Quantification of atmospheric carbon dioxide from the Geostationary Operational Environmental Satellite (GOES East)

Aaron Sonabend-W; Anna M. Michalak; Christopher Van Arsdale; John Platt; Sean Campbell

arxiv: 2605.23991 · v2 · pith:7KRBQURVnew · submitted 2026-05-17 · ⚛️ physics.ao-ph · astro-ph.EP· cs.LG

Quantification of atmospheric carbon dioxide from the Geostationary Operational Environmental Satellite (GOES East)

Aaron Sonabend-W , Sean Campbell , John Platt , Christopher Van Arsdale , Anna M. Michalak This is my paper

Pith reviewed 2026-06-30 18:59 UTC · model grok-4.3

classification ⚛️ physics.ao-ph astro-ph.EPcs.LG

keywords XCO2 estimationGOES-Eastneural networkssatellite remote sensingatmospheric carbon dioxideOCO-2TCCON

0 comments

The pith

A neural network estimates realistic column CO2 from geostationary satellite imagery.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Researchers built DeepXCO2 to infer dry-air column CO2 from GOES-East's frequent spectral observations. The network uses 16 bands, weather data, surface info, and geometry as inputs. Training relies on matches with OCO-2 and OCO-3 soundings. Tests on unseen years and TCCON sites confirm it tracks real XCO2 changes. Urban and farm case studies demonstrate its ability to spot local signals.

Core claim

The central claim is that a physics-guided neural network can produce XCO2 estimates from GOES-East data that match the variability seen in dedicated CO2 sensors and ground networks, despite lower precision.

What carries the argument

DeepXCO2: a single-pixel physics-guided neural network trained on collocated GOES-East and OCO observations to map spectral time series to XCO2.

If this is right

Enables 10-minute temporal sampling of XCO2 over the western hemisphere.
Reveals CO2 enhancements over cities and drawdown over agricultural areas.
Supplies a multi-year contiguous record for variability studies.
Supplements sparse high-precision measurements with high-frequency coverage.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Application to other geostationary platforms could extend coverage globally.
Useful for detecting short-term emission changes if biases are controlled.
Best suited for relative changes rather than absolute flux quantification.

Load-bearing premise

Collocated GOES-East and OCO-2/OCO-3 observations supply unbiased training targets that let the network generalize without systematic error.

What would settle it

Independent validation against aircraft profiles or additional TCCON sites during a strong local emission event would test if the reported enhancements are accurate in magnitude.

read the original abstract

There is a growing urgency to track greenhouse gasses with the resolution, precision and accuracy needed to support independent verification of $CO_2$ fluxes at local to global scales. The current generation of space-based sensors, however, only provides sparse observations in space and time. This challenge has fueled interest in the potential use of data from existing missions originally developed for other applications to infer global greenhouse gas variability. The Advanced Baseline Imager (ABI) onboard the Geostationary Operational Environmental Satellite (GOES-East), operational since 2017, provides full coverage of much of the western hemisphere at 10-minute intervals from geostationary orbit across 16 spectral channels at an approximately 2 km$^2$ spatial resolution. Here, we leverage this high spatial coverage and temporal revisit to develop Deep$XCO_2$, a single-pixel, physics-guided neural network to estimate dry-air column $CO_2$ mole fraction ($XCO_2$). Deep$XCO_2$ employs a time series of GOES-East's 16 spectral bands, ECMWF ERA5 lower tropospheric meteorology, MODIS surface reflectance, solar and satellite viewing geometry, and day of year. The network was trained on collocated GOES-East and OCO-2/OCO-3 observations. Deep$XCO_2$ is able to capture realistic $XCO_2$ variability when compared against a held-out year of OCO-2 and OCO-3 observations, and against observations from the TCCON network. We also present case studies illustrating the use of Deep$XCO_2$ to observe $XCO_2$ enhancements over urban areas and drawdown over agricultural regions. Overall, while the precision of GOES-East derived $XCO_2$ can never rival that of dedicated instruments, the unprecedented combination of contiguous geographic coverage, 10-minute temporal frequency, and multi-year record offers the potential to observe aspects of atmospheric $CO_2$ variability currently unseen from space.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper gives a concrete neural-net implementation for XCO2 from GOES-East that reaches TCCON and held-out OCO levels, but the training-target bias risk flagged in the stress test looks real and the abstract still withholds the actual error numbers.

read the letter

The main takeaway is that DeepXCO2 produces XCO2 fields at 10-minute cadence over the western hemisphere by training a single-pixel network on GOES ABI bands plus ERA5 and MODIS auxiliaries, with OCO-2/3 as targets. It shows plausible variability against a held-out OCO year and TCCON stations, plus a couple of urban-enhancement and agricultural-drawdown examples.

What is actually new is the specific mapping from GOES-East's 16 channels and fixed viewing geometry to column CO2; prior work on dedicated sensors does not cover this operational geostationary case. The physics-guided framing and the use of day-of-year plus viewing angles are reasonable choices that keep the model from being a pure black box.

The soft spot is the training setup. OCO retrievals carry their own forward-model and aerosol assumptions tuned to a different instrument and orbit. The inputs listed do not include explicit terms that correct for the spectral mismatch or the constant GOES viewing angle, so the network could be fitting OCO-specific artifacts rather than true atmospheric signals. The held-out OCO year and TCCON comparisons would not catch that because they sit in similar regimes. The abstract gives no RMSE, bias, or uncertainty numbers and no description of how collocations or data exclusions were handled, which makes it impossible to judge whether the reported realism is at the level needed for flux work.

This is the kind of paper that belongs in a methods journal or a remote-sensing venue. A reader who wants to test high-frequency XCO2 products from existing satellites will find the implementation useful to try; someone needing quantified precision for verification will need the missing error tables first. It deserves peer review because the core idea is testable and the data sources are public, but the review should focus on whether the bias concern is addressed in the full methods and results.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces DeepXCO2, a single-pixel physics-guided neural network that ingests GOES-East ABI 16-channel time series, ERA5 lower-tropospheric meteorology, MODIS surface reflectance, solar/satellite angles, and day-of-year to regress dry-air column CO2 mole fraction (XCO2). The network is trained on collocated GOES-East and OCO-2/OCO-3 observations; the central claim is that it reproduces realistic XCO2 variability on a held-out year of OCO-2/OCO-3 data and on TCCON ground-based measurements, while also enabling case studies of urban enhancements and agricultural drawdown at 10-minute cadence over the western hemisphere.

Significance. If the generalization holds, the work supplies the first geostationary, 10-minute, multi-year XCO2 product over a large continental domain from an existing operational sensor. This temporal density is unavailable from polar-orbiting sounders and could support studies of diurnal cycles, urban plumes, and regional flux constraints, even if the absolute precision remains coarser than dedicated CO2 missions.

major comments (2)

[Abstract / Methods (training)] Abstract and training description: the model is trained to regress XCO2 using OCO-2/OCO-3 retrievals as targets, yet the supplied inputs contain no explicit correction terms for the differences in spectral response functions, push-broom vs. fixed geostationary geometry, aerosol priors, or surface-pressure constraints between the two instruments. If the learned mapping therefore encodes OCO-specific retrieval artifacts, the held-out OCO year and TCCON comparisons (which share similar atmospheric and algorithmic regimes) will not detect the bias; this assumption is load-bearing for the claim that DeepXCO2 generalizes to GOES data.
[Abstract / Validation] Validation section: the abstract reports held-out validation and TCCON comparison but supplies no quantitative error metrics (RMSE, bias, correlation), uncertainty propagation, or explicit data-exclusion rules. Without these, it is impossible to judge whether post-hoc choices affect the central claim of realistic variability.

minor comments (1)

[Abstract] The abstract states that the model 'captures realistic XCO2 variability' without accompanying numerical values; adding at least one summary statistic (e.g., correlation or RMSE against TCCON) would make the claim more concrete.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed review. We address each major comment below, providing our response and indicating planned revisions to the manuscript.

read point-by-point responses

Referee: [Abstract / Methods (training)] Abstract and training description: the model is trained to regress XCO2 using OCO-2/OCO-3 retrievals as targets, yet the supplied inputs contain no explicit correction terms for the differences in spectral response functions, push-broom vs. fixed geostationary geometry, aerosol priors, or surface-pressure constraints between the two instruments. If the learned mapping therefore encodes OCO-specific retrieval artifacts, the held-out OCO year and TCCON comparisons (which share similar atmospheric and algorithmic regimes) will not detect the bias; this assumption is load-bearing for the claim that DeepXCO2 generalizes to GOES data.

Authors: We acknowledge the concern about potential encoding of OCO-specific artifacts. DeepXCO2 is trained as an empirical transfer function from GOES ABI time series to OCO XCO2 targets using collocated observations; the physics-guided inputs (ERA5 meteorology, viewing geometry, MODIS reflectance, and day-of-year) are intended to capture key atmospheric and observational differences. The held-out OCO test evaluates temporal generalization within the OCO regime, while TCCON provides an independent validation using a distinct ground-based technique. We will add a limitations subsection in the methods/discussion to explicitly address this assumption, the absence of explicit spectral corrections, and the role of the physics-guided features in mitigating instrument differences. revision: partial
Referee: [Abstract / Validation] Validation section: the abstract reports held-out validation and TCCON comparison but supplies no quantitative error metrics (RMSE, bias, correlation), uncertainty propagation, or explicit data-exclusion rules. Without these, it is impossible to judge whether post-hoc choices affect the central claim of realistic variability.

Authors: We agree that quantitative metrics and explicit exclusion rules are necessary for rigorous evaluation. The full manuscript reports RMSE, bias, and correlation values for the held-out OCO-2/OCO-3 year and TCCON sites in the validation section, along with data-exclusion criteria (quality flags and collocation thresholds). To address the comment, we will revise the abstract to include key quantitative metrics and ensure all exclusion rules and uncertainty details are clearly documented in the methods. revision: yes

Circularity Check

0 steps flagged

No significant circularity: derivation relies on external OCO-2/3 targets and independent TCCON validation

full rationale

The paper trains DeepXCO2 on collocated GOES-East inputs with OCO-2/OCO-3 XCO2 targets and evaluates against a held-out year of OCO observations plus independent TCCON ground truth. This structure keeps the central claim (realistic variability capture) anchored to external benchmarks rather than any self-referential fit, definition, or self-citation chain. No equations, ansatzes, or uniqueness theorems are shown that reduce the output to the inputs by construction. The derivation is therefore self-contained against external data sources.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Only the abstract is available, so the ledger is limited to the core modeling assumptions stated or implied there. The network weights constitute the primary fitted content; no new physical entities are postulated.

free parameters (1)

neural network weights and biases
Learned during supervised training on collocated GOES-OCO pairs; the central claim depends on these fitted values generalizing to new scenes.

axioms (1)

domain assumption GOES-East's 16 spectral bands plus ERA5 meteorology, MODIS reflectance, and geometry contain sufficient information to infer XCO2 after training on OCO targets
Implicit premise required for the supervised-learning approach to succeed; stated via the choice of inputs and training strategy.

pith-pipeline@v0.9.1-grok · 5924 in / 1359 out tokens · 30040 ms · 2026-06-30T18:59:35.202192+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

4 extracted references · 4 canonical work pages · 3 internal anchors

[1]

Aygün, E., Belyaeva, A., Comanici, G., Coram, M., Cui, H., Garrison, J., Kast, R.J., McLean, C.Y., Norgaard, P.C., Shamsi, Z., Smalling, D., Thompson, J., Venugopalan, S., Williams, B.P., He, C., Martinson, S., Plomecka, M., Wei, L., Zhou, Y., Zhu, Q., Abraham, M., Brand, E., Bulanova, A., Cardille, J.A., Co, C., Ellsworth, S., Joseph, G., Kane, M., Krueg...

work page internal anchor Pith review Pith/arXiv arXiv 2025
[2]

Bengio, Y. (2012). Practical recommendations for gradient-based training of deep architectures. In G. Montavon, G. B. Orr, & K.-R. Müller (Eds.), Neural networks: Tricks of the trade (2nd ed., pp. 437–478). Springer. https://doi.org/10.1007/978-3-642-35289-8_26 Buchwitz, M., de Beek, R., Burrows, J. P., Bovensmann, H., Warneke, T., Notholt, J., Meirink, J...

work page doi:10.1007/978-3-642-35289-8_26 2012
[3]

Network In Network

Springer, Berlin, Heidelberg. https://doi.org/10.1007/978-3-642-35289-8_3 Lin, M., Chen, Q., & Yan, S. (2013). Network in network. arXiv preprint arXiv:1312.4400 . https://doi.org/10.48550/arXiv.1312.4400 Maddy, E. S., Barnet, C. D., Goldberg, M., Sweeney, C., & Liu, X. (2008). CO ₂ retrievals from the Atmospheric Infrared Sounder: Methodology and validat...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1007/978-3-642-35289-8_3 2013
[4]

Ramachandran, P., Zoph, B., & Le, Q. V. (2017). Searching for activation functions. arXiv preprint arXiv:1710.05941 . https://doi.org/10.48550/arXiv.1710.05941 Science Computing Facility, J. P. L. (2017). OCO-2 Level 2 bias-corrected XCO2 and other select fields from the full-physics retrieval aggregated as daily files, Retrospective processing V11.2r [Da...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1710.05941 2017

[1] [1]

Aygün, E., Belyaeva, A., Comanici, G., Coram, M., Cui, H., Garrison, J., Kast, R.J., McLean, C.Y., Norgaard, P.C., Shamsi, Z., Smalling, D., Thompson, J., Venugopalan, S., Williams, B.P., He, C., Martinson, S., Plomecka, M., Wei, L., Zhou, Y., Zhu, Q., Abraham, M., Brand, E., Bulanova, A., Cardille, J.A., Co, C., Ellsworth, S., Joseph, G., Kane, M., Krueg...

work page internal anchor Pith review Pith/arXiv arXiv 2025

[2] [2]

Bengio, Y. (2012). Practical recommendations for gradient-based training of deep architectures. In G. Montavon, G. B. Orr, & K.-R. Müller (Eds.), Neural networks: Tricks of the trade (2nd ed., pp. 437–478). Springer. https://doi.org/10.1007/978-3-642-35289-8_26 Buchwitz, M., de Beek, R., Burrows, J. P., Bovensmann, H., Warneke, T., Notholt, J., Meirink, J...

work page doi:10.1007/978-3-642-35289-8_26 2012

[3] [3]

Network In Network

Springer, Berlin, Heidelberg. https://doi.org/10.1007/978-3-642-35289-8_3 Lin, M., Chen, Q., & Yan, S. (2013). Network in network. arXiv preprint arXiv:1312.4400 . https://doi.org/10.48550/arXiv.1312.4400 Maddy, E. S., Barnet, C. D., Goldberg, M., Sweeney, C., & Liu, X. (2008). CO ₂ retrievals from the Atmospheric Infrared Sounder: Methodology and validat...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1007/978-3-642-35289-8_3 2013

[4] [4]

Ramachandran, P., Zoph, B., & Le, Q. V. (2017). Searching for activation functions. arXiv preprint arXiv:1710.05941 . https://doi.org/10.48550/arXiv.1710.05941 Science Computing Facility, J. P. L. (2017). OCO-2 Level 2 bias-corrected XCO2 and other select fields from the full-physics retrieval aggregated as daily files, Retrospective processing V11.2r [Da...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1710.05941 2017