Observational Evidence for Two Modes of Coupling Between Sea Surface Temperatures, Tropospheric Temperature Profile, and Shortwave Cloud Radiative Effect in the Tropics
Abstract
Tropical average shortwave cloud radiative effect (SWCRE) anomalies observed by CERES/EBAF v4 are explained by observed average sea surface temperature (
) and the difference between the warmest 30% where deep convection occurs and
). Observed tropospheric temperatures show variations in boundary layer capping strength over time consistent with the evolution of SST#. The CERES/EBAF v4 data confirm that associated cloud fraction changes over the colder waters dominate SWCRE. This observational evidence for the “pattern effect” noted in General Circulation Model simulations suggests that SST# captures much of this effect. The observed sensitivities (dSWCRE/d
W·m−2·K−1, dSWCRE/dSST#≈−4.8W·m−2·K−1) largely reflect El Niño–Southern Oscillation. As El Niño develops,
increases and SST# decreases (both increasing SWCRE). Only after the El Niño peak, SST# increases and SWCRE decreases. SST# is also relevant for the tropical temperature trend profile controversy and the discrepancy between observed and modeled equatorial Pacific SST trends. Causality and implications for future climates are discussed.
Key Points
- Average SST and difference between the warmest waters and the average (SST#) control tropical tropospheric temperature profile and SWCRE
- Observational evidence for pattern effect (encapsulated by SST#) on climate sensitivity as previously noted in GCM simulations
- Observations driven by ENSO have lagged phase of SST# to Niño 3.4 index; forcing on SST# may change global average temperature
1 Introduction
Changes in the climate system are commonly normalized by the associated change in global average surface temperature change, and the widely used concept of climate sensitivity uses the definition of global average surface temperature change in response to a forcing. Similarly, climate feedbacks in General Circulation Models (GCMs) may be determined through a uniform increase in sea surface temperatures (SSTs; Cess & Potter, 1988). Indeed, coupled atmosphere-ocean GCMs predict a fairly homogeneous surface temperature change in particular over the oceans in response to radiative forcing. However, it is increasingly recognized that a range of observations suggest that additional degrees freedom need to be considered. Here, we argue that open questions regarding the cloud impact on climate sensitivity (the “pattern effect”, Andrews et al., 2018; Andrews & Webb, 2018; Ceppi & Gregory, 2017; Silvers et al., 2018; Stevens et al., 2016; Zhou et al., 2016) are related to the controversy surrounding the question whether the tropical tropospheric temperature trend profile follows moist adiabatic scaling (Flannaghan et al., 2014; Po-Chedley & Fu, 2012; Santer et al., 2005). That is, both depend on the question whether the temperature difference between the warmest regions with atmospheric deep convection and the tropical average remains constant under climate change. Flannaghan et al. (2014) show that since the 1980s, the Hurrell SST reconstruction (Hurrell et al., 2008) shows much stronger amplified warming of the warmest regions than the HadISST1 reconstruction (Rayner et al., 2003). Similarly, comparisons of SST trend differences focused on the cold Eastern Pacific reveal differences between coupled GCM simulations and observations, but the differences also depend on SST reconstruction and internal variability in the model simulations (e.g., Coats & Karnauskas, 2017; Zhou et al., 2016).
In the following, we show in section 2.1 that much of the observed variations in tropical average shortwave cloud radiative effect (SWCRE) as observed by CERES/EBAF v4 (Loeb et al., 2012, 2018) over the period March 2000 to March 2018 can be explained by just two parameters,
and SST#, where SST# is defined as the temperature of the warmest 30% minus the tropical average SST. SST# is a descriptor for the evenness of the SST distribution, and similar results may be obtained with other descriptors that reflect the degree of evenness. We use SST#4 because it is simple to calculate, and the SST field is the forcing for the aforementioned pattern effect. The emphasis on the warmest waters (rather than, say, coldest waters) facilitates physical intuition of the mechanism: The strong nonlocal constraint on local criticality arising from the weak temperature gradients in the tropical free troposphere (e.g., Emanuel et al., 1994; Polvani & Sobel, 2002; Sobel et al., 2002) allows the warmest regions with most deep convection to exert a strong control on the tropical average-free tropospheric temperature profile. By contrast, the boundary layer is in direct contact with the underlying surface, and consequently, the tropical average SST and the tropical average boundary layer over the ocean are tightly coupled. It follows that the tropical average boundary layer capping (reflecting the difference between average boundary layer and average-free tropospheric temperature) is related to the difference between the SST in the warmest regions with deep convection and the tropical average SST, and SST# is an index thereof that can be calculated from the SST field alone.
In sections 2.2-2.4 we provide observational support for the underlying mechanistic model connecting SST# and tropical cloudiness. Specifically, consistency is shown between variations in SWCRE and cloud cover as reported by CERES/EBAF v4, variations in tropical tropospheric stability in ERA-Interim reanalysis data (Dee et al., 2011), and the SSTs inferred from the CERES/EBAF v2.8 surface fluxes product (results with other SST reconstructions are very similar). The reduction to only two free parameters is facilitated by casting the problem in terms of frequency distributions rather than geographical patterns. Shifts of patterns in geographic space lead to large local anomalies of opposite sign that generally preclude conclusive attribution of the net effect (e.g., Radley et al., 2014). Sobel et al. (2002) stressed the importance of distinguishing convective from nonconvective regions and analyzed changes over time in the frequency distribution of SSTs. In section 2.3 we explain why our conclusions concerning SST frequency distributions differ from theirs.
The importance of the temperature difference between warmest regions and the average for low clouds is also emphasized by Miller (1997). Their model study seeks to predict the differential warming of the warmest regions following radiative forcing in order to predict the cloud radiative effect response and its impact on climate sensitivity. By contrast, the observational evidence presented here over the period covered by CERES/EBAF v4 is based on observed SST variations that are largely associated with El Niño–Southern Oscillation (ENSO; section 2.5). We therefore conclude in section 3 with a brief discussion of causality and forcing versus feedback and of implications for future and past climate states where the relation between SST# and
may be driven by processes other than the transient oscillations between El Niño and La Niña.
2 Results
All analyses are carried out with monthly mean data, and “tropical” refers to the latitude range 30° South to 30° North. The analysis of cloud radiative effect is restricted to the period March 2000 to March 2018 for which CERES/EBAF v4 data are currently available. The SWCRE is defined as the upwelling shortwave radiative flux under clear-sky conditions minus that under all sky conditions, at the top of the atmosphere. Thus, a negative SWCRE anomaly corresponds to an increase in reflected solar radiation (i.e., a cooling tendency) and vice versa. A simple 6-month running mean low-pass filtering is applied to reduce the noise of the monthly mean CERES/EBAF v4 data.
2.1 The Observed Tropical SWCRE Variations
Figure 1a shows the deseasonalized tropical average SWCRE as reported by CERES/EBAF v4. The figure further shows the results of two linear regressions: a linear regression of the observed SWCRE variations against
and a multiple linear regression against
and SST#. Adding SST# roughly doubles the explained variance in SWCRE compared to the linear regression against
only, giving a high correlation of r=0.8 (r2=0.64). Visual inspection shows that the most prominent variations in SWCRE are captured with the exception of the period 2003–2004. Also, we note that with the inclusion of SST# the SWCRE and the SST signals are largely in-phase, a point we will return to in section 3.
) scaled with slope from linear regression against SWCRE and the same for the multiple linear regression with
and SST# as predictors. (b) The time series of
(solid line) and SST# (dotted line) multiplied with their slopes dSWCRE/d
and dSWCRE/dSST# of the multiple linear regression. Correlation coefficients and slopes determined for low-pass filtered data (6-month running mean) as shown. Periods where the low-pass filtered Nino 3.4 index exceeds ±0.6 are highlighted in light red (El Niño situation) and light cyan (La Niña situation). SWCRE = shortwave cloud radiative effect; TOA = top of atmosphere; SST = sea surface temperature.The parameter SST# is a crude estimator for the subcloud moist static energy in the regions of deep convection relative to the tropical average. A separation of the domain according to some threshold applied to the midtropospheric vertical wind field (Bony & Dufresne, 2005; Zhou et al., 2016) or taking precipitation as an estimator of convective activity (Flannaghan et al., 2014; Fueglistaler et al., 2015; Sobel et al., 2002) may provide a more precise descriptor. However, the point here is that the relation with SWCRE is very robust and can be captured without additional information that may not be always available (e.g., rainfall distribution before the satellite era) and may suffer from observational (rainfall) and model biases (vertical wind field).
The two predictors
and SST# are not orthogonal by construction. Over the period studied here, their correlation is marginal (r≈0.07). Correspondingly, the sensitivity of SWCRE to
changes only slightly when introducing SST# to the linear regression. For the multiple linear regression, we get an increase in SWCRE (i.e., an increase in absorbed solar radiation) with average warming of
1.8 Wm−2K−1 and a decrease in SWCRE (i.e., a decrease in absorbed solar radiation) with an increase of SST# of dSWCRE/dSST#≈ -4.8 Wm−2K−1, as shown in Figure 1b. The sensitivity of SWCRE to tropical average SST of ≈1.8 W·m−2·K−1 agrees well with that to local SST as reported by Klein et al. (2017; their Figure 3b). The magnitude of dSWCRE/dSST# depends somewhat on the low-pass filtering applied. Over the period considered here, variations in SST# are predominantly associated with ENSO, and the 6-month timescale for the running mean is a compromise between noise reduction and preservation of the ENSO signal. The connection of
and SST# to ENSO clearly visible in Figure 1b is further discussed in section 2.5, and the question of causality between SWCRE,
, and SST# is discussed in section 3.
2.2 The Mechanism
The mechanism underlying the previous multiple linear regression builds on the expectation that the tropical average boundary layer temperature is closely tied to
, while the tropical average-free tropospheric temperature is closely tied to SSTs in the warmest regions with atmospheric deep convection (Flannaghan et al., 2014; Fueglistaler et al., 2015; Sobel et al., 2002; Zhang, 1993). To be precise, the relation is for subcloud moist entropy, but over the ocean the latter forms a fairly compact, monotonic relation with SSTs and we may use SST as proxy for subcloud moist entropy. The parameter SST# thus is expected to affect the tropical average boundary layer capping strength, which in turn is known to affect low-level cloudiness in particular over the colder oceanic regions (Bretherton, 2015; Klein & Hartmann, 1993; Klein et al., 2017; Myers & Norris, 2016; Qu et al., 2014, 2015, Wood & Bretherton, 2006; Wood, 2012).
We do not seek to directly relate the “boundary layer capping strength” to commonly used metrics such as the “lower-tropospheric stability” (Klein & Hartmann, 1993) or the “estimated inversion strength” (Wood & Bretherton, 2006). Rather, we simply assume that the stability is affected by changes in the temperature difference between boundary layer and free troposphere in a two-layer sense. By connecting the latter to SST#, we introduce a (global) closure assumption for the temperature above the boundary layer (typically at 700 hPa; see Klein & Hartmann, 1993) required by the above metrics of tropospheric stability. An anomalously high SST# thus corresponds to an anomalously strong boundary layer capping and consequently more clouds and a negative SWCRE as is observed in Figure 1. The sensitivity of the cloud field to
reflects stability changes within the boundary layer with temperature, as well as other bounday layer parameters related to surface temperature (such as moisture and radiative fluxes) known to affect the cloud field.
In the following, we demonstrate that the observed tropical average boundary layer capping strength as represented in the ERA-Interim reanalysis (Dee et al., 2011) provided by the European Centre for Medium-range Weather Forecast (similar results are obtained with MERRA-2 provided by NASA Rienecker, 2011) indeed follows SST# as determined from the CERES/EBAF v2.8 surface flux data product. Over the period 2000–2018, the differences between different SST reconstructions emphasized by Flannaghan et al. (2014) with respect to trends play a secondary role, and results (not shown) with the HadISST1 data (Rayner et al., 2003) are very similar.
Figure 2a shows that the observed tropical average tropospheric temperature and
are highly correlated, in line with the notion discussed in the introduction of a unique, compact relation between the average surface temperature and any other domain-averaged variable. However, indications of a residual with nonstationary phase are visible. In order to demonstrate that the tropical temperature profile indeed has a second mode in addition to coherent temperature variations over the entire depth of the troposphere, we subtract the expected scaling with average boundary layer temperature. That is, we perform a linear regression of the deseasonalized temperature profile against the boundary layer average (here taken to be 1,000–925 hPa) deseasonalized temperature. The resulting climatological mean scaling profile closely follows the moist adiabatic scaling (not shown). The resulting residual, that is, the departure from the climatologically expected scaling with the average boundary layer temperature, is shown in Figure 2b.
(blue) and tropospheric air (orange, 1,000- to 200-hPa average) temperature (in Kelvin); low-pass filtered with 6-month running mean. (b) Monthly mean residual of deseasonalized tropical average tropospheric temperature (in Kelvin) after subtracting the climatological mean scaling with tropical average boundary layer (1,000–925 hPa) temperature. The white contour of the color bar is centered at 0. (c) Tropical monthly mean SST (in Kelvin) distribution in equal area percentiles from coldest to warmest, deseaonalized and tropical average (blue line in panel a) subtracted. The white contour of the colorbar is centered at 0. (d) Time series of 6-month running means of SST# (blue, difference between warmest 30% and tropical average) and variations in boundary layer capping strength not related to
(orange, departure from climatological scaling as shown in panel b, averaged from 700 to 600 hPa). Scaling of capping strength given by linear regression against SST# (r=0.52).We note that above about 200 hPa these residuals are likely affected by the temperature shift in ERA-Interim caused by the introduction of COSMIC/GPS temperature data over the year 2006 (Fueglistaler et al., 2013). Also, we find that the same residuals determined from GCM simulations (not shown) are more coherent throughout the free troposphere than those in ERA-Interim shown in Figure 2b. These caveats likely rooted in observational uncertainties notwithstanding, the temperature residuals in ERA-Interim show distinct phases when the free troposphere is warmer or colder than expected from scaling with average boundary layer temperature. We average these residuals over the pressure range 700 to 600 hPa to obtain a robust estimate of the implied variation in boundary layer capping strength. Figure 2d shows (Figure 2c discussed below) that these variations in boundary layer capping strength not related to average SST are indeed correlated to SST# (r=0.52 for 6-month running mean low-pass filtered data). Using only the 700-hPa level that is used for the lower-tropospheric stability (Klein & Hartmann, 1993) and the estimated inversion strength (Wood & Bretherton, 2006) indexes yields similar results, with slightly lower correlations due to more noise.
2.3 The Leading Mode of Tropical SST Variations
Sobel et al. (2002) performed an Empirical Orthogonal Function (EOF) analysis of the observed SST frequency distribution time series and found that the leading mode is an approximately uniform temperature change. In contrast, our analysis emphasizes the importance of SST#, that is, nonuniform temperature changes. Inspection of the time series of the SST frequency distribution (Figure 2c) shows why the two studies arrive at different conclusions. In Figure 2c we describe the SST distribution in terms of the average SST in equal area bins upon sorting by SST. While less common than the standard frequency distribution, this representation is more straightforward to interpret. We then deseasonalize the data and further subtract the domain-averaged anomaly (shown in Figure 2a). The resulting residuals shown in Figure 2c thus reveal the departure from uniform warming or cooling, and SST# is the average of this residual over the warmest 30% of the total tropical ocean. We note prominent seesaws between warm and cold areas, for example, in the years 2007–2010 and 2016–2017, and it is these periods that give the correlation between SST# and SWCRE. While an EOF analysis in the percentile space employed here captures some of that nonuniform behavior (i.e., the leading mode is not a uniform temperature change), it is evident that a substantial fraction of variance does not follow a stationary phase relation between the area bins. For reasons beyond the scope of this paper, an EOF analysis in SST frequency distribution space as employed by Sobel et al. (2002) maps most of the departure from uniform warming onto several higher-order EOFs that individually explain only little variance.
2.4 Clouds Over the Colder Waters
Figure 3 confirms that the tropical average SWCRE variations shown in Figure 1 are dominated by variations in oceanic low-level clouds and that the variations in boundary layer capping strength induced by SST# affect their fractional area coverage. Figure 3a shows that the tropical average SWCRE largely follows the SWCRE in the “colder” 70% of the tropical oceans, with deviations noted in the early 2000s which also show less agreement between SWCRE and the prediction based on
and SST# (see Figure 1a). Figure 3b shows that over the colder oceanic regions SWCRE and the observed cloud fraction as reported by the CERES/EBAF v4 data agree well, with a sensitivity of SWCRE to cloud fraction of about −1 W/m2 per 1% increase in cloud fraction, in agreement with previously published values (e.g., Klein & Hartmann, 1993). The variations in cloud fraction over the colder oceanic regions in turn are also quite well explained by
and SST# (r=0.64).
and SST# (olive) yields a correlation of r=0.64. SWCRE = shortwave cloud radiative effect; TOA = top of atmosphere; SST = sea surface temperature.2.5 ENSO
Figure 1b shows that over the period of available observations, much of the variance in SWCRE,
, and SST# is related to ENSO. Broadly speaking, during the La Niña situations
is colder (increasing cloudiness), and during the El Niño situations
is higher (decreasing cloudiness). Conversely, over the course of the 2009/2010 and 2015/2016 El Niño situations SST# induces a transition from positive SWCRE (decreased cloudiness) to negative SWCRE (increased cloudiness) in the decay phase. The exceptionally strong El Niño of 1997/1998 shows an evolution of
and SST# similar to that of the 2009/2010 and 2015/2016 El Niño situations. While no CERES/EBAF v4 data are available for that period, we use the 1997/1998 El Niño due to its strong signal to exemplarily discuss the evolution of
and SST# and their impact on tropospheric temperatures.
Figure 4a shows the deseasonalized tropical average tropospheric temperature profile from January 1997 to January 1999. Figure 4c shows the Niño 3.4 index over the same period, peaking in December 1997 (El Niño) and subsequently reaching a local minimum (La Niña). Comparison of the Nino 3.4 index with the temperature profile (Figure 4a) shows that the well-established lag in tropospheric warming relative to Nino 3.4 (e.g., Kumar & Hoerling, 2003) is largely due to the signal in the free troposphere. The residual from expected scaling with average boundary layer shown in Figure 4b (as in Figure 2b) confirms the visual impression from Figure 4a that during the peak of Nino 3.4 the free troposphere is anomalously cold and afterward anomalously warm relative to the boundary layer. Figure 4c shows that the warming at 300 hPa is delayed compared to the Niño 3.4 index but is well explained (both in terms of phase and magnitude) by the SSTs (here taken from HadISST1) in the convective region determined either by SST# (red line) or precipitation-weighting (orange line; Flannaghan et al., 2014) of the HadISST1 data with rainfall from GPCP v2 (Huffman et al., 2001). Figure 4d shows that the evolution of SST# recovers the tropospheric temperature deviation from expected scaling with average boundary layer temperature shown in Figure 4b, with a weaker boundary layer capping at the onset of El Niño and a stronger boundary layer capping as El Niño decays. Finally, Figures 4e and 4f show the evolution of the SST distribution (in terms of average SST in equal area bins), with and without the domain average anomaly (i.e., Figure 4f corresponds to Figure 2c).
Thus, as an El Niño situation builds up,
and the boundary layer warm up in tandem. However, the warmest, convective regions of the ocean do not warm at the same rate. Consequently, during the peak El Niño situation SST# is anomalously cold which induces a weakening of the boundary layer capping strength. In this phase, both SST# and overline SST reduce the tropical average cloudiness (i.e., a positive SWCRE anomaly). During the subsequent decay of El Niño,
is still high (inducing reduced cloudiness) but SST# is now also large and correspondingly has an antagonistic effect on cloudiness.
3 Discussion and Outlook
We have demonstrated that in addition to the average SST, the difference between average and the warmest waters with deep convection in the atmosphere plays a key role for climate. This difference, for which we have used SST# as a proxy, controls variations in the tropical average boundary layer capping strength by controlling the average temperature difference between the boundary layer and overlying free troposphere. With low-level clouds particularly over the colder waters being sensitive to boundary layer capping strength, SST# exerts a control on tropical average albedo not related to average surface temperature. This simple mechanistic picture is based on casting the tropics in terms of a distribution of states grouped by SSTs. In this frame of reference, the tropical average shortwave cloud radiative effect is formulated as a function of two thermodynamic state variables (
and SST#) rather than the traditional geographic patterns with emphasis on East-West gradients in the equatorial Pacific and the Walker cell and Hadley cell circulations. Indeed, SST# may be a useful parameter to encapsulate the net effect of previously noted changes in these circulations and patterns (e.g., Coats & Karnauskas, 2017; Funk & Hoell, 2015; Meehl et al., 2013; Vecchi et al., 2006; Zhou et al., 2016).
In order to quantify the sensitivity of tropical cloud albedo, we have used an ordinary linear regression with the SST parameters as predictors. Hence, we have implicitly assumed that the cloud field is forced by
and SST#, rather than the cloud field forcing the SSTs. Lutsko (2018) argue that at ENSO periods, the decrease in cloudiness seen in CMIP5 models amplifies the surface temperature response. Our analysis shows that a substantial part of the initial decrease in cloudiness is due to the lagged warming of the warmest waters, thus initially decreasing the average boundary layer capping strength and hence cloudiness.
From the perspective of climate change, the question is whether
controls SST#, such that the effect of SST# on SWCRE would constitute a positive/negative feedback if SST# decreases/increases with an increase in
. The negative low cloud feedback proposed by Miller (1997) arises from their idealized model's prediction that dSST#/d
. The differences in trends of SST# between different reconstructions of observed SSTs thus induce not only uncertainty regarding the tropical temperature trend profile (Flannaghan et al., 2014) but also for climate sensitivity estimates and the question of causality between
and SST#. It is conceivable that SST# is not only dependent on
but that some process in the global climate system can act as a forcing on SST#. In this situation, the associated cloudiness change would force
, with a positive feedback (if the dSWCRE/d
observed here is not restricted to ENSO variability) from the cloud change following a change in
. Consequently, all else being equal, a climate state with more uniform tropical SSTs (smaller SST#) has fewer low-level clouds and hence has a higher average surface temperature.
Data Availability
All data used in this study are publicly available in repositories maintained by the data providers as listed in the Acknowledgments section. All data postprocessing steps are straightforward averaging and linear regressions as described in the text.
Acknowledgments
S. F. would like to thank the constructive reviews from anonymous reviewers and Marcia Baker and Levi Silvers for helpful feedback on an early version of this paper. S. F. acknowledges support from the National Science Fund Awards AGS-1417659 and AGS-1743753. S. F. thanks the following agencies for generously providing free access to their data: the European Centre for Medium-range Weather Forecast (ECMWF) for providing the ERA-Interim reanalysis data, the data used here were downloaded online (https://www.ecmwf.int/en/forecasts/datasets/archive-datasets/reanalysis-datasets/era-interim); the U.K. Met Office and the Hadley Centre for providing the HadISST data, downloaded online (https://www.metoffice.gov.uk/hadobs/hadisst); NASA for providing the MERRA-2 reanalysis data, the data used here were downloaded online (https://disc.gsfc.nasa.gov); and NASA for providing the CERES/EBAF v4 data, which was downloaded online (https://ceres.larc.nasa.gov/compare_products.php). The GPCP data were provided by the NASA/Goddard Space Flight Center's Mesoscale Atmospheric Processes Laboratory, which develops and computes the 1DD as a contribution to the GEWEX Global Precipitation Climatology Project. Data set accessed online (ftp://meso.gsfc.nasa.gov/pub/).
