SEARCH

SEARCH BY CITATION

Keywords:

  • Climate Forecast System Reanalysis;
  • precipitation

Abstract

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Data and Model Simulation
  5. 3. Results
  6. 4. Summary and Discussions
  7. Acknowledgments
  8. References
  9. Supporting Information

[1] Reanalysis products often show discontinuities in the time evolution of analyzed fields, particularly for the derived variables, for example, precipitation. In the recently completed Climate Forecast System Reanalysis (CFSR) at the National Centers for Environmental Prediction, a sharp increase in the global mean precipitation around 1998 has been noted. A similar abrupt change also appears in several other reanalyses that have been recently completed. This sharp increase coincides with the introduction of the Advanced TIROS Operational Vertical Sounder (ATOVS) data into the assimilation system. Based on model simulations to complement the CFSR, this study analyzes the reason behind the sudden jump in precipitation. The outputs from both the CFSR and a simulation with the CFSR's atmospheric component model forced with the observed sea surface temperature and carbon dioxide concentration are used to understand the precipitation abrupt change in the CFSR. The analysis indicates that the change results from (1) the tendency of the forecast (during the assimilation cycle) to drift toward its mean state that is cooler and drier, (2) the tendency of the assimilation to correct the first guess forecast to a wetter analysis, (3) an analysis that the period after 1998 is in an even wetter mean state due to the ingestion of the ATOVS data, and (4) the combination of previous three factors resulting in a abrupt change in precipitation during the assimilation around 1998. The results presented here may also provide the developers of the reanalysis systems further insights about the possible causes for the change in precipitation around 1998.

1. Introduction

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Data and Model Simulation
  5. 3. Results
  6. 4. Summary and Discussions
  7. Acknowledgments
  8. References
  9. Supporting Information

[2] Model based reanalysis products are valuable for climate monitoring and for advancing physical understanding of the climate system. Usage of reanalysis products extends from understanding climate trends and variability [Folland et al., 2009; L'Heureux et al., 2010] to validating climate model simulations [Luo et al., 2005; Goddard et al., 2006], initializing and verifying weather and climate predictions [Kanamitsu et al., 2002; Vitart, 2004; Saha et al., 2006; Chen et al., 2010], providing lateral boundary conditions for downscaling with regional models [Liang et al., 2004; Xie et al., 2007], and providing upper boundary conditions for simulations with ocean models [Auad et al., 2001; Waliser et al., 2003; Ayina et al., 2006], among others.

[3] Given the importance of reanalysis products, efforts are under way to improve reanalysis data sets, and newer generations of reanalysis are continually made available to the user community [Uppala et al., 2005; Onogi et al., 2005; Saha et al., 2010; Dee et al., 2011; Rienecker et al., 2011]. Improvements in the successive reanalysis efforts rely on availability of more historical observations resulting from the data mining efforts, improved models and data assimilation techniques, higher spatial resolution of assimilation systems, and inclusion of physically consistent coupled interactions between various components of the Earth system etc. Once the reanalysis data sets are released to the broad user community, extensive diagnostics provides useful guidance to the developers of the reanalysis about strengths and shortcomings of the reanalysis products.

[4] The use of reanalyses for climate studies, and reliable monitoring of the climate system requires a realistic representation of interannual variability as well as long-term trends. Although reanalyses are generated using a frozen data assimilation system to avoid discontinuities in reanalysis products that could be introduced if the assimilation system was continuously altered (as is the case for the real-time analysis system associated with the weather forecasts), changes in observational data with time are also consequential. To highlight evolutions in the observational data,Table 1 lists some major changes in observations used in the Climate Forecast System Reanalysis (CFSR) [Saha et al., 2010]. The changes include continued additions of observational platforms (for examples, Goddard Earth Observing System model (GEOS) radiances in 1997, ATOVS satellite observations in October 1998, and QuikSCAT winds in 2001).

Table 1. Some Major Changes in Observation Systems in CFSRa
YearMajor Changes
  • a

    The observed data includes TIROS Operational Vertical Sounder (TOVS), High-Resolution Infrared Sounder Unit (HIRS), microwave sounding unit (MSU), Stratospheric Sounder Unit (SSU), Geostationary Meteorological Satellite (GMS) from the Japanese Meteorological Agency (JMA), European Remote Sensing satellite (ERS) from the European Space Agency (ESA), Aircraft Communications Addressing and Reporting System (ACARS), Meteorological Aviation Report (METAR), Special Sensor Microwave Imager (SSMI), GEOS, Advanced TIROS Operational Vertical Sounder (ATOVS), advanced MSU (AMSU), Quick Scatterometer (QuikSCAT), mesoscale network of automated weather stations (MSONET), Atmospheric Infrared Sounder (AIRS), Advanced Scanning Radiometer–EOS (AMSR-E) from the National Space Development Agency of Japan (NSDA), NRL (Naval Research Laboratory) WindSat scatterometer (WINDSAT), and Meteorological Operation (MetOP).

1979TOVS (HIRS2, MSU, SSU)
1981ESA/ERS
1987JMA/GMS
1992ACARS
1997METAR, SSMIWND, GEOS
1998ATOVS (HIRS3, AMSU-a, AMSU-b)
2001NASA/QuikSCAT
2002MSONET
2004AIRS
2007NSDA/AMSR-E
2008NRL/WINDSAT
2009MetOp-A

[5] While a great number of changes occurred during the course of CFSR, our focus in this study is the temporal changes in the CFSR around 1998 when ATOVS data started to be assimilated. Compared to earlier satellite data sets, the ATOVS observations provide higher vertical resolution and denser horizontal coverage with many more channels.

[6] Interactions between changes in observational data and initial drift in the assimilation system (from the analysis to model's equilibrium state) can create unrealistic climate signals in various reanalysis fields. One such example of a sudden change in time evolution of global mean precipitation and evaporation in the CFSR is illustrated in Figure 1. Around 1998, the CFSR global mean precipitation (red curve in Figure 1a) has an abrupt increase while evaporation (red curve in Figure 1c) decreases dramatically. The changes in the mean precipitation and evaporation happen to be coincident with the advent of the ingestion of the ATOVS data into the assimilation system. A similar change in the characteristics of these variables is not unique to the CFSR but also appears in some other reanalyses, e.g., Japanese Reanalysis (JRA) [Onogi et al., 2005] and the National Aeronautics and Space Administration (NASA)'s Modern-Era Retrospective Analysis for Research and Applications (MERRA) [Rienecker et al., 2011].

image

Figure 1. Annual mean of global average of (a) precipitation (mm/d), (b) precipitable water (kg/m2), (c) evaporation (mm/d), and (d) temperature at 500 mb (degree) for Climate Forecast System Reanalysis (CFSR) (red lines), Twentieth Century Reanalysis (20CR) (blue lines), Modern-Era Retrospective Analysis for Research and Applications (MERRA) (green lines), ECMWF Reanalysis Interim (ERAI) (yellow lines), and Atmospheric Model Intercomparison Project (AMIP) (black lines). In Figure 1a, the lefty axis is for all data sets except for MERRA, which follows the scale on the right y axis.

Download figure to PowerPoint

[7] For the CFSR (and for other reanalyses data sets) it is an interesting question to understand what interactions during the assimilation cycle contributed to the change in precipitation and evaporation around 1998 in the CFSR. One such possibility is that the change around 1998 in Figure 1 is a consequence of an interaction between ingestion of the new observational data and the systematic biases in the assimilation system. The impacts of the assimilation system's mean state may vary with the input observations. In particular, with the enhanced availability of additional data, the reanalysis may then be constrained more by the observations leading to a sudden change in characteristics of different fields.

[8] Understanding of climate variations in the reanalysis that may be related to physical factors, such as evolution of external forcing, versus unphysical factors that may be an artifact of inhomogeneities in the observational data, is essential for an informed utilization of reanalysis products in climate monitoring, diagnosis, and predictions. Such an analysis is also important in the context of cataloging lessons learned that could be used in the future generation of reanalysis efforts to possibly rectify some of the issues. With this aim, the goal of our analysis is to investigate the sudden change in the characteristics of the atmospheric large-scale circulation and global energy and water balance in the CFSR around 1998. To provide a better understanding, an Atmospheric Model Intercomparison Project (AMIP) [Gates, 1992] simulation with the same atmospheric component of the CFSR is used to assess the realism of trends associated with the time evolving observed external forcing. We demonstrate that change in precipitation around 1998 can be explained by interaction between changes in the assimilated data, the analyzed state, and its interaction with the model's drift to its equilibrium state (i.e., the model bias) during the 6 h forecast cycle. Given the commonality of an abrupt change around 1998 among various reanalyses products, the present effort is also relevant for the understanding of similar issues in other reanalyses. We should emphasize that our focus is the dependence of the reanalysis on the model's systematic errors under the ever evolving observation techniques, using the changes around 1998 as a specific example. As the user of the reanalysis products generally does not have the wherewithal to quantify impacts of intricate nuances of the data assimilation procedures implemented as part of reanalysis, providing a detailed attribution for the abrupt change in the CFSR is beyond the scope of this study.

[9] The paper is divided into four sections. Section 2 describes the data and model simulations that will be used in this analysis. Section 3 presents the comparison between model simulation and the CFSR. Section 4 provides a summary and discussion.

2. Data and Model Simulation

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Data and Model Simulation
  5. 3. Results
  6. 4. Summary and Discussions
  7. Acknowledgments
  8. References
  9. Supporting Information

[10] We use monthly mean fields from January 1979 to December 2009 to analyze CFSR climate variability and trend. The first guess for CFSR is from a coupled atmosphere-ocean model consisting of the NCEP global forecast system (GFS) for the atmosphere and the Geophysical Fluid Dynamics Laboratory Modular Ocean Model version 4 (MOM4) for the ocean, including an interactive sea ice component. The atmospheric component (GFS) is run at a horizontal resolution of T382 (∼38 km) with 64 vertical levels extending from the surface to 0.26 hPa. The oceanic component (MOM4) uses 40 levels in the vertical, a zonal resolution of 0.5° and a meridional resolution of 0.25° between 10°S and 10°N, gradually increasing through the tropics until becoming fixed at 0.5° poleward of 30°S and 30°N.

[11] Atmospheric and oceanic analysis in the CFSR is made every 6 h. In each analysis cycle, in situ and satellite observations are assimilated with the first guess obtained from 6 h coupled forecast from the previous analysis to produce analyzed fields. The new analysis is then taken as the initial conditions for the model forecast integration needed by the next cycle. Accordingly, the accuracy of the analyzed fields depends on the model's forecast errors (including drift from the analysis) and the available observations.

[12] In the CFSR, the concentrations of CO2 and other trace gases together with changes in aerosols and solar variations are prescribed from observational data sets as a function of time and space. More details about the CFSR can be found in the work of Saha et al. [2010]. Some aspects of climate variability in the CFSR have been documented in a few previous studies, including oceanic variability [Xue et al., 2011], surface climate and variability [Wang et al., 2011], tropospheric variability [Chelliah et al., 2011], precipitation frequency and intensity characteristics [Higgins et al., 2010], and local drought features [Mo et al., 2011].

[13] In addition to the CFSR, monthly mean data from an AMIP simulation is also used. The AMIP simulation helps understand the systematic errors of the atmospheric model's equilibrium state toward which the forecast guess drifts during the assimilation cycle. The AMIP simulation is for 1979–2009 with the atmospheric component of the CFSR at a reduced horizontal resolution of T126 forced by the National Climate Data Center (NCDC) daily sea surface temperatures (SSTs) [Reynolds et al., 2007], and NCEP sea ice concentrations (SIC) [Grumbine, 2009]. Observed greenhouse gas concentrations, aerosols and solar variations are specified in the AMIP simulation as those in the CFSR. The CFSR assimilates both the NCDC SSTs and the NCEP SICs. Accordingly, the AMIP simulation provides a measure of the long-term variations in the CFSR that are due to the atmospheric model's response to the external forcings.

[14] The CFSR and the AMIP simulation are validated against several observational analyses. We use the CPC Merged Analysis of Precipitation (CMAP) [Xie and Arkin, 1997], outgoing longwave radiation (OLR) from the NOAA AVHRR [Liebmann and Smith, 1996], and the National Oceanographic Data Center (NODC) seasonal oceanic temperature [Levitus et al., 2009]. The NODC temperature analysis is at 16 levels ranging from the ocean surface to 700 m in depth on a global 1 × 1 degree grid.

[15] For a comparison with other reanalyses, temporal changes from MERRA, and ECMWF Reanalysis Interim (ERAI) [Dee et al., 2011] are also analyzed. Such a comparison allows an examination of the commonality of the long-term changes in reanalyses and dependence of temporal variations in various reanalysis products on assimilation systems. In addition, the Twentieth Century Reanalysis (20CR) [Compo et al., 2011], which only assimilates surface pressure but forced with observed sea surface conditions, is also used to analyze variations without the assimilation of atmospheric observations. The DOE/NCEP reanalysis 2 (R2) is used later for a comparison of surface winds.

3. Results

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Data and Model Simulation
  5. 3. Results
  6. 4. Summary and Discussions
  7. Acknowledgments
  8. References
  9. Supporting Information

[16] We begin with an analysis of the temporal evolution of global mean fields and time mean differences between two 11 year periods: 1987–1997 and 1999–2009. We then discuss the spatial distributions of the differences and their interactions with the systematic bias of the atmospheric model, followed by an example of the implications of the impacts on the ocean analysis that is part of the coupled assimilation system. The latter analysis provides an illustration how changes in the observational platform in the component of the system can affect the other parts of the analysis system.

[17] In Figure 1 time evolutions of global annual mean precipitation (Figure 1a) and evaporation (Figure 1c) are compared. For the CFSR, there is a sudden increase (decrease) in precipitation (evaporation) around 1998. These sudden changes correspond to the beginning of the use of ATOVS data in October 1998 (Table 1). For the AMIP simulation such dramatic changes do not exist. It is noted that there is another jump in CFSR precipitation in 2001. This jump may be related to the use of QuickSCAT surface wind data starting 2001 (Table 1). In this analysis we will focus on the changes in 1998 and the possible impacts of the ATOVS data.

[18] For precipitation, there is a good similarity in changes around 1998 between MERRA (green curves in Figure 1a) and CFSR. However, in contrast to CFSR and MERRA, the evolution in precipitation in ERAI shows a decrease around 1998. We will discuss possible reasons for the different behaviors in precipitation variations between CFSR and ERAI in section 4. Temporal evolutions in the 20CR (which only assimilates surface pressure data) are quite consistent with those in the AMIP simulation. This comparison indicates that the abrupt change in the precipitation time series may be because of the assimilation of the ATOVs data in the CFSR, MERRA, and ERAI. It is also interesting to note that around 2005 while CFSR and MERRA have a downward change in precipitation, ERAI shows an increase.

[19] A similar change in the CFSR precipitable water (PW) toward larger values also occurs (Figure 1b). Such a change can also be seen in the AMIP simulation, indicating that it is partly related to the evolution of the underlying SSTs as to be discussed next. However, the PW increase after 1998 in CFSR and MERRA are more sustained than that in other reanalyses and the AMIP simulation, suggesting the strong impacts of changes in the observation that are assimilated in CFSR and MERRA.

[20] While discrepancies are found in the time evolution for the hydrological variables, for 500 mb temperature (Figure 1d), apart from a mean offset, variations in the time series for CFSR, AMIP, and other reanalyses have a good agreement. We note that the mean offset between the CFSR and the AMIP is indicative of the model bias. For example, 500 mb temperature in the AMIP simulation is about 1 K colder than for the CFSR reanalysis, while PW is about 1.5–2 kg/m2 smaller, indicating a drier and colder equilibrium state of the atmospheric component of the CFSR. It is also noted that the ERAI has warmer temperatures at 500 hPa than CFSR before 1998. After 1998, 500 hPa temperatures are comparable between ERAI and CFSR.

[21] Time evolution in the AMIP simulation is solely driven by the evolution of the observed SSTs, and radiative fluxes related to greenhouse gases and solar forcing etc.. Some of the trends in the AMIP time series, and interannual variability, can be explained by the corresponding global annual mean SST time series (see Figure 2). A slight upward trend in tropospheric temperature (Figure 1d) can be related to a similar upward trend in the SST. This is consistent with earlier results that have documented the SST control on the temperature field [Kumar et al., 2004; Hoerling et al., 2008; Dommenget, 2009]. Such a control of SSTs also holds true for variations on interannual time scale. For example, an increase in temperature and PW in 1997 is related to the El Niño of 1997. We note that (1) the increase in SST in recent years may be due to a corresponding increase in the CO2, and (2) SST is more effective in communicating the changes in the CO2 to the atmosphere than the direct radiative influence of the variations in the CO2 [Yang et al., 2003; Deser and Phillips, 2009]. In the rest of this paper a reference to the influence of trends in SST is assumed to include the indirect contribution from changes in the CO2 forcing.

image

Figure 2. Time series of the annual mean of the sea surface temperature averaged over the global ocean between 60°S and 60°N for the period of 1979–2009 for the AMIP run.

Download figure to PowerPoint

[22] Differences between the CFSR and the AMIP are further summarized in Table 2 where values for the 11 year mean before and after 1998 are listed. In general, changes for the AMIP simulation are much smaller than those for the CFSR. Differences in the 11 year mean in both precipitation and evaporation in the AMIP are 0.03 mm/d, indicative of hydrological balance in the free model simulation. The balance between precipitation and evaporation also holds for individual 11 year average. Increase in PW from 1987–1997 to 1999–2009 is 0.47 kg/m2 in the AMIP simulation and is consistent with the increase in SST and of tropospheric temperature in the later period. Increased SST, by increasing the tropospheric temperature, and under the assumption of less varying relative humidity (RH) (see Figure 3), also increases the water holding capacity of the atmosphere, leading to an increase in the PW. We should point out that although the analysis of the AMIP results is based on a single simulation, for the temporal and spatial averages analyzed (e.g., annual means, global average), the statistical significance is unlikely to be an issue.

Table 2. The Global Averaged Precipitation, Evaporation, and Precipitable Water of CFSR and AMIP for the Periods of 1987–1997 and 1999–2009 and the Two Periods' Differences
 1987–19971999–2009Differences Between 1987–1997 and 1999–2009
  • a

    PW, precipitable water.

CFSR
Precipitation3.093.210.13
Evaporation2.912.87−0.04
PWa23.8524.901.05
 
AMIP
Precipitation3.253.280.03
Evaporation3.253.280.03
PWa22.1222.590.47
image

Figure 3. The zonal mean differences of (a) temperature, (b) relative humidity, (c) geopotential height, and (d) specific humidity between the periods of 1999–2009 and 1987–1997 for the AMIP run.

Download figure to PowerPoint

[23] Changes in the mean value over the two 11 year averages for the CFSR are much larger. Increase in precipitable water is almost double that for the AMIP (1.05 kg/m2 compared to 0.47 kg/m2). Much larger increase (about 4 times as much as in the AMIP simulation) is found for the CFSR precipitation. For evaporation, there is a decrease in the later period in CFSR while AMIP evaporation shows an increase. Further, for the AMIP there is no hydrological balance either for the individual 11 year averages, or for the difference between 11 year average values. We note that the hydrological imbalance (precipitation minus evaporation) of 0.34 mm/d for the 1999–2009 period (i.e., the period after the introduction of the ATOVs data) is much larger than 0.18 mm/d for 1987–1997 period with precipitation exceeding the evaporation in both periods. While there exists a common upward PW change in both CFSR and AMIP simulation, the differences in PW between CFSR and the AMIP simulation in the earlier period (23.85–22.12 = 1.73 kg/m2) is smaller than that in the later period (24.90–22.59 = 2.31 kg/m2), indicating that the reanalysis was closer to the AMIP (i.e., atmospheric models equilibrium state) during the earlier period when less amount of observed data is available to constrain the analysis.

[24] Differences in the zonal mean temperature, RH, geopotential height, and specific humidity between the two 11 year periods in the AMIP simulation are shown in Figure 3. Throughout most of the troposphere, there is an increase in temperature (Figure 3a) which is consistent with an increase in global mean SST. Increase in tropospheric temperature is capped by a decrease in temperature in the stratosphere, consistent with the radiative effects of the increase of CO2 concentration [Manabe and Wetherald, 1967]. Increase in tropospheric temperatures is also reflected in the zonal mean heights (Figure 3c). Changes in the RH between two periods are less than 2% (Figure 3b). On the other hand, there are significant changes in the specific humidity that are confined near the surface where water holding capacity is the largest (Figure 3d). Under the assumption of constant RH, increase in the lower tropospheric temperature due to an increase in SST can be used to explain an increase in the specific humidity. To summarize, for the AMIP simulation, changes in these variables are consistent with the expected influence of changes in the observed SST.

[25] Corresponding changes in the zonal mean for the 11 year average for the CFSR are shown in Figure 4. The most striking difference from the change in the AMIP is a large increase in specific humidity in the CFSR that is almost 6 times larger than that in the AMIP for the equatorial latitudes. Also confined at the equator, there is a marked increase in RH near 200 mb and is likely related to increase in precipitation, corresponding to a stronger convective activity which dominates the tropical precipitation in the CFSR (not shown), resulting in a larger outflow of humid air parcels at the top of cloud plumes. Increase in tropospheric temperatures in the CFSR is also larger than that in the AMIP simulation. Another marked difference from the AMIP simulation is a large increase in temperature throughout the troposphere north of 60°N, a feature also shared by other reanalyses (Figure 10). Reasons for this are not clear and are not pursued further in the context of the present analysis, the main purpose of which is an understanding for the abrupt change in precipitation after 1998 in the CFSR which is dominated by the contribution from the tropical precipitation.

image

Figure 4. As in Figure 3 but for CFSR.

Download figure to PowerPoint

[26] To explain the larger jump in the CFSR compared to the AMIP, we rely on (1) equilibrium state of the atmospheric model, (2) its difference from the analysis state, and (3) how the two interact during the assimilation cycle. Since both CFSR and AMIP simulations were produced with the same frozen atmospheric component and the SST in the CFSR is essentially the same as that in the AMIP simulation, their differences in time evolution are expected to result from the variations in the assimilated observations. However, the characteristics of the mean differences between CFSR and AMIP simulation may depend on the model physics which are responsible for the establishment of the dynamical, thermodynamical, and hydrological balances, and are unlikely to be the same for other assimilation systems (and their AMIP simulations).

[27] The mean state for the AMIP simulation is the model's mean atmospheric state in a long run well separated from the initial conditions at the start of the integration. Differences in model's mean state and the analysis mean state reflect the model bias relative to the analysis. Starting from an initial atmospheric analysis that differs from the model's mean (or the equilibrium) state, the drift during the initial forecast toward the mean state is often referred as the “spin up.” The initial model drift, or the spin up, has been used previously in understanding the onset of model biases [Klinker and Sardeshmukh, 1988] in that differences in the initial state and forecast state with a short lead time (which could be as short as an integration over a single time step) often resemble the model bias after a long-free integration.

[28] The bias of the atmospheric model used in the CFSR is quantified as the difference between the AMIP simulation and the CFSR analysis. Figure 5 shows the zonal mean difference between the CFSR and the AMIP for temperature and specific humidity averaged over 1987–1997 (Figures 5a and 5b) and 1999–2009 (Figures 5c and 5d). In general, the model mean state is colder and drier relative to the CFSR. This is true for almost all the latitudes and throughout the troposphere. Further, the drier bias is much larger for the 1999–2009 period (Figure 5d).

image

Figure 5. The zonal mean differences between CFSR and AMIP for (a) temperature during 1987–1997, (b) specific humidity during 1987–1997, (c) temperature during 1999–2009 and (d) specific humidity during 1999–2009.

Download figure to PowerPoint

[29] As discussed before, if the model's initial states are taken from the CFSR reanalysis, during the subsequent integration the model state will then drift to a colder and drier state. This happens during the 6 h forecast-assimilation cycle. In other words, starting from a particular analysis time (e.g., 12Z) with CFSR analyzed specific humidity as the initial state, during the subsequent 6 h forecast before the next analysis time, the forecast drifts toward a drier specific humidity state preferred by the model. At the end of the 6 h forecast, at the next analysis time (i.e., 18Z), the assimilation brings the specific humidity back to a moister atmospheric state as dictated by the insertion of observational data.

[30] A consequence of interaction between the model's tendency to drift toward its own equilibrium state and the assimilation to bring the analysis back to a moister state is that increments for specific humidity (and PW), defined as the resulting analysis after the assimilation minus the 6 h forecast guess, would be positive during the forecast-assimilation cycle. As to be shown later, this inference is indeed consistent with positive specific humidity and PW increments during the analysis.

[31] To understand the sudden jump in the CFSR precipitation after 1998, we make an assumption that a substantial increase in the amount of observational ATOVS data starting 1998 [Saha et al., 2010] resulted in more humid atmosphere, i.e., a higher specific humidity in the troposphere, and particularly in the lower troposphere. This assumption is supported by the analysis differences in the CFSR before and after 1998 (Figure 4), and results reported for the MERRA [Rienecker et al., 2011]. We also use two additional arguments, i.e., (1) during the 6 h forecast initial states from the CFSR drift toward a colder and drier model state, and (2) after 1998, analysis every 6 h brings the atmosphere back to a more humid atmosphere relative to the period before 1998, to explain the sudden jump in the CFSR precipitation after1998.

[32] Immediately after the CFSR analysis as the initial condition, and during the 6 h forecast before the next analysis, the extra initial moisture is shed as the precipitation as the forecast drifts to a drier state. And further, the shedding of moisture, and resulting precipitation, would be larger during the assimilation cycle after 1998 because of a larger specific humidity in the initial analysis.

[33] This argument will also imply that during the assimilation cycle, the analysis increments for the specific humidity, and PW, after 1998 would be higher as the analysis attempts to bring the atmospheric state after a 6 h forecast back to one dictated by the observations being assimilated. That this indeed happens is backed by comparing the analysis increments before and after 1998.

[34] Shown in Figure 6 is the time series of analysis PW increments averaged over the globe. The time series clearly shows a positive increment in PW for the entire analysis period, and a jump to a higher PW increment beginning October 1998 when the ATOVS observations started to be assimilated. The average of PW increment before 1998 is about 0.2 kg/m2, compared to the average value of ∼0.4 kg/m2 after 1998, which also corresponds to the large differences in the precipitation and evaporation before and after 1998. It is interesting to see that PW increments have a seasonal cycle with larger values in September.

image

Figure 6. Monthly mean of global average of precipitable water increment for CFSR.

Download figure to PowerPoint

[35] Zonal mean increments for CFSR temperatures and specific humidity averaged over an 11 year period before and after 1998 are shown in Figure 7. Temperature increments, which corrects for the model's forecast drift toward a colder equilibrium state, are quite similar between two periods. For the specific humidity, on the other hand, analysis increments after 1998 are much larger. This implies that assimilation cycle continuously brings back the 6 h forecast to more humid analyzed state, and this tendency is even larger for the atmospheric states after 1998, and further, the extra moisture is immediately shed out as precipitation, consistent with the jump in the CFSR precipitation after 1998. A reduction in evaporation after 1998 is likely due to a larger value of the near surface specific humidity leading to smaller differential in saturation specific humidity that corresponds to the underlying SST and the near surface atmospheric value, leading to a smaller evaporation as would be dictated by a simple bulk parameterization for the evaporation.

image

Figure 7. The zonal mean for CFSR for (a) temperature increment during the period of 1987–1997, (b) specific humidity increment during 1987–1997, (c) temperature increment during the period of 1999–2009, and (d) specific humidity increment during 1999–2009.

Download figure to PowerPoint

[36] The larger precipitation rate after 1998 in the CFSR is not a standalone feature, but also influences the large-scale atmospheric circulation. To illustrate this we give an example of changes in the spatial distribution of precipitation and low-level circulation between the CFSR and the AMIP before and after 1998 inFigure 8. The change in precipitation is also compared with the observed estimate based on the CMAP precipitation analysis, and with the OLR observations.

image

Figure 8. Precipitation and 10 m surface wind differences between the periods of 1999–2009 and 1987–1997 for (a) outgoing longwave radiation, (b) CFSR, (c) CPC Merged Analysis of Precipitation, and (d) AMIP.

Download figure to PowerPoint

[37] For the CMAP there is a reduction in the rainfall in the equatorial Pacific extending from the dateline to west coast of South America (Figure 8c). East of the dateline, this reduction in precipitation is also along the climatological position of the Intertropical Convergence Zone (ITCZ). In the southern hemisphere, the change in the precipitation represents a westward shift in the South Pacific Convergence Zone (SPCZ). West of the dateline there is an increase in precipitation over the maritime continent. The change in the tropical precipitation is confirmed from the differences in OLR which shows a similar pattern with opposite sign (Figure 8a).

[38] The spatial structure of changes in the CMAP precipitation is consistent with the change in the SST over the two periods in that the SST has warmed around the maritime continent and the warm pool, and has cooled in the eastern Pacific (not shown). To a certain extent SST changes as the cause for changes in the CMAP precipitation is confirmed by the AMIP simulation that also has reduction in rainfall over the equatorial Pacific, and a westward shift in the SPCZ as shown in Figure 8d. The discrepancy between the AMIP and the CMAP over the maritime continent are likely related to the biases in the AMIP simulation which produces too little precipitation over this region (not shown), resulting in small changes between the two periods.

[39] The changes in the precipitation in the CFSR, on the other hand, have a very different spatial structure in the tropical Pacific. There is an increase in precipitation over the ITCZ with no discernible westward shift in the SPCZ. An increase in the precipitation over the ITCZ for the 1999–2009 period is consistent with an increase in the specific humidity, and the “rich get richer” hypothesis in that areas of the large climatological precipitation get more humid, leading to a larger amount of precipitation [Chou and Neelin, 2004].

[40] Differences in precipitation between the CFSR and the AMIP also lead to dynamically consistent differences in the low-level winds as would be implied by the low-level convergence (divergence) associated with an increase (decrease) in the precipitation. For the AMIP there is an increase in easterlies at the equator near the date line associated with a decrease in precipitation. For the CFSR, on the other hand, there is more northerly flow feeding into the increase in precipitation along the ITCZ. There is broad region of surface convergence in the AMIP associated with the westward shift in the SPCZ, a change that is not well defined in the CFSR. Changes in the surface flow can also be seen in the differences in the equatorial overturning (or the Walker) circulation throughout the deep troposphere (not shown). This example demonstrates that an increase in the CFSR rainfall as a result of interplay between the model bias, and sudden changes in the characteristics of assimilated data not only affects the precipitation but also leads to changes in other dynamical fields.

[41] Some other independent evidence suggests that an increase in equatorial easterlies in 1999–2009, as a result of changes in the SST pattern, and associated precipitation is a more realistic feature of the observed trend. One such evidence we present is the change in the thermal structure of the subsurface ocean based on the NODC objective analysis of available thermal profiles (Figure 9d). The change in the subsurface oceanic temperature field for 1999–2009 relative to 1987–1997 indicate a subsurface warming in the western Pacific and a cooling in the eastern Pacific. This change is indicative of the deepening (shoaling) of the thermocline in the west (east) and leads to a steeper east-west thermocline tilt. The structure of the change in thermocline is consistent with an increase in the equatorial easterlies in the AMIP simulation between 150°E and 135°W (Figure 9b) that would lead to an increase in the upwelling in the eastern Pacific and convergence of warmer surface water in the western Pacific that downwells to the deeper ocean. The CFSR also produces easterlies around the dateline but with much weaker amplitude (Figure 9a), compared to R2 winds (Figure 9c) whose easterlies increase extend to 140°W.

image

Figure 9. Time series of the 10 m zonal wind difference averaged in the 2°S–2°N band for (a) CFSR, (b) AMIP, and (c) DOE/NCEP reanalysis 2, and (d) average temperature in the 2°S–2°N band for NODC temperature difference between the periods of 1999–2009 and 1987–1997.

Download figure to PowerPoint

4. Summary and Discussions

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Data and Model Simulation
  5. 3. Results
  6. 4. Summary and Discussions
  7. Acknowledgments
  8. References
  9. Supporting Information

[42] In this paper a sudden increase around 1998 in precipitation from the CFSR reanalysis was investigated. The analysis relied on an AMIP simulation forced with the observed SSTs. It was demonstrated that a sudden increase in precipitation after 1998 was due to interplay between the bias of the assimilation model and nonstationarity in the ingestion of observed data. We argued that starting from the analyzed state, during the subsequent 6 h forecast cycle the model drifts toward its mean state that is colder and drier. As part of this drift additional moisture is precipitated out. The shedding of the analyzed moisture is much more after 1998 (i.e., an increase in precipitation) because of a wetter analysis as governed by the inclusion of the ATOVS data. That the large jump in precipitation rate around 1998 is related to the spin down of the initial moisture is confirmed by an increase in analysis increments for specific humidity in the lower troposphere and also the precipitable water. The analysis also illustrates that a change in precipitation is also associated with other dynamical fields that respond to precipitation, e.g., surface wind.

[43] Impacts of the evolution of observation systems, and the role of model bias on the reanalyses products can be examined following different approaches, including data denial experiments in which certain observational data are excluded from the assimilation [Rienecker et al., 2011], or from the analysis of different assimilation systems with the use of same observations. Our approach demonstrates another helpful tool that utilizes AMIP simulation with the atmospheric model that is part of the assimilation system. Knowing the model bias, and the state toward which the analysis will drift to during the 6 h forecast, provided an important link in understanding changes in precipitation for the CFSR. We argued that impacts of observations on the reanalyses not only depend on the evolution of observations but also on their interactions with the systematic errors of the model used for the reanalyses. Following this approach, we provide hypothesis for the changes in precipitation in MERRA and ERAI that can be tested with the availability of the AMIP simulation for respective models.

[44] Differences between reanalyses precipitation in Figure 1 show a decrease around 1998 in global mean precipitation from ERAI (Figure 1a), and an increase for the MEERA, with change in the ERAI (MERRA) being opposite (similar) to an increase in CFSR. Assuming that model bias for MERRA is the same as for the CFSR, i.e., the model drifts toward a cooler and drier state, a sudden increase in rainfall in MERRA around 1998 will follow the same reasoning as for the CFSR. Indeed, changes in specific humidity changes in MERRA (Figure 10d) before and after 1998 appear to be quite similar to that for the CFSR (Figure 4d) with large moistening in the lower troposphere in the tropics.

image

Figure 10. Zonal mean differences in (a) ERAI temperature, (b) ERAI specific humidity, (c) MERRA temperature, and (d) MERRA specific humidity between the periods of 1999–2009 and 1987–1997. Units are kelvin for Figures 10a and 10c and grams per kilogram for Figures 10b and 10d.

Download figure to PowerPoint

[45] Causes for the contrasting precipitation change around 1998 in ERAI compared to CFSR and MERRA are harder to interpret. A complicating factor is the difference in the use of the satellite data. For example, AMSU-A window channels (1–3, 15) were not assimilated in ERAI but were assimilated in both CFSR and MERRA, and therefore, ERAI may not have an inherent tendency toward a moister lower troposphere after 1998. Sensitivity experiments reported byRienecker et al. [2011]implied that the inclusion of the AMSU-A data, which “MERRA responds more to,” caused the increase in precipitation. One could also attempt to understand a decrease in ERAI precipitation based on the model having a warm bias toward which it drifts to during the assimilation cycle. If the model has a warm bias, a warmer analyzed state after 1998 compared to the period before 1998 (Figure 10a) may require less precipitation (and corresponding heating) for the 6 h forecast to drift to a warmer mean state. Although the hypothesized causes for changes in the precipitation utilize different mechanisms—in the case of MERRA and CFSR a wetter analysis after 1998 drifting to a drier model state leading to an increase in precipitation, while for ERAI a warmer atmosphere after 1998 requiring less precipitation for the model to drift to a warmer mean state—both rely on initial model drift to their respective model climatology. Another interesting aspect of the precipitation across different reanalyses is the differences in its mean value that ranges from 3.2 mm/d for the CFSR to ∼2.8 mm/d for the ERAI and MEERA (after 1998). Such differences are likely due to differences in radiative cooling for the atmosphere in the respective analysis, and for which heating related to precipitation provides the balancing component [Allen and Ingram, 2002; Yang et al., 2003; Lambert and Allen, 2009].

[46] Beyond our speculation that can be easily tested based on the availability of AMIP simulation for the respective analysis, there are a large number of differences between the assimilation systems that users of the reanalyses products are generally unaware of. Therefore, a more complete understanding of the causes for the different trends around 1998 in ERAI, CFSR, and MERRA requires further studies to examine the impacts of different aspects, including the inclusion or exclusion of certain observations, use of different assimilation algorithms, and model bias etc. For example, the cold and dry bias of the CFSR model is consistent with negative bias of cloud amount and excessive downward solar radiation at the surface [Wang et al., 2011]. Another possible source of spurious changes (and differences) in the reanalyses may be the bias correction for the satellite radiances. The CFSR computes satellite scan-angle-dependent bias as well as an air-mass-dependent bias which is estimated using a variational bias correction scheme [Derber and Wu, 1998; Saha et al., 2010]. Additional studies are needed to determine if the bias correction scheme was effective in removing errors contained in the ATOVS data.

[47] We conclude our analysis with some thoughts. Although a reduction in model bias will greatly help reduce artificial changes in reanalysis, whether they can be removed all together in the presence of changing data platforms is an open question. It may be possible to reduce artificial changes further with data assimilation procedures other than the conventional assimilation-forecast cycle that are used for the numerical weather prediction system. This should be an interesting topic to pursue for the data assimilation and climate community. Our results also point to the need for maintaining a hierarchy of assimilation systems and simulations that could be used diagnose the realism of climate trends in observational record, and also to understand various nuances of data assimilation systems. Further, given the critical dependence of the quality of the reanalysis on the accuracy of the observational data, it is important to examine the impact of individual observations. In this aspect, data denial experiments in which certain observational data, such as the ATOVS observations, are excluded from the assimilation will help understand their impacts on the improvement or degradation of the quality of the reanalysis. Also, the bias corrections to the satellite data and increments applied to the analyzed fields are important to understand the performance of the reanalysis system. Making these fields available by the operational centers producing the reanalyses will enhance participation of the user community in the development of future reanalyses efforts.

Acknowledgments

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Data and Model Simulation
  5. 3. Results
  6. 4. Summary and Discussions
  7. Acknowledgments
  8. References
  9. Supporting Information

[48] We thank the two internal reviewers at CPC, who improved the final version of the manuscript. Useful comments and thoughtful reviews by the three reviewers are appreciated.

References

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Data and Model Simulation
  5. 3. Results
  6. 4. Summary and Discussions
  7. Acknowledgments
  8. References
  9. Supporting Information

Supporting Information

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Data and Model Simulation
  5. 3. Results
  6. 4. Summary and Discussions
  7. Acknowledgments
  8. References
  9. Supporting Information
FilenameFormatSizeDescription
jgrd17885-sup-0001-t01.txtplain text document1KTab-delimited Table 1.
jgrd17885-sup-0002-t02.txtplain text document0KTab-delimited Table 2.

Please note: Wiley Blackwell is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.