CSRFormer: An Instantaneous Global-ocean Clear-sky Radiative Flux Dataset Derived From CERES

Zheng, Boyang; Zhu, Yannian; Cao, Yang; Huang, Kang-En; Liu, Jihu; Wang, Yichuan; Shi, Jun; Wei, Yicheng; Rosenfeld, Daniel; Zhou, Chen; Wang, Minghuai

doi:10.5194/essd-2026-360

Preprints

https://doi.org/10.5194/essd-2026-360

Preprints

11 Aug 2026

| 11 Aug 2026

Status: this preprint is currently under review for the journal ESSD.

CSRFormer: An Instantaneous Global-ocean Clear-sky Radiative Flux Dataset Derived From CERES

Boyang Zheng, Yannian Zhu, Yang Cao, Kang-En Huang, Jihu Liu, Yichuan Wang, Jun Shi, Yicheng Wei, Daniel Rosenfeld, Chen Zhou, and Minghuai Wang

Abstract. Accurate estimates of the Cloud Radiative Effect (CRE) require clear-sky radiative fluxes that are temporally consistent with the corresponding all-sky observations. Because passive satellite sensors cannot directly observe the sub-cloud clear-sky state, instantaneous clear-sky fluxes in cloudy regions are commonly approximated using radiative transfer calculations or temporally interpolated products, both of which introduce limitations. Here we present CSRFormer, a physically constrained deep-learning framework for estimating instantaneous clear-sky shortwave and longwave radiative fluxes over the global ocean. CSRFormer uses ERA5 atmospheric profiles, MERRA-2 aerosol properties, and NOAA sea surface temperature as inputs, and is trained against CERES Single Scanner Footprint (SSF) clear-sky observations. Over independent clear-sky samples, the model shows strong agreement with CERES SSF, with R² values of 0.81 for shortwave flux and 0.99 for longwave flux, corresponding to estimated uncertainties (RMSE) of ~3.2 W m⁻² and ~2.0 W m⁻², respectively. Because the predictions are matched to the observation time of the CERES overpass, the resulting dataset avoids the temporal mismatch inherent in hourly mean products. Application of CSRFormer to cloudy scenes provides an observation-trained estimate of the theoretical clear-sky reference state beneath clouds for CRE analyses over the global ocean. The dataset is therefore a useful framework for examining the radiative effects of clouds and their covariation with cloud properties at the satellite overpass time. The CSRFormer v1.0 dataset is available at https://doi.org/10.5281/zenodo.20046058.

Received: 09 May 2026 – Discussion started: 11 Aug 2026

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

Boyang Zheng, Yannian Zhu, Yang Cao, Kang-En Huang, Jihu Liu, Yichuan Wang, Jun Shi, Yicheng Wei, Daniel Rosenfeld, Chen Zhou, and Minghuai Wang

Status: open (until 05 Oct 2026)

Post a comment Subscribe to comment alert

RC1:
'Comment on essd-2026-360', Anonymous Referee #1, 31 Aug 2026 reply
Summary
CSRFormer is a physically-constrained transformer that reconstructs instantaneous clear-sky TOA SW/LW flux over ocean from ERA5 profiles, MERRA-2 aerosols, and OISST, trained against CERES SSF. The pitch is clean and useful: give CRE studies a clear-sky reference matched to the actual overpass time, instead of RT-modeled (CRS) or interpolated (SYN1deg). Validation against SSF is strong and holds up reasonably on the independent 2018 test year, the CRS/SYN1deg/ERA5/MERRA-2 intercomparisons are a nice addition, and the cloudy/clear consistency check plus cloud-type CRE decomposition show real effort to validate something that's fundamentally unobservable (clear sky under a cloud). Solid dataset paper, squarely in ESSD's scope.
Major comments
The 7:3 train/val split in 3.1 isn't described in enough detail — is it random at the sample level, or blocked by day/orbit? If it's a random split, footprints from the same overpass are strongly autocorrelated (similar AOD, humidity, SST across a swath), so some of that validation performance is probably leakage rather than genuine generalization. That would also explain the pattern in the results: SW R² drops from 0.93 (val) to 0.81 (2018 test) while LW barely moves (0.99 → 0.99) — SW is more sensitive to synoptic-scale aerosol structure, which is exactly what a random split would leak across train/val. Worth clarifying how the split was actually done, and if it is sample-level random, a day-blocked validation run (even on a subset) would give a more honest read on how much the val numbers overstate true skill.

I'd like to see more evidence that the transformer itself is doing the work, not just the physical constraint in Eq. 1. LW is pinned to σT_skin⁴ with SST also fed in directly, and Fig. 5 shows SST as by far the dominant LW predictor — which tracks with the paper's own citations (Allan et al., 2007; Koll and Cronin, 2018) that clear-sky OLR is already close to linear in surface temperature and column water vapor. Given that, my guess is a good chunk of the R²=0.99 is just the T⁴ relationship doing the heavy lifting, not the attention mechanism resolving the 37+72-level profiles — which is supposed to be the paper's main advance over the MLP in Zheng et al. (2025). A simple baseline comparison (σT_skin⁴ minus a plain SST/TCWV regression for ΔLW, no transformer) using data already in hand would make the case for the architecture much more convincing.

Minor comments
3.2: "trained with a learning rate of in a 5×10⁻⁵" reads garbled — please check.

Eq. 2 weights (α = 7, 3, 2, 0.5, 3): a sentence on how these were set (tuned vs. manual) would help anyone reusing the loss design.

Fig. A6a/b: ERA5 SW comparison gives R² = −2.06 and −2.40 — a negative R² this large is striking enough to merit one explicit sentence confirming it's not a units/product-definition mismatch, alongside the existing physical explanation (idealized albedo, climatological aerosol).

The archived NetCDF (§4.5) carries flux + clear-sky fraction only. Since error clearly varies by regime (aerosol load, cloud-edge "brightening"), a per-footprint uncertainty or QC flag would make v1.0 more usable for downstream CRE error propagation.

Reply
Citation: https://doi.org/10.5194/essd-2026-360-RC1
RC2: 'Comment on essd-2026-360', Anonymous Referee #2, 18 Sep 2026 reply

Summary:

The motivation of the paper is to estimate clear-sky fluxes during partly cloudy and overcast conditions for the observed CERES instantaneous footprint all-sky fluxes.
However, only clear-sky regions that have an area extent of at least a 20 km are used to estimate the clear-sky fluxes for partly cloudy and overcast footprint conditions. The population of clear-sky footprints is ~2% of all footprints and are not evenly distributed over the ocean as they occur over mostly the sub-tropics and rarely over the southern oceans and the ITCZ. The atmospheric profile conditions vary considerably if the clear-sky conditions are greater than 20-km than clear-sky conditions next to cloud edges, surrounded by clouds, or overcast conditions with no clear-sky. The near TOA clear-sky fluxes near clouds are colder and brighter than large domain clear-sky conditions (Eytan et al 2025). The clear-sky conditions within mostly cloudy conditions are less sampled than for dryer profiles. The clear-sky CSRFormer fluxes are then applied to overcast conditions, where the atmosphere is more humid and have greater aerosol reflectivity, for which it was not trained on and must extrapolate over. Maybe I do not fully grasp the capability of the CSRFormer methodology. Can a paragraph be added to address this.

Eytan, E., Gristey, J. J., & Feingold, G. (2025). The net radiative effect of the Ill-Defined clear-Sky in the vicinity of clouds. AGU Advances, 6, e2024AV001407. https://doi.org/10.1029/2024AV001407

If the goal of the Cloud Radiative Effect (CRE) is to derive the clear-sky flux under overcast conditions or the impact of instantaneously removing all clouds over the Earth, where the atmosphere still retains the cloudy atmospheric profile. I do not see how these unsampled clear-sky events can be estimated using CERES SSF L2 20-km clear sky fluxes that are based on clear-sky atmospheric profiles. It would be interesting to see a frequency plot similar to Fig. 4 of all SSF footprints during the training period.

Comments

Line 265. It must be remembered that the CERES products do not use the Goddard cloud properties or mask but are based on the CERES MODIS cloud retrievals designed to have as few non retrievals as possible. The cloud retrievals designate the scene type required to convert nearly all footprint radiances into fluxes. The Goddard cloud retrievals are mostly from high confidence clear-sky events disregarding the number of no retrievals from the cloud mask. From the CERES web site under references for the SSF L2 product
https://ceres.larc.nasa.gov/data/documentation/#ssf

Trepte, Q. Z., P. Minnis, S. Sun-Mack, C. R. Yost, Y. Chen, Z. Jin, F.-L. Chang, W. L. Smith, Jr., K. M. Bedka, and T. L. Chee, 2019: Global cloud detection for CERES Edition 4 using Terra and Aqua MODIS data. IEEE Trans. Geosci. Remote Sens., 57, 9410-9449, doi: 10.1109/TGRS.2019.2926620.

Line 171 Where the clear-sky footprints for CSRFormer using both the Goddard and the CERES MODIS cloud mask? How many additional CERES clear-sky footprints were rejected due to the Goddard cloud mask?

Figure 5.
I appreciate the authors including the important input features analysis, to see if the inputs align with physics. It is good to see the AOD is a primary component in clear-sky non-polar ocean SW fluxes. I wonder why SZA is a distant second. I believe this is due to the Aqua 1:30 sun-synch orbit where the SZA are constrained by LT and a bit by season. But why would SZA and cos(SZA) be treated separately, is this not redundant? I see wind speed contributions are small but impact the probability of glint. I see the glint is not considered. I see that SST is included in the SW contributions, is this to determine the near surface water vapor absorption along with specific or relative humidity? For the clear-sky LW, how is upper tropospheric humidity not considered in the feature importance?

Figure 6. I would agree that the CSRFormer algorithm would achieve a near zero bias to the 20-km clear-sky observations, which is not the case for untuned Fu-Liou computed fluxes utilized by the CERES project, since the CSRFormer algorithm was trained with the observed clear-sky fluxes. Do the CSRFormer input datasets that differ from the CERES input dataset significantly improve the consistency of the CSRFormer and observed clear-sky fluxes of is most of the impact from CSRFormer algorithm?

Figure 7. The figure caption says PDF of all-sky conditions, yet a) and b) indicate for clear-sky fluxes.

Section 4.3.2 and Fig. 9 and 10. What SYN1deg clear-sky fluxes are compared here as well as only during Terra/Aqua overpass times? Please make this clear in the text and in the Fig. 9 and 10 captions
If it is the SYN1deg clear-sky observed flux (not computed). The SYN1deg product clear-sky observed fluxes are strictly based on CERES instrument observations and are temporally interpolated using albedo models and linear interpolation between 12-hourly CERES observations over oceans. These are based on the same 20-km SSF footprint but spatially averaged over 1° region. If the comparisons are made only at the CERES observation times from the SYN1deg product (not specifically mentioned in the text), then it is either based on the SSF L2 observed or the temporally interpolated. The temporally interpolate fluxes assume the same atmospheric conditions as during the observations as stated in the text (line 465).

Conclusions. The authors trained 2019 and 2020 to compare against 2019 and 2020. If the authors were to process the whole 22-year SSF record is the training for 2019 and 2020 sufficient to apply to 2000 with confidence? Would CSRFormer fluxes capture the trends in the input datasets?

Reply

Citation: https://doi.org/10.5194/essd-2026-360-RC2

Boyang Zheng, Yannian Zhu, Yang Cao, Kang-En Huang, Jihu Liu, Yichuan Wang, Jun Shi, Yicheng Wei, Daniel Rosenfeld, Chen Zhou, and Minghuai Wang

Data sets

CSRFormer: An Instantaneous Global-ocean Clear-sky Radiative Flux Dataset Derived From CERES B. Zheng et al. https://doi.org/10.5281/zenodo.20046058

Model code and software

CSRFormer B. Zheng et al. https://github.com/Boya928861918/CSRFormer

Boyang Zheng, Yannian Zhu, Yang Cao, Kang-En Huang, Jihu Liu, Yichuan Wang, Jun Shi, Yicheng Wei, Daniel Rosenfeld, Chen Zhou, and Minghuai Wang

Viewed

Total article views: 256 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
154	56	46	256	29	62

HTML: 154
PDF: 56
XML: 46
Total: 256
BibTeX: 29
EndNote: 62

Views and downloads (calculated since 11 Aug 2026)

Month	HTML	PDF	XML	Total
Aug 2026	68	34	5	107
Sep 2026	86	22	41	149

Cumulative views and downloads (calculated since 11 Aug 2026)

Month	HTML	PDF	XML	Total
Aug 2026	68	34	5	107
Sep 2026	86	22	41	149

Viewed (geographical distribution)

Total article views: 199 (including HTML, PDF, and XML) Thereof 199 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 21 Sep 2026

Short summary

We created a global dataset of clear‑sky energy at the top of the atmosphere over the ocean that matches real satellite overpass times. Knowing how much energy the Earth reflects and emits under cloud‑free conditions helps assess how clouds influence climate. We trained a model with satellite measurements and atmospheric data to estimate clear‑sky energy fluxes at observation times. This dataset can support studies of atmospheric processes and improve comparisons between models and observations.


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