The ENSO-driven bias in the assessment of long-term cloud feedback to global warming

Liu, Huan; Koren, Ilan; Altaratz, Orit; Mu, Shutian

doi:https://doi.org/10.5194/egusphere-2025-2574

Preprints

https://doi.org/10.5194/egusphere-2025-2574

Preprints

13 Jun 2025

| 13 Jun 2025

The ENSO-driven bias in the assessment of long-term cloud feedback to global warming

Huan Liu, Ilan Koren, Orit Altaratz, and Shutian Mu

Abstract. Accurately assessing the cloud feedback to global warming is essential for producing reliable climate projections. Linear regression analysis is a widely used method for this purpose, offering a straightforward approach for examining the relationship between cloud radiative effects and global mean surface temperature. However, the El Niño–Southern Oscillation (ENSO) can introduce a significant bias in these estimations, which is often overlooked due to ENSO’s relatively short periodicity. Using 72 years of reanalysis data and 150 years of simulations by 12 global climate models, this study demonstrates that ENSO can produce a bias of comparable magnitude to the estimated cloud feedback, over decades and even centuries. By providing a detailed spatial and temporal analysis of this bias, our findings underscore the importance of accounting for and removing the ENSO’s influence to improve the accuracy of cloud feedback assessment in the context of global warming.

Received: 01 Jun 2025 – Discussion started: 13 Jun 2025

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Huan Liu, Ilan Koren, Orit Altaratz, and Shutian Mu

Status: final response (author comments only)

RC1:
'Comment on egusphere-2025-2574', Anonymous Referee #1, 01 Jul 2025
This study investigates the influence of ENSO on cloud feedback estimates under global warming, using a regression-based de-ENSO method. Based on reanalysis observations and GCM simulations, the authors find that ENSO variability biases estimates of long-term cloud feedback to global warming both in the historical record and the abrupt4xCO2 experiment, with large impacts on regional scales.
Overall, the paper addresses an interesting topic. However, I have several major concerns regarding the methodology and the main findings. Some of them may arise from a lack of clarity in the Method section. I strongly suggest that the authors improve the clarity of the manuscript, particularly by providing a clearer and more detailed explanation of the main framework.
Major comments on the methodology:
1. My main question regarding the method: Is the ONI (ENSO timeseries) detrended? This is critical because the proposed linear framework (Eq. 1)
Y = a x time + b x ONI + c,
aims to separate the long-term trends (the first term, a x time) from ENSO-related variability (the second term, b x ONI). If the ONI itself contains a long-term trend, this could lead to double-counting of changes in the targeted variable (CRE or GMST) that are associated with tropical mean-state SST changes. I’m wondering if this may be the case in the abrupt4xCO2 analysis (see more on this below). In its current form, the manuscript doesn’t clearly state whether the ONI has been detrended. The text mentions that the ONI was bandpass-filtered, but the filtering timescale was not specified (it only says “to remove ONI variations beyond ENSO’s typical periodicities”). Also, Line 106 “it retains the ENSO-induced long-term trend effect” seems to suggest the ONI trend is retained. Moreover, Lines 134-136, which discuss results from GCMs, also suggest that no detrending is applied to the ONI timeseries before the decomposition. In either case, more clarification is needed. If the ONI is indeed not detrended, I’m concerned about the linearity of this method and would appreciate the authors’ comments.
2. My second methodological concern is related to the residual term (c in Eq. 1): How large is its contribution? Is it sufficiently small that one can justify focusing only on the first two terms, as done in the paper? Figs. 2bc show that the sum of GMST variance explained by the first two terms is notably less than 1, and can be even smaller than 0.5 (depending on the period). This again raises questions about whether this linear decomposition is appropriate and whether the unexplained residual term undermines the interpretation of the results.
3. Finally, Figure 1 shows that ENSO’s influence on long-term changes decreases with time. Given this, it is unclear to me why the authors choose to focus on an arbitrary 40-yr period (1982-2021) throughout the paper, especially since the ERA5 reanalysis data used in this paper spans from 1950-2021.
Overall, I think the paper suffers from a lack of clarity in the Method section and would benefit from substantial revision and clarification.
Major comments on the Results:
1. CRE decomposition in Fig. 3
According to Eq. 1, the linear framework decomposes total CRE variations into two components: (1) a linear trend term (a x time) and (2) the portion associated with ENSO variability (b x ONI). In Fig. 3, however, term (1) is interpreted as the CRE change driven by the warming trend, which I find difficult to justify. It assumes that the trend in CRE is equivalent to the CRE response to long-term warming, which may not be valid. This issue may arise from ambiguity in Eq. 1. Specifically, what is the unit of the coefficient a? Is it the trend unit of the targeted variable (e.g. for CRE, it would be W/m2/year), or is it a regression coefficient with respect to long-term global warming (W/m2/K)? If it’s the former (seems more likely based on Eq. 1), I do not think it can be interpreted as “CER due to warming trend”. Either way, this concern highlights a fundamental confusion in the framework that needs to be clarified.
2. ENSO-related biases in 4xCO2 experiment
I was quite surprised by the (really) large ENSO-related biases in cloud feedback estimates from the 150-yr abrupt 4xCO2 simulations (Figs. 6, 7), considering (1) the long timescales (150-yr) of the experiment and (2) the potential high signal-to-noise ratio in this strong forcing scenario. It again raises concerns related to the methodological question of whether the ONI has been detrended. If the ONI timeseries contains a strong linear trend in this case, the trend could actually reflect forced mean-state changes rather than ENSO variability. In that case, the current method might be attributing part of the long-term signal to ENSO, thus overestimating the ENSO-related contributions by double-counting tropical Pacific SST trends.
To address this, I suggest the authors show the timeseries of GSMT and global-mean CRE over the course of the simulations, either for each model or one representative model. It should include both the full variations as well as their decomposed components (the long-term trends and ENSO-related variations). This would allow us to directly asses the evolution and relative magnitude of the two terms, and to verify whether the ENSO related signals are not being mixed with the global warming trend.

Specific comments
Figure 1: The filtered ONI timeseries appears to have removed much of the high-frequency variability rather than the low-frequency variability?

Line 85: Why not use SST (a readily available variable in GCM output) to compute the ONI to be consistent with observations?

Line 105-106: what “delayed components of ENSO-related variations” are referred to here?

Line 135: Does this mean that 9 out of 12 GCMs actually show a significant ENSO/ONI trend over this period? If so, does this linear decomposition still hold?

Line 143-144: This sentence is very confusing and unclear, I do not understand what is meant here. Please consider rephrasing.

Line 149: Based on Fig. 2b, the variance in GSMT explained by the trend over a randomly-selected 40yr period (other than 1982-2021) can be as low as 0.3. Similarly, ENSO’s contribution (R²_ONI) is up to 0.1. This suggests that more than 50% of the total GMST variance could come from the residual term. If so, this linear decomposition doesn’t seem to work well and may not accurately reproduce the original variance.

Line 165: What exactly is meant by “covariations between clouds and the warming trend”? As noted in my major concerns, is this essentially just the trend in CRE?

Line 169: While it’s true that ENSO has a relatively small impact on the GMST during this period, it may have a more notable impact on regional surface temperature variations (e.g. in the tropical Pacific). If so, this statement may be unfair. For a more solid comparison, I suggest the authors show spatial maps of surface temperature variance explained by the warming trend and by ENSO (similar to Fig. 3 but for TS instead of CRE). In addition, it would also be helpful to show the global-mean CRE timeseries along with its decompositions into the linear trend and ENSO-related component (similar to Fig. 1 but for global-mean CRE instead of GMST).

Related to Fig. 3: how much of the regional variance is associated with the residual term? It would be informative to provide another row of panels showing the contribution of the residual term (c in Eq. 1)
Citation: https://doi.org/10.5194/egusphere-2025-2574-RC1
RC2:
'Comment on egusphere-2025-2574', Anonymous Referee #2, 05 Aug 2025
Overview
This paper provides a novel, straightforward framework for assessing the separate influences of (1) externally forced long-term trends and (2) natural climate variability on regression-based estimates of climate feedbacks. The authors show that by removing ENSO from observed and modeled estimates of surface temperature and the cloud-radiative effect (CRE), the resulting local cloud feedbacks (estimated by regressing spatially-resolved CRE against global-mean surface temperature) are notably distinct from the feedbacks obtained without removing ENSO.
The "de-ENSO" methodology proposed by the authors appears robust and their results appear physically sound. However, the implications of their results for previous estimates of observed and modeled cloud feedbacks are not yet clear, and their recommendation that future related research adopt this procedure is not yet fully justified. The "relative bias" and "ENSO effect minimal time" metrics proposed by the authors may also not be robust, while other results from the de-ENSO methodology that may be of broader interest to the climate dynamics community are missing. I therefore recommend reconsideration after substantial revisions to the figures and text addressing the below concerns.
Major suggestions
I suggest the authors remove the "relative bias" metrics in Figure 4, and remove Figures 3 and 7 altogether. Since the denominator in either metric may closely approach zero, the robustness and interpretation of the results is unclear. The yellow colorbar in Figure 3 illustrates this issue: While cases where ENSO explains from 1 to 100 times the variance in CRE compared to long-term trends are shown very clearly, cases where the long-term trend explains from 1 to 100 times the variance in CRE compared to ENSO are hidden in pale yellow. This overemphasizes the impact of ENSO relative to long-term trends. The metrics in Figures 4 and 7 are even more vulnerable to this issue, since the local cloud feedback changes sign across different regions. The motivation behind normalizing by the local feedback may also rest on a common misconception regarding climate feedback regressions. That is, regression slopes of 0 W m-2 K-1 may indicate physically meaningful feedback values rather than "unsuccessful" results. In the feedback context, the strongly negative Planck feedback is the reference value, while 0 W m-2 K-1 indicates that other processes are counteracting the Planck feedback. The 0 W m-2 K-1 result is also not necessarily highly uncertain, since the uncertainty of the regression slope depends only on the variance in the residuals, which can still be arbitrarily small (e.g., constant CRE with rising temperature).

I suggest the authors remove Figure 5 (the CMIP-based "ENSO effect minimal time"), then either (1) remove the ERA5-based "ENSO effect minimal time" in Figure 4, or (2) replace this metric with an alternative metric based on the "absolute bias". The robustness of the current metric is unclear, since it depends on the uncertain relative bias term (see above). As an example for an alternative metric, the authors could pick a reasonable precision threshold (e.g., 0.1 W m-2 K-1) and show the average number of years required until the absolute value of the "absolute bias" falls and remains below the threshold.

I suggest the authors only use ERA5 to (1) illustrate the robustness and physical interpretation of the de-ENSO methodology (Figures 2B and 2C), and optionally (2) estimate the "ENSO effect minimal time" (Figure 4, bottom row; see above). Beyond this, I recommend the authors replace the ENSO bias estimates in the top row of Figure 4 with results obtained from observational data rather than a reanalysis product. The results should be much more robust, since ERA5 estimates of cloud-radiative effect (CRE) are significantly biased compared to satellite-based estimates of CRE (e.g., Loeb et al. 2022, DOI 10.1029/2022JD036686). The results should also be more directly relevant to the climate dynamics community, since a large number of recent studies use satellite observations to estimate climate feedbacks (more than cited here). To estimate CRE, the authors could use the energy-balanced-and-filled (EBAF) CERES product (e.g., He et al. 2021, DOI 10.1029/2020GL092309; Davis et al. 2024, DOI 10.1029/2024GL112774), a combination of CERES and ERBE (e.g., Uribe et al. 2024, DOI 10.5194/acp-24-13371-2024), or optionally estimate clear-sky fluxes from ERA5 (e.g., Dessler and Loeb 2013, DOI 10.1002/jgrd.50199). To estimate surface temperature, a more direct observational data set like HadCRUT5 or GISTEMP4 could be used. The full available record should also be used instead of the 1982-2021 example period.

I suggest the authors add two rows above the "absolute bias" term in Figure 4: The first row showing local CRE feedbacks before the de-ENSO procedure, the second row showing local CRE feedbacks after the de-ENSO procedure (note these should be based on satellite observations rather than ERA5; see above). The "absolute bias" can then be understood visually as the difference between the first and second rows. Each row should also use the same blue-red colorbar and (if possible) the same minimum and maximum colorbar values. This will give a qualitative picture of the sign and relative magnitude of each term across regions. Without these results, it is difficult to contextualize the importance of "ENSO biases" and their possible impact on the interpretation of local feedback processes (e.g., over the Southern Ocean or in the subtropical stratocumulus regions).

In most studies, local climate feedbacks are used to interpret the physical and regional processes contributing to global climate feedbacks. Thus, while "ENSO biases" in local feedbacks may affect this interpretation, any biases in the global feedbacks themselves may be more directly relevant to the climate dynamics community. I therefore suggest the authors add a new table or bar-plot after Figure 4, showing satellite-based estimates of (1) global CRE feedbacks before the de-ENSO procedure, (2) global CRE feedbacks after the de-ENSO procedure, and (3) the difference between these terms (i.e., the global-average ENSO bias). Note that since the least-squares linear regression slope Sum[Y'X']/Sum[X'^2] is a linear operator on Y, these terms should be equivalent to the global average of each panel in Figure 4. Similarly, I suggest the authors add a new table or bar-plot after Figure 6, showing CMIP6 estimates of the global-average ENSO bias. To further address recent literature, the authors may also wish to explore "ENSO biases" in the short-term (typically years 1-20; Andrews et al. 2015, DOI 10.1175/JCLI-D-14-00545.1) and long-term (years 21-150) components of the 4xCO2 response. But this last suggestion is not critical.

Previous studies have quantified the "feedbacks" associated with (primarily ENSO-driven) internal variability by regressing observed radiative flux against surface temperature after subtracting the long-term trend from each term (e.g., Zhou et al. 2015, DOI 10.1002/2015GL066698; Dessler and Forster 2018, DOI 10.1029/2018JD028481; Lutsko 2018, DOI 10.1029/2018GL079236; Davis et al. 2024, DOI 10.1029/2024GL112774). These "interannual feedbacks" may be similar to the "ENSO bias" term used in this paper -- but the referenced papers frame them as metrics for a different physical process rather than a bias, and the referenced papers show the "interannual feedback" is itself related to the long-term climate feedback across CMIP models. I therefore suggest the authors use more neutral language for the "ENSO bias" term, e.g. "ENSO contribution" or "ENSO adjustment". For added relevance, the authors may also wish to compare their local or global-average "ENSO bias" results with results from these papers.

Minor suggestions
All paragraphs: Please add vertical space or indentation before each paragraph. Currently it is a bit difficult to differentiate separate paragraphs.

Lines 112, 106, 117: Please re-format the numbered equations to follow ACP style guidelines (horizontal centering on separate lines, with empty space above and below, and equation numbers in parentheses on the right-hand side).

Lines 178, 179, 182, 217, 219, 226, 232: I suggest replacing the term "absolute bias" with e.g. "ENSO contribution" or "ENSO adjustment" (see above).

Lines 38, 91, 92, 97, 103, 105, 115, 174, 192, 245, 249: The term "de-ENSO" is grammatically unusual. I suggest replacing "de-ENSO method" on the referenced lines with "ENSO-correction method", or consider not naming the method at all (e.g., on line 38, "regression-based de-ENSO method" can be replaced with "regression-based method", since it is clear from the subsequent clause that this method removes ENSO). The subscript "deENSO" used in equations could then be replaced with e.g. "trend" (since the method seeks to capture the trend component), or an asterisk or prime superscript denoting an anomaly (since each de-ENSO result is a residual with respect to the ENSO-fit).

Lines 17-19: The formatting used to describe each CRE term is unusual. I suggest replacing with "shortwave cloud-radiative effect", "longwave cloud-radiative effect" and "net cloud-radiative effect".

Lines 54-57: The formatting used to describe each radiative flux term is unusual. I suggest replacing with "net top-of-atmosphere (TOA) shortwave flux", "TOA longwave flux", "TOA clear-sky shortwave flux", and "TOA clear-sky longwave flux". The additional information in parentheses can be deleted (see below).

Lines 54-57, Lines 71-72, Lines 78-79: I don't think it's necessary to spell out the variable names used in the ERA5 and CMIP6 data files (i.e., TSR, TSRC, TTR, TTRC, tas, rsut, rsutcs, rlut, rlutcs). Tracking them all is a bit confusing, and the relevant variables in each data set should be clear from your descriptions. I suggest deleting the abbreviations and replacing with the descriptions suggested above when referencing these quantities.

Lines 66, 110, 122, 127, 148, 157, 161, 175, 178, 212: The date format "MM.YYYY" may not follow ACP style guidelines. I suggest either spelling out the calendar month (e.g. January 1950 to December 2021) or using 3-character abbreviations (e.g. Jan. 1950 to Dec. 2021).

Lines 14, 26, 36, 149, 169 (twice), 208, 210, 234: The phrase "the ENSO" is unusual, since acronyms are typically used without definite articles. Please replace instances of "the ENSO" on the referenced lines with "ENSO".

Lines 50-64, Lines 76-85: The items (1) and (2) should be formatted as a numbered list. The sentence introducing the numbered list can also be shorter and less specific, e.g. "For each data set, our analysis is based on the following two-step approach:".

Lines 61-63: The description of the variant label "r1i1f1p1" can be deleted and replaced with a reference to Eyring et al. 2016 (as in the following sentence).

Lines 85-87: The weighting methodology and details here are unnecessary. The authors can closely approximate grid cell area using the product of the cosine of the central latitude (in radians) with the longitude- and latitude-widths of the cell (only required if they vary in space). Plotting the cosine weights against the exact arc length weights should reveal very close agreement up to grid cell widths outside the range used by CMIP6 models.

Additional suggestions
There are a number of other grammatical and typographical errors throughout the text that should be addressed before re-submission. Some examples and suggested corrections:
Line 10: "in these estimations" -> "in these estimates".

Line 20: "climate predictions" -> "climate change projections" or "projections of climate change".

Line 23: There is an extra space after the comma following "natural climate variability".

Line 45: "Based on which, the Oceanic Niño Index (ONI) is derived for measuring" -> "For each dataset, we derive the Oceanic Niño Index (ONI) to measure"

Line 54: "sea surface temperature" -> "sea-surface temperature"

Line 61: "usethe" -> "use the"

Line 67: "is a baseline experiment of the [...] experiments" -> "is a [...] experiment"

Line 68: "immediate climate response" -> "climate response" (the forcing is immediate, but the response is studied

over decades and centuries)

Line 79: "Global Mean Surface Temperature" -> "global-mean surface temperature" (upper case should be reserved for proper nouns)

Line 112: "OLS correlation slope" -> "OLS regression slope"

Line 125: "marks" -> "indicates"

Line 140: "Of course, " can be deleted.

Line 145: "As shown, " can be deleted.

Line 150: "Please note that " can be deleted.

Line 149: The dash after "ENSO" should be removed.

Line 151: "get similar results" -> "found similar results".

Line 158: The comma after "ENSO" should be removed.

Line 168: "It's clear that, " can be deleted.

Line 175: "presents" -> "shows"

Line 190: "an almost opposite one" -> "almost opposite changes"

Line 195: "As mentioned in" -> "As shown by".

Line 195: "To quantify it" -> "To quantify the impact".

Line 197: "introduce the concept of" -> "using a metric we call"

Line 198: Commas surrounding "for which" can be deleted.

Line 221: "on one hand" can be deleted.

Line 221: "on the other hand" -> "However" (new sentence).

Line 225: "between the 12 models, GCMs like" -> "between the 12 models. For example,"

Line 233: "As discussed before" -> "As discussed above".

Line 235: "Current GCMs present" -> "many GCMs have".
Citation: https://doi.org/10.5194/egusphere-2025-2574-RC2

Huan Liu, Ilan Koren, Orit Altaratz, and Shutian Mu

Viewed

Total article views: 377 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
296	66	15	377	9	23

HTML: 296
PDF: 66
XML: 15
Total: 377
BibTeX: 9
EndNote: 23

Views and downloads (calculated since 13 Jun 2025)

Month	HTML	PDF	XML	Total
Jun 2025	98	21	9	128
Jul 2025	55	14	3	72
Aug 2025	104	30	3	137
Sep 2025	39	1	0	40

Cumulative views and downloads (calculated since 13 Jun 2025)

Month	HTML	PDF	XML	Total
Jun 2025	98	21	9	128
Jul 2025	55	14	3	72
Aug 2025	104	30	3	137
Sep 2025	39	1	0	40

Viewed (geographical distribution)

Total article views: 374 (including HTML, PDF, and XML) Thereof 374 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 06 Sep 2025

Short summary

Clouds play a crucial role in Earth's climate by reflecting sunlight and trapping heat. Understanding how clouds respond to global warming (cloud feedback) is essential for climate change. However, the natural climate variability, like ENSO, can distort these estimates. Relying on long-term reanalysis data and simulations, this study finds that ENSO with a typical periodicity of 2–7 years can introduce a significant bias on cloud feedback estimates on even decadal to century time scales.


Total:	0
HTML:	0
PDF:	0
XML:	0