1. Introduction
There have been several recent reanalysis projects producing data over decadal periods using modern weather forecast analysis systems. They provide us with the best available four-dimensional, homogeneous datasets for studying the dynamics of the atmosphere, for short-term climate studies of the current climate, and for possible validation of climate general circulation models (GCMs). Since each reanalysis uses a fixed assimilation–model combination throughout, though the assimilation–model combinations differ among the reanalyses, they are free of secular biases often introduced into operational systems with the introduction of new assimilation methods or changes to the model formulation. Changes in the observational network is a separate issue. It is to be expected that the individual techniques and methodologies will leave a distinct imprint on the resulting reanalysis products. This raises the issue of whether these imprints are negligible or are misleading artifacts of the true atmospheric state.
If we are to have confidence in the use of reanalysis data for the above-mentioned studies, it is of interest to see how they intercompare and to identify any discrepancies and their causes. This has been done to some degree for large-scale aspects of the data (Higgins et al. 1996; Trenberth et al. 2001). However, it is also important to examine any discrepancies at the smaller spatial and temporal scales represented by synoptic-scale weather systems to provide some quantification of the uncertainty in their representation. This is particularly important if we wish to study changes in storminess or extreme events, for example. It is also important if we wish to use reanalyses for validating the synoptic behavior of a GCM in intercomparison projects such as the second Atmospheric Model Intercomparison Project (AMIP II; Gates 1992).
In a recent study of the Northern Hemisphere (NH) wintertime storm tracks, Hoskins and Hodges (2002, hereafter HH) performed an analysis of the synoptic-scale activity using the first European Centre for Medium-Range Weather Forecasts (ECMWF) reanalysis data. A wider range of fields than is normally used at several tropospheric levels was explored using the system-centered approach of identifying synoptic features, tracking them, and generating a wide range of statistical diagnostics. The results presented in HH will form the basis of the comparisons for the storm tracks presented in this paper with the analysis extended to other reanalysis datasets. The tracking technique has also recently been used by Thorncroft and Hodges (2001) to diagnose the nature of African easterly waves over west Africa and the tropical Atlantic. This analysis will be extended here to include easterly wave (EW) activity throughout the whole Tropics for the NH summer period May–October (MJJASO).
The reanalyses studied in this paper are the ECMWF 15-yr Reanalysis (ERA15; Gibson et al. 1997); the National Centers for Environmental Prediction–National Center for Atmospheric Research (NCEP–NCAR) 40-yr Reanalysis (Kalnay et al. 1996) limited to the period 1979–96; a repeat of the NCEP–NCAR reanalysis by the U.S. Department of Energy (NCEP–DOE; Kanamitsu et al. 1999), which fixed some of the problems in the first reanalysis and updated the physical parameterizations; and finally the 14-yr reanalysis of the Data Assimilation Office (DAO) of the National Aeronautics and Space Administration (NASA) Goddard Space Flight Center, the Goddard Earth Observing System first system (GEOS-1; Schubert et al. 1993). Additionally, we will also make use of some of the ECMWF operational analyses for the period 1994–2001 and some preliminary data for several years from the ERA40 reanalysis, which is a repeat reanalysis performed by ECMWF with a more sophisticated data assimilation system.
2. Reanalyses background and data description
The objective of an operational meteorological analysis system is to produce a sufficiently accurate atmospheric state, which is also in balance for model specifications, from which a forecast can be made. To accomplish this the system synthesizes a short-term (typically 6 h) model forecast with all the observations valid for the analysis time window. The analysis determines how the scattered and disparate observations are combined with the model forecast to produce an atmospheric state strongly constrained by the observations but acceptable for a model initial state. The quality and quantity of the observations can be highly inhomogeneous in space and time, but the assimilation system uses the model to propagate information previously assimilated into the data-sparse regions in a dynamically and physically consistent manner. For some regions of the globe where the observations are infrequent or widely spread, it is to be expected that the model and specifications of boundary forcing will impose a signature, characteristic of their biases, on the final analysis product. In addition, even in data-rich regions bias can be imposed on an analysis by the biases in the data. A brief description of each of the systems used to produced the data for this study is given next. A summary of the configuration of the differing systems is given in Table 1.
The ERA15 dataset was produced by ECMWF for the period 1979–93 based on the operational scheme current at the end of 1995 (apart from the horizontal resolution). Full details of the analysis system used and the data production for ERA15 are available in Gibson et al. (1997). The salient features are the use of a spectral model with T106 truncation and 31 hybrid levels in the vertical together with an intermittent optimal interpolation (OI) assimilation scheme. The OI technique employs modeled spatial and temporal covariances and statistical relations to interpolate the observations to the analysis grid. Although optimal in a statistical sense, the technique can result in the introduction of discontinuities between separate analysis times, particularly at the subplanetary scales typical of synoptic systems. The study reported here has also used the ECMWF operational analyses for 1994–2001 to compare, at least over a limited time period, the effects of improvements in the analysis system used by ECMWF with the introduction of 3D and 4D variational data assimilation (3DVAR and 4DVAR) as well as the impact of the increased horizontal and vertical resolution. Other changes to the operational system that could have an impact in this study include the change to 12-hourly assimilation cycles (since September 2000) with the intermediate 6 h filled with the forecast. Further details of the operational system are given in ECMWF (1999). Variational schemes provide a more powerful means to enforce both temporal and spatial consistency by using the forecast model equations and underlying model physics to constrain the assimilation of observations. However, the full model physics are often simplified to prevent convergence problems during the minimization process (Mahfouf 1999). These adiabatic linear models, while suitable at midlatitudes due to the dominanace of the dynamics, are less valid in the Tropics and work is being carried out to rectify this situation (Mahfouf 1999). For variational schemes spatial consistency is enforced such that surface observations can affect changes in the entire atmospheric column, whereas with OI the observations at various levels could be decoupled.
The NCEP reanalyses provide an interesting contrast to the ERA15 reanalysis, and include both the original NCEP–NCAR reanalysis (Kalnay et al. 1996) and the NCEP–DOE reanalysis (Kanamitsu et al. 1999). The model is also spectral but is integrated at the lower resolution of T62 with 28 levels in the vertical. However, a 3DVAR scheme is used throughout. The original period covered by the NCEP–NCAR reanalysis is 1957–96, but in this study we only use the data from the period 1979–96 to exclude the major data inhomogeneity resulting from the introduction of satellite data and to restrict the study period to one similar to the other reanalyses. The NCEP–DOE reanalysis is a repeat of the NCEP–NCAR reanalysis, initially for the period 1979–95, using the same system but correcting several analysis problems and updating some of the physical parameterizations in the model. One problem involved the use of Australian bogus pressure data called PAOBs. These data are estimates of sea level pressure produced by operational meteorologists at the Australian Bureau of Meteorology from satellite data. They are used to augment the observational database over the data-poor Southern Ocean. Some PAOB locations were misnavigated by 180° of longitude in the original NCEP–NCAR reanalysis. Additionally, in the NCEP–NCAR reanalysis the snow cover analysis, the humidity diffusion, the oceanic albedo, the cloud tuning parameters, and the snowmelt were incorrectly specified for the whole length of the period (Kanamitsu et al. 1999). These problems were all corrected for NCEP–DOE. Updates in NCEP–DOE to the model parameterizations included the planetary boundary layer, shortwave radiation, and clouds. In addition, computation of soil wetness did not use the model precipitation, but rather the estimated values from Xie and Arkin (1997).
The GEOS-1 reanalysis (Schubert et al. 1993) is of interest because it uses a different GCM–assimilation system combination. The GCM is a finite-difference model on an Arakawa-C grid with a horizontal and vertical resolution of 2° × 2.5° and 20 levels, respectively. The analysis system uses a form of OI called incremental update assimilation (IUA), which introduces information from the observations at every model time step. This is done by performing an analysis from the original forecast after which the model is rerun with the integration being nudged toward the analyzed field every time step. This is effective in enforcing temporal consistency, particularly when assimilating asynoptic data.
Even over the 20-yr period being considered here the observing system is not constant with the introduction of new types of observations and a decrease in the availability of some others. The quality of the data may also vary with the introduction of more accurate instruments and according to how the data are treated: for example, some data require bias correction. This raises the interesting question of the sensitivity of the reanalyses to the observing system. So far only a few limited studies have been performed, for example, by Bengtsson and Kållberg (1981).
3. Analysis methodology
The analysis method used in this study identifies synoptic-scale features, tracks them, and derives statistical diagnostics from the track ensembles. This approach directly yields information on the propagation and evolution of atmospheric features more clearly than from the more traditionally used Eulerian methods such as bandpass-filtered variances. The tracking information also permits a direct assessment of the assimilation methods with respect to the coherence of the propagation of weather systems. This form of analysis has its origins in the early manual studies based on daily synoptic charts (Klein 1957; Whittaker and Horn 1984). While it might be interesting to compare automated results with these manual studies, in practice this can at best only be done qualitatively because the manual studies do not cover the same period as our analysis period, are only based on mean sea level pressure (MSLP), and use different methodologies for the statistical estimation. In addition, the manual tracking depends to some extent on subjective judgement, although this can be useful in complex synoptic situations.
The tracking analysis system used here is that of Hodges (1996, 1999), which performs the tracking and statistical analysis directly on the unit sphere. Working on the sphere reduces the likelihood of introducing bias into the tracking and statistics and makes the parameters required for the tracking and statistical estimation independent of any projection. All the processing is performed using 6-hourly data at a spectral resolution of T42 with the spectral coefficients being tapered to reduce the Gibbs effect (Hoskins and Sardeshmukh 1984). This resolution is sufficient for the identification of synoptic features as extrema and provides some smoothing of the noisier fields such as vorticity. It also allows the same spatial scale of a feature to be identified in each reanalysis. While this smoothing of the data may make weaker systems harder to identify, particularly in fields such as MSLP, this is partially offset by the planetary wave filtering described next.
For the analysis of the midlatitude storm track activity the planetary scales are first removed for total wavenumbers n ≤ 5 at each time step. This approach was used in HH and is discussed in more detail in Anderson et al. (2003). This is essential for the identification of synoptic systems in those fields that have large meridional gradients at the planetary scale, which can mask the smaller-scale synoptic features. It is also useful for other fields such as MSLP, which can be dominated by large planetary scales. For some other smaller-scale fields, for example, relative vorticity, it has little effect (Anderson et al. 2003).
Although a wide range of fields has been explored, the discussion will concentrate on the relative vorticity in the lower and upper troposphere and will be contrasted with results obtained using the MSLP field. Only anomaly signs relevant to cyclonic activity are discussed, that is, positive (negative) for the NH (SH) vorticity and negative for MSLP. Relative vorticity is the preferred field because smaller-scale systems such as typified by Mediterranean storms are more easily identified, it is less influenced by the large-scale background flow, and it does not depend to any large extent on below-ground extrapolation techniques for the levels considered here. In a geostrophic sense, vorticity focuses on the small-spatial-scale end of the synoptic range so that many more systems are identified than in the MSLP, for example, which focuses on the large-scale end. It is important that the vorticity be computed in a consistent way before removing the planetary scales. When there are missing data in the lower troposphere due to the intersection of pressure levels with the orography, these must first be interpolated or extrapolated. This is best done by first interpolating or extrapolating the wind fields and then computing the vorticity. Interpolating the vorticity field itself can result in spurious planetary wave structures due to high values of vorticity near steep orography such as around the Antarctic plateau. The EW activity is also diagnosed at T42 spatial resolution but without the removal of the planetary waves.
The system trajectories are determined for each seasonal period and then filtered to remove short-lived (less than 2-days lifetime) and semistationary (total displacement distance less than 10° as measured along a great circle) systems. The combined seasonal ensembles are then used for the statistical estimation. Although a wide range of statistics are computed, the discussion of the storm tracks will focus on the track density and mean system intensity distributions for the NH and SH winters, December–February (DJF) and June–July (JJA), respectively. For the EW only the track density is considered for the NH MJJASO period. The track density indicates the spatial distribution of systems while the mean intensity indicates the mean strength of systems in different regions. For the presentation of the track density statistics, the track density is scaled to number density per month per unit area as in HH, where the unit area is equivalent to a 5° spherical cap (∼106 K m2).
4. Storm tracks in the lower troposphere
In this section we present results comparing the various reanalyses for the NH and SH storm tracks based on relative vorticity at 850 hPA (ξ850) and MSLP. The comparison will be made on the basis of the spatial statistics and, in addition, by a direct track-by-track comparison of the track ensembles. Throughout, we will arbitrarily use the results from the ERA15 reanalysis as our base dataset. This does not affect any of the deductions made from our results.
a. NH comparison of reanalyses
1) The ξ850 and MSLP climatologies
The climatological track density and mean intensity statistics for the NH, ξ850, are shown in Figs. 1a,c, respectively, for ERA15. These differ little from the results shown in HH for the combined ERA15 and operational analyses. The track density shows the two oceanic storm tracks as a region of high density stretching across both ocean basins with the Atlantic storm track showing evidence of continuance into western Europe. The Pacific storm track extends from east Asia coming to an abrupt end on the upslope side of the coastal mountains of British Columbia: lysis density and mean decay rates (not shown) indicate strong cyclolysis in the Gulf of Alaska–Vancouver region. Further analysis by HH showed the Pacific storm track to be the result of the contributions from two source regions. One group of storms is initiated in east Asia, tracks to the northeast, and mostly occludes and decays in the central North Pacific region. The second group has its origin in a secondary development region in the mid-Pacific with cyclones occluding in the upstream Gulf of Alaska region. Also apparent are the Siberian and Mediterranean candidates for storm tracks also discussed by HH. Cyclones in these regions are generally smaller scale and weaker than their oceanic counterparts as indicated by the mean intensity (Fig. 1c). Also apparent is the high level of activity over the continent of North America associated with strong genesis and growth rates (see HH) in the lee of the Rockies. The MSLP results shown in Figs. 1b,d are consistent with this ξ850 picture of the cyclone activity in the NH but with lower levels of activity, particularly in the Mediterranean, Siberia, and the lee of the Rockies, reflecting the larger-scale nature of the features identified in the MSLP field and the smaller scale of the features found in these regions. For further discussion of the cyclone distributions for these two fields, refer to HH.
2) Comparisons of reanalyses for ξ850 and MSLP
Figure 2 shows the difference in the track densities for cyclonic ξ850 features between ERA15 and the other reanalyses, as well as between NCEP–NCAR and NCEP–DOE reanalyses. The differences in the density of storms in the NH are quite small, typically less than 2–4 month−1 (per unit area). In the two main storm track regions, the NCEP–NCAR and NCEP–DOE reanalyses are generally lower in activity than in ERA15 except for some regions near 30°N such as the Sargasso Sea and the end of the Mediterranean storm track (Iran, Afghanistan). These regions are generally low in track density so that small differences may carry less confidence compared to the main storm track, high-density regions. For the GEOS-1 reanalysis there appears to be less in the way of a systematic difference from ERA15, except near 30°N on the flanks of the storm tracks, particularly in the Pacific. Also apparent for GEOS-1 is a large difference in the track density in ξ850 in the lee of the Rockies, again related to small-scale systems that are generated there. This may be related to differences in the orographic representation between ERA15 and GEOS-1, which are mentioned below. The MSLP (not shown) shows little of these differences indicating that they are mostly due to weak (cf. Fig. 1d), small-scale systems.
The differences between the mean intensities for ξ850 are shown in Fig. 3. The overall impression is that ERA15 has systems that are more intense in the ocean basins and weaker around the major orography than the other reanalyses. The differences in intensities in the ocean basins are most pronounced for NCEP–NCAR and particularly for NCEP–DOE (both are integrated at a lower resolution than ERA15 with the latter showing a systematic weakening of systems relative to NCEP–NCAR. However, differences among all the reanalyses are relatively small, being mostly less than 0.5 × 10−5 s−1 (compared with typical mean intensities of 5–7 × 10−5 s−1 relative to the removed background). For GEOS-1 the difference in intensity in the ocean basins shows less systematic differences from ERA15, perhaps reflecting the nominally similar horizontal integration resolution of GEOS-1 and ERA15. Not all the differences can be attributable to resolution as is indicated by the comparison between NCEP–NCAR and NCEP–DOE. The differences around the major orography between ERA15 and the other reanalyses are also unlikely to be related to resolution and could be attributable to the use of different extrapolation methods below orography. However, it should be noted that ERA15 uses the subgrid-scale orographic parameterization of Lott and Miller (1997), which includes a more realistic representation of the mountain drag and produces a more realistic distribution of precipitation in regions of high orography than the old envelop orography scheme. The fact that the other reanalyses, which all use a form of mean or smoothed orography, are systematically stronger than ERA15 suggests that the orographic parameterization in ERA15 may be the reason for these differences, rather than the use of different methods of extrapolation.
The differences between the reanalyses for MSLP in the NH show relatively smaller and much less systematic differences in the track density (not shown). However, there are several regions where there are differences between ERA15 and the other reanalyses. These are the Icelandic region, the Denmark Straits, the Gulf of Alaska, Cape Hatteras, the eastern Mediterranean, and the lee of the Rockies. These are all regions of cyclogenesis or cyclolysis. It will be shown later that the majority of the differences between the track ensembles occur for the weakest systems and these differences in regions of genesis or lysis reflect the fact that the weaker systems have a tendency to be shorter lived than more intense systems.
The mean intensity differences for MSLP (not shown) are consistent with those for ξ850 in that ERA15 appears stronger in terms of the mean intensities in the main NH ocean basins and with NCEP–NCAR being systematically stronger than NCEP–DOE. The consistency of the comparison for ξ850 and MSLP between NCEP–NCAR and NCEP–DOE gives some confidence in the suggestion that the changes made to some of the physical parameterizations have had an effect on the intensities of the storms. For GEOS-1, whereas the Atlantic does show ERA15 to have more intense systems, in the Pacific this is less clear. This was also the case to some extent for ξ850. MSLP shows little sign of the differences around the orography seen in ξ850, indicating that these differences are mostly attributable to smaller-scale systems that are not identified in the MSLP field.
3) Direct comparison of track ensembles
Since the reanalyses should be simulating the same meteorological events for the same analysis period, an alternative comparison between the reanalyses, which is independent of the spatial statistical fields already considered, is to directly compare the track ensembles. To do this the data periods must be identical for any pair of reanalyses, so only the winters for 1979–93 are considered for the ERA15–NCEP comparisons, 1980–93 for the ERA15–GEOS-1 comparison. Two approaches are described. The first calculates the distribution of some attribute of the storms: here the mean intensity of each storm is used. These are binned and the distributions compared for the different reanalyses.
In the second approach a direct comparison in both space and time is performed of the individual members of the track ensembles. The study of Blender (2000), which compared cyclone tracking at different space and time resolutions using a combined spatiotemporal, measure is an elegant example of this methodology. Here, a simpler approach is considered using separate temporal and spatial measures, since the periods, spatial resolution, and temporal sampling at which the tracking analysis is performed are identical. Each track in the first ensemble is compared with those in the second by first finding the tracks in the second ensemble that overlap in time with the track in the first ensemble. If the number of points that overlap is greater than or equal to 60% of the number of points in the tracks, this is considered a possible good match in time. The percentage is computed as χ = 100[2nm/(n1 + n2)], where nm is the number of points that match in time, with n1 and n2 the number of points in the track from ensembles 1 and 2, respectively. For the tracks that satisfy the constraint in time, the mean separation on the unit sphere is computed from those points that match using the geodesic distance measure, Dm = 1/nm
Figures 4a,b show the results for the distributions of system mean intensity, based on the first direct comparison method, for each reanalysis for ξ850 and MSLP, respectively. This shows that all four reanalyses produce very similar distributions. For ξ850 the peak of the distribution occurs between 2.0 and 4.0 × 10−5 s−1 and then tails off to a maximum value of about 10.0 × 10−5 s−1. In general, ERA15 has two–four fewer systems per month near the peak in the distribution. This difference is not a large difference and may not be significant. In the high-intensity tails of the distributions, again differences are small, although NCEP–NCAR and NCEP–DOE appear to be consistently lower than ERA15 and GEOS-1. In fact, if we compare these distributions with those in the SH for ξ850 (Fig. 4c), this behavior is seen to be very similar but more pronounced. This suggests a systematic difference driven by the model and physics since it is less prevalent in the NH where the greater number of observations provide more constraint on the model. This could be tested by considering AMIP-type integrations of the individual models. For MSLP the NH results confirm those for ξ850, although the numbers of contributing systems are much lower than for ξ850, reflecting the larger-scale nature of the MSLP features.
Figure 5 shows the mean track intensity distributions, based on using the second comparison method for ξ850. This shows distributions both for the tracks that do and do not match. For those that do match the total number of systems must be identical between any pair of reanalyses, although the distributions may be different. For those systems that do not match there will, in general, be different numbers between any pair of reanalyses. The overall impression is that those tracks that match between the different reanalyses are typically those with the larger mean intensities and those that do not match tend to be those with the weaker mean intensities. For those tracks that do match, the distributions are very similar for all the reanalyses, although for ERA15 there is a slight shift to higher mean intensities relative to each of the other reanalyses. This is perhaps most apparent for NCEP–DOE and is consistent with the results presented previously for the mean intensity spatial statistics in section 4a(2), as well as the mean track intensity distributions shown in Fig. 4. This intensity shift is even more evident for MSLP (not shown). The comparison of NCEP–NCAR and NCEP–DOE also shows this feature, consistent with the spatial statistical results of mean intensity differences (Fig. 3d), perhaps pointing to the changes made to the NCEP–NCAR physics for NCEP–DOE as a cause. Another general impression is that the distributions for those systems that do match are quite broad compared with those that do not match. This indicates that the common observational database used by the reanalyses clearly resolves the more intense systems while the weaker systems are more dependent on the details of how the observations have been assimilated. The GEOS-1 reanalyses appear to have many more weak systems that are not matched than the other reanalyses. Some of these may be model-generated storms. This is also consistent with the spatial statistics for track density differences shown previously, in particular in the Pacific.
A summary of these results is given in Table 2. For MSLP, in general, about 70% of tracks between the different pairs of ensembles match well. This percentage is a little lower for ERA15–GEOS-1 (65%) and higher for NCEP–NCAR and NCEP–DOE (85%). This is also reflected in the results for ξ850, where the percentage of the number of tracks that now match is around 65% but the total number of tracks is more than double that for MSLP. Varying the matching parameters, for example, relaxing the mean separation distance threshold to 2.0°, changes these values to about 78% for MSLP and 70% for ξ850. This change in the number of tracks that match may indicate differences in location for some tracks, although as we relax this constraint further the possibility of matching tracks that are not the same systems becomes more prevalent. Relaxing the temporal overlap criterion also results in an increase in the number of matches; for example, relaxing this constraint to 50% results in the number of systems that match for MSLP changing to 74%, and for ξ850 to 67%, indicating some uncertainty in the temporal coherence of some systems. This can happen if the system intensity falls below the intensity threshold used for identification (1.0 × 10−5 s−1 for ξ850 and 4.0 hPa for MSLP relative to the removed background) or because no extremum can be found. Thus, the systems that do not match appear to be a mixture of those with significant location differences between the reanalyses, those with different temporal coherence, those that are resolution dependent, and perhaps some that are model generated.
Using the relaxed matching criteria, in particular, the mean separation threshold, allows the uncertainty in the location and intensities to be further quantified. Figures 6a,b show frequency distributions for point-by-point intensity differences between matching tracks and their mean separation distances for ξ850. Similar results are obtained for MSLP. A broad distribution in the intensity differences will indicate greater uncertainty in the systems' intensities between the reanalyses, while the “flatter” the distribution of mean seperation distances the greater uncertainty there is in the position of systems. The intensity difference distributions (Fig. 6a) reflect much of the results already discussed, with the comparison of ERA15 with the NCEP–NCAR reanalyses showing a bias to stronger systems in ERA15, indicated by the peak in the relevant distributions occurring at −0.5 × 10−5 s−1. For GEOS-1 the peak shows no particular bias with respect to ERA15, while the NCEP–NCAR–NCEP–DOE comparison shows a weak bias toward stronger systems for NCEP–NCAR. These results are all consistent with the previous discussion. Of more interest is the distribution for the mean separation distances (Fig. 6b). This shows that there is a strong bias toward small separation distances for those systems that match for all the reanalysis comparisons in the NH, indicated by the rapid fall in the distributions to a minimum around 0.5° mean separation for each of the comparisons. This highlights the choice of 0.5° as a matching criteria to be a good one. The long but low-valued tail indicates that some of those systems that do not match for the stricter criteria are systems that have variable locations between the reanalyses.
The previous analysis indicates that there is uncertainty in the day-to-day weather, but does this uncertainty impact on the climatologies of the weather systems? This can be answered to some extent by noticing that the results discussed in this section indicate that there is a set of systems that is consistently represented in MSLP or ξ850 across all the reanalyses with a background of less well-represented systems that are typically quite weak and small scale and with possible differences in location and temporal coherence between the reanalyses. To further explore this, the spatial distributions of the systems that do and do not match for the stricter matching criteria have been recomputed. The result of differencing the mean intensity statistics for those tracks that match (not shown) is very similar to those for MSLP and ξ850 (Fig. 3), where all the tracks were included. This provides us with confidence in the results for ξ850 and for MSLP discussed in section 4a(2) for the mean intensity (which did not necessarily use identical periods) and indicates that the tracks that do not match have relatively little impact on the mean attributes of the storms in the NH.
The spatial distributions of the systems that do not match are also of interest (though not shown) at least for ξ850. For ERA15, these systems are distributed throughout the two main ocean storm tracks as well as the Mediterranean storm track and over North America with the highest values of track density in the regions in which smaller-scale systems often occur, such as the Pacific secondary development region (HH), the Mediterranean storm track stretching through the Middle East, and the Icelandic region. The end of the Mediterranean storm track appears particularly sensitive with very little agreement between the systems, which tend to be quite weak, from the different reanalyses. For NCEP–NCAR and NCEP–DOE the relatively small number of systems that do not match occur predominately on the flanks of the Pacific storm track and at the end of the Mediterranean storm track. For GEOS-1 the Pacific stands out as a region with many systems that are inconsistent with ERA15. However, for the Atlantic and other regions the agreement is similar to the other analyses. Some of this behavior has been highlighted by the track density difference plots shown in section 4a(2). The general impression is that this level of day-to-day weather uncertainty is unimportant to the storm climatologies in the NH, although it may have more of an impact on the diabatic processes that could then be climatologically important; for example, the storms at the end of the Mediterranean storm track that have a poor correspondance among the reanalyses, though weak, are the only source of rainfall to the arid regions of Afghanistan and Iraq.
b. SH comparison of reanalyses
1) The ξ850 and MSLP climatologies
The ERA15 climatological track density and mean intensity statistics for the SH, ξ850, are given in Figs. 7a,c, respectively. This shows the type of distribution of storms seen in previous studies (e.g., Sinclair and Watterson 1999). A region of high track density encircles the SH between latitudes 45° and 65°S, with very high activity levels between 135° and 180°E longitude and slightly lower levels extending on either side. The maximum mean intensity (Fig. 7c) is equatorward and upstream of this peak in the track density in a band extending from 45°W to 140°E, with the peak in intensity occurring between 40° and 100°E. Interestingly, most of the coastal region around Antarctica is a region of cyclolysis and strong decay as indicated by the lysis density and mean growth and decay statistics (not shown), with very strong lysis and decay occurring just upstream of the peak in track density. The region of the peak in track density shows a reduction in cyclolysis and growth in the mean. The latter may result from some secondary development associated with the upstream decaying systems. This impression is further enhanced by a maximum in cyclogenesis (not shown) in this region. In the western part of the SH between 90° and 180°W there is also a region of enhanced mean intensity although of less intensity than in the eastern part of the SH. The contrast between the eastern and western parts of the SH was previously highlighted by Murray and Simmonds (1991).
The MSLP statistics shown in Figs. 7b,d are consistent with those of the ξ850, though showing the greatest track density quite close to the Antarctic coast south of 60°S in all sectors. The mean intensities are also consistent with ξ850 showing the strongest systems to occur in the eastern part of the SH. Murray and Simmonds (1991) attribute this behavior to the stronger temperature gradient in this region. Examination of the mean growth and decay rates (not shown) indicates slightly higher values in the Indian Ocean sector between 30° and 60°S for MSLP but not for ξ850. The reduced level of activity in the MSLP at latitudes lower than 60°S compared with ξ850 indicates that many of the systems identified in this region in ξ850 are probably relatively small scale. However, the bias in activity toward the Antarctic coast in MSLP may also be due to a residual influence of planetary scales not removed by the n ≤ 5 planetary-scale filter. Removing the n ≤ 7 or 10 planetary scales from the MSLP may reveal more of the activity seen in ξ850 at the lower latitudes. This requires further analysis similar to that performed by Anderson et al. (2003). Some indication that this might be a useful approach was seen by Anderson et al. (2003) in the NH where removing n ≤ 7 or 10 allowed more of the Mediterranean activity to be identified in the MSLP field. This further highlights the efficacy of using ξ850 for tracking studies.
2) Comparisons of reanalyses for ξ850 and MSLP
Figure 8 shows the difference in track density for ξ850 for the different reanalyses. It is not surprising to see greater variation in the differences between the different reanalyses in the SH than in the NH due to the relatively fewer available observations to assimilate in the SH. This is particularly the case for surface and radiosonde observations. In consequence, the analyses are more dependent on the model and physical parameterizations than in the NH. Once again there are some coherent patterns of difference between the reanalyses. Both NCEP–NCAR and NCEP–DOE reanalyses appear to be systematically less active than ERA15 poleward of 50°S and more active equatorward of this, particularly NCEP–DOE in the Indian and Pacific Ocean sectors. The NCEP–NCAR–NCEP–DOE difference does indeed show the NCEP–DOE reanalysis to be more active in the region 30°–60°S. This is also seen in the MSLP field, though to a lesser extent. The source of these differences in track density between NCEP–NCAR and NCEP–DOE is difficult to determine considering the number of changes made for the NCEP–DOE reanalysis. It may result from fixing the misplaced PAOBs or from changes to the physical parameterizations. Studies at the Climate Prediction Center into the sensitivity of the NCEP–NCAR analysis to the misplaced PAOBs indicate that the largest differences for synoptic-scale features are seen poleward of 40°S over the oceans in the winter months, and decrease rapidly with height. It was found that synoptic features could be shifted in amplitude and/or phase between assimilations with and without corrected PAOBs. Case studies showed that the impact on individual weather systems was variable, with some unaffected while others could be severely affected. Perhaps the most striking comparison is between ERA15 and GEOS-1, which demonstrates that GEOS-1 is consistently more active in most regions of the SH storm track, particularly the Indian Ocean. The nature of this extra activity will be explored later.
In the MSLP field (not shown) a somewhat different perspective on the differences in track density are seen, with ERA15 being consistently more active close to the Antarctic coast and the other analyses being significantly more active in a band around 60°S. The storm track in ERA15 is slightly narrower than in the other reanalyses.
The mean intensity differences for ξ850 are shown in Fig. 9. ERA15 appears consistently more intense in the mean throughout the storm track regions, particularly for the NCEP–NCAR and NCEP–DOE reanalyses with ERA15 being 0.5–1.0 × 10−5 greater in intensity in the mean. This is greater than the difference seen in the NH and is likely to be related not just to differences in resolution of the model integrations but also to the physical parameterizations. The same differences in intensity are also found for GEOS-1 compared with ERA15 except in the Pacific part of the storm track region, which shows some regions where GEOS-1 is more intense in the mean. As was highlighted in the NH, ERA15 appears consistently weaker in the mean intensity near orography, notably near Antarctica. This is particularly striking in the comparison with GEOS-1. For the NCEP–NCAR–NCEP–DOE comparison the mean intensities for ξ850 appear consistently stronger in the storm track regions in NCEP–NCAR than in NCEP–DOE. This is consistent with what was seen in the Northern Hemisphere, indicating that it is changes to the physical parameterizations that are responsible rather than the fixed PAOBs and other aspects.
For MSLP the mean intensities (not shown) indicate that ERA15 is significantly more intense that the other reanalyses, particularly in the main storm track region south of 60°S; ERA15 shows very large differences close to the Antarctic coast, varying from ∼5 hPa with NCEP–NCAR to ∼15 hPa with GEOS-1. Away from the Antarctic coast the behavior is consistent with that seen for ξ850. Also similar to ξ850, the mean intensity is higher for GEOS-1 in the Pacific. Finally, the difference between NCEP–NCAR and NCEP–DOE mean intensity is also similar to that seen for ξ850, with NCEP being typically 1–5 hPa more intense in the main storm track region south of 60°S. The large differences seen close to the Antarctic coast in both the track density and, more particularly, the mean intensity for MSLP perhaps points to the residual influence of the circumpolar pressure trough not removed by the planetary wave filter. Further evidence for this comes from ξ850, which does not show this behavior, indicating its large-scale nature. This could be tested by extending the planetary wave filter to 7 or 10 wavenumbers.
3) Direct comparison of track ensembles
Figures 4c,d show the distributions of mean intensity for the storms in each reanalysis in the SH for ξ850 and MSLP, respectively. These distributions now begin to show obvious differences from their counterparts in the NH. For ξ850 it is evident that ERA15 and GEOS-1 are quite similar at the high-intensity end of the distribution but that GEOS-1 has many more weaker systems. For NCEP–NCAR and NCEP–DOE there are more systems around the peak in the distributions and their tails fall away more rapidly. For the MSLP mean intensities the four distributions now differ significantly from each other and from those in the NH. ERA15 has a bimodal distribution extending to much higher mean intensity values than the other reanalyses. The NCEP–NCAR reanalysis distribution has a similar character but with a much narrower distribution. This bimodal distribution is not seen in the NCEP–DOE or GEOS-1 distributions. However, this behavior in the MSLP is consistent with the results discussed above. The nature of the bimodal distribution for ERA15 has been explored further by partitioning the tracks according to their mean intensity into two groups for high- and low-intensity systems using the dip in the distribution (22 hPa) as the separation point. Reconstructing the spatial statistics for each group shows the high-intensity systems to predominate at high latitudes in the main storm track region around Antarctica, while the low-intensity systems predominate at low latitudes in the Pacific and Atlantic. The fact that this bimodal behavior is weaker for NCEP–NCAR and does not exist for NCEP–DOE or GEOS-1 may point to differences with the circumpolar pressure trough already discussed.
The direct comparison of the track ensembles, shown in Fig. 10 for ξ850, indicates much poorer results than in the NH with, in general, almost as many bad matches as good matches. Indeed, for GEOS-1 and ERA15 there is a larger number of bad matches than good matches. Also, the distributions for the bad matches are now much broader, including more of the intense systems. The results are equally as striking for MSLP.
A further summary of these results is shown in Table 3. This further confirms the differences between the various reanalyses in the SH. While at the large-scale end of the synoptic range highlighted by MSLP about ∼60% of the systems in ERA15, NCEP–NCAR, and NCEP–DOE match, the number is less than 50% for the small-scale end of the synoptic range highlighted by ξ850. For the ERA15–GEOS-1 comparison the number of matches is considerably worse. Separating those systems that do and do not match and recomputing the spatial statistics is probably more contentious in the SH due to the large differences already highlighted.
The fact that there are so many poor matches may reflect more severe problems with location than was seen in the NH, rather than temporal coherence, although this will still occur for the weaker systems. This is indicated by relaxing the constraints, as previously described. It is also possible that more of the systems are model generated, due to the lower level of constraint provided by the available observations, and so have no correspondence between reanalyses. Figures 6c,d show the ξ850 frequency distributions for the point-by-point intensity differences and the mean separation distances as discussed previously for the NH. The intensity distributions show the same form of biases as seen in the NH, but the distributions are now broader, highlighting a greater uncertainty in the intensities that contribute to the differences seen in the spatial intensity statistic. Of perhaps more interest is the separation distance distributions (Fig. 6d). The distributions are now much flatter with higher values in the tail, indicating a much larger degree of uncertainty in the locations of systems. This is particularly striking for the ERA15–GEOS-1 comparison and partially explains the poor number of matches found for the stricter criteria.
The larger differences seen in the SH in the spatial statistics, as well as the greater level of uncertainty in location and intensity and the greater possibility of model-generated storms, mean that the use of the reanalyses for GCM validation, for example, requires more care than in the NH. Any studies should perhaps use more than one reanalysis to put the uncertainty in context.
5. Storm tracks in the upper troposphere
In this section the comparison of the various reanalyses in the upper troposphere are considered using cyclonic 250-hPa vorticity (ξ250) in the NH and SH to identify activity usually associated with pressure troughs. The 500-hPa vorticity (ξ500) will also briefly be considered. Other fields used in HH such as potential vorticity (PV) on a 330-K isentropic surface and potential temperature on a PV = 2 pvu surface could not be used as they are not available from all the reanalyses, although they could be computed given the model level data.
Figure 11 shows the track density for ξ250 in the NH for ERA15, the ECMWF operational analyses (1994–2001), and the NCEP–NCAR and GEOS-1 reanalyses. The pattern of the track density distributions looks very similar, but the level of activity for the ERA15 reanalysis differs significantly from that of the other analyses. The distributions for the ECMWF operational analyses and the NCEP–NCAR and GEOS-1 reanalyses are qualitatively very similar. These show a spiral of disturbances starting in the subtropical Atlantic, then extending through the Mediterranean and into the east Asian jet entrance region in a very narrow band before extending across the Pacific, continuing across North America and into the NH storm track. As pointed out by HH this spiraling band of upper-tropospheric disturbances can amplify through the depth of the troposphere where conditions are favorable such as in the usual baroclinic regions of the east coast of North America and Asia. There is also significant activity over the Siberian storm track region and a second track possibly related to this to the north of the Himalayas that also links up with the more southern track in the east Asian jet entrance region. The impression is of a split track to the north and south of the Himalayas. This is more apparent for ξ500 where the orography has more effect (Figs. 12c,d).
The differences between the ERA15 track density and those for the other analyses are quite dramatic, particularly in the western Pacific subtropical jet region. This is due to a considerable proportion of broken tracks that are removed by the track filtering due to their lack of propagation coherence. This, in turn, appears to be the result of the way the relatively sparse upper-tropospheric data are assimilated in ERA15, resulting in multiple centers (even at T42) and large jumps in displacement and direction when observations occur. Relaxing the constraints used in the optimization procedure for the tracking, that is, the maximum displacement distance and track smoothness, does not resolve this problem. The fact that this does not occur for the ECMWF operational analyses and the NCEP–NCAR and GEOS-1 reanalyses, which use more sophisticated assimilation systems, strongly suggests that this is a problem with the use of a limited form of OI in ERA15. In GEOS-1, IUA is used (Pfaendtner et al. 1995) as discussed in section 2, resulting in smoother fields. In this diagnostic the performance of this method appears to be comparable with the variational schemes, but it is more computationally expensive than the standard OI due to the need to rerun the forecast. The fact that the IUA approach gives comparable results with the variational schemes highlights the importance of the temporal as well as the spatial interpolation of observations. The ECMWF operational analyses give slightly higher track densities than the NCEP–NCAR and GEOS-1 results, probably due to the significantly higher integration resolution of the operational analyses of T213/L31, more recently T319/L50, and currently T511/L60.
The problem with the coherence of ERA15 upper-tropospheric disturbances is most pronounced in the smaller-scale fields such as vorticity and PV. It is less of a problem in larger-scale fields such as potential temperature, although it still exists to a lesser extent. Preliminary studies of ECMWF 40-yr Reanalysis (ERA40) data, which uses a 3DVAR system, shows again the efficacy of using a more sophisticated assimilation scheme with improvements in all upper-tropospheric fields but in particular in the small-scale fields.
In other seasons this problem is much reduced in ERA15. For example, the summer activity (JJA) at the 250-hPa level indicates much better agreement between ERA15 and NCEP–NCAR (cf. Figs. 12a,b, respectively). Hence, this problem appears to occur when the subtropical jets are at their strongest, with large upper-tropospheric winds and shear. Limited case studies performed as part of this study have shown filamentary structures associated with anomalies, particularly in the vicinity of the subtropical jets, so that poor use of the available observations in a relatively high-resolution model may result in a noisy representation of the systems.
To further explore the vertical extent of the problem with ERA15, the 500-hPa relative vorticity (ξ500) has also been analyzed using the same methodology. Figures 12c,d show the track density results for the DJF period for ERA15 and NCEP–NCAR, respectively. At this level the two results are very similar and reflect interesting aspects seen at both the 850- and 250-hPa levels but with a generally higher level of activity than at either of these levels. For example, the subtropical activity through the Mediterranean and into east Asia appears now to have more activity than at 250 hPa and to definitely deviate to the south of the Himalayas. This similarity between the analyses indicates that the problems experienced with ERA15 occur above 500 hPa, and perhaps supports the decoupling of the influence of observations at different levels mentioned previously.
A similar situation occurs in the SH winter (JJA) 250-hPa ξ250 cyclonic activity (not shown). The activity in the Australian subtropical jet region is better defined in the ECMWF operational analysis and both the NCEP–NCAR and GEOS-1 reanalyses than in ERA15.
6. Tropical easterly wave activity
The tropical EWs considered here are synoptic off-equatorial systems in the NH summer (MJJASO), typically with wavelengths ∼2000–3000 km. Analysis of such systems is a challenge, since in general they occur in regions in which available observations are sparse even over land, so that there is greater potential for the imprint of the model to appear in the results. They are also generally very much weaker than midlatitude synoptic systems unless they develop into tropical storms. This has implications for the analysis of EWs since they may be particularly sensitive to the model and assimilation methodology. For example, Slingo et al. (1994) showed the waves to be sensitive to the convective parameterization used in the model. Here, the unfiltered 850-hPa relative vorticity is used to identify and track these systems. However, by focusing on the one level and identifying the waves as vorticity maxima, only a particular aspect of the EW is being identified. It should be noted that only the strongest waves are being considered due to the nature of the current tracking scheme and the fact that only those systems that have positive vorticity extrema with values greater that 5 × 10−6 s−1 are identified. This approach was successfully used by Thorncroft and Hodges (2001) to explore EW statistics in the West African and tropical Atlantic region using the ERA15 data extended with operational analyses. In this study their work is extended to encompass the whole tropical and subtropical region between 0° and 40°N. This allows recurving tropical cyclones to be tracked.
Figures 13a–d show the results for the track density for the various reanalyses. They show that synoptic systems are found in most of the tropical region for this season, including West Africa and the tropical Atlantic (Reed et al. 1988; Thorncroft and Hodges 2001), the Caribbean (Riehl 1945; Molinari et al. 1997), the east Pacific (Zehnder et al. 1999), the central and west Pacific (Reed and Recker 1971), and the Indian Ocean (Annamalai et al. 1999). However, it is likely that the nature of the EWs in these regions will vary from region to region, consistent with the different ambient conditions in which they develop. For example, the presence of marked low-level baroclinicity over West Africa results in very different EW structures (Pytharoulis and Thorncroft 1999) compared with the west Pacific (Reed and Recker 1971). The African region is of particular interest as it is the largest region of land in the NH Tropics that is dominated by EWs and provides an interesting contrast with the predominately oceanic activity of the Atlantic and Pacific.
Following Thorncroft and Hodges (2001) the Atlantic EWs are discussed first. Figure 13a shows the results for ERA15 plus post-ERA operational analyses (ERA+). This shows two storm tracks, one poleward of 15°N over West Africa associated with little rainfall and a second located in the general region of the intertropical convergence zone (ITCZ) over the land and the ocean and associated with the main rainy zone. In general, the storm track over the land in the ITCZ region appears quite weak at the 850-hPa level. The main activity in this region occurs at the African easterly jet level (∼700–600 hPa; Thorncroft and Hodges 2001). This dual storm track has been commented on elsewhere (Reed et al. 1988) and is consistent with the strong influence of low-level baroclinicity over the land (Pytharoulis and Thorncroft 1999) and the more important role of moist convection over the ocean. In the analysis of Thorncroft and Hodges (2001) it is commented that the two tracks appear to be separate with very little indication of systems on the more northern track merging with the equatorward track (although this does appear to occur occasionally). This has consequences for tropical cyclogenesis through the expected influence on the vorticity and humidity structure of the EW precursor.
The NCEP–NCAR results (Fig. 13b) also show two storm tracks in this region though they appear to be less separate than in ERA+. This may indicate the influence of the different models used in ERA+ and NCEP–NCAR or perhaps the different resolutions used. The level of activity in this region is somewhat lower than for ERA+ both over the land and the Atlantic Ocean. For the NCEP–DOE activity in this region (Fig. 13c) the situation is very different, with the activity over the land poleward of 15°N very much weaker and the activity in the ITCZ over the ocean showing significantly increased activity. There is also some indication of EWs occurring farther eastward over the land east of Lake Chad, suggesting more coherent systems in the vicinity of the ITCZ over the land. The large differences between NCEP–DOE and the other reanalyses is apparent not just over West Africa and the tropical Atlantic but throughout the Tropics and can probably be attributed to the changes made to the physical parameterizations of the NCEP–NCAR model for NCEP–DOE as discussed below. The impression from the GEOS-1 results (Fig. 13d) is that the northerly track over West Africa is continuous with the more southerly oceanic track, the ITCZ activity over the land being very weak.
For the western Atlantic region ERA+ and NCEP–NCAR show a similar pattern of activity, with the storm track appearing to split west of the Caribbean. There is a hint of the recurving systems in the eastern Caribbean associated with tropical cyclones. Other systems extend across the northern coast of South America and across Central America into the eastern Pacific. Also apparent is the lower level of activity in the Caribbean Sea, south of Cuba. The continuation of tracks from the African coast into the Caribbean proposed by some studies is hinted at by ERA+, suggesting that resolution may be important for tracking the vorticity extrema across the Atlantic, although 850 hPa may not be the best level to see this. The NCEP–DOE activity in this region shows increased activity, probably associated with the changes in parameterizations in the NCEP–NCAR model and with less indication of the recurving systems in the eastern Caribbean seen in ERA+ and NCEP–NCAR. The GEOS-1 results have lower Venezuelan coast activity than in the other reanalyses, possibly due to differences in the local physics but also possibly due to a lower likelihood of African EWs reaching this region. This, together with the higher density and more continuous track seen in the eastern Caribbean, suggests that most systems that reach this region recurve. The reason why some systems recurve has often been given as the steering influence of the Bermuda high and upper-level low pressure regions. It would be of interest to explore these aspects further in the reanalyses.
In the eastern tropical Pacific both ERA+ and NCEP–NCAR show the well-defined northwest–southeast (NW–SE)-oriented storm track consistent with the known activity in this region. NCEP–DOE also shows this but with much higher levels of activity as previously mentioned. GEOS-1 has a similar structure to that for NCEP–NCAR and ERA+ but is considerably weaker. In the central and western Pacific, as with the previous regions, NCEP–NCAR and ERA+ are very similar, with the storm track recurving at the western end and some activity extending toward Vietnam. NCEP–DOE is considerably more active, with a peak in activity in the central Pacific. However, in the western Pacific the behavior is now similar to that of ERA+ and NCEP–NCAR, possibly due to the influence of the larger number of observations in the western Pacific. GEOS-1 appears much weaker in the central Pacific than either NCEP–NCAR or ERA+, though in the western Pacific it is comparable. Again this may reflect the greater number of observations available in the western Pacific. The region with the least activity covers the Indian Ocean, the Bay of Bengal, and the Arabian Sea. However, there is some weak activity seen in the Bay of Bengal in ERA+ and NCEP–NCAR, In particular ERA+ gives some indication of the propagation of depressions along the monsoon trough over India (Annamalai et al. 1999) but this is less apparent for NCEP–NCAR.
The results presented here for the tropical EW activity have raised several issues with respect to the sensitivity of the waves to observations, assimilation method, resolution, and the model parameterizations. The fact that the observations can be sparsely distributed leads to a likelihood that the model physics will have a large impact on the intensity of the analyzed waves and, hence, their detection. This has been seen in the differences between the NCEP–NCAR and NCEP–DOE levels of activity, particularly in the Atlantic and Pacific Oceans. The fact that this is a model dependency problem is apparent if we consider the NCEP model integration for AMIP II shown in Fig. 13e. It should be noted that the contour scale has been changed to accommodate the significant increase in the number of systems identified. This shows high levels of activity throughout the Atlantic and Pacific Oceans. In the Atlantic the activity extends from Africa through the Caribbean and into the eastern Pacific, with the eastern Pacific activity appearing more zonal than in the reanalyses. The southern Carribean is now an active one in contrast with the reanalyses. There is also significant activity stretching from the western Pacific into the Bay of Bengal. Studies have shown that the EWs are sensitive to the convective parameterization (Slingo et al. 1994). However, the differences in the convective parameterization between NCEP–NCAR and NCEP–DOE were just minor tuning of some of the parameters. The biggest parameterization change that could have an impact on the EWs was the introduction of the new planetary boundary layer (PBL) scheme of Hong and Pan (1996). This was partly motivated by trying to increase the efficiency of the supply of moisture to the convective scheme (M. Kanamitsu 2002, personal communication). This may explain the large increase in EW activity over the oceans, with the intensification of the waves making them more visible to the tracking scheme. It could also explain the increase in the wave activity over Africa in the ITCZ region.
The study of Thorncroft and Hodges (2001) also raised the issue of how good the ERA15 reanalysis is with respect to coherently propagating waves from the perspective of identifying them as vorticity centers. As pointed out in the previous discussion of midlatitude storm tracks, there can be some sensitivity, especially for weak systems, to the assimilation system in data-sparse regions. This might also be expected to be the same here. Certainly if the activity in ERA15 is compared with that found in the operational analyses, the coherence of EWs appears to be much better in the operational analyses. The EWs also appear to be more coherent in the limited ERA40 data considered here, probably associated with higher resolution (T159) and the use of 3DVAR. The move to even higher resolution in the operational analyses and 4DVAR appears to allow even better identification of the waves.
7. Conclusions
A comparison of four different reanalysis datasets has been performed based on feature tracking analysis methods. In general, all the reanalyses give similar results for the lower-tropospheric NH storm tracks, with most differences in the distribution of storms being small and associated with relatively small spatial-scale systems. These differences are most marked where small-scale systems are often found, such as in the secondary development regions of the Pacific and Atlantic and also over North America and the Mediterranean. In particular, there is very little agreement at the end of the Mediterranean storm track through the Middle East, which is relatively low in track density compared with the rest of the Mediterranean and dominated by very weak systems.
For the NH mean system intensity, differences are relatively small but with ERA15 appearing to have slightly more intense systems. This is systematically seen in both ocean basins for both MSLP and ξ850, though perhaps less clearly for GEOS-1. This is probably related to the spatial integration resolution. For ξ850 it is also apparent that ERA15 is less intense near major orography. This is possibly due to the use of a subgrid-scale orographic parameterization in EAR15 whereas the other analyses used a mean orography. The fact that this is consistent between the different reanalyses and hemispheres lends credence to this explanation rather than the use of different methods of extrapolation of ξ850 into orographic regions.
The comparison between NCEP–NCAR and NCEP–DOE in the NH showed no significant systematic differences in the track densities but did show NCEP–NCAR to be systematically more intense than NCEP–DOE in both MSLP and ξ850. Since the resolutions of NCEP–NCAR and NCEP–DOE are identical this is most likely to be associated with changes in physical parameterizations. This hypothesis is also confirmed by this behavior being seen in both hemispheres. It is, however, difficult to attribute these differences between NCEP–NCAR and NCEP–DOE to any one particular change in parameterization or boundary forcing.
Direct comparison of the track ensembles in the NH shows that most of the systems that compare well between the reanalyses are those with mean intensities in the moderate to large range, with those that do not match predominately at the weak end of the range. The general impression is of a set of significant systems that compare well for all the reanalyses plus a background of weak intensity systems that do not compare well but that are relatively unimportant to the storm climatologies. It is perhaps often not realized that even in the well-observed NH there is uncertainty in the description of day-to-day weather. Here, this uncertainty has been quantified by exploring the differences in location and intensity between the reanalyses. This analysis has confirmed the impression that much of the uncertainty is small in relation to the storm climatology.
In the SH the comparison between the different reanalyses highlights the greater uncertainty in the weather system representation. The comparison of the distributions for track density and mean intensity shows that there are significant climatological differences between the reanalyses as a result of the model characteristics being less constrained by the sparser available observations. The degree of the uncertainty in the SH is highlighted by the distribution for intensity differences and separation distances. These show much greater uncertainty in both intensity and location than in the NH. There is also the uncertainty that since the models are less constrained in the SH, model-generated storms can contribute to the statistics.
The fact that the correspondence between the different reanalyses is more variable in some regions or levels highlights the observational uncertainty inherent in the data. The general impression is that while the lower troposphere in the NH is well observed, elsewhere this is not the case. This varying degree of uncertainty does not make the reanalyses unusable in these regions, for example, in GCM validation. It does, however, mean that the quantitative use of the reanalyses in these regions at the weather system scale will necessarily carry less significance and have larger error bars due to the observational uncertainty. It also highlights the fact that reanalyses should not be used individually without some quantification of the uncertainty that studies such as this one provide. In the same vein, when there is uncertainty it is difficult to quantify which reanalysis is closer to reality due to the lack of a truly independent dataset. Such a dataset could be provided by the independent use of the extensive satellite data available since the late 1970s to identify weather systems. However, this presents many difficulties related to sampling issues and the detection of the weather systems, although work is being carried out in this area.
The consistency of the results in the NH for the lower-tropospheric fields provides us with some confidence that any of the reanalyses can be used to validate climate models at the synoptic scale. In the upper troposphere, however, results show that the identification of coherently propagating systems can depend on how the data assimilation is performed. These problems appear most acute above 500 hPa in both the NH and SH winters in ERA15. The problem appears less pronounced during the summer periods when wind speeds and shears are lower. Some of the problems can be associated with the form that upper-level disturbances take in the presence of strong winds and shear. However, more sophisticated data assimilation systems such as 3DVAR and 4DVAR can provide greater vertical consistency and appear to provide a better representation of upper-tropospheric disturbances as coherently propagating systems. Additionally, the fact that GEOS-1 is comparable to the 3DVAR and 4DVAR schemes in the upper troposphere suggests that temporal interpolation is also an important aspect.
The study of the tropical EW activity has highlighted the sensitivity of the EW activity to the observing system, the model physics, and the model integration resolution. In particular, the model physics can have a significant impact on the representation of the EWs in the data-sparse regions of the Tropics. This is indicated by the significant increase in EW activity seen throughout the Tropics in the NCEP–DOE reanalysis and the AMIP integration of the NCEP model. This appears to be attributable to the change to a new boundary layer parameterization for NCEP–DOE and AMIP–NCEP, which aims to increase the efficiency of moisture supply to the atmosphere and which clearly affects the intensity of EWs in the ITCZ region, particularly in the oceanic regions. The tropical EW behavior should also be contrasted with the behavior seen in the NH and SH storm tracks where the NCEP–DOE reanalysis is systematically weaker in terms of the mean intensity than for NCEP–NCAR, although this difference is not particularly large. This highlights the fact that tropical and extratropical systems may exhibit different sensitivities to changes in parameterizations. Issues not addressed here are the intensity and longevity of cyclones identified in the reanalyses. This will be explored in the future, in particular when the newer reanalyses from ERA40 and the Japanese 25-yr Reanalysis Project become available. Also of interest are the annual cycle and interannual variability; these will also be studied using these reanalyses. It is clear that there is great difficulty in drawing any conclusions on the performance of a GCM in simulating tropical weather systems using comparisons with the reanalyses due to the large uncertainty in the EW representation of the reanalyses and the difficulties in identifying the EWs.
One issue not addressed in this paper concerns significance and whether confidence levels can be computed. In practice, this is difficult due to the nature of the data: the storm trajectories represent data that are spatially and serially correlated. This means that classic parametric methods cannot be used. The alternative is to use nonparametric methods such as the bootstrap method (sampling with or without replacement; Effron and Tibshirani 1994) to compute a sampling distribution. However, this requires a large number of samples and, at present, is computationally prohibitive. Methods to speed up the statistical estimation are being considered, which, allied with computer resource management software such as CONDOR (Basney and Livny 1999), may make this possible in the future.
Acknowledgments
The authors would like to thank ECMWF for making the ERA15, operational analyses, and ERA40 data available; the Climate Diagnostics Center for supplying the NCEP–NCAR data; PCMDI for supplying the NCEP–DOE data; and EOSDIS for supplying the GEOS-1 data. The authors would also like to thank Dr. Adrian Simmons of ECMWF and the reviewers for their constructive comments.
REFERENCES
Anderson, D., K. I. Hodges, and B. J. Hoskins, 2003: Sensitivity of feature-based analysis methods of storm tracks to the form of background field removal. Mon. Wea. Rev., 131 , 565–573.
Annamalai, H., J. M. Slingo, K. R. Sperber, and K. Hodges, 1999: The mean evolution and variability of the Asian summer monsoon: Comparison of ECMWF and NCEP–NCAR reanalyses. Mon. Wea. Rev., 127 , 1157–1186.
Basney, J., and M. Livny, 1999: Deploying a high throughput computing cluster. Architecture and Systems, Vol. 1, High Performance Cluster Computing, R. Buyya, Ed., Prentice Hall PTR, 116–134.
Bengtsson, L., and P. Kållberg, 1981: Numerical simulation—Assessment of FGGE data with regard to their assimilation in a global data set. Adv. Space Res., 1 , 165–187.
Blender, R., 2000: Cyclone tracking in different spatial and temporal resolutions. Mon. Wea. Rev., 128 , 377–384.
Chou, M. D., 1992: A solar radiation model for use in climate studies. J. Atmos. Sci., 49 , 762–772.
ECMWF, 1999: The description of the evolution of the ECMWF forecasting system and corresponding archive. Tech. Rep., 143 pp. [Available from ECMWF, Shinfield Park, Reading, Berkshire, RG2 9AX, United Kingdom.].
Effron, B., and R. J. Tibshirani, 1994: An Introduction to the Bootstrap. Monographs on Statistics and Applied Probability, No. 57, Chapman & Hall/CRC, 436 pp.
Fels, S. B., and M. D. Schwarztkopf, 1975: The simplified exchange approximation: A new method for radiative transfer calculations. J. Atmos. Sci., 32 , 1475–1488.
Gates, W. L., 1992: AMIP: The Atmospheric Model Intercomparison Project. Bull. Amer. Meteor. Soc., 73 , 1962–1970.
Gibson, J., P. Kållberg, S. Uppala, A. Nomura, A. Hernandez, and E. Serrano, 1997: ERA description. ECMWF Re-Analysis Final Rep. Series 1, 71 pp.
Harshvardhan, D. A. Randell, and T. G. Corsetti, 1987: A fast radiation parameterization for atmospheric circulations models. J. Geophys. Res., 92 , 1009–1016.
Higgins, R. W., Y. Yao, M. Chelliah, W. Ebisuzaki, J. E. Janowiak, C. F. Ropelewski, and R. E. Kistler, 1996: Intercomparison of the NCEP/NCAR and the NASA/DAO Reanalyses (1985–1993). NCEP/Climate Prediction Center ATLAS No. 2, U.S. Department of Commerce, 169 pp.
Hodges, K. I., 1996: Spherical nonparametric estimators applied to the UGAMP model integration for AMIP. Mon. Wea. Rev., 124 , 2914–2932.
Hodges, K. I., 1999: Adaptive constraints for feature tracking. Mon. Wea. Rev., 127 , 1362–1373.
Hong, S. Y., and H. L. Pan, 1996: Nonlocal boundary layer vertical diffusion in a medium-range forecast model. Mon. Wea. Rev., 124 , 2322–2339.
Hoskins, B. J., and P. D. Sardeshmukh, 1984: Spectral smoothing on the sphere. Mon. Wea. Rev., 112 , 2524–2529.
Hoskins, B. J., and K. I. Hodges, 2002: New perspectives on the Northern Hemisphere winter storm tracks. J. Atmos. Sci., 59 , 1041–1061.
Kalnay, E., and Coauthors. 1996: The NCEP/NCAR 40-Yr Reanalysis Project. Bull. Amer. Meteor. Soc., 77 , 437–471.
Kanamitsu, M., W. Ebisuzaki, J. Woolen, J. Potter, and M. Fiorino, 1999: An overview of reanalysis—2. Proc. Second Int. Conf. on Reanalyses, Reading, United Kingdom, WCRP.
Klein, W. H., 1957: Principal tracks and mean frequencies of cyclones and anticyclones in the Northern Hemisphere. Res. Paper 40, U.S. Weather Bureau, 60 pp.
Lacis, A. A., and J. E. Hansen, 1974: A parameterization for the absorption of solar radiation in the earth's atmosphere. J. Atmos. Sci., 31 , 118–133.
Lott, F., and M. J. Miller, 1997: A new subgrid-scale orographic drag parameterization: Its formulation and testing. Quart. J. Roy. Meteor. Soc., 123 , 101–127.
Mahfouf, J-F., 1999: Influence of physical processes on the tangent-linear approximation. Tellus, 51A , 147–166.
Mahrt, L., and K. L. Pan, 1984: A two layer model of soil hydrology. Bound.-Layer Meteor., 29 , 1–20.
Molinari, J., D. Knight, M. Dickinson, D. Vollaro, and S. Skubis, 1997: Potential vorticity, easterly waves, and eastern Pacific tropical cyclogenesis. Mon. Wea. Rev., 125 , 2699–2708.
Moorthi, S., and M. J. Suarez, 1992: Relaxed Arakawa-Schubert: A parameterization of moist convection for general circulation models. Mon. Wea. Rev., 120 , 978–1002.
Morcrette, J. J., 1991: Radiation and cloud radiative properties in the European Centre for Medium-Range Weather Forecasts forecasting system. J. Geophys. Res., 96 , 9121–9132.
Murray, R. J., and I. Simmonds, 1991: A numerical scheme for tracking cyclone centres from digital data. Part 2. Application to January and July GCM simulations. Aust. Meteor. Mag., 39 , 167–180.
Pan, H. L., and L. Mahrt, 1987: Interaction between soil hydrology and boundary layer development. Bound.-Layer Meteor., 38 , 185–220.
Pfaendtner, J., S. Bloom, D. Lamich, M. Seablom, M. Sienkiewicz, J. Stobie, and A. da Silva, 1995: Documentation of the Goddard Earth Observing System (GEOS) Data Assimilation System—Version 1. Technical Report Series on Global Modeling and Data Assimilation, Vol. 4, NASA Tech. Memo. 104606, 44 pp.
Pytharoulis, I., and C. D. Thorncroft, 1999: The low-level structure of African easterly waves in 1995. Mon. Wea. Rev., 127 , 2266–2280.
Reed, R. J., and E. E. Recker, 1971: Structure and properties of synoptic-scale wave disturbances in the equatorial western Pacific. J. Atmos. Sci., 28 , 1117–1133.
Reed, R. J., A. Hollingsworth, W. A. Heckley, and F. Delsol, 1988: An evaluation of the performance of the ECMWF operational system in analyzing and forecasting easterly wave disturbances over Africa and the tropical Atlantic. Mon. Wea. Rev., 116 , 824–865.
Riehl, H., 1945: Waves in the easterlies and the polar front in the tropics. Miscellaneous Rep. 17, Department of Meteorology, University of Chicago, 79 pp.
Schubert, S. D., R. B. Rood, and J. Pfaendtner, 1993: An assimilated dataset for earth science applications. Bull. Amer. Meteor. Soc., 74 , 2331–2342.
Sinclair, M. R., and I. G. Watterson, 1999: Objective assesment of extratropical weather systems in simulated climates. J. Climate, 12 , 3467–3485.
Slingo, J., and Coauthors. 1994: Mean climate and transience in the tropics of the UGAMP GCM: Sensitivity to convective parametrization. Quart. J. Roy. Meteor. Soc., 120 , 881–922.
Thorncroft, C., and K. I. Hodges, 2001: African easterly wave variability and its relationship to Atlantic tropical cyclone activity. J. Climate, 14 , 1166–1179.
Tiedtke, M., 1989: A comprehensive mass flux scheme for cumulus parameterization in large-scale models. Mon. Wea. Rev., 117 , 1779–1800.
Trenberth, K. E., D. P. Stepaniak, J. W. Hurrell, and M. Fiorino, 2001: Quality of reanalyses in the Tropics. J. Climate, 14 , 1499–1510.
Whittaker, L. M., and L. H. Horn, 1984: Northern Hemispheric extratropical cyclone activity for four midseason months. J. Climatol., 4 , 297–310.
Xie, P., and P. A. Arkin, 1997: Global precipitation: A 17-yr monthly analysis based on gauge observations, satellite estimates, and numerical model outputs. Bull. Amer. Meteor. Soc., 78 , 2539–2558.
Zehnder, J. A., D. Powell, and D. Ropp, 1999: The interaction of easterly waves, orography, and the intertropical convergence zone in the genesis of eastern Pacific tropical cyclones. Mon. Wea. Rev., 127 , 1566–1585.
The NH DJF ξ850 and MSLP cyclonic climatologies: (a) track density and (c) mean intensity for ξ850; (b) track density and (d) mean intensity for MSLP. Track densities are in units of number density per month per unit area as defined in the text; mean intensities for ξ850 are in units of 10−5 s−1 and for MSLP in units of hPa, relative to the removed background. Mean intensities have been suppressed where the track density is below 1 month−1 per unit area
Citation: Monthly Weather Review 131, 9; 10.1175/1520-0493(2003)131<2012:ACORRD>2.0.CO;2
Difference plots of NH DJF cyclonic ξ850 track density: (a) NCEP–NCAR–ERA15, (b) NCEP–DOE–ERA15, (c) GEOS-1–ERA15, and (d) NCEP–NCAR–NCEP–DOE. Track densities in units of number density per month per unit area as defined in the text.
Citation: Monthly Weather Review 131, 9; 10.1175/1520-0493(2003)131<2012:ACORRD>2.0.CO;2
As in Fig. 2, but for mean intensity. Mean intensity is suppressed where the maximum in the track density between two reanalyses is less than 1 month−1 per unit area. Mean intensities in units of 10−5 s−1 relative to the removed background
Citation: Monthly Weather Review 131, 9; 10.1175/1520-0493(2003)131<2012:ACORRD>2.0.CO;2
Distributions of cyclonic mean intensity for the four reanalyses for (a) NH DJF ξ850, (b) NH DJF MSLP, (c) SH JJA ξ850, and (d) SH JJA MSLP. Mean intensities in units of 10−5 s−1 and hPa for ξ850 and MSLP, respectively, both relative to the background removed state
Citation: Monthly Weather Review 131, 9; 10.1175/1520-0493(2003)131<2012:ACORRD>2.0.CO;2
Distributions of mean intensity for the comparison of the four reanalysis ensembles for the NH DJF cyclonic ξ850 for both the tracks that do and do not match: (a) ERA15–NCEP–NCAR, (b) ERA15–NCEP–DOE, (c) ERA15–GEOS-1, and (d) NCEP–NCAR–NCEP–DOE. Mean intensities in units of 10−5 s−1 relative to the background removed state
Citation: Monthly Weather Review 131, 9; 10.1175/1520-0493(2003)131<2012:ACORRD>2.0.CO;2
Cyclonic ξ850 frequency distributions of (a), (c) point-by-point intensity differences and (b), (d) mean separation differences for the comparison of track ensembles in the NH and SH, respectively
Citation: Monthly Weather Review 131, 9; 10.1175/1520-0493(2003)131<2012:ACORRD>2.0.CO;2
The SH JJA ξ850 and MSLP cyclonic climatology: (a) track density and (c) mean intensity for ξ850; (b) track density and (d) mean intensity for MSLP. Track densities in units of number density per month per unit area as defined in the text; mean intensities for ξ850 in units of 10−5 s−1 and for MSLP in units of hPa, relative to the removed background. Mean intensities have been suppressed where the track density is below 1 month−1 per unit area
Citation: Monthly Weather Review 131, 9; 10.1175/1520-0493(2003)131<2012:ACORRD>2.0.CO;2
As in Fig. 2, but for SH JJA
Citation: Monthly Weather Review 131, 9; 10.1175/1520-0493(2003)131<2012:ACORRD>2.0.CO;2
As in Fig. 3, but for SH JJA
Citation: Monthly Weather Review 131, 9; 10.1175/1520-0493(2003)131<2012:ACORRD>2.0.CO;2
As in Fig. 5, but for SH JJA
Citation: Monthly Weather Review 131, 9; 10.1175/1520-0493(2003)131<2012:ACORRD>2.0.CO;2
Track density distributions for NH DJF cyclonic ξ250 for (a) ERA15, (b) ECMWF operational analyses, (c) NCEP–NCAR, and (d) GEOS-1. Track density units are number density per month per unit area as defined in the text.
Citation: Monthly Weather Review 131, 9; 10.1175/1520-0493(2003)131<2012:ACORRD>2.0.CO;2
Track density distributions for cyclonic relative vorticity in the NH at 250 and 500 hPa: (a) ERA15, ξ250, JJA; (b) NCEP–NCAR, ξ250, JJA; (c) ERA15, ξ500, DJF; and (d) NCEP–NCAR, ξ500, DJF. Track density units are number density per month per unit area as defined in the text.
Citation: Monthly Weather Review 131, 9; 10.1175/1520-0493(2003)131<2012:ACORRD>2.0.CO;2
Track density distributions for EWs: (a) ERA15 + operational analyses, (b) NCEP–NCAR, (c) NCEP–DOE, (d) GEOS-1, and (e) AMIP–NCEP. Track density units are number density per month per unit area as defined in the text
Citation: Monthly Weather Review 131, 9; 10.1175/1520-0493(2003)131<2012:ACORRD>2.0.CO;2
Summary of the various assimilation system, data, and model combinations. Note that the data sections only describe the data types and those used from 1979 onward. Abbreviations and acronyms defined as follows: intermittent OI (int. OI), first guess at the appropriate time (FGAT), incremental update assimilation (IUA), normal mode initialization (NMI), semi-Lagrangian (SL), finite difference (FD), triangular truncation (T***), grid (G), retrieval (ret.), orography (orog.), direct radiance assimilation (DRA), short wave (SW), long wave (LW), National Environmental Satellite, Data and Information Service (NESDIS), Australian sea level bogus data (PAOB), the Alpine Experiment (ALPEX), the First Global Atmospheric Research Programme (GARP) Global Experiment (FGGE), Tropical Ocean and Global Atmosphere Coupled Ocean–Atmosphere Reponse Experiment (TOGA COARE), Global Sea-Ice and SST data (GISST), Operational Vertical Sounder and the Television Infrared Operational Satellite (TOVS/ATOVS)
Summary of NH direct track ensemble matching statistics, percentage of tracks that match, and the total number of tracks per month for each ensemble
As in Table 2, but for SH
