The Tropical Atmospheric El Niño Signal in Satellite Precipitation Data and a Global Climate Model

a. Rainfall products

Three monthly gridded products (version 6) from the TRMM satellite (Kummerow et al. 2000), covering the period December 1997–June 2005, are included in this study: TRMM 3A12 (Kummerow et al. 2001) from the TRMM Microwave Imager (TMI), TRMM 3A25–6A (Iguchi et al. 2000) from the Precipitation Radar (PR), and TRMM convective–stratiform heating (CSH; Tao et al. 2001) from the PR. TMI data include surface total rain rate, stratiform–convective partitioning, and latent heating profiles at 0.5° × 0.5° resolution covering ±40° latitude. PR data include surface total rain rate and stratiform–convective partitioning at 0.5° × 0.5° resolution with similar area coverage as TMI but fewer samples per unit area due to its narrower swath. Although these two products are based on two different data sources and retrieval algorithms, their rainfall rates are highly correlated, with TMI values being slightly higher than those from PR (Kummerow et al. 2000). CSH provides monthly 0.5° × 0.5° latent heating profiles derived from surface convective and stratiform rain rates.

Several caveats about the TRMM rainfall rates should be noted. Tropical mean ENSO rainfall anomalies differed significantly in earlier versions of the TMI and PR data (Berg et al. 2002; Robertson et al. 2003), but discrepancies have been reduced in version 6. In addition, the TRMM satellite’s orbit was boosted in August 2001 to extend its lifetime. The larger footprints decrease the detection of weak rainfall and thus increase the PR conditional mean rain rate slightly (Shimizu et al. 2003), while the effect on TMI appears to be a more substantial increase in high rain-rate areas (De Moss and Bowman 2007). Since we focus only on regional components of change that correlate with an ENSO SST index, we expect these problems to have little effect on our results.

The retrieval methods of rain type (convective versus stratiform) for TMI and PR are different as well. The TMI retrievals are indirect, based on texture in the horizontal brightness temperature gradient and the 85.5-GHz polarization signatures (Olson et al. 2001). The PR retrievals are based on the radar reflectivity profile, in which a bright band indicates the stratiform region (Awaka et al. 1997). This is a direct physically based method, though the horizontal texture of the reflectivity field is also used when ambiguity exists. The separation of convective and stratiform precipitation plays an important role in correctly estimating the latent heating profiles in the Tropics (Houze 1982, 1997): convective heating is concentrated in the lower troposphere, where water vapor concentrations and thus condensation rates in cumulus updrafts are greatest, while stratiform heating peaks in the middle and upper troposphere where anvils form above the freezing level. The TMI-based latent heating profiles are derived using a technique described by Olson et al. (1999, 2006). The CSH algorithm of Tao et al. (1993) utilizes the stratiform fraction of rain and lookup tables simulated from a cloud-resolving model or estimated from a diagnostic budget study.

The Advanced Microwave Scanning Radiometer–Earth Observing System (AMSR-E) instrument on the EOS Aqua satellite provides passive microwave measurements of terrestrial, oceanic, and atmospheric parameters twice daily covering up to ±70° latitude. The level-3 rainfall accumulation product (AE_RnGd) used in this study is the monthly averaged rainfall accumulation over ocean and land at 5° × 5° resolution for the period June 2002–June 2005 (Wilheit et al. 2003; Adler et al. 2004).

b. Other TRMM products

Other atmospheric parameters over ocean from TMI are available from Remote Sensing Systems (RSS) online (see http://www.ssmi.com/tmi/tmi_browse.html). The monthly gridded TMI ocean product (version 3a) includes SST, column water vapor (CWV), and surface wind speed (Wentz and Meissner 2000). The resolution is 0.25° × 0.25° for the same area and temporal coverage as TRMM 3A12.

c. Other datasets

Monthly mean vertical profiles of pressure vertical velocity (omega) and temperature through June 2005 are obtained from the National Centers for Environmental Prediction–National Center for Atmospheric Research (NCEP–NCAR) reanalysis (Kalnay et al. 1996) at 2.5° × 2.5° resolution. Outgoing longwave radiation (OLR) at the top of the atmosphere (TOA) through December 2004 is obtained from the International Satellite Cloud Climatology Project radiative flux dataset (ISCCP-FD; Zhang et al. 2004).

a. Precipitation

The annual mean precipitation distribution from TMI, PR, and the GCM is plotted in Fig. 3. There are some discrepancies between the TMI and PR annual mean precipitation both in the spatial patterns and zonal mean distributions: TMI rain rates generally exceed those of PR, and TMI maxima over central Africa, the Himalayas, and the Amazon basin are missing or only weakly present in PR, which is more reliable over land. But in general, the differences are small compared to those between the GCM and either dataset. The simulated Pacific intertropical convergence zone (ITCZ) and South Pacific convergence zone (SPCZ) are too broad and their rainfall rates too intense. There is also an unrealistically strong rainfall maximum in the tropical west Atlantic and a spurious maximum over the north Indian Ocean. The precipitation maximum over the Amazon is underestimated, while a spurious peak over the Himalayas not seen by PR is simulated. The seasonal mean precipitation (not shown) has similar patterns of deficiency in the GCM in that there is stronger rainfall than observed over most ocean areas and weaker over some land areas. Over the Maritime Continent (MC), the GCM produces more (less) rainfall in the north (south) part during boreal winter and spring, and behaves in the opposite sense during boreal summer and fall.

Figure 4 compares the evolution of seasonal ENSO anomalies in precipitation for TMI (Fig. 4a) and the GCM (Fig. 4b). The ENSO-induced precipitation patterns are characterized by a positive anomaly over the central and eastern Pacific, and a negative anomaly over the western Pacific and the MC during El Niño, and vice versa during La Niña (Fig. 4a). Positive Pacific rainfall anomalies are less extensive than SST anomalies, however. Rainfall decreases during the warm phase in the climatological ITCZ and SPCZ locations, including the east Pacific ITCZ where SST warms. Positive IO rainfall anomalies occur during the warm phase, but only in the western part of the basin, while SST warms over most of the IO. Although the GCM simulation shows a similar pattern of El Niño– and La Niña–like anomalies, the magnitude is weaker than observed in most areas (Fig. 4b). Furthermore, the anomaly pattern is more zonally oriented than observed, suggesting that the model’s Walker cell perturbation is underestimated, especially over the MC and IO regions. The GCM also misses the observed precipitation changes in the southeastern United States during the 1997/98 ENSO event. Finally, tropical anomalies of both signs generally weaken a season too early compared to those observed, for example, MAM in 1998, 2000, and 2003 in Fig. 4b.

Figures 5a,b show the correlation between TMI and PR monthly gridbox precipitation anomalies within the dataset domain for El Niño (N34 > 0.5) and La Niña (N34 < −0.5) periods over the period December 1997–June 2005. The two datasets are highly correlated (correlation coefficient R = 0.9) but with PR anomalies only 80%–85% as strong as TMI anomalies. The largest differences typically occur in the central equatorial Pacific. For the June 2002–June 2005 period that overlaps AMSR-E, which contained only a single weak ENSO event, we use positive and negative values of N34 instead to separate El Niño and La Niña. The resulting PR–TMI correlation (Figs. 5c,d) is lower (R = 0.7), and the PR signal only about 2/3 the strength of the TMI signal. In part this may be due to the ambiguity in isolating ENSO over the shorter time interval, but it may also reflect the greater disparity in the two products at weaker rain rates where TMI and PR disagree the most (Kummerow et al. 2001), since this period contained only moderate ENSO anomalies. The GCM (not shown), whose ENSO anomalies are displaced relative to those observed in some locations, is correlated more weakly with TMI (R = 0.5), and its anomalies are only ∼40% as large.

Figures 5e,f compare AMSR-E and TMI anomalies over this same period. The strengths of AMSR-E and TMI ENSO anomalies are more similar than those of PR and TMI, though the correlation is slightly worse. The latter is due primarily to a small number of grid boxes, mostly over the Himalayas, where the AMSR-E monthly product using the Wilheit et al. (2003) algorithm diagnoses large anomalies while the TMI algorithm of Kummerow et al. (2001) does not. Excluding these, Figs. 5e,f suggest that the sparse diurnal sampling of AMSR-E does not seriously degrade its ENSO precipitation signal.

Figure 6 shows the extreme correlation at least lag (see appendix) of precipitation with N34 and corresponding lag–lead relationship maps for PR (Fig. 6a), TMI (Fig. 6b), and the GCM (Fig. 6c). Both TMI and PR are negatively correlated with N34 over the eastern IO, the MC, the SPCZ, the Amazon basin, and the west equatorial Atlantic Ocean, and positively correlated over the central–east equatorial Pacific Ocean (upper panels in Fig. 6a and Fig. 6b). The local central–east Pacific precipitation anomaly tends to lag the evolution of Pacific SST by 1–2 months except right at the equator (lower panels in Fig. 6a and Fig. 6b), suggesting that low-level moisture convergence anomalies must develop before precipitation is affected. The observed precipitation anomaly in the MC region leads the Pacific SST anomaly, consistent with the finding of Curtis and Adler (2000). This implies that subsidence anomalies due to weakening trade winds may determine MC behavior as much as the delayed remote response to resulting east Pacific SST anomalies. The GCM simulates reasonably well the observed positive correlation between equatorial Pacific precipitation anomalies and SST, and the negative correlation in the SPCZ and Atlantic. However, the strong negative (positive) correlations observed in the MC (IO) regions are missing (misplaced) in the model. The missing MC response, if driven by trade wind rather than SST anomalies at ENSO initiation, may be a limitation of atmosphere-only GCM simulations in which the surface wind can respond to but not affect the prescribed SST.

b. Stratiform–convective partitioning and latent heating

The release of latent heat in precipitating areas is a major driving force to the large-scale circulation in the Tropics. The profile of latent heating is associated with the partitioning of stratiform versus convective rainfall (Houze 1982, 1997). In a convective dominated region, heating peaks in the lower troposphere, while it is concentrated in the upper troposphere (with evaporative cooling below) in the stratiform anvil region. Changes in the stratiform rainfall fraction (SRF) and corresponding latent heating profile during El Niño events have been reported by Schumacher and Houze (2003), who show an enhanced near-equatorial trans-Pacific SRF gradient during El Niño.

The two TRMM SRF products show different features both in the mean and ENSO anomaly fields. For the mean field (Fig. 7, left), the largest discrepancies are found over land and over the east Pacific ITCZ. (Large apparent PR–TMI discrepancies over the Peruvian and Namibian marine stratocumulus regions are not associated with deep convective systems and are irrelevant to this discussion.) In general, PR stratiform fraction over ocean is ∼40%–60% and that for TMI is ∼10% smaller in heavily precipitating regions. Over land, the PR SRF is somewhat smaller and the TMI SRF is considerably smaller. The GCM simulation has a general pattern that agrees with PR, but with much larger SRF than PR and TMI have over land (where GCM mean rainfall is deficient), and much smaller SRF over most ocean areas (∼10%–20%), symptomatic of the absence of a mesoscale updraft parameterization in the model (Donner et al. 2001). Corresponding to those differences in SRF between the GCM and data over ocean regions, the GCM produces more convective precipitation, and thus latent heating peaks at a much lower altitude with a much stronger magnitude compared to the TRMM latent heating profile (see below and section 6).

In the right column of Fig. 7, the 1998 December–February (DJF) SRF ENSO anomaly is plotted for PR (Fig. 7d), TMI (Fig. 7e), and the GCM (Fig. 7f). The sign of the SRF ENSO anomaly is opposite in TMI and PR data. PR has a negative SRF anomaly over the MC region, the SPCZ, and the subtropical Pacific in the Northern Hemisphere, and a strong positive anomaly over the central/east Pacific (Schumacher and Houze 2003). TMI instead shows a positive 1997/98 ENSO anomaly in the subtropical Pacific in both hemispheres, and weak negative changes over the MC region and the central equatorial Pacific Ocean. Similar patterns exist during the 2002/03 El Niño (not shown), but the area covered and anomaly magnitude are smaller. The GCM’s SRF ENSO anomaly pattern resembles the less reliable TMI estimate but is stronger over the central Pacific.

Mean and ENSO 1998 DJF anomalies in the altitude of peak latent heating (ALPH) for the CSH product and the GCM are shown in Fig. 8. The GCM plots are shown for the deep convective component of heating only, since shallow nonprecipitating clouds are not detected by PR and do not contribute to the CSH profiles. The PR retrieval peaks in the middle troposphere in convecting regions, higher in the west Pacific warm pool and over the tropical continents than in the east Pacific ITCZ. ENSO anomalies in the PR-based ALPH are fairly consistent with the PR SRF anomalies in the central Pacific (ALPH rises where SRF increases during the warm phase), the SPCZ, and the MC (ALPH and SRF both decrease). We also examined the TMI-based latent heating profiles (not shown), which show a higher ALPH over most ocean areas than the PR-based product. The TMI ENSO perturbation latent heating is also noisier than the PR anomaly. The patterns in the GCM’s ALPH are similar to those in CSH, but its latent heating peaks at a lower altitude than that of CSH over the oceans and most continental regions, and at higher altitudes over a few tropical land locations. The GCM’s ALPH also increases in the central/east Pacific during the warm phase despite its SRF decreasing there, suggesting that it is determined by an upward shift in the convective heating profile as rain rate increases rather than a change in the weighting between stratiform and convective contributions. The big decrease of ALPH during El Niño in the MC, however, is missing from the GCM.

The DJF latent heating profiles from shallow convective, deep convective, and stratiform clouds in the GCM are shown in Fig. 9. The deep convective heating peaks near 600 mb but has significant heating down to 800 mb. The latter feature is consistent with the GCM’s underestimated downdraft mass flux (Xie et al. 2002). Shallow convection and heating from low-level stratiform cloud further depress the level of the overall heating. The upper-level stratiform heating shows the effect of the absent mesoscale updraft—rather than a strong upper-troposphere heating peak with evaporative cooling below, the model exhibits only slight net heating above the 300-mb level and a deep layer of weak evaporative cooling below that.

c. Other hydrological fields

In this subsection, we explore the atmospheric responses to El Niño in several other components of the tropical water and energy cycle: CWV, OLR (upward positive), and middle-tropospheric temperature (TT). In the subtropical descending branch of the Hadley cell, CWV is the primary influence on OLR and is also an indicator of the strength of the descending motion. In the ITCZ, OLR is controlled by cumulus anvil and thinner cirrus clouds and is thus diagnostic of the ice water path produced by convection and large-scale ascent. The warming of tropospheric temperature and its spread across the entire tropical belt via propagating Kelvin and Rossby waves has been proposed as an ENSO teleconnection mechanism for remote tropical surface warming (Chiang and Lintner 2005).

The GCM CWV mean field is almost uniformly lower than that observed by TMI (by ∼12%) throughout the Tropics (not shown). This corresponds to a GCM OLR mean value larger than the ISCCP mean (by ∼9.5 W m⁻²). The mean middle-tropospheric temperature at 400 mb is reasonably simulated in the GCM. Over the Pacific and Indian Oceans, the CWV anomaly has the same sign and similar patterns as precipitation, while the OLR anomaly has the opposite sign but similar patterns (Figs. 10a,b). We are especially interested in the CWV and OLR ENSO signal over the subtropical subsidence regions (i.e., 15°–30°S and 15°–30°N). Figure 11 shows scatterplots of the monthly 500-mb omega anomaly versus CWV anomaly for December 1997 to June 2005 from the NCEP–NCAR reanalysis and TMI (top panel) and the GCM (bottom panel). The response of the model’s subtropical CWV to ENSO Hadley circulation anomalies is too weak by ∼40%. This may be due to its underestimate of SRF, which weakens the remote circulation response (Schumacher et al. 2004) and thus subsidence drying. The model’s weaker-than-observed negative OLR anomaly in the central–east equatorial Pacific in Fig. 10b is diagnostic of an underestimated ENSO increase in cumulus condensate detrainment and/or anvil ice production, also consistent with the model’s SRF problem. Details of the vertical structure of the response over different regions will be discussed in section 6.

The ENSO-induced anomaly in NCEP–NCAR reanalysis and GCM TT are compared in Fig. 10c. The signal is dominated by a TT warming response to El Niño through most of the tropical belt both in NCEP–NCAR and the GCM, consistent with previous studies (e.g., Yulaeva and Wallace 1994). NCEP–NCAR shows the classic Gill (1980)-type pattern of response to a surface heating anomaly. The initial eastern equatorial Pacific warming generates a Kelvin wave that propagates to the east, while Rossby waves on either side of the equator propagate the signal to the west Pacific. This is proposed by Chiang and Sobel (2002) and Chiang and Lintner (2005) as a mechanism for ENSO impacts in the remote Tropics. Although the GISS GCM produces TT warming across the Pacific, the Gill pattern is less obvious. The Kelvin wave remote influence over Africa and the west IO is weaker. Some features of the Rossby wave pattern do not propagate sufficiently far westward—the primary warm anomaly barely extends past the date line, and cooling anomalies over east Asia and west Australia–southeast IO are mostly missing in the GCM.

a. Walker circulation

We first consider the mean and DJF 1998 anomaly behavior of the GCM in the MC and CEEP regions, which constitute the primary upwelling branch of the equatorial Walker cell during normal and ENSO conditions, respectively. Figures 12 and 13 show observed and simulated vertical profiles of latent heating, vertical velocity, static stability, and the two contributions to the adiabatic heating/cooling anomaly in (1) for each region, respectively. Tables 1 and 2 summarize the precipitation and precipitation-type characteristics of the model relative to that observed for each region. (Reanalysis fields in these figures are not totally constrained by observations, but we assume their gross features on seasonal or greater time scales to have some information content.)

Although the GCM overestimates climatological precipitation in the MC relative to both TRMM products (Table 1), its latent heating peaks at lower levels (Fig. 12a) due to its small SRF (23%, versus 43% in PR; Table 1), weak downdraft, and low-level shallow convective and stratiform cloud. As a result, the model’s mean vertical velocity is slightly greater than that in the NCEP–NCAR reanalysis at low levels but considerably weaker in the middle and upper troposphere (Fig. 12b). For the same reason, the temperature profile is too unstable in the middle and upper troposphere throughout the equatorial region (Figs. 12c and 13c). In the CEEP, mean subsidence in the model is too weak at all levels (Fig. 13b). These results are broadly consistent with those of Schumacher et al. (2004, their Fig. 6). The GCM ENSO latent heating anomaly in the CEEP is also centered at low levels, while the TRMM heating anomaly is concentrated at upper levels (Fig. 13a). In fact the GCM SRF anomaly is negative while that observed by PR is positive (Table 2). The remote ENSO subsidence response in the MC is too weak (Fig. 12b) in the model. The vertical velocity anomaly is the primary balance for the diabatic heating anomaly in the Pacific both in the model and observations (Figs. 12d and 13d). Thus, the deficiencies in the model’s remote tropospheric temperature response (Fig. 10c) can be traced to the weakness of its remote subsidence response to ENSO SST perturbations.

This may be understood in the framework of the Gill (1980) model. The remote response depends on the propagation of Kelvin and Rossby waves away from the anomalous heat source. The propagation speed increases with increasing static stability. Since the GCM’s troposphere is too unstable due to weak stratiform heating at upper levels and weak downdraft cooling at lower levels, its wave responses propagate too slowly and thus weaken by dissipation over shorter distances. The resulting weakness of the GCM’s Walker circulation may also explain the overly zonal orientation of its mean and anomalous precipitation fields in the MC and the presence of a primarily north–south rather than west–east dipole ENSO pattern in the IO (Fig. 4).

Another possibility is that the model dissipation is too strong, especially given its use of a cumulus momentum mixing parameterization that neglects cloud-scale pressure gradients. Clarke and Kim (2005) emphasize the role of weak damping in producing a zonally symmetric TT response to ENSO. We performed a sensitivity test in which momentum mixing was suppressed in the convection scheme. The resulting ENSO anomalies are shown in Fig. 14. Eliminating momentum mixing has a beneficial effect on several fields: Spurious positive (negative) ENSO CWV anomalies over the MC (western IO) present in the control simulation are now eliminated. The TT field now shows more of the Gill-like pattern seen in the reanalysis, and cold anomalies are now present over east Asia and Australia. However, negative MC precipitation anomalies are still not produced, and the magnitude of the CEEP ENSO response is still too weak; these features are more likely tied to the deficient upper-level heating.

b. Hadley circulation

The STNH over the Pacific Ocean is the location of the descending branch of the Hadley cell in boreal winter. The excess mean precipitation in the GCM throughout the equatorial region (Tables 1 and 2) implies a meridional gradient of diabatic heating that is too large, and thus a Hadley cell that is too strong, resulting in excess subtropical subsidence (Fig. 15; Table 3). This explains why the GCM underestimates the mean CWV and overestimates the mean OLR in this region (Table 3). On the other hand, the GCM’s lack of stratiform heating and negative SRF ENSO anomaly in the CEEP create a weak subtropical Hadley cell response (Fig. 15), especially in the middle and upper troposphere, so it is no surprise that the model’s CWV and OLR ENSO anomalies are both weaker than observed (Table 3).

In summary, the GISS GCM’s inability to simulate the ENSO remote atmospheric response is mainly a consequence of its underestimates of stratiform rain fraction and downdraft mass flux, and its excessive heating from shallow convective and low stratiform clouds, which result in diabatic heating concentrated in the lower troposphere and a weak upper-tropospheric gradient of anomalous heating. This has potential consequences for the GCM’s estimates of cloud and water vapor feedback in a long-term climate change. Tropical convective cloud feedback depends on changes in the balance between vertical motions, ice production, and sedimentation; the results of this study call into question the model’s ability to simulate these balances. Furthermore, water vapor feedback depends on changes in subtropical Hadley cell subsidence, since it is in these dry regions that OLR is most sensitive to changes in water vapor. The GCM’s weak subtropical ENSO drying anomaly does not necessarily imply that its water vapor feedback is overestimated, since the sense of the Hadley cell change may differ in response to external forcing; it does, however, suggest that the magnitude of water vapor feedback should be considered somewhat more uncertain than might be concluded from the good agreement among current climate models, since most models underestimate SRF.

Together with evidence from the differences in other related fields, we are led to speculate that the lack of a mesoscale updraft, which causes the underestimate of SRF, the weakness of the convective downdraft in the GCM, and the lower-level heating from shallow convection and stratiform cloud, are mainly responsible for its weaker ENSO atmospheric response. For example, the differences between simulated and observed ENSO SRF and ALPH anomalies indicate that lack of a mesoscale updraft limits the GCM’s ability to have anvils respond to ENSO independently of the parent convection. Mesoscale updraft parameterizations are still rare in GCMs (see Donner et al. 2001 for an exception), so this might be a fruitful area for model improvement across the full spectrum of GCMs. No GCM yet produces a reasonable downdraft mass flux (Xie et al. 2002), and some GCMs still have no downdraft parameterization at all; this should also be a priority for model development. In our GCM these weaknesses combine to give an unrealistic heating profile. In fact, it is possible to get a reasonable heating profile without one or both of these parameterization features (Tiedtke 1989; Gregory and Rowntree 1990; Hack 1994; Donner et al. 2001). However, this must be the result of unrealistic updraft mass fluxes or cumulus microphysics, or compensations in other parts of the model, as evidenced by the difficulty these models have in simultaneously getting the cumulus heating and moistening–drying profiles correct. The excess low cloud in the GCM is probably symptomatic of deficiencies in a boundary layer scheme based on dry conserved variables. We also showed that weakening the dissipation associated with cumulus momentum mixing improves some aspects of the remote ENSO response in the simulation. Parameterizations that include the effect of cloud-scale pressure gradients on cumulus momentum transport are available (e.g., Gregory et al. 1997) and should be tested as well. Some of the problems seen in the GISS GCM, for example, its overly zonal response to ENSO (Fig. 4), appear to be a feature of many current GCMs (Dai 2006).

Finally, our comparisons of the various satellite precipitation products have implications both for the use of current products and planning for future missions. Our study makes clear that convective–stratiform partitioning algorithms based on passive microwave observations, while capable of producing reasonable mean geographical distributions, are not yet reliable for studying variability. The planned Global Precipitation Mission (GPM) will rely on a constellation of satellites to provide global coverage on synoptic time scales, but only the core satellite will carry a radar. If GPM is to realize its potential for monitoring convective systems over their life cycle, as the mix of convective and stratiform rainfall changes, further development of passive microwave algorithms, using the TRMM PR as a training device, will be necessary.