5. 1. Anomalies in analysed temperatures
Figure 7 shows time series from 1979 to 2012 of monthly anomalies in global-mean temperature from ERA-Interim (solid lines) and ERA-40 (dotted lines) at a set of tropospheric and stratospheric levels. Shading denotes corresponding values from JRA-55. Anomalies are relative to 1979–2001, when ERA-Interim, ERA-40 and JRA-55 all overlap.
Figure 7. Anomalies in monthly and globally averaged temperatures (K) relative to 1979–2001 from ERA-Interim (black lines), ERA-40 (dotted) and JRA-55 (shading), at the indicated tropospheric and stratospheric levels. Common vertical scales are used for levels from 2 m to 100 hPa, from 70 to 10 hPa, and for 5 and 1 hPa.
Download figure to PowerPoint
ERA-Interim shows net warming over the period at tropospheric levels up to 300 hPa. Superimposed on this is substantial shorter-term variability, of which the peak associated with the 1997/1998 El Niño is the most prominent feature. The trend and variability increase coherently with height from the surface to 300 hPa, as discussed further for the Tropics in section 9. The 200 and 100 hPa near-tropopause levels are characterised by little trend but considerable variability. Short-term warming peaks associated with the volcanic eruptions of El Chichón in 1982 and Pinatubo in 1991 are evident in the lower to middle stratosphere, with cooling in steps around these peaks, and some subsequent warming. Cooling increases strongly with height in the upper stratosphere, where there are discontinuities in ERA-Interim associated with the initial introduction of SSU-3 data in mid-1979, the change from NOAA-7 to NOAA-9 SSU-3 data in early 1985, and the introduction of AMSU-A data in mid-1998. Several of the features shown in Figure 7 are familiar from other plots of time series of global temperature from reanalyses or direct analyses of data from specific types of instrument, as shown for example in the annual Bulletin of the American Meteorological Society articles on the state of the climate (e.g. Blunden and Arndt, 2013). Much of what follows in this article is devoted to establishing the reliance that can be placed on the details of what is provided by ERA-Interim and other datasets, and to identifying where future improvement is expected.
Consistency between ERA-Interim and its predecessor, ERA-40, varies quite considerably with height. Agreement between the two reanalyses is generally good at 2 m, in the middle to upper troposphere and in the lower to middle stratosphere. It is poorer in the lower troposphere, where ERA-40 suffered from problematic assimilation of data from several HIRS channels from 1989 to 1997 (Uppala et al., 2005). It is poorer also near the tropopause, particularly early in the period, and deteriorates above 30 hPa, where ERA-40 exhibits many more discontinuities associated with changes in satellite data.
JRA-55 is also very close to ERA-Interim at many levels, more so than ERA-40 in general. In the lower troposphere it exhibits larger short-term warming associated with the 1997/1998 El Niño than ERA-Interim, being closer to ERA-40 for this particular feature, and its lower-tropospheric warming trend over the period as a whole is somewhat larger than that of ERA-Interim. At 100 hPa, low-frequency variability is similar for JRA-55 and ERA-Interim, but there is poorer agreement for annual variability. Moreover, ERA-Interim shifts quite substantially to warmer values in late 2006. This is linked to the assimilation of GPSRO data that correct a cold bias at the tropical tropopause in ERA-Interim, as discussed below for the fit of ERA-Interim to radiosonde data and in section 6 when comparing ERA-Interim with MERRA. Conversely, JRA-55 is cooled a little in the tropical upper troposphere by assimilating GPSRO data, as this provides additional correction of a warm model bias that is insufficiently constrained by the observations assimilated in earlier years.
JRA-55 agrees especially well with ERA-Interim at 70 and 50 hPa, and also at 30 and 20 hPa until around the late 1990s, but does not warm to the extent that ERA-Interim does beyond then at the latter two levels. It generally cools more over time than ERA-Interim at 10 hPa. The use of full variational bias correction to the uppermost satellite sounding channels in JRA-55, something not done in ERA-Interim as discussed in section 5.4, avoids the discontinuities seen in ERA-Interim at 5 and 1 hPa. However, the relatively large cooling rates at these levels in the periods either side of discontinuities in ERA-Interim are of similar magnitude to those in JRA-55.
5. 2. Fits to radiosonde temperature data
Figure 8 shows times series of the differences between assimilated (bias-adjusted) radiosonde observations and corresponding values from the ERA-Interim analyses and background forecasts. Differences are evaluated at observation points during the assimilation process, and the plotted values are from a database of averages taken over all values for layers centred on standard pressure levels, without area weighting. Values are also shown (using shading) for the ERA-40 background forecasts. The vertical axes are chosen to display detail for ERA-Interim; the larger ERA-40 values often exceed the axis limits.
Figure 8. Monthly mean ERA-Interim observation-minus-analysis (black lines) and observation-minus-background (grey lines) differences for radiosonde temperatures (K) at the indicated tropospheric and stratospheric standard and nearby significant levels. Shading denotes the corresponding ERA-40 background differences. Averages are taken globally over all assimilated data. No area weighting is applied.
Download figure to PowerPoint
The ERA-Interim analyses fit the radiosonde data particularly closely in the lower and middle troposphere, especially for the first 20 years. The fit is to within about 0.1 K or better at 500 hPa and below for the period as a whole. The background forecasts at 500 and 700 hPa show almost no drift in boreal winter, but warm by 0.1–0.2 K in boreal summer. The distribution of ascents (Figure 3) is such that these averages are dominated by the contributions from the land masses of the Northern Hemisphere. Analyses and background forecasts are biased slightly cold relative to the radiosondes at 850 hPa. Later in the period, analysis and to a lesser extent background values warm slightly relative to the radiosondes up to around 2005, and then cool. The corresponding lower- and middle-tropospheric background fits for ERA-40 are much poorer prior to the 1997 change in HIRS usage, and generally vary more over time than the ERA-Interim fits.
A much more marked warming of analysis and background occurs in the upper troposphere, especially at 200 hPa, beginning in the late 1990s. This is associated with assimilating increasing amounts of warm-biased temperature data from commercial aircraft (Dee and Uppala, 2009). Implementation of a variational bias correction scheme for aircraft data in the operational ECMWF assimilation system in November 2011 (Isaksen et al., 2012) significantly improved the mean fits to both radiosonde and GPSRO data, and paves the way for better future reanalyses.
The impact of assimilating substantial amounts of GPSRO data from late 2006 onwards (Poli et al., 2010) is seen most clearly at 100 hPa. Here the fits to radiosonde data are quite stable in earlier years, with analyses biased cold relative to the assimilated radiosonde data and a drift to colder values of around 0.2 K in the course of the background forecasts. GPSRO provides a weight of data in later years that brings the analysis into much closer agreement with the radiosonde data.
Several features of the fits to radiosonde data in the lower to middle stratosphere merit comment. The first concerns the trend over the period. This is generally downwards, from values for analyses and background that are positive. This corresponds to analysis and background values that are colder than the radiosonde values, more so early than later in the period. The downward trend is much larger for the ERA-40 background than for ERA-Interim, although the two reanalyses themselves vary similarly in the lower stratosphere. This is consistent with ERA-Interim's use of radiosonde data that have been subject to much stronger homogenisation than those used in ERA-40. It is also consistent with Haimberger et al.'s (2012) finding that ERA-Interim cools less in the lower stratosphere than their RAOBCORE and RICH sets of homogenised radiosonde data, a result they ascribe in part to remaining uncorrected breaks in the radiosonde time series and in part to the warming late in the period in ERA-Interim from assimilating GPSRO data. Before then, the implication is that the background forecast and assimilated satellite data draw ERA-Interim away from too tightly fitting the homogenised radiosonde temperatures in the lower stratosphere.
Figure 8 also shows a spike in the fit to radiosonde temperatures from 20 to 200 hPa in 1979, and poorer lower stratospheric fits for periods immediately following the El Chichón and Pinatubo eruptions. The early spike occurs at a time of sharp shifts (illustrated later) in the bias adjustments made to radiances from the instruments flown on the TIROS-N satellite, which occur when data from a second satellite, NOAA-6, begin to be assimilated. The poorer fits following the volcanic eruptions are due to absence of the associated warming by stratospheric aerosols in the background model. As discussed by Dee and Uppala (2009) and illustrated later in this article, a fraction of the warming signal in the radiances is taken up by the variational bias adjustments of the data from channels sensitive to the lower to middle stratosphere. Much of the signal survives however: at 50 and 70 hPa the radiosonde–analysis difference increases by less than 10% of the analysed warming following the Pinatubo eruption. At 30 hPa, where there are fewer radiosonde data, the underestimation of warming is somewhat larger. ERA-40, which also lacked aerosol heating but used a more static radiance bias adjustment scheme, exhibits slightly stronger warmings than ERA-Interim.
Included in Figure 8 are the mean fits to all radiosonde temperature data received from 7 hPa and above. Data counts vary both within each year and over the longer term, ranging from below 2000 for several months between 1979 and 1985 to above 10 000 for several months from 2004 to 2007. The numbers are generally small enough for the observations to have little direct effect on global-mean analyses, even though the analyses may fit the data locally. Variations in spatial coverage and data quality also do not make for straightforward interpretation. It is striking nevertheless that the mean fits of background forecasts to these high-level radiosonde data improve substantially over time, becoming generally within 1 K or less over the past ten years or so. Root-mean-square fits also improve significantly. This most likely indicates beneficial assimilation of upper-level satellite data whose quality and quantity increase over time.
5. 3. Regional variations
Figure 9 shows regional-mean analysis and background fits to radiosonde data for 300, 500 and 850 hPa. Seasonal variations are more marked for such averages, and are suppressed here by applying 12-month running means so that longer-term variations can be seen more easily. Observation numbers for the Arctic and Antarctic are low, so variations over time for these regions must be regarded with caution. Nevertheless, the fits for the Arctic, where the largest lower-tropospheric warming occurs, are within 0.2 K throughout at 850 hPa and 500 hPa. There is generally little long-term trend in the fits for the other regions shown, especially for North America, Europe and Asia, although at 300 hPa the effect of assimilating increasing amounts of warm-biased aircraft data from 1999 onwards is evident for North America and to a lesser extent Europe. Many regions show a slight dip in the time series around 2005, as seen in Figure 8 for the global average. Overall, these results give confidence in the extent to which ERA-Interim captures the main temperature changes over the period, especially in the lower to middle troposphere over the land masses of the Northern Hemisphere that are relatively well covered by the radiosonde network.
Figure 9. Twelve-month running means of observation-minus-analysis (a), (c), (e) and observation-minus-background (b), (d), (f) differences for radiosonde temperatures (K) at (a), (b) the 300 hPa standard level and significant levels between 275 and 350 hPa, (c), (d) the 500 hPa standard level and significant levels between 450 and 600 hPa, and (e), (f) the 850 hPa standard level and significant levels between 775 and 887 hPa. Averages are taken without area weighting for all assimilated measurements from the Arctic (70°N–90°N), North America (10°N–70°N; 170°W–20°W), Europe (30°N–70°N; 20°W–60°E), Asia (Eq-70°N; 60°E–170°W), South America (50°S–10°N; 90°W–20°W), Africa (40°S–30°N; 20°W–60°E), Australasia (50°S–Eq; 110°E–170°W) and the Antarctic (90°S–70°S).
Download figure to PowerPoint
The analysed changes in decadal-mean temperature from 1981–1990 to 2001–2010 are shown in map form in Figure 10 for levels from 2 m to 100 hPa. There is general consistency in the patterns of change over land from the surface to the middle troposphere, with a decrease in the strength and extent of warming with increasing height, most evident in the fall-off in warming over northeast Canada and Greenland, and the expansion of the region of weak warming or cooling over Russia. In contrast, the band of weak near-surface warming or cooling stretching from western Canada to the south-eastern USA is overlain by stronger warming. More extensive warming over Australia at 850 hPa than at 2 m may be further evidence of the analysis problem for surface data surmised from the comparison with HadCRUT4 in section 4. The 850 hPa analysis over Antarctica does not show the marked localised warming and cooling seen at 2 m. It should be recalled that the latter is based on a separate analysis of synoptic surface observations not used in the upper-air analysis (Simmons et al., 2004).
Figure 10. Changes in decadal-mean temperature (K) from (1981–1990) to (2001–2010) at (a) 700 hPa, (c) 850 hPa and (e) 2 m, and at (b) 100 hPa, (d) 300 hPa and (f) 500 hPa. Grey shading denotes where the 850 and 700 hPa surfaces lie beneath the orography of the assimilating model.
Download figure to PowerPoint
Warming increases with height in the lower troposphere over sea, on average, in contrast to what happens over land. Changes over sea are nevertheless not as coherent in the vertical as over land between the surface, 850 and 700 hPa. This may be due in part to assimilating satellite data that compensate in the lower troposphere for the cooling shift in SST discussed earlier, but there is warming southwest of the USA, west of central Africa and between Australia and southern Africa that is quite pronounced at 850 hPa but not at either 2 m or 700 hPa. Warming over the high Arctic is much weaker at 850 hPa than 2 m, but a little stronger at 700 hPa than 850 hPa. Here it must be noted that the observational control brought about by assimilation of satellite radiance data is limited, and that MERRA (compared further in section 6) shows stronger warming at 850 hPa than ERA-Interim.
Warming is quite marked in the Tropics at 300 hPa, as discussed further in section 9. For this level the overestimation over North America by several tenths of a Kelvin from assimilating warm-biased aircraft data must be recalled. The extratropics at 100 hPa mostly show the cooling characteristic of the lower stratosphere, but warming still occurs over the North Polar cap. Values in the Tropics must be regarded with caution, as will be discussed further in the context of the comparisons with MERRA.
5. 4. Fits to radiance data from satellites
The assimilation of satellite radiance data exercises large-scale control on the temperature analyses. As discussed by Dee and Uppala (2009) and illustrated further below, ERA-Interim draws very closely to radiances that are fully adjusted for perceived biases, and growth of error in the background forecasts is generally slow. Key to the reliance that can be placed on long-term temperature variations from reanalysis is the reliance that can be placed on the bias adjustments made to radiance data. These adjustments are dependent on such factors as the coverage and homogenisation of anchoring radiosonde data, biases in the assimilating model or the fast radiative transfer calculations used to derive background equivalents of the measured radiances, major changes in observational coverage that cause background errors to shift, such as the introduction of GPSRO data, and the extent to which data have been subject to separate inter-satellite cross-calibration prior to assimilation.
The fits to bias-adjusted satellite data and the bias estimates are presented below for averages over all used data, or later for all used data from the tropical belt from 20°N to 20°S, without taking account of spatial variations in observation density. At its simplest, the density of soundings from polar orbit is highest near the Poles and lowest in the Tropics, with some 30% of the SSU soundings coming from poleward of 60° latitude. Use of data from lower-sounding channels depends on surface type and height, and on the presence of cloud in the case of infrared radiances and of precipitation or high liquid-water content in the case of microwave data. Thus about three times more data are used from the stratospheric HIRS-2 channel than from the tropospheric HIRS-5 channel, whilst about 50% more data are used for the stratospheric AMSU-A-10 channel than for the tropospheric AMSU-A-6 channel. With AMSU-A data available from as many as five satellites late in the period, these data are thinned prior to assimilation to avoid giving them excessive weight. This happens predominantly near the Poles.
Figure 11 shows time series of the global analysis fits for selected groups of channels from the various sounding instruments. Corresponding background fits are shown in Figure 12 and the bias estimates (the values that are subtracted from the unadjusted radiances) are shown in Figure 13. Archived monthly-mean values are plotted, and for each of these figures the value for the first month in which data from a particular satellite were used has been discarded, as it can be unrepresentative if based on only a few assimilation cycles while the variational bias scheme was determining the general level of adjustment. Also not shown are SSU-1 and SSU-2 values for NOAA-8 for a short period in 1985 when data reappear in the record, as data numbers were small for much of the time. Figure 1 provides the key that links the segments of plotted data with the satellites from which the data originated.
Figure 11. Monthly-mean observation-minus-analysis differences in brightness temperature (K) for selected channels from the SSU, HIRS, MSU, AMSU-A and AIRS sounding instruments. Observations have been adjusted for the estimated biases shown later in Figure 13. For the high-resolution AIRS data, results from channels in bands (20–40 and 201–221) from which data were assimilated have been amalgamated. The channels selected span from (a) the upper stratosphere to (d) the middle troposphere, as discussed further in the text.
Download figure to PowerPoint
Figure 13. As Figure 11, but for corresponding estimates of biases in the data from each channel or group of channels.
Download figure to PowerPoint
Data from only the two highest-sounding channels shown in Figure 11, SSU-3 and AMSU-A-14 (a), were not completely adjusted for bias. All radiances from these two channels received an adjustment for scan angle, but only the SSU-3 data from August 1998 (when AMSU-A-14 data started to be assimilated) received full adjustment. Before then, analysis equivalents of the SSU-3 brightness temperatures are mostly warmer than the observations, though by less than 0.2 K, and quite stable over time apart from instances prior to 1989 when data from new instruments begin to be assimilated. The analysis draws quite closely to the AMSU-A-14 radiances from NOAA-15 when it is the only satellite flying the instrument, and closer still to the fully adjusted SSU-3 radiances. Later, the analysis compromises between AMSU-A-14 brightness temperatures that differ by several tenths of a K from one instrument to another.
Figure 11(b) and (c) relate to channels that primarily sound the stratosphere, but for which full bias adjustment is applied. Here the fits in brightness temperature are mostly to within less than 0.01 K, and are particularly close for the microwave and AIRS instruments. Values are amalgamated from several neighbouring channels for AIRS, but fits for individual channels are similarly good. SSU-1 and SSU-2 fits are generally a little noisy, while those for HIRS-2 vary from instrument to instrument, being similar to the microwave instruments in some cases but much worse in others, especially for the NOAA-18 data assimilated from mid-2005 to early 2010. The standard deviation of the analysis fit for the HIRS-2 on NOAA-18 is larger than for any other satellite1 ; values for HIRS-2 vary from satellite to satellite from about 0.2 to 0.8 K, whereas corresponding values for SSU-1 lie mostly in the range from 0.4 to 0.6 K. Aside from varying levels of instrumental noise, it is known that the HIRS spectral response functions differ significantly from satellite to satellite and some of the functions specified from pre-launch measurement are subject to significant error (Cao et al., 2009; Shi and Bates, 2011). Revised spectral response functions are available in a new release (version 11: Saunders et al., 2013) of the RTTOV fast radiative transfer model. This provides a ready route to improvement of future reanalyses, as previous versions of RTTOV are currently used in both the ERA and the JRA assimilation systems.
Figure 11(d) shows analysis fits for a set of tropospheric sounding channels. Whilst still close, with model-equivalent brightness temperatures within 0.02 K of the bias-adjusted observations, these fits are nevertheless poorer than for the bias-adjusted stratospheric channels. Two features are of note. The first is the clear distinction between values for the microwave (MSU and AMSU-A) and infrared (HIRS and AIRS) instruments, with better fits for the infrared. This most likely reflects where data are assimilated for the two types of instrument, as the fits for the microwave instruments include cloudy areas where background temperature errors are likely to be larger than in the clear-sky regions where the data from both microwave and infrared instruments are assimilated. The second is the cooling and subsequent warming of the analysis relative to the observations that occurs for both infrared and microwave data starting around the year 2000. This upward bowing of the observation–analysis curves is opposite in sign to the bowing in the lower-tropospheric radiosonde fits seen in Figure 8. It comes primarily from the Tropics: the 20°N to 20°S mean fit for brightness temperature for the tropospheric sounding channels reaches a maximum around the beginning of 2007 that is about 0.03 K higher than typical pre-2000 values. It is also more pronounced in the corresponding fits of the background forecast presented in Figure 12. Further discussion is given in section 5.7.
Figure 12 shows that the background forecast cools relative to the tropospheric sounding radiances, although differences in brightness temperature are mainly below 0.05 K, as they are for the lower-sounding stratospheric channels. Differences increase with height in the stratosphere, with warm background biases predominant at higher levels. They are relatively large for SSU data between 1985 and 1998. The unadjusted data for this period from SSU-3 on the NOAA-9, -11 and -14 platforms are more inconsistent with the background model than either the SSU-3 data from earlier satellites or the AMSU-A-14 data from later ones. The background forecasts accordingly warm more between 1985 and 1998 in the upper stratosphere due to assimilation of incompatible SSU-3 data in a region where no direct control is provided by radiosonde data. The background fits to SSU-2, SSU-1 and HIRS-2 are affected by this to varying degrees, as the weighting functions for these channels extend into the upper stratosphere. Figure 12 shows that the background fit for HIRS-2 varies from satellite to satellite during this period. The satellites for which the fit is poorer are NOAA-9, -11 and -14. The HIRS-2 fit is thus poorer for the satellites that carry an SSU instrument than for those (NOAA-10 and -12) that do not. As NOAA-9, -11 and -14 flew in ‘afternoon’ orbits separated geographically from the ‘morning’ orbits of NOAA-10 and -12, and as the background forecast evolves over the 12 h assimilation window and has a significant large-scale tidal component in the upper stratosphere, the forecast is sampled differently by the HIRS instruments on one set of satellites than by those on the other set.
The change in character of the SSU-3 data near the start of 1985 can be linked with the higher mean pressures of the cells of carbon dioxide in the later instruments carried on NOAA-9, -11 and -14. This was not accounted for in the version of RTTOV used to compute background equivalents of the measured radiances in ERA-Interim. Kobayashi et al. (2009) identified the problem and showed that improvement can indeed be brought about by taking cell-pressure differences into account in the radiative transfer modelling.
Figure 12 also shows a slight drift in the observation–background differences for AMSU-A-10 and AMSU-A-12 late in the period. This indicates some contention between the background forecasts and drifts in the bias adjustments made over this period. These drifts are discussed later.
5. 5. Radiance bias adjustments
The estimates of bias used to make the adjustments to brightness temperatures are shown in Figure 13. They are generally much larger than the analysis and background differences discussed above, and the range of the adjustments applied is similar to or larger than the range of analysed climatic variations. One thus looks in the first instance for estimates that vary little over time and represent primarily the adjustments (or calibrations) needed to produce continuity of values from one satellite to another for a particular instrument and channel.
Oscillations and drifts of values are nevertheless to be expected. One source of longer-term variation is the changing solar heating of a satellite whose orbit drifts, as illustrated by Dee and Uppala (2009) in the case of the adjustment for MSU-2 on NOAA-14, the last of the MSU segments plotted (from 1995 to 2006) in Figure 13(d). Another is the use of fixed CO2 in modelling the measured infrared radiances. Insofar as other observations and the prescribed SSTs control the tropospheric warming trend, a downward drift in the bias estimates for tropospheric-sounding CO2-absorption channels such as HIRS-6 or the selected AIRS channels should be inferred by the variational bias adjustment scheme, because the decrease in observed brightness temperatures due to increasing CO2 (Chung and Soden, 2010) is not matched by the background equivalents derived using fixed CO2 in RTTOV. In practice, however, these infrared instruments provide data that are part of the mix that determines the trend, so use of fixed CO2 in this component of the data assimilation system is a factor that likely contributes to some underestimation of tropospheric warming.
Dee and Uppala (2009) identified drift of possibly instrumental origin in bias estimates for the tropospheric sounding channels of the early AMSU-A instruments, which can be seen in Figure 13(d). Conversely, there is no significant drift in the corresponding estimates for the later instruments on NOAA-18, Metop-A and NOAA-19. Lu and Bell (2014) argue that significant drift in measurements from channels 6–8 of the earlier AMSU-A instruments, but not from higher-sounding channels, arises from in-orbit changes in the frequencies that are sensed by the instruments, which are actively stabilised for channels 9–14 but not 6–8. Drift in bias estimates may also occur due to changes in the sampling of diurnally varying model bias as orbits drift, and cyclical shorter-term variations can arise from annual or quasi-biennial variation in systematic model bias.
Notwithstanding the difficulties in interpretation caused by such effects, a number of conclusions can be drawn from Figure 13. The scan-angle corrections applied to SSU-3 prior to August 1998 and to the AMSU-A-14 data thereafter are small, especially for AMSU. The ∼2 K bias estimate for SSU-3 brightness temperatures in the period of overlap with AMSU is very similar for the two satellites supplying SSU data at the time, NOAA-11 and NOAA-14. The bias adjustments applied to SSU-2 and AMSU-A-12 are mostly quite stable over time and less than 1 K in magnitude. One exception is the sharply increasing bias estimated for NOAA-7 from mid-1981 onwards, which is not shown fully due to the choice of axis limit but grows to almost 8 K by the time data cease to be assimilated early in 1985. This is consistent with the known high leakage of CO2 from the pressurized cell for this particular instrument and channel; much smaller shifts for other instruments may be due to leakage of water vapour (Nash and Saunders, 2013). The small downward trend from around 2003 onwards in the AMSU-A-12 bias estimates is discussed later in the context of differences between ERA-Interim and MERRA. The jump in the bias estimates early in 2007 for one AMSU instrument is due to recalibration by EUMETSAT of Metop-A radiances, which were assimilated from the beginning of 2007, prior to the data being formally declared operational by EUMETSAT.
The bias estimates for the channels that sound the lower to middle stratosphere are stable over time for much of the period apart from inter-satellite differences, which in the case of SSU-1 can again be linked to the cell-pressure differences discussed by Kobayashi et al. (2009). Variations do occur from time to time that are common to all types of instrument and thus indicative of issues in the data assimilation. Bias estimates for SSU-1, HIRS-2 and MSU-4 all rise in the second half of 1991 and then fall back by mid-1993 due to the absence of background-model warming from the Pinatubo eruption. The estimates for SSU-1 and HIRS-2 (and SSU-2) drop in 1998 when full bias correction is applied to SSU-3, but eventually recover. Those for HIRS-2, AMSU-A-10 and AIRS-20…40 shift to accommodate changed bias in the background forecast due to the assimilation of GPSRO data, mostly over about two years from late 2006 when the amount of occultation data increases substantially.
Particular discussion is needed for the initial period up to mid-1979 when TIROS-N was the only source of sounding data used in ERA-Interim. The bias adjustment to the SSU-2 data from this satellite shifts substantially when data subsequently become available from NOAA-6, making the adjustments for the two instruments almost identical. At the same time, shifts occur in the bias estimates shown for SSU-1, MSU-2 and MSU-4, and radiosonde data fits for the lower stratosphere deteriorate as shown earlier in Figure 8. These changes may be related to limitations in the data provided by TIROS-N: no data were assimilated from its SSU-3, HIRS-4 and HIRS-5 channels due to instrument problems. Absence of SSU-3 data from TIROS-N means that one of the key anchors of the variational bias-adjustment scheme was not operating for the first six or so months of 1979. This likely explains the shift in bias adjustment of the stratospheric-sounding channels and poorer fits to radiosonde data. For the troposphere, the number of assimilated soundings increased substantially when NOAA-6 started providing HIRS-4 and -5 data as well as additional data for the channels for which TIROS-N provided data. This change appears to have been sufficient to shift the bias adjustment of the TIROS-N MSU-2 data. Shifts also occurred at the same time in the adjustments applied to the lower-sounding TIROS-N channels not included in Figure 13. Although 1979 is often taken as the starting point for calculating trends over the modern satellite era, a start at least one year later would be appropriate given the deficiencies in TIROS-N data.
The plots discussed above are from 1989 onwards for the main ERA-Interim production stream and from 1979 to 1988 for the second production stream. It is nevertheless hard to discern in them any mismatches from 1988 to 1989 in the plotted data fits and bias estimates. The second production stream was in fact continued until the end of 1989, and Figure 14 presents examples of overlaps, showing bias estimates for HIRS-2 on NOAA-10, and SSU-1 and MSU-4 on NOAA-11. Differences are very small indeed by the end of 1989, but more evident in the first few months of the year. The main production stream started at the beginning of December 1988; Figure 14 and other evidence suggest that a warm-up period of three or more months would have been preferable. It is clear nevertheless that the variational bias adjustment scheme is robust to a change in starting point, which is important in practice as reanalyses are time-consuming to produce and typically executed in a number of parallel streams in order to speed up production.
Figure 14. Examples of bias estimates from the original production stream (solid lines) from 1989 onwards and from the later production stream (dotted lines) for 1979–1988, which was extended through 1989 to study overlap.
Download figure to PowerPoint
5. 6. Bias-adjusted stratospheric radiance data
It is instructive also to examine the time series of the bias-adjusted brightness temperatures themselves. Figure 15 displays HIRS-2, SSU-1 and AMSU-A-11 values, using a 12-month running mean to remove pronounced seasonal cycles. It must be recalled that these plots are averages over all soundings, not estimates of true global-average temperatures. Also, the three channels have different weighting functions, though the values for HIRS-2 and AMSU-A-11 differ by only around 1 K. To ease comparison, 5.5 K is subtracted from the values for SSU-1 as this channel is more sensitive to the warmer temperatures found higher in the stratosphere. Values also may absorb some bias from the background forecasts or the radiative transfer used to compute background equivalents of the measured radiances.
Figure 15. Twelve-month running means of bias-adjusted brightness temperatures (K) from HIRS-2 (black), SSU-1 (dark grey) and AMSU-A-11 (light grey). Averages are over all soundings, and 5.5 K is subtracted from SSU-1 values.
Download figure to PowerPoint
Nevertheless, the variations over time seen in Figure 15 can be clearly identified in the height-resolved reanalysis values shown as monthly anomalies in Figure 7 (and later as 12-month running means in Figure 17). They are also similar for the three types of instrument. There is, however, a quite distinct mismatch of the absolute values for HIRS-2 for some instruments, with values from NOAA-11 (which come in two segments, one in the early 1990s and one late in the decade) lying above those for NOAA-10, NOAA-12 and later satellites. The mismatch can occur because channel-2 spectral response functions differ from one HIRS instrument to another, but may include an effect of bias adjustments that compensate for poorly specifying the functions in the radiative transfer modelling.
5. 7. Interpretation of the variation in lower-tropospheric data fits
Figure 16(a) shows time series of 12-month running means of analysis increments, the analysis–background differences, for temperature at 850, 700 and 500 hPa. The increments are calculated from complete model fields and are true global averages. The background is the average of forecasts at 6 h and 12 h range made twice-daily from 0000 and 1200 UTC. The data assimilation generally increases background model temperatures that are perceived to be biased cold. It has already been noted that this is generally the case for the near-surface over land, and that the background is biased warm in the mid-troposphere relative to overwhelmingly land-based radiosondes, but biased cold relative to the globally sampling tropospheric soundings from satellite. The warming increment reaches a maximum around 2005 and shows a bowed form similar to that seen in the analysis and background fits to satellite data, and thus opposite in sign to the otherwise similar feature seen in the radiosonde fits.
Figure 16. Twelve-month running means of (a) the global-mean ERA-Interim temperature (K) analysis increment at 850, 700 and 500 hPa, (b) the contributions to the global-mean atmospheric humidity budget (mm day−1) of ERA-Interim from the difference between precipitation and evaporation averaged over land and from the difference between evaporation and precipitation averaged over sea (based on values from twice-daily 12 h forecasts), and (c) precipitation (mm day−1) over land from ERA-Interim and from the GPCC Full Data Reanalysis version 6.0 up to the end of 2010 and the GPCC Monitoring Product for 2011 and 2012; ERA-Interim values are reduced by 0.2 mm day−1 for clearer comparison with GPCC.
Download figure to PowerPoint
Figure 17. Twelve-month running mean anomalies in global-mean temperature (K) from ERA-Interim (solid), MERRA (dotted) and ERA-40 (shaded) at selected levels. Anomalies are calculated in each case relative to the ERA-Interim monthly means for 1981–2010. The global averages for 700 hPa exclude points for which MERRA does not provide values due to the presence of orography. ERA-40 results are not shown at 10 and 20 hPa as differences from ERA-Interim and MERRA are large at these levels, and the change in axis needed to display them would detract from the differences between ERA-Interim and MERRA.
Download figure to PowerPoint
Evidence related to the increasing and then decreasing cold bias in background forecasts over sea is presented in the Figure 16(b). It shows 12-month running means of the contributions to the global-mean atmospheric humidity budget from the differences between precipitation and evaporation averaged over land and between evaporation and precipitation averaged over sea, from the 12 h forecasts (which are not generally in hydrological balance). The values for land are quite stable over time, but the excess of evaporation over precipitation increases over sea from 1991, reaching a broad maximum between 1999 and 2006, followed by a quite rapid subsequent decrease. This variation over time is comparable with that of the counts of 1D-Var retrievals from rain-affected SSMI radiances shown in Figure 2. The early version of the scheme for assimilating rain-affected radiances used in ERA-Interim had a pronounced and erroneous drying effect (Geer et al., 2008; Dee et al., 2011a), with a consequent reduction in latent heating in background forecasts and likely additional radiative effects, particularly in the Tropics. Background temperatures also shift over sea because of the 2001 shift to cooler SST analyses.
From the above it can be inferred that although the background forecast model is held fixed in ERA-Interim, the biases in background temperature change due to changes in the humidity analysis and the externally provided SST analyses. Background cooling over sea is counteracted by warming analysis increments, primarily from satellite radiances. A data assimilation system in which there are systematic warming increments can be viewed as a free-running model in which an additional heating is added to represent the systematic effects of the increments, with implications for the circulation that maintains balance (Uppala et al., 2005; Fueglistaler et al., 2009). In the case of warming tropospheric increments over sea, a balancing thermally driven circulation would be set up, with descent and thus warming of the background forecasts over land. This would in turn be partly compensated by cooling increments from assimilating radiosonde data and a reduction in the warming increment from analysing synoptic screen-level temperatures. The converse happens if the background forecasts warm over sea. This is a plausible explanation for the difference in sign of the bowing seen in the data-fit time series for radiosondes and satellite radiances, and the dip centred around 2005 in the increments shown in Figure 4 for the near-surface analysis.
Precipitation would be expected to decrease over land in ERA-Interim when moisture supply from the ocean is reduced by the drying of marine air that results from assimilating rain-affected radiances and from shifting to cooler SSTs that give less evaporation. Additional suppression of rainfall would be expected due to the descent over land forced by systematic warming increments over sea. Precipitation over land would increase when drying of the marine atmosphere is replaced by moistening.
This is indeed what is found. Figure 16(c) compares time series of 12-month running means of precipitation over land from ERA-Interim and from the Global Precipitation Climatology Centre (GPCC: Becker et al., 2013). To counter inhomogeneity over time in the distribution of gauge data available for analysis by GPCC, the land-average for both ERA-Interim and GPCC has been made by averaging only over the 2938 1° grid squares for which GPCC had access to data from at least one station for every month of the period. The resulting time series for GPCC shows a slight trend for precipitation to increase over time and maxima that are presumably related to increased moisture supply from the ocean associated with the El Niño events of 1982–1983, 1997–1998 and 2009–2010. The successive 12 h ERA-Interim forecasts produce more precipitation than analysed by GPCC, but the discrepancy varies over time. It is about 0.3 mm day−1 on average for the first decade or more before declining slowly at first and then more quickly to reach about 0.2 mm day−1 from around 2004 to 2007. It then increases slightly. Notwithstanding this, ERA-Interim reproduces the shorter-term variability in the GPCC mean very closely, in line with the agreement reported for regional averages by Simmons et al. (2010). Implications for soil-moisture trends are discussed by Albergel et al. (2013).