source in the tropics (Molod et al., 2015). The GEOS-5 forecast model does not produce a QBO before these changes are implemented but does produce a QBO afterwards. Starting with the September 2009 version (Cycle 35r3), the ECMWF IFS includes the non-orographic gravity wave drag parametrization proposed by Scinocca (2003). The version of the IFS model used for ERA-20C (the first ECMWF reanalysis to include this parametrization) produces a QBO (Hersbach et al., 2015), but with a shorter period than observed and a weak semi-annual oscillation (SAO). CFSv2 includes a non-orographic gravity wave parametrization that considers stationary gravity waves generated by deep convection (Chun and Baik, 1998; Saha et al., 2014); this parametrization was not included in CFSR. Despite the inclusion of this parametrization, the CFSv2 forecast model does not produce a QBO. The ozone parametrizations used in current reanalysis systems are introduced and discussed in Sect. 6.1.

Forecast model representations of horizontal and vertical diffusion have strong influences on tracer transport and thermodynamic structure, particularly near the tropopause (Flannaghan and Fueglistaler, 2011, 2014). All of the models using spectral dynamical cores include implicit linear horizontal diffusion. These implicit representations are second order (R1, R2, and 20CR), fourth order (ERA-40, ERA-Interim, ERA-20C, JRA-25, and JRA-55), or eighth order (CFSR) in spectral space. Horizontal diffusion along model sigma layers in R1 causes the occurrence of spurious “spectral precipitation”, particularly in mountainous areas at high latitudes (Kanamitsu et al., 2002). A special precipitation product was produced for R1 to address this issue, which is greatly reduced in R2. The finite-volume dynamical cores used for MERRA and MERRA-2 do not include implicit diffusion, so an explicit formulation is required. Both MERRA and MERRA-2 include explicit second-order horizontal divergence damping with a dimensionless coefficient of 0.0075 below the sponge layer. MERRA-2 also includes a second-order Smagorinsky divergence damping with a dimensionless coefficient of 0.2 that was not applied in MERRA. The approaches to horizontal diffusion used in reanalysis schemes have been discussed in detail by Jablonowski and Williamson (2011). Approaches to vertical diffusion in the free atmosphere (above the boundary layer) are all based on the local Richardson number. ERA-40, ERA-Interim, and ERA-20C use the revised Louis scheme (Louis, 1979; Beljaars, 1995; Flannaghan and Fueglistaler, 2011), while JRA-25 and JRA-55 use the level-2 turbulence closure proposed by Mellor and Yamada (1974). R1, R2, CFSR, and 20CR use the local K closure proposed by Louis et al. (1982) in the free troposphere, but with different specifications of the background diffusion coefficients. Background diffusion coefficients in R1 and R2 are uniform throughout the atmosphere, which results in very strong vertical mixing across the tropical tropopause (Wright and Fueglistaler, 2013). By contrast, background diffusion coefficients in CFSR and 20CR decay exponentially with height from a surface value of 1 m2 s-1 (Saha et al., 2010). Free-tropospheric vertical diffusion in MERRA and MERRA-2 is also parametrized based on the Louis et al. (1982) scheme, with diffusion coefficients based on the gradient Richardson number; however, a tuning parameter applied in MERRA severely suppressed turbulent mixing at pressures less than ∼ 900 hPa. That restriction has been removed in MERRA-2, but diffusion coefficients are still usually quite small in the free atmosphere. Parametrizations of vertical diffusion within the boundary layer vary more widely amongst reanalysis systems. These parametrizations are documented in Chapter 2 of the planned S-RIP reports, but we do not discuss them here.

Boundary and other specified conditions may be regarded as “externally supplied forcings” to the forecast model. These conditions include all elements of the reanalysis system that are not taken from the forecast model or data assimilation but are used to produce the outputs. The factors that can be considered “external” vary somewhat among reanalyses because the forecast and assimilation components have provided a progressively larger fraction of the inputs for the forecast model as reanalysis systems have increased in sophistication. For example, while most of the reanalyses are run with specified SSTs and sea ice concentrations, CFSR and CFSv2 are coupled atmosphere–ocean–sea-ice reanalysis systems. SST and sea ice lower boundary conditions for the CFSR and CFSv2 atmospheric models are therefore generated by coupled ocean and sea ice models (although temperatures at the atmosphere–ocean interface are relaxed every 6 hours to separate SST analyses like those used by other reanalysis systems; Saha et al., 2010). Table 4 lists the SST and sea ice analyses used by the reanalysis systems. Several reanalyses use different SSTs and sea ice concentrations for different time periods, which can lead to temporal discontinuities in reanalysis products (e.g. Simmons et al., 2010). Bosilovich et al. (2015; their Sect. 8a) further discuss these discontinuities and the steps taken in MERRA-2 to limit them and provide a cursory graphical intercomparison of the SST fields used in the production of several recent reanalyses. Ozone is another prime example of a quantity that may either be internally generated or externally imposed, with particular relevance to SPARC studies. The treatment of ozone in these reanalysis systems is discussed in Sect. 6.1.

The treatment of aerosols and other trace gases also differs: MERRA-2 assimilates aerosol optical depths and uses these analysed aerosol fields for radiation calculations (Randles et al., 2016), while most other systems use climatological aerosol fields. Different aerosol climatologies are used in ERA-40 (Tanré et al., 1984), ERA-Interim (Tegen et al., 1997), JRA-25 and JRA-55 (WMO, 1986), MERRA (Colarco et al., 2010), and CFSR and 20CR (Koepke et al., 1997). ERA-20C uses decadally varying monthly aerosol fields prepared for CMIP5 (Lamarque et al., 2010; van Vuuren et al., 2011; Hersbach et al., 2015), while R1 and R2 neglect the role of aerosols altogether. Of the systems using prescribed aerosol fields, only CFSR, 20CR, and ERA-20C adjust them to account for the effects of volcanic eruptions. Therefore, in the majority of reanalyses, the volcanic response in many dynamical and chemical variables is entirely due to the influences of assimilated observations. MERRA-2 aerosol analyses, which are produced using the GOCART model (Chin et al., 2002) and the Goddard Aerosol Assimilation System (Buchard et al., 2015; Randles et al., 2016), track the evolution of black and organic carbon, dust, sea salt, and sulfates. These analyses are supported by the assimilation of bias-corrected AOD at 550 nm from a variety of remote-sensing platforms, including AVHRR (1980–2000), MODIS instruments on the Terra (2000–present) and Aqua (2002–present) satellites, MISR over bright surfaces (2000–2014), and the ground-based AERONET (1999–2014). Analysed aerosols, including volcanic aerosols, interact with the MERRA-2 meteorological state via direct radiative coupling.

The assumptions governing greenhouse gas concentrations also vary widely. For example, the treatment of CO2 ranges from assumptions of constant global mean CO2 (330 ppmv in R1; 350 ppmv in R2; 375 ppmv in JRA-25) to a linear trend extrapolated from observed 1990 values (ERA-40 and ERA-Interim) to various permutations of historical observations and future emissions scenarios (all other systems). Climatological values of several other trace gases, including CH4, N2O, major chlorofluorocarbons (CFCs), and occasionally hydrochlorofluorocarbons (HCFCs), are used in most of the reanalysis systems but are not included in R1, R2, and JRA-25. Radiatively active trace gases (including CO2) are generally assumed to be globally well-mixed within the atmosphere, with a few notable exceptions. First, ERA-20C applies a rescaling to CMIP5 recommended values that varies by latitude, height, month, and species (see Hersbach et al., 2015, for details). Second, distributions of CH4, N2O, CFCs, and HCFCs used in MERRA and MERRA-2 are based on steady-state monthly mean climatologies generated using a two-dimensional chemistry transport model; these climatologies vary by latitude, height, and month, but trace gas concentrations do not change from year to year. Third, CFSR and 20CR (after 1955) use monthly 15∘ × 15∘ gridded estimates of CO2 derived from historical WMO Global Atmospheric Watch observations. Figure 4a and b show temporal variations in prescribed values of CO2 and CH4. For simplicity, the base values of CO2 and CH4 (before rescaling) are shown for ERA-20C, while values of CH4 for MERRA and MERRA-2 have been calculated using a mass- and area-weighted integral between 1000 and 288 hPa (seasonal variability in tropospheric methane is small in this climatology and is therefore omitted in the integrated estimate shown in Fig. 4b). The temporal evolution of greenhouse gas concentrations prescribed in these reanalysis systems is provided in the Supplement to this paper.

The representation of solar radiation at the top of the atmosphere (TOA) also varies by reanalysis. Most reanalyses assume a constant total solar irradiance (TSI) of 1365 W m-2 (R2, JRA-25, JRA-55, and MERRA), 1367.4 W m-2 (R1), or 1370 W m-2 (ERA-40 and ERA-Interim). These reanalyses therefore do not explicitly account for the ∼ 11-year solar cycle in the radiative calculations, although the influences of this cycle may be introduced into the reanalysis via the assimilated observations (or in some cases via boundary conditions; see also Simmons et al., 2014). MERRA-2 and ERA-20C use TSI variations provided for CMIP5 historical simulations by the SPARC SOLARIS working group (Lean, 2000; Wang et al., 2005), with the Total Irradiance Monitor (TIM) correction applied. These variations account for solar cycle changes through mid-2008 and repeat the final cycle (April 1996–June 2008) thereafter, with magnitudes ranging from 1360.2 to 1362.7 W m-2 between 1900 and 2008 and from 1360.6 to 1362.5 W m-2 between 1980 and 2008. CFSR and 20CR use annual average TSI variations ranging from 1365.7 to 1367.0 W m-2 based on data prepared by Huug van den Dool (personal communication, 2006). The solar cycle before 1944 is repeated backwards for 20CR (e.g. insolation for 1943 is the same as that for 1954, that for 1942 is the same as that for 1953, and so on) and the solar cycle after 2006 is repeated forwards in a similar manner for both CFSR and 20CR. A programming error in ERA-40 and ERA-Interim artificially increased the effective TSI by about 2 W m-2 relative to the specified value (so that the effective TSI is ∼ 1372 W m-2 rather than 1370 W m-2). Dee et al. (2011) reported that the impact of this error is mainly expressed as a warm bias of approximately 1 K in the upper stratosphere; systematic errors in other regions are negligible. Figure 4c shows temporal variations in prescribed values of TSI from 1979 through 2015. To better highlight key differences among the reanalyses, seasonal variations resulting from the eccentricity of Earth's orbit around the sun are omitted from the figure. These seasonal variations have peak-to-peak amplitudes of ∼ 6.8 %, approximately 1 order of magnitude larger than the maximum difference among TSI estimates shown in Fig. 4c. The temporal evolution of TSI values prescribed in these reanalysis systems is provided in the Supplement to this paper.

Radiosondes provide high vertical resolution profiles of temperature, horizontal wind, and humidity worldwide, although most radiosonde stations are located in the Northern Hemisphere at middle and high latitudes over land. The typical vertical coverage of radiosonde data extends from the surface up to 30–10 hPa for temperature and winds and from the surface up to 300–200 hPa for humidity. The main source of systematic errors in radiosonde temperature measurements stems from the effects of solar radiative heating and (to a lesser extent) infrared cooling on the temperature sensor (Nash et al., 2011). This issue, which is sometimes called the “radiation error”, can cause pronounced warm biases in raw daytime stratospheric measurements. These biases may be corrected onsite in the ground-data-receiving system before reporting, and further corrections may be applied at each reanalysis centre before assimilation. The major issue with radiosonde humidity measurements is that the sensor response has historically been too slow at low temperatures (Nash et al., 2011). Radiosonde observations of humidity at pressures less than about 300 hPa are therefore often unreported and/or excluded from the assimilation. Recent advances in radiosonde instrumentation are beginning to improve this situation, although operational radiosondes remain unable to provide accurate estimates of humidity in the stratosphere. Other issues include frequent (and often undocumented) changes in radiosonde instrumentation and observing methods at radiosonde stations, which may cause jumps in the time series of temperature, relative humidity, and winds. Several “homogenization” activities for radiosonde temperature data exist to support climate monitoring and trend analyses, in which observations from different launch sites and instrument suites are post-processed to remove biases, drifts, and jumps in the data record. Although some of these activities have been conducted independently of reanalysis activities (e.g. Sherwood, 2007), others (notably RAOBCORE; Haimberger et al., 2008, 2012) have been conducted with reanalysis applications in mind. One or more versions of RAOBCORE are used in ERA-Interim (v1.3), MERRA and MERRA-2 (v1.4 through 2005), and JRA-55 (v1.4 through 2005; v1.5 thereafter). Further efforts on data rescue, reprocessing, homogenization, and uncertainty evaluation by the broader research community are likely to be an essential part of the next generation of reanalyses (e.g. ACRE and GRUAN; Allan et al., 2011; Bodeker et al., 2016).

Figure 6 shows timelines of conventional data sources assimilated by ERA-Interim, JRA-55, MERRA, MERRA-2, and CFSR. These timelines are quite consistent among modern full-input reanalyses (as well as the conventional-input JRA-55C). All of the reanalysis systems listed in Table 1 assimilate records of surface pressure from manned and automated weather stations, ships, and buoys, while all but 20CR assimilate at least some records of surface winds over oceans. All but ERA-Interim, ERA-20C, 20CR, and JRA-55C assimilated synthetic surface pressure data for the Southern Hemisphere (PAOBS). PAOBS are subjective analyses of surface pressure produced by the Australian Bureau of Meteorology based on available observations and temporal continuity, which are used to compensate for the scarcity of direct observations in the Southern Hemisphere. The influence of these data in reanalysis systems has waned in recent years, as the availability of direct observations covering the Southern Hemisphere has expanded. All of the full-input reanalyses and JRA-55C assimilate upper-air observations made by radiosondes, dropsondes, and wind profilers. JRA-25, JRA-55, and JRA-55C assimilate wind speed profiles in tropical cyclones, while 20CR assimilates records of tropical cyclone central pressures. CFSR uses the NCEP tropical storm relocation package (Liu et al., 1999) to relocate tropical storm vortices to observed locations.

Although different reanalysis systems may assimilate (or exclude via quality control) different subsets of conventional data within the broader categories shown in Fig. 6, conventional data are often shared among reanalysis centres. Major quality control criteria for conventional data include checks for completeness, physical and climatological consistency, and duplicate reports. Data may also be filtered using locally compiled blacklists or blacklists acquired from other data providers and reanalysis centres. Detailed intercomparisons of the conventional data or quality control criteria used in each reanalysis are beyond the scope of this review; however, four of the reanalyses that assimilate upper-air observations (ERA-40, ERA-Interim, JRA-25, JRA-55, and JRA-55C) use the ERA-40 ingest as a starting point, and the ERA-40 ingest has much in common with the conventional data archives used by NCEP (R1, R2, and CFSR) and the NASA GMAO (MERRA and MERRA-2). More recent updates in data holdings at ECMWF, JMA, GMAO, and NCEP rely heavily on near-real-time data gathered from the WMO Global Telecommunication System (GTS), which also contributes to the use of a largely (but not completely) common set of conventional data among reanalysis systems.

Measurements made by aircraft, such as the Aircraft Meteorological Data Relay (AMDAR) collection, are influential inputs in many atmospheric analyses and reanalyses (Petersen, 2016). Horizontal wind data from aircraft are assimilated in all of the reanalysis systems except JRA-55C, ERA-20C, and 20CR, while temperature data from aircraft are assimilated in all of the reanalysis systems except JRA-25, JRA-55 (and JRA-55C), ERA-20C, and 20CR. In principle, aircraft data were assimilated from the outset by ERA-40 (September 1957; Uppala et al., 2005), JRA-55 (January 1958; Kobayashi et al., 2015), and R1 (January 1958; Kalnay et al., 1996), although many of the data from these early years do not meet the necessary standards for assimilation. The volume of aircraft data suitable for assimilation increased substantially after January 1973 (Uppala et al., 2005; Kobayashi et al., 2015). Aircraft temperature data have been reported to have a warm bias with respect to radiosonde observations (Ballish and Kumar, 2008). This type of discrepancy among ingested data sources can have important impacts on the analysis. For example, Rienecker et al. (2011) and Simmons et al. (2014) have shown that an increase in the magnitude of the temperature bias at 300 hPa in MERRA with respect to radiosondes in the middle to late 1990s coincides with a large increase in the number of aircraft observations assimilated by the system and that differences in temperature trends at 200 hPa between MERRA and ERA-Interim reflect the different impacts of aircraft temperatures in these two reanalysis systems. MERRA-2 applies adaptive bias corrections to AMDAR observations that may help to reduce the uncertainties associated with assimilating these data (Bosilovich et al., 2015): after each analysis step the updated bias is estimated as a weighted running mean of the aircraft observation increments from preceding analysis times. These adaptive bias corrections are calculated and applied for each aircraft tail number in the database separately.

Operational satellite radiance measurements provide constraints for temperature and moisture with more homogeneous spatial coverage than radiosondes but with coarser vertical resolutions and deep vertical weighting functions (Fig. 7). All of the full-input reanalyses described in this paper (i.e. excluding the conventional-input reanalysis JRA-55C and the surface-input reanalyses ERA-20C and 20CR) assimilate some form of satellite measurements after the coverage of those measurements expanded in the 1970s. The earliest satellite data assimilated by these systems are VTPR radiances, which were assimilated by ERA-40 and JRA-55 from January 1973 through late 1978 (ERA-40) or early 1979 (JRA-55). The most prevalent satellite data assimilated by reanalyses are observations made by microwave and infrared sounders in the TOVS suite (October 1978–October 2006 on multiple satellites) and the ATOVS suite (October 1998 to present on multiple satellites), which influence the global temperature and moisture analyses produced by all of the full-input reanalysis systems described in this paper. The TOVS suite included SSU, MSU, and HIRS. The ATOVS suite includes the AMSU-A, AMSU-B, and more recent versions of HIRS. R1 and R2 assimilate temperature retrievals from these instruments (e.g. Reale, 2001), while the other full-input reanalysis systems assimilate radiance data directly. The assimilation of satellite radiances requires the use of a radiative transfer scheme, which differs from the radiative transfer scheme used in the forecast model. Several systems use one or more versions of RTTOV, including RTTOV-5 (ERA-40), RTTOV-6 (TOVS radiances in JRA-25), RTTOV-7 (ERA-Interim and ATOVS radiances in JRA-25), and RTTOV-9 (JRA-55). MERRA uses the GLATOVS model for assimilating SSU radiances and the CRTM for assimilating all other radiances. MERRA-2 and CFSR use the CRTM for all radiances.

Figure 8 shows timelines of satellite radiances assimilated by ERA-Interim, JRA-55, MERRA, MERRA-2, and CFSR. In addition to the TOVS and ATOVS suites, several reanalysis systems also assimilate radiances from AIRS (ERA-Interim, MERRA, MERRA-2, and CFSR). MERRA-2 and CFSR assimilate hyperspectral radiances from IASI, while MERRA-2 also assimilates radiances from the hyperspectral sounder CrIS and the most recent generation of microwave sounder ATMS.

Raw radiance data often include drifts and jumps due to orbital drift, inaccurate calibration offsets, long-term trends in atmospheric CO2, and other issues (Zou et al., 2006, 2014; Simmons et al., 2014). Moreover, the overlap periods between successive instruments are often short, which complicates efforts to adjust for these issues. The biases associated with these drifts and jumps can propagate into the reanalysis fields. For example, Rienecker et al. (2011) speculated that artificial annual cycles emerge in upper-stratospheric temperature in MERRA because variations in atmospheric CO2 are not considered in the GLATOVS radiative transfer model used to assimilate SSU radiances (these issues have been corrected in MERRA-2, which uses the CRTM to assimilate SSU radiances). Discontinuities associated with the TOVS-to-ATOVS transition show up in several aspects of the reanalysis but are particularly pronounced in the upper stratosphere and lower mesosphere (e.g. Simmons et al., 2014). The strong influence of the TOVS-to-ATOVS transition in the middle atmosphere is mainly attributable to the substantial improvement in vertical resolution involved in switching from SSU (Fig. 7a) to AMSU-A (Fig. 7b). A further example of the influence of ATOVS is provided by the cold bias (∼ 2 K) in middle-stratospheric temperatures that persisted in JRA-25 between 1979 and 1998 (Onogi et al., 2007). This feature resulted from a known cold bias in the radiative transfer model used by JRA-25. SSU had only three channels sensitive to stratospheric temperature, too few to correct the model bias, whereas the assimilation of the higher-resolution AMSU-A radiances starting in 1998 was sufficient to effectively correct the model bias. Discontinuities associated with the TOVS-to-ATOVS transition will be discussed in more detail in the planned S-RIP reports.

Post-launch inter-satellite calibration (or “homogenization”) efforts by the satellite remote-sensing community (such as the WMO GSICS; Goldberg et al., 2011) have substantially reduced inter-satellite differences in several cases, including MSU (Zou et al., 2006), AMSU-A (Zou and Wang, 2011), and SSU (Zou et al., 2014). This type of inter-satellite calibration has historically been performed by reanalysis systems internally via bias correction terms applied during the data assimilation step. In older reanalyses that assimilated satellite radiances, such as ERA-40 and JRA-25, bias corrections were often (but not always) based on a fixed regression that spanned the lifetime of the instrument (Uppala et al., 2005; Onogi et al., 2007; Sakamoto and Christy, 2009). This approach, which occasionally required the reanalysis to be interrupted for manual retuning of bias correction terms, has been replaced by adaptive (or variational) bias correction schemes in recent reanalysis systems. Adaptive bias corrections for satellite radiances are based on differences between observed radiances and expected radiances calculated from model-generated background states. Some early implementations of adaptive bias corrections, such as that applied to TOVS data in JRA-25, left the reanalysis vulnerable to jumps and drifts inherited from the assimilated radiances (Sakamoto and Christy, 2009). These problems are addressed in many recent reanalysis systems by defining observational “anchors” that are regarded as unbiased and are therefore allowed to contribute directly to the background state (Dee, 2005). A key example is the use of homogenized radiosonde data (Sect. 5.1) to anchor bias corrections for satellite radiances (e.g. Auligné et al., 2007). Versions of this approach have been implemented in ERA-Interim, JRA-55, MERRA, and MERRA-2. Global Navigation Satellite System radio occultation (GNSS-RO) observations (Sect. 5.3) are also useful for anchoring bias corrections (e.g. Poli et al., 2010) and have been used in this capacity in ERA-Interim, JRA-55, and MERRA-2; however, GNSS-RO data are only available after May 2001. The approach used in CFSR and CFSv2 (Derber and Wu, 1998; Saha et al., 2010) differs, in that anchor observations are not used to adjust the background state prior to the assimilation of satellite radiances. Instead, initial bias corrections are determined for each new satellite instrument via a 3-month spin-up assimilation and then allowed to evolve slowly. The effects of satellite-specific drifts and jumps are kept small by assigning very low weights to the most recent biases between the observed and expected radiances and by accounting for known historical variations in satellite performance as catalogued by multiple research centres. It is therefore not strictly necessary for satellite data to be homogenized prior to their assimilation in a reanalysis system, although it is beneficial for the system to assimilate data with biases that are as small as possible.

The use of externally homogenized satellite radiances has been found to improve some aspects of recent reanalyses. For example, homogenized MSU data (Zou et al., 2006) assimilated by CFSR, MERRA, and MERRA-2 have been found to improve temporal consistency in bias correction patterns (Rienecker et al., 2011) and may have helped MERRA to produce a more realistic stratospheric temperature response than ERA-Interim following the eruption of Mount Pinatubo (Simmons et al., 2014). Homogenized radiance data may be even more effective in eliminating artificial drifts and jumps in the analysis state in situations where conventional data are unavailable or insufficient to provide a reference for satellite bias correction, such as SSU in the middle and upper stratosphere. Homogenized satellite radiance time series only represent a relatively small fraction of the satellite data ingested by current reanalysis systems (many of which do not assimilate homogenized data at all); however, the availability of homogenized satellite radiance time series is increasing and these data are likely to become more influential in future reanalysis efforts.

Quality control checks for satellite radiance data vary by satellite sensor and reanalysis system, and also often differ for different spectral channels of the same sensor. Commonly used criteria include the exclusion of radiances affected by the presence of clouds or rain and of radiances measured over certain types of surfaces (e.g. land or ocean, snow/ice, high terrain). The information assimilated from a given sensor may also differ among reanalyses. For example, MERRA assimilates channels 1 through 15 from AMSU-A, while CFSR assimilates channels 1 through 13 and channel 15 (excluding channel 14); JRA-55 and MERRA-2 assimilate channels 4 through 14 (excluding channels 1 through 3 and channel 15); JRA-25 assimilates channels 4 through 13 (excluding channels 1 through 3, channel 14, and channel 15); and ERA-Interim assimilates channels 5 through 14 (excluding channels 1 through 4 and channel 15). Such differences are not unique to AMSU-A. Detailed discussion of the satellite radiance quality control criteria applied in modern reanalysis systems is beyond the scope of this paper; additional information can be found in the publications listed in Table 1.

In addition to the satellite sounding data discussed above, atmospheric motion vector (AMV) data derived from geostationary and polar-orbiting satellite images have relatively large influences on reanalysis fields in the upper troposphere and lower stratosphere, as do satellite ozone retrievals and other satellite-derived quantities. Figures 9 and 10 show timelines of AMVs and other non-radiance satellite data assimilated by ERA-Interim, JRA-55, MERRA, MERRA-2, and CFSR. AMVs derived from satellite imagery are assimilated by all of the full-input reanalysis systems, although the data sources and temporal coverage differ. All of the full-input systems except for R1 and R2 also assimilate satellite observations of ocean surface winds (scatterometers, SSM/I, and SSMIS), but again the data sources and temporal coverage vary substantially among different systems.

ERA-Interim, JRA-55, MERRA-2, and CFSR assimilate data from GNSS-RO instruments (also referred to in the literature as GPS-RO). These data are assimilated in the form of bending angles or refractivity at the tangent point, rather than temperature or water vapour retrievals. GNSS-RO occultations are based on radio waves that are calibrated against on-board atomic clocks and are therefore exceptionally stable both in time and across satellite platforms (Poli et al., 2010). The resulting retrievals have small random errors (equivalent to ∼ 1 K) and very small systematic errors (less than ∼ 0.2 K). GNSS-RO data can therefore be assimilated without bias corrections, thereby serving as anchors that facilitate the derivation of more accurate bias corrections for other data sources. These data are particularly useful for constraining the vertical structure of the lower and middle stratosphere (Bonavita, 2014; Polavarapu and Pulido, 2017). MERRA-2 also assimilates temperature retrievals from Aura MLS but only at pressures less than 5 hPa (Bosilovich et al., 2015). This choice produces discontinuities in some upper-level MERRA-2 products that coincide with the first assimilation of MLS temperature retrievals in August 2004 but provides valuable constraints for the reanalysis in the upper stratosphere and lower mesosphere (USLM) during the Aura mission.

All of the reanalysis systems assimilate satellite ozone measurements except for ERA-20C, R1, R2, 20CR (and JRA-55 and ERA-40 before 1978), and all except for R1 and R2 include some form of prognostic ozone parametrization and analysis. However, some of the reanalysis systems that assimilate ozone and produce ozone analyses (notably ERA-40, ERA-Interim, and ERA-20C) nevertheless use climatological ozone fields for radiation calculations in the forecast model, rather than their respective ozone analyses, because interactions between radiation and prognostic ozone have been found to amplify errors in the analysed temperature fields (Dethof and Hólm, 2004; Dee et al., 2011). Even amongst reanalyses that use climatological fields the characteristics of these fields may differ. For example, while ERA-40 and ERA-Interim use zonal mean monthly mean fields with a repeating annual cycle (Fortuin and Langematz, 1995), ERA-20C uses monthly three-dimensional fields that evolve in time (Cionni et al., 2011).

The prognostic ozone models used in the reanalysis systems are listed in Table 3. The ozone parametrization in ERA-40 is an update of the Cariolle and Déqué (1986) scheme as described by Dethof and Hólm (2004). Recent upgrades of the scheme (Cariolle and Teyssèdre, 2007) are included in ERA-Interim (Dragani, 2011) and ERA-20C. The scheme is a linearization of the ozone continuity equation, where the linear coefficients have been computed using an external 2-D photochemical model. This 2-D model does not include heterogeneous chemistry, but the parametrization includes an ad hoc ozone destruction term to account for the chemical loss due to polar stratospheric clouds. The ozone parametrizations in CFSR, CFSv2, and 20CR are based on the scheme proposed by McCormack et al. (2006). This scheme also assumes linear relaxation toward a zonal mean monthly mean photochemical equilibrium state, but with a different representation of chemical loss rates and independent derivations of the reference state and model parameters. MERRA and MERRA-2 diagnose ozone using a diurnally and height-varying empirical relationship between ozone and prognostic odd oxygen (Rienecker et al., 2008). This partitioning only applies at pressures greater than 1 hPa; at lower pressures (higher altitudes) all odd oxygen is assumed to be ozone. The odd-oxygen scheme includes tracer advection and climatological monthly mean zonal mean production and loss rates, with production rates tuned for agreement with satellite-based ozone climatologies. This scheme differs from the linearized relaxation schemes discussed above in that it specifies production and loss rates directly, rather than diagnosing them as functions of the deviation from a specified reference state. JRA-25 and JRA-55 (after 1979) use a fundamentally different approach, in which daily three-dimensional ozone concentrations are estimated using the MRI-CCM1 offline CTM (Shibata et al., 2005) and then nudged to satellite observations of total column ozone (TCO). The version of the CTM used for JRA-55 is slightly modified from that used for JRA-25, with 68 vertical levels rather than 45. JRA-55 does not include prognostic ozone variations before 1979. The treatment of ozone in JRA-55C matches that in JRA-55 exactly both before and after 1979.

Assimilated satellite observations of ozone vary widely among reanalysis systems that produce ozone analysis products (Fig. 10). For example, JRA-25 and JRA-55 only use retrievals of TCO for nudging daily offline CTM output toward observations, while ERA-40, ERA-Interim, MERRA, MERRA-2, and CFSR assimilate both ozone profiles and TCO. MERRA and CFSR only assimilate ozone retrievals (both TCO and profiles, with relatively coarse vertical resolution) from SBUV and SBUV/2, while ERA-40 assimilated coarse vertical resolution profiles from SBUV and TCO from TOMS starting in January 1979. MERRA-2 assimilates ozone retrievals from SBUV and SBUV/2 (both TCO and profiles) through September 2004 and assimilates OMI TCO and Aura MLS profiles thereafter. Estimates of TCO from SBUV and SBUV/2 assimilated by MERRA and MERRA-2 are calculated as the mass-weighted sum of layer values, rather than direct retrievals of TCO. ERA-Interim assimilates retrievals from a wide variety of satellite instruments, including some with relatively high vertical resolution. Ozone-sonde measurements are not assimilated in current reanalysis systems but are often used for validation.

The assimilation of radiosonde and satellite observations of humidity fields is problematic in the upper troposphere and above, where water vapour mixing ratios are very low. The impact of saturation means that humidity probability density functions are often highly non-Gaussian (Ingleby et al., 2013). These issues are particularly pronounced in the vicinity of the tropopause, where sharp temperature gradients complicate the calculation and application of bias corrections for humidity variables during the assimilation step. Reanalysis systems therefore often only assimilate observations of water vapour provided by radiosondes and/or microwave and infrared sounders (usually in the form of radiances) from the surface up to a specified upper bound, which is typically between ∼ 300 and ∼ 100 hPa. In regions of the atmosphere that lie above this upper bound (i.e. the uppermost troposphere and stratosphere), the water vapour field is typically determined by the forecast model alone. In this case, estimates of water vapour in the stratosphere are determined by some combination of transport from below, turbulent mixing, and dehydration in the vicinity of the tropical cold point tropopause (e.g. Fueglistaler et al., 2009).

The ECMWF reanalyses considered here do not allow adjustments to the water vapour field due to data assimilation in the stratosphere (above the diagnosed tropopause) but include simple parametrizations of methane oxidation (Dethof, 2003; Monge-Sanz et al., 2013). ERA-Interim and ERA-20C also include a parametrization that allows supersaturation with respect to ice in the cloud-free portions of grid cells with temperatures less than 250 K, which yields substantial increases in relative humidity in the upper troposphere and stratospheric polar cap in ERA-Interim relative to ERA-40 (Dee et al., 2011). JRA-25 and JRA-55 do not assimilate observations of humidity at pressures smaller than 100 hPa and set the vertical correlations of humidity background errors to zero at pressures smaller than 50 hPa (JRA-25) or 5 hPa (JRA-55) to prevent spurious analysis increments at higher levels. JRA-25 assumes a constant mixing ratio of 2.5 ppmv in the stratosphere for radiation calculations, while JRA-55 uses an annual mean climatology derived from HALOE and UARS MLS measurements made during 1991–1997. MERRA and MERRA-2 tightly constrain stratospheric water vapour to a specified profile, which is based on zonal mean monthly climatologies from HALOE and Aura MLS (Jiang et al., 2010; Rienecker et al., 2011). Water vapour above the tropopause does not undergo physically meaningful variations in MERRA or MERRA-2. Neither R1 nor R2 assimilates satellite humidity retrievals, and R1 does not provide analyses of moisture variables at pressures smaller than 300 hPa. CFSR only assimilates radiosonde humidities at pressures of 250 hPa and larger, although no upper altitude limit is assigned to assimilated GNSS-RO data. CFSR and 20CR provide moisture variables in the stratosphere, but dehydration processes in the tropopause layer may yield negative values. These negative values are artificially replaced by very small positive values for the radiation calculations but are not replaced in the analysis.

In general, reanalyses do not provide physically meaningful estimates of water vapour above the tropopause, although it should be noted that observational data sets used for comparison to the models have their own rather large biases in this region (Hegglin et al., 2013). Given the importance of water vapour in the upper troposphere and lower stratosphere (UTLS) for radiative forcing (e.g. de Forster and Shine, 2002; Randel et al., 2007; Gettelmann et al., 2011; Riese et al., 2012), large biases in the representation of the water vapour gradients across the tropopause and in the lower stratosphere may lead to non-negligible radiative and dynamical impacts in reanalysis systems. The magnitude of these impacts in the different reanalyses is not yet quantified but is under investigation within the S-RIP. Regardless, we emphasize in no uncertain terms that reanalysis humidity products in the upper troposphere and stratosphere should be used only with extreme caution.