Robert Kistler^*, Eugenia Kalnay^*,1, William Collins^*, Richard Reynolds^*, Suranjana Saha^*, Glenn White^*, John Woollen^*, Yuejian Zhu^*, Muthuvel Chelliah⁺, Wesley Ebisuzaki⁺, Masao Kanamitsu⁺, Vernon Kousky⁺, Kingtse Mo⁺, Chester Ropelewski^+, Huug vanden Dool⁺, Roy Jenne^#, Dennis Joseph^#, and Michael Fiorino^&

* Environmental Modeling Center, ⁺ Climate Prediction Center

National Centers for Environmental Prediction

Washington DC 20233

1 Corresponding author

Additional Affiliation:

School of Meteorology, University of Oklahoma, Norman, OK 73019

The NCEP and NCAR have cooperated in a project (denoted "Reanalysis") to produce a 50-year (1948-1997) record of global analyses of atmospheric fields in support of the needs of the research and climate monitoring communities. This effort involved the recovery of land surface, ship, rawinsonde, pibal, aircraft, satellite and other data, quality controlling and assimilating these data with a data assimilation system which is kept unchanged over the reanalysis period. This eliminated perceived climate jumps associated with changes in the data assimilation system, although it is still affected by changes in the observing systems. In particular, during the earliest decade (1948-1957), upper air data observations were made 3 hours later than the current main synoptic times (e.g., 03UTC), and primarily in the Northern Hemisphere, so that the reanalysis is less reliable than for the latter 40 years. The reanalysis system continues to be used with current data in real time (Climate Data Assimilation System or CDAS), so that its products are available from 1948 to the present.

The NCEP/NCAR 50-year reanalysis uses a frozen modern global data assimilation system, and a data base as complete as possible. The data assimilation (3D-VAR) and the global spectral model are identical to the global system implemented operationally at NMC on January 1995, except that the horizontal resolution is T62 (about 210 km). The database has been enhanced with many observations not available in real time for operational use provided by different countries and organizations and gathered mostly at NCAR. Different types of output archives have been created to satisfy different user needs, including one CD-ROM containing many weather and climate NCEP/NCAR Reanalysis products for each year. A special CD-ROM, containing 50 years of selected monthly and climatological data from the NCEP/NCAR Reanalysis, is included in BAMS with this issue. Reanalysis information and selected output is also available online by Internet from several organizations (wesley.wwb.noaa.gov/reanalysis.html).

There are two major outputs from the reanalysis: gridded fields and observations. Gridded output variables are classified into four classes, depending on the degree to which they are influenced by the observations and/or the model. For example "C" variables (such as precipitation and surface fluxes) are completely determined by the model during the data assimilation, and should be used with caution. Nevertheless, a comparison of these variables with observations and with several climatologies shows that they generally contain considerable useful information. Eight-day "reforecasts" have also been produced every five days and are very useful for predictability studies and for monitoring the quality of the observing systems. The second major output of this project provides, for the first time, an archive of 5 decades of observations encoded into a common format (denoted BUFR), including additional information relevant to their quality (meta data), such as the 6 hour forecast interpolated to the observation location.

The last 40-years of reanalysis (1958-1996) were completed in the fall of 1997, and the reanalysis for the previous decade (useful only in the NH) in the summer of 1998. The continuation of the reanalysis into the future through the identical Climate Data Assimilation System (CDAS) carried out with 3-days delay, allows researchers to reliably compare recent anomalies with those in earlier decades, and it has been notably useful during the 1997-98 El Niño episode. Since changes in the observing systems inevitably produce perceived changes in the climate, a parallel reanalysis without satellite data was generated for the FGGE year (1979), when major new satellite observing systems were introduced in order to assess the impact of such changes on the reanalysis.

In this paper we present examples of applications of the reanalysis, show the impact of observing systems and quality control on 50-years of forecasts from the reanalysis and compare some of the products with those of the other two major reanalyses (ECMWF and NASA/DAO). We also provide information about problems and errors that were inadvertently introduced during the execution and their impact on the reanalysis. Some of these problems were corrected in a shorter reanalysis for 1979 onwards being performed in collaboration with the Department of Energy. Finally, we discuss NCEP's future plans, which, if supported, currently call for an updated global reanalysis using a state-of-the-art system every eight to ten years, alternating with a Regional Reanalysis over North America with much higher resolution. Future reanalyses will be greatly facilitated by the quality-controlled, comprehensive observational database created by the present reanalysis.

The National Centers for Environmental Prediction/National Center for Atmospheric Research (NCEP/NCAR) Reanalysis Project is described in detail in Kalnay et al (1996), and updated documentation and information about data access and problems can be found in the Reanalysis Home Page (http://

wesley.wwb.noaa.gov/reanalysis.html). The Project started in 1989 at NCEP (formerly known as the National Meteorological Center or NMC) with the initial goal of just building a "Climate Data Assimilation System" (CDAS) which would not be affected by the changes introduced by improvements to the numerical weather prediction operational systems. The changes in the NCEP operational systems are documented in Kalnay et al (1998).

The CDAS Advisory Panel suggested in 1990 that CDAS would be much more useful if carried out in conjunction with a long-term reanalysis. NCEP contacted NCAR, which has developed a comprehensive archive of atmospheric and surface data, to explore the possibility of a joint project to perform a very long reanalysis. NCAR enthusiastically agreed and was also supportive of the idea of starting as far back as the International Geophysical Year (1957-1958) and to gather the data to perform a long-term reanalysis using a frozen state-of-the-art system. The support of NOAA (the National Weather Service, NWS, and the Office of Global Programs, OGP) and of the National Science Foundation (NSF) was essential to carrying out the project.

The NCEP/NCAR Reanalysis system was designed, developed and implemented during 1990-1994. The early design was discussed at the Reanalysis Workshop held at NCEP in April 1991 (Kalnay and Jenne, 1991). The reanalysis system required a completely different design from NCEP operations: the goal was to perform one month of reanalysis per day in order to carry out a 40-year reanalysis in just a few years. Several new systems which were developed for this project and later ported to NCEP operations, including a new BUFR archive designed to keep track of additional information about each observation ("meta data") and an advanced quality control (QC) system. The model was also upgraded, and was identical (except for having half the horizontal resolution) to the system that became operational on 25 January 1995. The most difficult task, made easier by the collaboration with NCAR, was to assimilate data that came from many different sources, with very different formats, and to quality control them (Woollen and Zhu, 1997, Ebisuzaki, 1997, Kistler et al, 1997).

The actual reanalysis started in June 1994. CDAS (performed soon after real time with a system identical to the reanalysis) became operational in 1995, and since then it has been extensively used for climate monitoring by the Climate Prediction Center and many other groups. The period 1968-1996 was completed in early 1997. The reanalysis period 1957-1967, was started in July 1997 and finished on 13 October 1997. This completed the 40 years originally planned. Following the recommendation of the Advisory Panel, the decade 1948-1957 was also reanalyzed during 1998, although these oldest data presented many additional problems, such as a different observing schedule and coverage primarily over NH land. This (combined with CDAS which extends it to the present) completes more than 50 years of reanalysis.

There are two major products of the reanalysis. The first is the 4-dimensional gridded fields of the global atmosphere, the output most widely used by weather and climate researchers. It also includes 8-days "reforecasts" performed every 5 days. The second product is the Binary Universal Format Representation (BUFR) archive of the atmospheric and surface observations for the 5 decades of reanalysis. The BUFR archive includes, for each observation, additional useful information about each observation, such as the value of the 6hr forecast, analysis, and a climatology, collocated with each observation, as well as quality control decisions (meta data). In the process of preparation, and during the reanalysis, many errors were identified in the data archives, and many were corrected. This information, and the meta data included in the BUFR archives, will be invaluable for carrying out improvements in the next phase of reanalysis planned by both NCEP and the European Centre for Medium-range Weather Prediction (ECMWF).

In this paper we briefly describe the data distribution (Section 2), review the data sources (Section 3), and present examples of applications to climate and numerical weather prediction (Sections 4 and 5). We also discuss problems and errors discovered and their impact (Section 6), present comparisons with other reanalyses (Section 7), and plans for the future (Section 8).

From the beginning of the project, a special effort was made to ensure the widest possible distribution of the products to researchers through the use of tapes, Internet, and CD-ROMs. NCAR (Data Support Section, headed by Roy Jenne) is the main distributor of tapes and CD-ROMs. The data is also available from NOAA's Climate Diagnostics Center (CDC), from NOAA's Climate Data Center (NCDC), and from NCEP's Reanalysis Home Page. Their Internet addresses can be linked from the NCEP Reanalysis address (wesley.wwb.noaa.gov/reanalysis.html). Tens of terabytes of reanalysis data have already been distributed through tapes, or downloaded from the Internet by researchers from all over the world.

Following the suggestion of the Advisory Panel, a special effort has been made to create CD-ROMs that are readable from both workstations and personal computers (PC), containing fields that satisfy the needs of the majority of the users. These CD-ROMs include easy-to-use demonstration interactive data displays developed by Wesley Ebisuzaki based on the use of the Graphic Analysis and Display System (GrADS) created by Brian Doty at the Center for Ocean Land and Atmosphere interactions (COLA). A special CD-ROM with 13 years of early reanalysis results was included with the Bulletin of the American Meteorological Society (AMS) March 1996 issue, and was distributed to the 13,000 members of the AMS. The present article is also accompanied by a 40-year reanalysis CD-ROM similar to the one the one issued with the AMS Bulletin of March 1996, but with 40-years (1958-1997) of monthly means of reanalysis and estimates of precipitation derived from satellite and in situ observations. Selected monthly fields for the earlier decade (1948-1957) are also included, as are data distribution information. Glenn White further enhanced the interactive displays. The content of the attached CD-ROM is listed in Appendix 1.

In addition, "one-per-year" CD-ROMs with a large number of fields on pressure, sigma, and isentropic coordinates are also available form NCAR. At the time of this writing more than 30 years have been created and the rest of the 50 years should be available in the near future (as announced on the web page). Following a suggestion of John M. Wallace, Glenn White prepared special CD-ROMS: 1) "Global Synoptic 79-96", with winds, heights, vertical velocity at 00Z, and 2) "NH Synoptic 79-96", with winds, heights, vertical velocity, precipitable water, precipitation, isentropic potential vorticity at both 00Z and 12Z. Additional CD-ROMs will include 3) CD-ROMs containing forecasts (8-days long, every 5 days), and 4) rawinsonde mandatory pressure data with some meta data. Others may be prepared if the research community suggests their need.

The reanalysis system, described in more detail in Kalnay et al (1996) includes the NCEP global spectral model operational in 1995, with 28 "sigma" vertical levels and a horizontal triangular truncation of 62 waves, equivalent to about 210km. The analysis scheme is a 3-dimensional variational (3D-Var) scheme cast in spectral space denoted Spectral Statistical Interpolation (Parrish and Derber, 1992). It uses upper air observations of temperature, horizontal wind and specific humidity (Bob???), and surface observations of Bob????. The quality control systems are discussed in the next subsections.

The collection and consolidation of meteorological observations was a major thrust of this project. In this subsection we present background information and graphic inventories of some of the of the source archives. The reanalysis observation event archive is described, and an inventory shown of quality controlled observations used in the reanalysis assimilation (Woollen and Zhu, 1997).

Observed data are obviously the most essential component of atmospheric analyses. NCAR supplied most of the observations to the NCEP/NCAR reanalysis project (Jenne and Woollen, 1994). An enormous effort at NCAR went into the "archeology" required to make the historical archives scheduled for delivery to the project suitable for processing in a modern assimilation system. Additional source archives were obtained from the Japanese Meteorological Agency (JMA), the National Environmental Satellite Data Information Service (NESDIS), and the European Centre for Medium-Range Weather Forecasts (ECMWF). The archives, as received at NCEP, contained a variety of formats and contents.

A two step process at NCEP was developed to prepare observations for assimilation. The first step translates them into the Binary Universal Format Representation (BUFR), filters out unused data, and sorts the observations into six hour synoptic sets, by report type and source. These are referred to as the BUFR source archives. The second processing stage consolidates all report types into monolithic synoptic files, ready for assimilation, with duplicates removed within and across the different sources. These files constitute the BUFR events archive, as additional processing "events" (meta data) are recorded into them during the assimilation.

In what follows, we first list most of the archive sources incorporated into the reanalysis project. Then we present brief inventories of some of the BUFR source archives generated, as well as temperature, wind, and surface pressure observations from the BUFR events files which passed the quality control (Woollen et. al., 1994), and were assimilated.

The following table lists source archives that account for the bulk of the source archive data received at NCEP for the reanalysis project. It also includes a mnemonic name and a brief description of their origin. All sets, with the exception of the FGGE, ECM, and Japan Meteorological Agency (JMA) data, came to NCEP from NCAR. The mnemonics are used as legends in the graphs of the inventories in Section 3.1.b.

Table 1: Mnemonic and description of the major sources of data used in the reanalysis.

Table 2: Inventory of sources found in the special data, listed in descending order of report counts.

Several peculiar communications episodes were discovered in the processing of the data, and some remain unexplained. One of the most remarkable is a rather recent very long episode associated with the transmission of the data through GTS. Radiosonde bulletins are coded for the GTS in separate sections for mass and winds, mandatory or significant levels, and troposheric or stratospheric profiles. These "parts" are identified by character headers (such as TTAA) and must be assembled by the receiver of the GTS data to form complete observation profiles. For a period of 27 months from Aug 1989 through Sep 1991, there was a noticeable drop in inventories of total radiosonde/pibal levels from the NCEP operational archive, suggesting that the NMC decoders of the time had either lost, or never received, some of the bulletin parts. The significant level wind data (PPBB and PPDD) decoded at NMC was missing reports from the entire world, except for North America and China. We believe the NMC decoders never received the data for the following reason. A comparison with the ECMWF archive for the same period, revealed that ECMWF received data complementary to NMC: ECMWF had the missing PPBB/PPDD wind reports from the rest of the world, but they were missing instead the North America and China data (WMO blocks 50-59, 70-79)! Comparison of the rest of the report contents showed very good agreement between the two centers. In the NCEP reanalysis the data set was completed for this period, but this problem emphasizes the need to perform a merge of the data from NCEP and ECMWF for future reanalyses.

The total number of radiosondes used by the Reanalysis reflects mostly how many were operationally available at NCEP (ex NMC). At the main observation times (Fig. 3.5), there was a general increase from about 15,000 per month in 1958 to a maximum of about 22,000 per month in 1990. Since then, there has been a decrease to about 18,000 per month, mostly due to the reduction from about 5,500 to 2,500 observations per month at 00 and 12 UTC for the former USSR (WMO blocks 20-38). The counts of reports received at 06 and 18 UTC are both much lower and variable than at the main observation times. The largest number was about 5,000 per month in early 1970, increasing from about 250 per month in late 1969. The increase was also due mostly to former USSR stations.

Some of the changes to radiosonde counts from individual countries, while not always explainable, are nevertheless of interest. Most regions show a rapid increase from late 1957, and regions including the former USSR, Taiwan, Korea and Japan show a rapid and unexplained decrease in May 1963. For various regions, and for different report times, there appears an annual variation to the report count. The variation may be fairly regular for a decade or more and then disappear. One place with a large annual variation, not unexpectedly, is Antarctica. In the mid-1970's the count varied from about 100 to 200 at 12 UTC, with a maximum around December.

Some larger regions, such as Western and Eastern Europe, show jumps up and down in the radiosonde count, probably related to political or economical changes. In general, the jumps have left Western Europe with about the same count of around 2,000 per month at the main observation times in 1997 as in 1958, while the count in Eastern Europe has decreased from 2,500 in 1958 to about 500 in 1997. Central America, WMO blocks 76 and 78, showed a step increase from about 100 to 600 per month at the main observation times in January 1979. A few regions have shown a fairly steady increase in the number of radiosonde observations over the years. The following numbers are monthly counts for the main observation times. India has increased from 500 to 900; South Africa has increased from 200 to 350; at 00 UTC the Pacific region (WMO blocks 91, 96-98) has increased from less than 200 to 1100; and China has increased from about 1,500 to 4,000, but then decreased since 1990 to about 3,800.

The CQCHT used in the Reanalysis processing makes corrections to several different kinds of errors when they are detected (Collins and Gandin, 1990). All of these errors are accompanied by hydrostatic inconsistencies between the reported heights and temperatures. The complex of hydrostatic residuals--the difference between a layer thickness calculated from heights or hydrostatically from the temperatures--is used to suggest corrections. The suggested correction is only applied if it is corroborated by residuals (differences between expected and reported value) from the other checks made by CQCHT, namely, increment check, vertical optimal interpolation (OI) check, and horizontal OI check. Of the corrections made, the most accurate are an isolated height correction or isolated temperature correction, referred to as type 1 and type 2 error corrections respectively. In what follows we discuss mostly the numbers of these errors for the period 1958 to 1997.

Since these variations are similar for all regions, it is clear that they were introduced by changes in either WMO encoding (followed by all countries) or by the data processing at NCEP. In fact, both dates correspond to major operational changes at NCEP.

On January 1, 1973, the encoding format was changed, and the data was first operationally stored at NCEP in Office Note 29 (ON29) format. This was the first modern meteorological data archival system, and was the model for the eventual BUFR format adopted by the WMO. However, it is likely that the introduction of a new format and new decoders produced most of the problems. The decrease in type 1 and 2 errors in the subsequent two decades was partly due to improvements to the upper-air decoder.

In the next section, we show that despite the fact that there were many more errors introduced with the changes of 1973, the forecast skill from the Reanalysis forecasts showed a remarkable improvement starting this year. This is presumably because the new data format and the modern nature of ON29 included sufficient data to avoid errors introduced with the previous format. It is important to note that the Reanalysis was carried out using a modern QC system, including the Complex Quality Control which corrected a large portion of the errors in the rawinsondes heights and temperatures, and an Optimal Interpolation based QC (OIQC) which monitored all other data, including rawinsondes winds. A comparison with long operational scores (Shuman, 1989, Kalnay et al, 1998) indicates that, in contrast to the Reanalysis results, the operational scores actually were actually significantly poorer in 1973, and that it took three years to recover to the 1972 level. Since the operational forecasts did not have the benefit of the current QC system, these results provide strong circumstantial evidence of a major positive impact from the QC on the rawinsonde and other data, an important additional benefit of the reanalysis.

In late 1996 there was another operational change at NCEP, which was incorporated in the Reanalysis processing beginning in March 1997. A CRAY J-90 computer--a change from MVS to Unix operating systems replaced the NAS 9000 series computers used for data processing, necessitating the rewriting of almost all decoders, including the upper-air decoder. Also introduced at the same time was the operational use of the data in BUFR format, but that change did not introduce the increase in error types 1 and 2 in March 1997 since the Reanalysis had been using this new format from the beginning. We can conclude that any major change is likely to be accompanied by undesirable side effects that take time to work out, but the magnitude of the "side effect" of the ON29 introduction was unexpectedly large, and only the Reanalysis was able to extract the benefit of the new formats.

There are also events that cause a short-term large increase in a particular error type. The magnitude of these events makes it unlikely that they are random. One such example is the peak of over 800 errors per month at the top level for Western Europe (WMO blocks 1-10, 16) in mid-1987. In adjacent months, the count is about 50 errors per month. Since this peak only occurs for one region, it is likely due to an error within that particular region. A similar event happened with the data from Taiwan/Korea/Japan (WMO blocks 45-47) in August 1980. Errors due to possibly both height and temperature at the same level jumped from about 10 to 1,100 and back again within 5 months. This error was local and the cause is not known. There was a similar jump in this same error type for the U.S., peaking at 2,500 errors in July 1983. Another short-term peak of error at the top level occurred in 1987. In this case, the increase is found for most regions. The global monthly sum of this type of error jumped from about 1,100 to 5,500 and back. This error is probably due to a processing or decoder change (error) within the U.S National Weather Service, which was corrected fairly soon.

In a few cases, the cause for an increase in the number of a particular error type is known. For example, in the mid-1980's Canada began automated processing and entry of radiosonde data into communication lines, using mini-computers. The processing code was included in the computers' Read-Only-Memory. This code placed the radiosonde observations at the nearest hPa, after solving hydrostatically for heights. Above 100 hPa, and most noticeably at the top reporting level, this led to hydrostatic errors of up to a few hundred meters. The number of hydrostatic errors at the top level increased from a very small number to 300-600 per month. NCEP's first complex quality control program for radiosonde heights and temperatures, during its testing for implementation, discovered this problem in 1988. The Canadian Weather Service was informed of this error which was soon after corrected (not shown).

On June 1 1957, with the beginning of the International Geophysical Year (IGY), the World Meteorological Organization (WMO) made several major changes. This included shifting the upper air observing times from 03UTC, 09UTC, 15UTC and 21UTC to 00UTC, 06UTC, 12UTC, and 18UTC respectively, the same major synoptic times already used for the surface observations.

Given this change, we decided to perform the NCEP/NCAR Reanalysis at the observing times for upper air data for the period 1948-June 1957 (03UTC, 09UTC, 15UTC, 21UTC). However, in order to facilitate user comparisons with the post-1957 reanalysis, the 3-hour forecasts are saved as well (at 06UTC, 12UTC, 18UTC and 00UTC). The diagnostic files were maintained at 06UTC, 12UTC, 18UTC and 00UTC for continuity with the following 40 years of reanalysis. In order to maintain the philosophy of a constant data assimilation/forecast system throughout the reanalysis, no attempt was made to modify the forecast errors statistics used in the 3-dimensional variational data assimilation system (3D-Var) used in the Reanalysis (Parrish and Derber, 1992). Since the forecasts for this early stage are of poorer quality, this decision implies that in the pre 1958 period, the information from the data was not optimally extracted: the forecasts were given relatively more weight than optimal. This also affects the apparent skill of the reforecasts, since they are 3 hours longer than after 1957 (Bob). When a long reanalysis is performed a second time, the information already gathered in the first reanalysis will allow to reconsider the pre-1958 period and estimate the forecast error covariances as they evolve with changes in the observing systems.

In order to provide the user with information about the data availability for different periods of the Reanalysis, we have prepared detailed monthly data density maps available at the NCEP Reanalysis home page, and monthly summaries are also available in the attached CD-ROM. The observations are counted in every 2.5 deg latitude-longitude box, and in the following figures they are also zonally averaged. Fig. 3.8 shows the number of rawinsonde and pibal observations from 1946 to 1997, indicating a major increase in the NH mid-latitudes in 1948. It was this observation that led us to extend back the reanalysis to 1948. In the SH the first upper air observations (in Australia **??**) started also around 1948, and in 1967 there was a major increase. Fig. 3.9 shows the distribution of aircraft data currently mostly around 250 hPa. The special observations taken in the FGGE year of 1979 are also apparent. The surface observation density is shown in Fig. 3.10, with major increases in the SH around 1950 and in 1979. (**Glenn: I think your plot allows only a max contour of 120**). The satellite data coverage (GW4) is relatively uniform in space and time starting in 1979 with the FGGE experiment. More detailed information about data distribution is available in the accompanying CD-ROM.

The information on data distribution is very valuable in assessing the reliability of the reanalysis. However, the data distribution alone does not take into account the ability of 4-dimensional data assimilation systems to transport information from data rich to data sparse regions. As a result, as will be shown by the forecast results in the next section, the disparity between the quality of the reanalyses in the NH and those in the SH before the advent of satellite data in 1979 is considerably smaller than the data density alone would suggest.

There were two major changes in the observing systems in the last 50 years. The first one took place during the period 1948 - 1957, when the NH upper air network was first established, and culminated in the International Geophysical Year of 1957-58. The NH network was relatively stable after that time, but the tropics and SH were very poorly observed. The second major addition took place with the First Global GARP (Global Atmospheric Research Project) Experiment (FGGE), run from Dec 1978 - Nov 1979, which introduced several innovative observing systems relying on spacecraft sensing and communication to provide unprecedented global observation coverage and timely data receipt. During and after FGGE, many satellite data impact tests were executed, and generally reported strongly positive results in the Southern Hemisphere, but little impact in the Northern Hemisphere. (See Mo et al., 1995 for a summary.)

The question addressed in this section is "How well does a present-day assimilation system fare when posed with the observational database of the pre-FGGE period?"

One measure of the quality of an assimilation system is the fidelity with which it represents the observations it is attempting to assimilate, i.e. how well do the analysis and guess "fit" the data. Among the tools created to monitor the reanalysis are profiles of tropospheric and stratospheric mean and rms differences of the guess and analysis with the raobs (referred to as ADPUPA in the NCEP system). These are summarized over time, and subdivided into latitude bands representing the northern (30N-60N) and southern (30S-60S) extratropics and the tropics (30N-30S). The four panels of Figure 4.1 compare these profiles in both the troposphere (1000-100mb, bottom panels) and stratosphere (100-10mb, top panels) for averages over years 1958 (left panels) and 1996 (right panels), for temperatures and vector winds. The solid profiles are the fits to the 6-hour forecast, the dashed, fits to the analyses.

In the stratosphere the fit to the winds improve between 1958 and 1996 by about 1 m/sec in the NH, and 2-4 m/sec in the SH, but only by about 0.5 m/sec in the tropics. The fit to raob temperatures is actually worse, due to the fact that the analyses fit better the more abundant satellite retrievals, which do not have the vertical resolution needed to resolve the tropopause. In the troposphere, the improvements to the raob fits are large in the SH extra tropics, 2-4 m/sec in the winds and 1-2K in the temperatures. In the NH extratropics, the improvements are much smaller. In the tropics there is a significant improvement in the near the tropopause, presumably, from the use of cloud tracked winds.

Fig. 4.2 shows the geographical distribution of the 6hr forecast rms increment (analysis minus first guess) of 500 hPa heights for 1958 and 1996. In 1996 the increments are rather small, uniform, depending mostly on latitude. In the tropics they are about 5m, and in the NH extratropics generally between 5 and 10m except over the Himalayas where they reach more than 15m. In the SH extratropics the 1996 increments range between 10-18m and are rather uniformly distributed. In 1958, by contrast, the increments were largest in regions with rawinsondes, downstream of data sparse regions, indicating large forecast errors, and almost zero over the South Pacific and South Indian oceans indicating lack of data to update the first guess in these regions. The contrast is particularly clear in the SH extratropics, where rms increments over land were generally 10-15m in 1996, and up to 35m over high latitudes rawinsonde stations in 1958.

In the day-to-day analysis increments over North America (Fig. 4.3), we see that the short range forecast errors have an enormous variability in the synoptic time scales of 2-3 days, a situation virtually identical in 1958 and in 1996, except that the errors were about 10% larger in 1958. These dynamically driven "errors of the day" should be taken into account in operational systems and are equally present today as they were with the observing systems of 40 years ago!

In the reanalysis testing phase, a SAT-NOSAT impact test was run and reported (Mo et al., 1995), covering only the period Jul-Aug 1985. While the results were encouraging, extrapolating the results to the pre-FGGE period was problematic. During the execution of the reanalysis we took advantage of transition periods to extend the SAT-NOSAT tests by reanalyzing the whole year of 1979 without satellite data (NOSAT). This provides a benchmark to assess the impact of the change in observing systems that started in 1979 upon the reanalysis.

Fig. 4.5 shows the monthly average 200hPa meridional wind for January and February 1979 from the analysis using satellite data ("SAT"), and for a "NOSAT" reanalysis, which simulates the observing systems prior to the advent of satellite data. The meridional circulation which varies substantially from month to month, was very anomalous over South America in January 1979, indicating the presence of large-amplitude, short-wave stationary Rossby waves discovered during FGGE (Kalnay et al, 1983). It is apparent that these anomalous waves were detected equally well even in the absence of satellite data. Similar agreements of monthly anomalies are obtained for most fields, with the exception of the regions above 200hPa and/or South of 60S, where the agreements of fields with and without satellite data are poorer (Mo et al, 1995).

The impact of more recent improvements in the observing systems is more gradual and rather small. In the winter scores (not shown) there is a very significant increase in skill in the NH after FGGE for the forecast range 3 days and beyond. This can be attributed to positive satellite data impact on oceanic analyses, with the downstream effects taking several days to have an influence.

Fig. 4.7 also shows that in the SH, the impact of improvements in the observing systems is much clearer than in the NH. In the Reanalysis the AC for the 5-day forecasts increases from less than 50% before the advent of satellite data to well over 60%. The impact that the continuous improvement of NESDIS TOVS retrievals had on the SH forecasts is also very apparent. The AC for the period before 1958 are indicated with dots to emphasize that this period is less reliable because of the change in observing schedule (which lengthens the forecast by 3 hr), the lack of observations in the SH, and the less than optimal analysis. In the NH a much lower AC reflects these factors in the earliest period. In the SH, the spurious very high AC simply reflects the fact that, without data, the analysis is essentially given by the forecast.

Many climate and weather prediction studies have already used the Reanalysis data and contributed to an assessment of its ability to capture anomalies in the atmospheric circulation. Wallace et al. (1998), for example, provided a comparison of the ENSO signal in precipitation, tropospheric temperature and surface stress estimated from COADS and MSU data and from the NCEP/NCAR Reanalysis. In this section we show a few climate applications and raise caveats that should be considered.

The 40-year (1958-1997) NCEP/NCAR reanalysis are essentially free of the inhomogeneities due to changes in model resolution and physics that plagued other data sets, such as the Climate Diagnostics Data Base (CDDB), previously used for real-time climate monitoring. Continuity of the reanalysis system into the current climate evolution has been maintained by running the same data assimilation system operationally since 1995 (CDAS).

Other climate monitoring products that have been developed or improved using the CDAS/reanalysis archive include: zonally averaged zonal wind in the stratosphere to monitor the quasi-biennial oscillation (QBO), east-west vertical cross-sections depicting the mean and anomalous divergent zonal circulation in the tropics, north-south vertical cross-sections of mean and anomalous divergent meridional circulation in the vicinity of the entrance and exit regions of the North Pacific jetstream, and mean and anomalous velocity potential and divergent component of the wind.

Fig. 5.6, on the cover of this AMS Bulletin, shows the monthly zonal mean of the zonal wind at the Equator for the 50 years of reanalysis. The QBO is quite apparent, although in the first decade the available rawinsonde observations were not enough to determine its strength, probably because the model forecast was given too much weight for that decade (see Section 3.3). Because of the sparsity of equatorial rawinsondes, the amplitudes are somewhat underestimated even in later years, but the reanalysis provides for the first time a long global indication of the timing and 3-D structure of the QBO. (see also Pawson and Fiorino, 1998)

Recently, Chelliah and Ropelewski (1999) did an extensive analysis of the tropospheric temperatures and decadal trends(for the decade of the 1980's and early 1990's) from all three reanalyses (NCEP/NCAR, ECMWF and DAO) as compared to the same from MSU and discussed the utility and uncertainties in them, in the context of global climate change detection.

Fig. 5.11a shows the climatological annual cycle in surface pressure. The NH pressure is lower than the SH because it has more land, although the average land height is larger in the SH than in the NH. Globally, the surface pressure is lowest in Dec/Jan (984.74) and highest in August (985.19), a difference of 0.45 hPa. The figure also indicates that the two hemispheres exchange mass in monsoon-like fashion, so that throughout the annual cycle, the warmer hemisphere has lower pressure by 1 or 2 mbar. A comparison with Fig. 7.3 of Peixoto and Oort (1992), shows qualitative good agreement, but larger global amplitude and smaller hemispheric amplitude. The variable moisture loading in the atmosphere, which peaks in July when the global mean temperature is highest, causes this annual variation in global mean surface pressure (Peixoto and Oort 1992, Van den Dool and Saha 1993, Trenberth and Guillemot 1994). The Reanalysis global mean net Evaporation minus Precipitation is fairly consistent with the time derivative of global mean pressure (Van den Dool et al 1995), and agrees closely with Trenberth and Guillemot (1994). A similar but more pronounced monsoonal exchange, dominated by the NH, takes place between land and ocean so that pressure is lower over land when it is warmer compared to the ocean (Fig. 5.11b).

In this section we briefly review some of the uncorrected problems that have been uncovered in the reanalysis. These problems, and many other errors that were corrected in time, were discovered through internal NCEP monitoring and by outside users who had access to early results. Some of the problems were inevitable, such as those due to changes in the observing systems or to model deficiencies whose improvement is a long-term project. Some were mistakes that were corrected once they were discovered (a.k.a. bugs), but when they affected periods longer than a few months, it was not possible to rerun the reanalysis with the corrected version. We have tried to make the users aware of these problems, and more detailed information is available in the Reanalysis home page. Many problems were also discovered in the observations themselves, and both corrected and uncorrected problems were reported back to NCAR, so that future Reanalyses will benefit from this a priori knowledge. The meta data accumulated in the BUFR archive will also be very useful in this respect.

Many factors affect the accuracy of the analysis. One of the most important is the fact that the model-derived products are obtained from very short-range forecasts, when the problem of spin-up of the moisture cycle, due to differences between the model and the real atmosphere, is highest. The use of a 3D-Var analysis scheme, which includes linear balance as part of the analysis cost function, and not as a separate step, reduced to a large degree problems of spin-up, but they are still not negligible. For this reason, we have advised caution in the use of model produced variables classified as "C", such as surface fluxes and precipitation (see Kalnay et al, 1996 for a complete classification).

The "A" variables (e.g., upper air temperature and wind) are determined primarily by the observations. Both the model and observations influence variables classified as "B". They include all moisture variables and variables near the surface. Table 3 gives a summary of the observation types actually used by the assimilation system. When the model is improved, variables "B" and especially "C" are also improved. For example, Figs. 6.1 and 6.2 show the changes in the surface temperature and precipitation between Reanalysis and an experimental reanalysis using an updated version of the forecast model (see Section 8). There are few observed humidity measurements in the upper tropical oceanic troposphere, and this region is sensitive to the parameterized tropical convection.  Fig. 6.3 shows the difference between the 300 mb relative humidity between Reanalysis and the experimental analyses which uses a different convection scheme.

As discussed in previous sections, the density of the observing systems and its changes also affects the accuracy and consistency of the reanalysis. The two major changes in observing systems, the increase of the upper air network during 1948-1958, and the introduction of the satellite observing system in 1979, also affected the reanalyzed climatology. In a data sparse area, even the change of a single observation can result in a significant bias (Fig. 5.5).

The previous paragraphs described some of the problems inevitable in all data assimilation systems.  In addition there were human errors made in the assimilation, mostly concerned with input data processing which were not corrected because of lack of time to repeat the period of reanalysis affected by the error. During 1974-1994, binary snow cover corresponding to 1973 was used by mistake.  This error has its largest impact near the surface over regions where one mask is snow-free and the other is snow covered.  The effects of an incorrect snow cover are typically local and only a few degrees, although differences of up to 5 degrees have been found in comparison with an experimental reanalysis.  An examination of the various snow masks suggests that North America has the most impact in transition seasons (October in particular), less in winter and the least in the summer. 

An important but inevitable variant of this problem occurs for the years when observed snow cover was not available, prior to 1973 in the NH, and throughout the reanalysis in the SH.  In the Reanalysis we used climatological snow cover in the SH, and modeled it in the NH for 1948-1972.

Sea-level pressure PAOBS are the product of Australian analysts who estimate the sea-level pressure using satellite data, conventional data and time continuity for the data-poor Southern ocean.  PAOBS are used in the current NCEP operational analyses but with low weights (the observation errors for PAOBS are assumed to be **hPa compared to (Bob) ** for stations), and are not used at all at ECMWF. Unfortunately, in the NCEP/NCAR Reanalysis the use of a different convention for longitude led to a shift of 180^o in the use of the data for 1979-1992.

The original reanalysis was repeated for 1979 with correctly located PAOBS, allowing an assessment of the impact of this error, which turned out to be relatively small for three separate reasons: 1) The weights given to the PAOBS are small compared to other surface observations. 2) Due to geostrophic adjustment the assimilation system does not "believe" surface pressure observations, especially in the tropics. The sea level pressure is changed in the analysis to agree with the PAOBS, but this change tends to quickly disappear during the 6-hour forecast. 3) The PAOBS with largest differences with the first guess were eliminated by the OIQC. The comparison with the corrected analysis led to the following conclusions: (A) The NH was not affected at all. (B) The SH was significantly affected only poleward of 40S. (C) The largest differences were close to the surface and decreased rapidly with height. (D) Differences were small on the global scale but significant on the synoptic scale. (E) Differences decreased rapidly as the time scale went from synoptic to monthly (because of 3). (F) Geopotential quadratic quantities of the type are affected in the monthly means, but cross products like or are not affected). (G)The RMS difference in the 500 hPa heights were of similar magnitude as the difference between the NCEP and ECMWF operational analyses south of 40S (a measure of uncertainty in the analysis). (H) The RMS difference in the 850 hPa temperature was smaller than the RMS difference between the NCEP and ECMWF operational analyses.  In summary, SH studies using monthly mean data should not be adversely affected (except for quadratic perturbations of the pressure or geopotential height). Studies of synoptic-scale features south of 40S are affected by the addition of an error which is of a magnitude comparable to the uncertainty of the analyses. This unfortunate error, which affects the reanalysis from 1980 to 1992, was corrected in the AMIP reanalysis (Section 8).

Problems in the forecast model can adversely affect some variables.  For example, the forecast model had a formulation of the horizontal moisture diffusion which caused moisture convergence leading to unreasonable snowfall over high latitude valleys in the winter ("spectral snow", see Fig. 7.9). This problem, discussed in more detail in the home page, is present in the "PRATE" field, but has been corrected a posteriori in the "XPRATE" model precipitation. However, moisture fluxes cannot be corrected a posteriori. Another problem occurred in the sensible heat flux parameterization, which allowed surface sensible heat flux to go to zero if the surface wind vanished. As a result, the surface temperature could occasionally have unrealistically high values.  This parameterization was corrected during the course of the reanalysis.

In summary, the tropospheric winds heights and temperatures are the best analyzed fields.  The other fields are more affected by the changes in observation systems over time as well as the inaccuracies in the model parameterizations.  Human errors have also been made through some of the periods of the reanalysis. Comparisons with observed data and other reanalyses are necessary to determine the reliability of the various fields, and are presented in the next section.

Reanalysis has provided a long, self-consistent estimation of the interchange of momentum, energy and moisture between the surface of the earth and the top of the atmosphere. There is no "ground truth" for most of these fluxes, since they are not directly measured and have to be estimated indirectly from observations and sometimes significantly tuned to ensure net energy balance. In this section we compare monthly mean top of the atmosphere (TOA) radiative fluxes and precipitation as well as surface fluxes from the three reanalyses to each other and to independent estimates.

Da Silva et al. (1994) carefully updated estimations of oceanic heat fluxes from several decades of carefully corrected COADS ship observations. As was the case with previous estimates, the result did not produce a reasonable global heat balance, reflecting at least in part the paucity of COADS data over most of the oceans outside the mid-latitude Northern Hemisphere. Other COADS-based air-sea flux climatologies display similar imbalances. To obtain a reasonable heat balance, da Silva et al. (1994) mathematically tuned the fluxes, effectively increasing the evaporation 15% and decreasing the net shortwave 7% globally. Similarly, Campbell and vonder Haar (1980) found a positive imbalance in the net heating of 9.2W/m², and their "correction" was to increase both reflected solar radiation and outgoing long-wave radiation by 2.5% (Peixoto and Oort, 1992).

Table 4 shows global mean components of the surface energy balance for the three reanalyses for 1981-92, as well as da Silva et al. (1994) untuned and tuned air-sea fluxes averaged over 1981-92, and net shortwave estimates by Darnell et al. (1992) and net longwave estimates of Gupta et al. (1992) averaged over July 1983-June 1991, based on satellite observations of radiation, ISCCP cloud estimates and radiative transfer codes.. *****Glenn: please create table*****

The ERA fluxes shown are from twice-daily 12-24 hr forecasts; GEOS and NCEP fluxes are from the four times daily 0-6 hr forecasts used as the first guess in the analysis cycle. The ERA hydrological cycle intensifies ("spins up") between the 0-6 hr and 12-24 hr forecasts; the global hydrological cycle and surface energy budget are in better balance in the 12-24 hr ERA forecasts than in the 0-6 hr ERA forecasts. The system used in the NCEP reanalysis appears to have less global spin-up than the system used in the ERA reanalysis. ERA's global evaporation is the most different from the untuned COADS estimate, but agrees the best with the tuned estimate. For NSW, the COADS untuned estimate displays the best agreement with the satellite estimate; GEOS displays considerable higher radiative fluxes and lower evaporation than the other estimates. NCEP and ERA are nearly in balance over the global ocean; GEOS and untuned COADS produce a substantial net heat flux into the ocean.

Comparison of the reanalyses with da Silva et al. (1994) reveal similar patterns in long-term means and in annual cycles. Temporal correlations of the estimated evaporation from da Silva et al. (1994) with the NCEP/NCAR reanalysis (not shown) are highest wherever the COADS observations are most abundant (such as over the Northern Hemisphere oceans between 20 and 70N, and off the west coast of northern Africa). The reanalyses include other sources of data over the ocean besides ship reports, and can interpolate and extrapolate other sources of data from land and from other levels in the atmosphere. Operational forecasts display nearly as much skill in the Southern Hemisphere as in the Northern Hemisphere, indicating that modern data assimilation systems do produce accurate daily analyses of the Southern Hemisphere. The differences in correlation between COADS data rich and data poor regions are much less evident in correlations of the reanalyses with each other or with independent estimates of fluxes and precipitation based on satellite data. This suggests that the asymmetry is due to the relative lack of ship observations outside 20-60N and that the usefulness of COADS in defining interannual variability may be largely limited to the Northern Hemisphere mid-latitudes.

A similar pattern of high temporal correlation with da Silva et al. (1994) over regions with high COADS observation density, and low correlation in COADS data sparse regions is observed for evaporation, surface stress and net heat flux for the three reanalyses. The correlations for surface fluxes are summarized in Tables 5 through 7. Tables 5 and 6 show that correlations with da Silva et al. (1994) net heat flux and zonal surface stress are virtually the same for NCEP and ECMWF and a little lower for GEOS. With respect to their geographical distribution (not shown), anomalies in these fields from the three reanalyses correlate well with each other over the oceans except near the poles and the equator; evaporation anomalies from the three reanalyses agree poorly with each other over land. At the international dateline on the equator, ERA evaporation is significantly lower after 1987 than before, a change not seen in the other two reanalyses (White and da Silva, 1998). Evidence of a similar change has been found in other fields from the ERA reanalysis (Stendel and Arpe, 1997), and in fig 7.11. The reanalyses zonal surface stress anomalies (Table 6) display better agreement with da Silva et al. (1994) than they do in net heat flux anomalies (Table 5), especially in the tropics.

The ECMWF reanalysis used a prognostic cloud parameterization tuned to ISCCP clouds, also used in calculating satellite-based estimate of surface NSW by Darnell et al. (1992). The NCEP/NCAR reanalysis used a diagnostic cloud scheme tuned to Air Force nephanalyses, which produce lower cloud amounts than the ISCCP estimates. The surface albedo over the oceans in the NCEP/NCAR reanalysis was too high (da Silva and White, 1995).

It has been suggested that satellite estimates of surface NSW may be more reliable than other global estimates of surface NSW, although satellite estimates of NSW can differ markedly from point surface measurements of NSW (White, 1996a). The time-mean surface NSW estimate from da Silva et al. (1994) agrees better with the satellite estimate than do the reanalyses, despite the simplicity of the COADS parameterization of NSW. The time-mean surface NSW from ERA and GEOS shows little evidence of the influence of low-level oceanic stratus clouds; NCEP's NSW shows more evidence of the influence of low-level stratus cloud (White and da Silva, 1998). The annual cycle in surface NSW in the reanalyses is well correlated with the annual cycle in the satellite estimate, generally exceeding 0.9 away from the equator where the annual cycle is small (not shown). The anomaly correlations of the reanalyses with the satellite estimates of surface NSW (Table 7) are much lower, with the majority of points showing a correlation of less than 0.6. GEOS correlations are lower than the other reanalyses and NCEP are slightly higher than ECMWF except over the tropical ocean. ECMWF has considerably more variability in monthly anomalies of surface NSW in the tropics than does the satellite estimate; NCEP has somewhat less than the satellite estimate.

Table 5: Temporal correlation of surface net flux from the reanalyses with da Silva et al. (1994) tuned estimate averaged over different oceanic regions for 1981-92.

Table 6: Temporal correlation of monthly mean anomalies in zonal surface stress from the reanalyses with da Silva et al. (1994) averaged over different oceanic regions for 1981-92.

Table 7: Temporal correlations of monthly mean anomalies in surface net shortwave from the reanalyses with Darnell and Staylor (1994) averaged over different regions for July 1983-June 1991.

At the top of the atmosphere (TOA) ERBE observations of both short and long wave radiation can be compared with the reanalyses. Fig. 7.1 shows that the reanalyses all have less NSW than ERBE over the tropical oceans, while GEOS has more than ERBE outside the tropics. Geographical distributions of the differences in NSW radiation are similar at the TOA and at the surface (not shown). Since satellites observe TOA NSW directly, the similarity lends credence to the satellite-based estimate of surface NSW. As can be seen in Table 8, ECMWF's annual cycle in TOA NSW agrees slightly better with the annual cycle in ERBE NSW than NCEP's annual cycle whereas NCEP's anomaly correlations with ERBE tend to exceed ERA's.

Fig. 7.2 compares the time mean TOA outgoing longwave radiation (OLR) from the three reanalyses with ERBE observations for 1985-89. NCEP is closest to ERBE, while ERA's estimate is too high and GEOS too low in the tropics and too high outside the tropics. Fig. 7.3 compares the standard deviation of monthly anomalies from the annual cycle. ERA and GEOS display too much variability in the tropics; NCEP has too much variability in mid-latitudes and less over the tropical oceans. Table 9 shows the anomaly correlations of TOA OLR observed by ERBE during 1985-89 with the reanalyses. Correlations of the reanalyses anomalies with ERBE anomalies are considerably higher for TOA OLR than for TOA NSW (cf. Table 7.5).

Table 8: Temporal correlations of monthly means over a) land and b) ocean and of monthly mean anomalies over c) land and d) ocean in TOA net shortwave from the reanalyses with ERBE averaged over different regions for 1985-89.

Table 9: Temporal correlations of monthly means over a) land and b) ocean and of monthly mean anomalies over c) land and d) ocean in TOA OLR from the reanalyses with ERBE averaged over different regions for 1985-89.

Stendel and Arpe (1997) thoroughly evaluated the hydrological cycle in the three reanalyses. Janowiak et al. (1998) displayed the correlation of precipitation anomalies estimated by GPCP with monthly mean precipitation anomalies from the NCEP reanalysis for 1988-95. Anomaly correlations of monthly mean precipitation estimated by Xie-Arkin with the ERA, GEOS and NCEP reanalyses during 1981-92 display similar patterns. Higher correlations appear over land areas where raingauges and radiosondes are generally more abundant. Low correlations appear over convective regions and over land areas where raingauges and radiosondes are relatively sparse and the reanalyses and independent precipitation estimates therefore are all the more uncertain.

Table 10 presents temporal correlation of precipitation with Xie-Arkin estimates, for the full fields of precipitation, including the annual cycle (a, b), and for the anomalies (c, d). ERA has the highest correlations except over tropical and Southern Hemisphere land areas where GEOS has the highest correlations. RMS differences from the monthly means of Xie-Arkin are shown in Table 10. Of the three reanalyses, ERA has the lowest RMS difference over the Northern Hemisphere continents and the extratropical oceans, but the largest RMS difference over the tropics, especially over land. NCEP has the lowest RMS differences in the tropics and the largest in the Northern Hemisphere. Also shown are comparisons between GPCP and Xie-Arkin estimates for the years 1988-94. Since GPCP and Xie-Arkin use similar methods of estimating precipitation, they should be regarded as a lower estimate of the uncertainty in Xie-Arkin.

Table 10: Temporal correlations of monthly means over a) land and b) ocean and of monthly mean anomalies over c) land and d) ocean in precipitation from the reanalyses with Xie-Arkin averaged over different regions for 1981-92. Also shown are correlations of monthly means from GPCP with Xie-Arkin for 1988-94.

Table 11: RMS difference in monthly means over a) land and b) ocean in precipitation between the reanalyses and Xie-Arkin averaged over different regions for 1981-92. Also shown are RMS differences in monthly means between GPCP and Xie-Arkin for 1988-94.

4. Summary of surface and TOA comparisons

Surface and TOA fluxes from the NCEP and ERA reanalyses generally show very similar levels of agreement with independent estimates; GEOS fluxes show somewhat less agreement. The annual cycle in ERA tends to agree slightly better with independent estimates than NCEP's annual cycle whereas NCEP's anomalies agree slightly better in general with independent estimates. ERA frequently displays better agreement with independent estimates over the tropical oceans; NCEP often displays better agreement over the tropical continents. NCEP appears to underestimate variability over the tropical oceans; ERA appears to overestimate variability in the tropics.

ERA appears to be somewhat better in depicting monthly precipitation patterns than NCEP or GEOS; NCEP may be somewhat better than GEOS over the oceans. All three reanalyses have more rainfall over the tropical continents than the independent estimates; ERA also exceeds the independent estimates over the tropical oceans.

The reanalyses appear to have large biases in TOA and surface NSW as well as in precipitation. Whether the biases in NSW reflect primarily problems in the parameterization of cloudiness or problems in the hydrological cycle is not clear. Why the reanalyses TOA OLR show better agreement with ERBE measurements than do the reanalyses TOA NSW is also not clear.

As shown in the previous subsection, intercomparison is a powerful tool in assessing the reliability of the reanalyses. The availability of two complementary reanalyses and has proven invaluable to both centers, particularly for finding errors. In this section we compare the seasonal mean climate during the ERA-15 period of 1979 - 1993 to give a measure of uncertainty in the reanalyses in climate scales. We also examine decadal linear trends in the common 15 years of the NCEP and ERA-15 reanalyses and show that both large and subtle changes in the observing system can create "noise" modes that can be large as the sought after signal.

In Kalnay et al (1996) we heuristically classified reanalysis variables into three classes: 1) "A" variables that are directly observed (e.g., wind); 2) "B" variables that are influenced by observations, but also by the model (e.g., moisture); and 3) "C" variables that are wholly determined by the model (e.g., precipitation). The purpose of this section is to quantify this notion for climate averaged quantities. We will see that this can change the classification of variables. For example, although the 3-D field of meridional velocity is classified as "A", its zonal average is the purely divergent mean meridional circulation, which is strongly influenced by the model parameterizations, making this a field of the type "B" or even "C".

While the 15-year ERA period is somewhat short, it has the advantage that the observing system is relatively constant (satellite data available) and had improving quality (e.g., fit of the analysis to radiosondes, see Uppala, 1997). Differences between the seasonal means from the two reanalyses are caused by differences in the assimilation global model and methods of analysis, and are thus a form of uncertainty. These differences do not directly account for uncertainty due to error in the underlying observations, but they are a practical measure of the state of our knowledge and should be used when applying reanalysis to climate problems.

We only consider the December, January, February (DJF) season, and examine means and year-to-year or interannual variability. The variability is of high interest to ENSO studies and seasonal prediction. We use the pooled total temporal variability for scaling the differences between the reanalyses, much like a t-test, and express the scaled difference as a percentage. The scaled difference shows areas such as the tropics where small differences may be actually very significant.

1. Comparisons of "A" and "B" variables

There are, however, some important differences. The dryness in Western Brazil in ERA-15 was due to an imposed linkage between soil moisture and surface specific humidity and the assimilation of surface pressure (Kallberg, 1997). The surface pressure observations (20 or so) that were responsible for this dryness were "blacklisted" starting in 1987. Thus, an imposed change in the ERA-15 observing system, to correct the desertification, resulted in a large swing in the precipitation (Fiorino et al., 1999). Over central Africa, as in Brazil, the NCEP precipitation is also more realistic. There are large differences in the Indian Ocean and Western Pacific where ERA has lighter rain with several "hot spots" compared to CMAP and NCEP. The ERA has lighter, more realistic precipitation in the subtropical ridges. NCEP captures better the dry shadows west of the SH continents, but it is too dry over Australia, and ERA has a more realistic South Pacific Convergence Zone.

Reanalysis allows an easy calculation of long-term trends in high-interest variables such as free atmosphere air temperature (e.g., Chelliah et al, 1998, Santer et al., 1999). However, as pointed out repeatedly, changes in the observing systems within the reanalysis may obscure climate changes (e.g., Basist et al, 1997, Fiorino, 1999). This climate noise has to be assessed and perhaps corrected before unbiased climate assessments can be made. Agreement between two reanalyses in the climate trend is obviously a necessary but not sufficient condition in order to have confidence in climate trends.

We discussed the many different data origins and their processing into a single BUFR format, including the generation of meta data that can be useful to assess their quality. This information should be very valuable for future reanalyses. An mysterious episode was uncovered in the Global Telecommunications System data distribution in the 1990's that lasted more than two years, with NCEP and ECMWF receiving complementary components of part of the rawinsonde data. Although this problem was uncovered and corrected during the NCEP reanalysis, it emphasizes the need to consolidate the data base even further. The reanalysis quality control system provides useful monitoring output on the upper air observations, and has been shown to be responsible for the improvement of the "reforecasts" after 1973, at a time in which the change of coding resulted in a deterioration of operational forecasts.

Although the reanalysis system was essentially unchanged during the 50 years processed, there were two major changes in the observing system. The first took place during 1948-1957, when the upper air network was being established and when the schedule of the rawinsondes was shifted by 3 hours, and the second in 1979 when the operational use of satellite observations was introduced. We assessed the impact of these changes on the reanalysis and concluded that the first decade is much less reliable than the last four. Furthermore, the introduction of satellite data in 1979 resulted in a significant change in the climatology, especially above 200 hPa and south of 50S, suggesting that a climatology based on the years 1979-present day is most reliable. The impact of satellite data on the 8-day "reforecasts" was also assessed.

Several climate applications were presented, including examples of operational climate monitoring. The agreement between the Southern Oscillation Index and the Reanalysis SOI allows the creation of an equatorial SOI that better represents the Southern Oscillation. Comparisons with independent surface data sets are show fairly good agreement and suggest that lack of agreement may be useful for tracking changes in the observing stations environment. Again, we showed that changes in the observing system, including a single isolated station, can have some effect on the local apparent climate.

In addition to the inevitable problems associated with changes in the observing system and model deficiencies whose corrections is a long term project, some human errors were detected in the course of the project. Many errors were detected and corrected in time to repeat short reanalysis periods. However, some errors were not detected until long periods were processed. We reviewed these errors and their consequences. The most serious error was that synthetic observations of sea level pressure made by Australia (PAOBS) in the Southern Hemisphere oceans were shifted in longitude by 180^o. The results affect synoptic patterns south of 40S. Because of geostrophic adjustment and quality control, comparisons with and without the error show that, with the exception of height covariances, the effect of the error on monthly means and covariance fields is negligible. A second very serious error was the use of the 1973 snow cover for several years. A third problems was the presence of a horizontal diffusion formulation that resulted in a "spectral valley snow" at cold high latitudes (corrected a posteriori in the XPRATE field). All these errors have been corrected in an a Reanalysis being performed in collaboration with the DoE for the period 1979-1998 (Masao Kanamitsu, pers. comm. 1998). This intermediate reanalysis project has also implemented changes to the model parameterizations that resulted in somewhat improved precipitation fields.

The availability of reanalyses from ECMWF and NASA/DAO allows intercomparisons to estimate of the reliability of the results. A comparison of the model derived fluxes of heat, momentum and evaporation at the surface and top of the atmosphere with ground-based climatologies indicates that the reanalyses have generally similar degree of agreement. The "observed climatologies" themselves are estimates some of which have been tuned to satisfy energy balances, so that differences among these estimates are in general not much smaller than those of the reanalyses. A comparison of seasonal means between NCEP and ECMWF reanalyses shows excellent agreement in heights and zonally averaged zonal velocity ("A" variables). However the meridional velocity, when zonally averaged, is purely divergent and therefore model dependent (a "B" or "C" variable). As a result the difference in the mean meridional circulation between the reanalyses, although small, is comparable to the interannual variability. The same is true for the humidity. The precipitation fields are overall comparable and realistic for both systems, although each has significant problems in specific regions. A comparison of decadal trends over the common 15 year period of reanalysis shows agreement in the stratospheric cooling but a low level change in the tropics that took place in 1987 the ECMWF reanalysis is not present in the NCEP or the DAO systems. Agreement in such longer trends is a necessary (but not sufficient) condition for confidence in assessments of climate change from reanalyses.

NCEP's future reanalysis plans, if supported, call for an updated global reanalysis using a state-of-the-art system every eight to ten years, and a maintenance of a CDAS allowing analysis of current climate anomalies. Within this time scale major improvements in the operational global system should take place, making the previous reanalysis obsolete, and justifying such major effort. Future reanalyses will be greatly facilitated by the quality-controlled, comprehensive observational database created by the present reanalysis, so that the development and execution of new global reanalyses should be completed in 4-5 years. In the meantime, it has been suggested that a Regional Reanalysis over North America would be particularly useful. There is a current plan for Regional Reanalysis developed during a Workshop on Regional Reanalysis that took place in Norman, OK during March 1998, sponsored by NOAA's Office of Global Programs (OGP), NCEP and the University of Oklahoma. The plan calls for the use the operational mesoscale Eta model driven by global reanalysis boundary conditions, at about 30km resolution, more resolution than would be currently possible with a global system, and much higher than the 210km used in the first global reanalysis. One major addition of the regional reanalysis is the 3D Var assimilation of radiances as well as the assimilation of "observed" precipitation. The latter should have the effect of forcing a very realistic hydrologic cycle during the reanalysis. If the Regional Reanalysis system currently under development is successful, an execution phase will start in 2001. A new global reanalysis would then follow around 2004. The benefits derived from reanalysis to the research and operational communities have been so important that it seems justifiable to support such a continued project at NCEP as part of the operational mission.

Fig. 3.1: Inventories of rawinsonde/pilot ballon observations, some of the most fragmented in terms of number of different sources.

Fig. 3.2: Same as Fig. 3.1 but for aircraft data.

Fig. 3.3: Same as Fig. 3.1 but for surface land observations.

Fig. 3.4: Inventory of individual observations of either temperature, wind, or surface pressure that passed the quality control and were assimilated.

Fig. 3.5: Total number of radiosondes used by the Reanalysis at the main observation times (Fig. 3.5).

Fig. 3.6: Time variation in the number of type 1 and 2 hydrostatic errors.

Fig. 4.1: 4 panel fits to data (ADPUPA). Bob

Fig. 4.2: Fig. 4.2 shows the geographical distribution of the 6hr forecast rms increment (analysis minus first guess) of 500 hPa heights for 1958 and 1996. Bob (from EK et al)

Fig. 4.3: Variability in the day-to-day analysis increments over North America dominated by synoptic time scales of 2-3 days both in 1958 and in 1996. Bob. (from EK et al)

Fig. 4.4: Anomaly correlation decay with time averaged over 73 predictions (one every 5 days) in 1979 for the SAT and NOSAT runs as verified with the SAT analyses for 1979. Bob's paper

Fig. 4.5: Monthly average 200hPa meridional wind for January and February 1979 from the analysis using satellite data ("SAT"), and for a "NOSAT" reanalysis. Suru

Fig. 4.6: Annual average of the rms difference between the forecasts and the verifying analyses, for days 1 to 8 for the NH (a) and SH (b). Bob

Fig. 4.7: Annually averaged 5-day anomaly correlation (AC) for the 50 years of reanalysis forecasts (full lines), as well as the operational scores (dashed lines) which are available only for the last decade. Bob. Are the pre-58 3 hr longer? Could you do that part with dots, not full lines?

Fig. 4.8: Comparison of the AC pre-FGGE (1958) and post-FGGE (1997). Bob's Fig. 3

Fig. 5.1: Comparison of the standard Tahiti-Darwin SOI and a corresponding index based on reanalysis pressure values at the nearest grid points to Tahiti (12.5S, 17.5E) and Darwin (17.5S, 150W) (Fig. 5.1). Vern K

Fig. 5.2: Equatorial Southern Oscilla Indext (Vern K)

Fig. 5.3: Cross-section of globally averaged temperature anomalies showing the impact of the satellite observing system introduced in 1979. Vern K

Fig. 5.4: Spatial distribution of this bias (Fig. 5.4) shows that it is greatest over the Southern Hemisphere oceans, which are regions lacking conventional radiosonde observations. Vern K

Fig. 5.5: From the early 1960s through 1981 a weather ship (4YP) was located near 50N, 145W. The time series of relative humidity at selected levels (Fig. 5.5) shows a wet bias for the period after the soundings were discontinued in 1981, especially for above 700 hPa. Vern K

Fig. 5.6: (Cover of the current AMS Bulletin). Monthly zonal mean of the zonal wind at the Equator for the 50 years of reanalysis above 100 hPa, showing how the 50-year reanalysis captured the qausi-biennial oscillation (QBO). Suru

Fig. 5.7: Maps of temperature anomalies from NCEP/NCAR for January 1958, 1979 and 1996 and interpolated to a 5 degree grid.

Fig. 5.8: Same as Fig. 5.7 but from the Jones (199*) surface analysis. Muthu

Fig. 5.9: Time series of global and tropical mean surface monthly temperature anomalies from NCEP/NCAR and from Jones (199*). Muthu

Fig. 5.10: Comparison of surface temperature anomalies from the Shanghai Observatory (121.9 long., 31.4 lat.) with the reanalysis estimated at the closest grid at 2.5^o by 2.5^o resolution, which is an "ocean" point. Eugenia

Fig. 5.11: Climatological annual cycle of surface pressure (mb) over 1965-1998. Upper panel: Globe, NH and SH. Lower panel: Globe, Land and Ocean. In the lower panel the values for Land and Ocean are offset by an amount so as to equal the annual mean global mean pressure.

Fig. 6.1: Changes in the surface temperature between Reanalysis and an experimental reanalysis using an updated version of the forecast model used in the NCEP AMIP reanalysis (section 8). Wesley

Fig. 6.2: Changes in the precipitation between Reanalysis and an experimental reanalysis using an updated version of the forecast model used in the NCEP AMIP reanalysis (section 8). Wesley

Fig. 6.3: Difference between the 300 mb relative humidity between Reanalysis and the experimental analyses which uses a different convection scheme. (skip?) Wesley

Fig. 7.1: Correlation of evaporation from da Silva et al. (1992) and from the NCEP/NCAR reanalysis for 1981-92 for (top) monthly means, (middle) the monthly mean annual cycle averaged over the 12 years, and (bottom) monthly mean anomalies from the annual cycles. Contours 0, .4, .6, .8, .9, .95.

Fig. 7.2: Zonal mean TOA net shortwave over land (top) and ocean (bottom) averaged over 1985-89. The dotted curve indicates the estimate from ERBE, the solid line the ERA estimate, the line with circles NCEP, and the line with crosses GEOS.

Fig. 7.3: Zonal mean TOA OLR averaged over 1985-89. The dotted curve indicates the estimate from ERBE, the solid line the ERA estimate, the line with circles NCEP, and the line with crosses GEOS.

Fig. 7.4: Zonal mean standard deviation of monthly mean anomalies from the annual cycle of TOA OLR over land (top) and ocean (bottom) averaged over 1985-89. The dotted curve indicates the estimate from ERBE, the solid line the ERA estimate, the line with circles NCEP, and the line with crosses GEOS.

Fig. 7.5: Zonal mean precipitation over land (top) and ocean (bottom) averaged over 1988-92. The dotted curve indicates the Xie-Arkin estimate, the dotted curve with crosses the GPCP estimate, the solid line the ERA estimate, the short dashes indicate NCEP, and the long dashes GEOS.

Fig. 7.6: Intercomparison of the mean DJF 500 hPa heights during ERA-15 period January, 1979 through February, 1994. The NCEP - ERA-15 difference, and the difference scaled by the total temporal variability (in %), are displayed in the bottom two panels.

Fig. 7.7: as in Fig. 7.6 except for zonally averaged zonal wind.

Fig. 7.8: As in Fig. 7.6 except zonally averaged meridional wind.

Fig. 7.9: Intercomparison of the mean DJF precipitation (top panels) against an analysis of observations and the difference during ERA-15 period January 1979 through February, 1994.

Fig. 7.10: as in Fig. 1 except 850 hPa interannual variance of temperature. The ratio of the interannual variance to the total variance is shown in the bottom panels.

Fig. 7.11: Intercomparison of the decadal linear trend [degC /10 years] in zonal average temperature anomaly during ERA-15 period January,

1979 through February, 1994.

Table 1: Mnemonic and description of the major sources of data used in the reanalysis.

Table 2: Inventory of sources found in the special data, listed in descending order of report counts.

Table 3: gives a summary of the observation types actually used by the assimilation system. Bob and Wesley

Table 4: shows global mean components of the surface energy balance for the three reanalyses for 1981-92, as well as da Silva et al. (1994) untuned and tuned air-sea fluxes averaged over 1981-92, and net shortwave estimates by Darnell et al. (1992) and net longwave estimates of Gupta et al. (1992) averaged over July 1983-June 1991, based on satellite observations of radiation, ISCCP cloud estimates and radiative transfer codes.. *****Glenn: please create table*****

Table 5: Temporal correlation of surface net flux from the reanalyses with da Silva et al. (1994) tuned estimate averaged over different oceanic regions for 1981-92.

Table 6: Temporal correlation of monthly mean anomalies in zonal surface stress from the reanalyses with da Silva et al. (1994) averaged over different oceanic regions for 1981-92.

Table 7: Temporal correlations of monthly mean anomalies in surface net shortwave from the reanalyses with Darnell and Staylor (1994) averaged over different regions for July 1983-June 1991.

Table 8: Temporal correlations of monthly means over a) land and b) ocean and of monthly mean anomalies over c) land and d) ocean in TOA net shortwave from the reanalyses with ERBE averaged over different regions for 1985-89.

Table 9: Temporal correlations of monthly means over a) land and b) ocean and of monthly mean anomalies over c) land and d) ocean in TOA OLR from the reanalyses with ERBE averaged over different regions for 1985-89.

Table 10: Temporal correlations of monthly means over a) land and b) ocean and of monthly mean anomalies over c) land and d) ocean in precipitation from the reanalyses with Xie-Arkin averaged over different regions for 1981-92. Also shown are correlations of monthly means from GPCP with Xie-Arkin for 1988-94.

Table 11: RMS difference in monthly means over a) land and b) ocean in precipitation between the reanalyses and Xie-Arkin averaged over different regions for 1981-92. Also shown are RMS differences in monthly means between GPCP and Xie-Arkin for 1988-94.

Appendix: Contents of the NCEP/NCAR Climate CD-ROM

a) For 1958-97: monthly means of

latent heat flux

net long wave radiation at the surface

net solar radiation at the surface

precipitation 

sensible heat flux

sfc pressure

sfc stress

skin temperature

2 m specific humidity and temperature

10 m winds

mean sea level pressure

TOA OLR

net solar radiation at TOA

precipitable water

height at 1000, 925,850, 700, 500,300,250,200,100,70,50,30,20

temperature at 925, 850, 700, 500,300,250,200,100

winds at 925 850 700 500 300 250 200 100 50 20

specific humidity at 925, 850, 700, 500, 300

vertical velocity at 925 850 700 500 300 200

b) 12- month climatology of these fields averaged over 1979-97.

c) For 1948-57: monthly means of

hgt 700 and 500

temperature 700

winds 200 and 850

skin temperature

sea level pressure

d)Observations: Monthly mean estimates of

Outgoing Long-wave Radiation (NOAA, 1979-97)

Precipitation (Xie-Arkin, 1979-97)

Precipitation for the US (Climate division data, 1948-97)

2 m temperature for the US (Climate division data, 1948-97)

e) Number of observations of 6 different types available for every month

from 1946-1997.

f) Demonstration program showing a few sample maps and an interactive program that will display the user to choose many combinations of variables, periods and geographical areas.

g) The GrADS software, sample scripts, the BAMS 96 paper and this paper.

Basist, A., and M. Chelliah, 1996: Comparison of tropospheric temperatures

derived from NCEP/NCAR reanalysis, NCEP operational analysis,

and the microwave sounding unit. Bull.Amer.Meteor.Soc., 78,

1431-1447.

Chelliah, M., 1990: The global climate for June-August 1989: A season of

near normal conditions in the tropical Pacific. J.Climate,3,138-162.

Chelliah, M. and C.F. Ropelewski, 1999: Reanalysis-based tropospheric

temperature estimates: Uncertainties in the context of global climate

change detection. Under review, J. Climate.

Collins (1998)

Collins and Gandin (1990)

Darnell, W.L., W.F. Staylor, S.K. Gupta, N.A. Ritchey, and A.C. Wilber, 1992: Seasonal variation of surface radiation budget derived from International Satellite Cloud Climatology Project C1 Data. J. Geophys. Res., 97, 15741-15760.

Da Silva, A., C.C. Young, and S. Levitus, 1994: Atlas of Surface Marine Data 1994 Vol. I: Algorithms and Procedures. NOAA Atlas NESDIS 6, U.S. Dept. Of Commerce, Washington, D.C., 83 pp.

--- and G. White, 1995: A comparison of surface marine fluxes from GEOS-1/DAS and NMC reanalyses. Proc. of the Workshop on the GEOS-1 Five-Year Assimilation, NASA Tech. Memo. 104606, Vol. 7, Goddard Space Flight Ctr., Greenbelt, MD, 129-136.

Derber and Wu, 1997

Ebisuzaki, 1997

Gupta, S., W. Darnell, and A. Wilber, 1992: A parameterization of long wave surface radiation from satellite data: Recent improvements. J. Appl. Meteor., 31, 1361-1367.

Janowiak, J.E., A. Gruber, C. R. Kondragunta, R. E. Livezey, and G. J. Huffman, 1998: A comparison of the NCEP/NCAR reanalysis precipitation and the GPCP raingauge-satellite combined data set with observational error considerations. J. Climate, in press.

Kalnay and Jenne 1991

Kalnay et al, 1983

Kalnay et al, 1996

Kalnay et al, 1998

Kistler et al, 1997

Parrish and Derber, 1992

Peixoto, J.P. and A. H. Oort 1992: Physics of climate. American Institute of Physics, New York. 520pp. ISBN 0-88318-711-6.

Stendahl, M., and K. Arpe, 1997: Evaluation of the hydrological cycle in reanalyses and observations. Max-Planck-Institut fur Meteorologie, Hamburg, Germany, 52 pp.

Trenberth, K. E. and C. J. Guillemot ,1994: The total mass of the

atmosphere. J. Geoph. Res, 99, D11, 23079-23088.

Van den Dool, H. M., and S. Saha, 1993: On the seasonal redistribution of mass in a 10-yr GCM run. J. Climate, 6, 22-30.

Van den Dool, H. M., J. Schemm and S. Saha, 1995: On the climatological annual cycle in global mean atmospheric pressure and hydrology as revealed by NCEP/NCAC Reanalysis. Proceedings of the 20-th climate Diagnostics Workshop, Seattle, October 23-27, 1995, 366-369.

Wallace, JM, EM Rassmusson, TP Mitchell, VE Kousky, ES Sarachik and H von Storch, 1998: On the structure and evolution of ENSO-related climate variability in the tropical Pacific: Lessons from TOGA. J. Geophys. Res, Vol 103, pp 14241-14250.

White, G., ed., 1996a: WCRP Workshop on Air-Sea Flux Fields for Forcing Ocean Models and Validating GCMs. Reading, England, 14-17 Oct. 1995. World Meteorological Organization, Geneva, Switzerland, WMO/TD-No. 762, 184 pp.

---, 1996b: Fluxes from the operational analysis/forecast systems and the NCEP/NCAR reanalysis.

WCRP Workshop on Air-Sea Flux Fields for Forcing Ocean Models and Validating GCMs. Reading, England, 14-17 Oct. 1995. World Meteorological Organization, Geneva, Switzerland, WMO/TD-No. 762, 131-137.

--- and A. da Silva, 1998: An intercomparison of surface marine fluxes from GEOS-1/DAS, ECMWF/ERA and NCEP/NCAR reanalyses. 9th Conf. On Interaction of the Sea and Atmosphere, 11-16 Jan, 1998, Phoenix, Az. Amer. Meteor. Soc., Boston, MA, 20-23.

Woollen et. al., 1994

Woollen and Zhu, 1997

World Climate Research Programme, 1990: Global Precipitation Climatology Project: Implementation and data management plan. WMO/TD - No. 367, WMO, Geneva, 47 pp.

Xie, P., and P. A. Arkin, 1996: Analyses of global monthly precipitation using gauge observations, satellite estimates and numerical model predictions. J. Climate, 9, 840-858.

--- and ---, 1997: Global precipitation: a 17-year monthly analysis based on gauge observations, satellite estimates and numerical model outputs. Bull. Amer. Meteor. Soc., in press.