WorksComputer Science
Emulating physics-based XCO2 retrievals across OCO-2 viewing modes with compact neural networks
Abstract
Operational full-physics retrieval of the column-averaged dry-air mole fraction of carbon dioxide from satellite spectra is computationally demanding. We use singular value decomposition to compress spectra from the Orbiting Carbon Observatory-2 separately by band. Compact multilayer perceptrons use the resulting spectral scores, prior information, and viewing geometry to predict the difference between the full-physics retrieval and its operational prior. Adding the predicted difference to the prior yields an estimate of the retrieved mole fraction. The models were trained on 2015-2019 observations and validated on 2020 observations. In 2021-2025 tests, the mean annual root-mean-square errors relative to the Atmospheric Carbon Dioxide Observations from Space retrieval product were 0.89, 1.35, and 1.43 parts per million for ocean glint, land glint, and nadir observations, respectively. On East Asian nadir observations screened using the published criteria for the spectral Transformer, our model yielded a lower mean annual error than reported for that model in 2020-2022. These results show that compact neural networks using compressed spectra can efficiently emulate the full-physics retrieval across multiple viewing modes and years.
Introduction
Atmospheric CO observations are needed to constrain the distribution and evolution of carbon sources and sinks. Atmospheric concentrations integrate anthropogenic emissions, terrestrial and oceanic carbon exchange, and transport, so their spatial and temporal variations provide complementary constraints on the global carbon budget [1]. Spaceborne measurements of column-averaged dry-air CO mole fraction (XCO) extend this information beyond the coverage of surface networks, provided that small atmospheric signals can be distinguished from variations in observing conditions [2].
The Orbiting Carbon Observatory-2 (OCO-2) measures reflected sunlight in the O-A, weak-CO, and strong-CO bands [3]. Operational full-physics retrievals estimate atmospheric and surface states by iteratively matching radiative-transfer simulations to these spectra [4], [5]. This physically constrained approach accounts for absorption, scattering, and surface reflection, but repeated forward-model evaluations impose a substantial computational cost.
Machine learning provides a fast alternative by mapping measured radiances directly to atmospheric CO. Early neural-network studies estimated tropical midtropospheric CO from NOAA satellite observations [6], whereas OCO-2 applications target full-column XCO. David et al. [7] demonstrated empirical XCO retrieval from land-nadir spectra using atmospheric-model estimates as training targets, and Bréon et al. [8] extended this approach to land-glint observations, showing that excluding the weak-CO band reduced date- and location-related spectral shortcuts and recovered emission-plume enhancements absent from the training targets. Both studies split contemporary observations into training and test subsets within the same multi-year period, using alternating even and odd months [7] or random partitions [8], leaving multi-year transfer to future observation years untested.
Predicting XCO in subsequent years is challenging because rising atmospheric CO and changing observing conditions shift observations beyond the historical training distribution. Xie et al. [9] augmented East Asian training observations with simulated spectra spanning higher CO concentrations, whereas Chen et al. [10] used a spectral Transformer (SpT) to predict the residual relative to the operational XCO prior with periodic fine-tuning. A remaining question is how much model capacity spectral retrieval requires. Although OCO-2 spectra contain thousands of channels, correlated absorption lines concentrate most spectral variance into a low-dimensional subspace. Because large spectral encoders increase computational cost and the risk of overfitting to training-period conditions, we test whether linear spectral compression preserves the information needed for accurate retrieval with a compact network.
Surface reflectance and viewing geometry introduce additional radiative variability across OCO-2 observation modes. OCO-2 collects nadir (ND) and glint (GL) soundings along orbital tracks that cross both land and ocean [3]. Over water, directional reflection near the solar glint spot provides a strong shortwave-infrared signal, whereas diffuse nadir reflection is weak [11], [5]. Over land, heterogeneous surface reflectance, topography, and aerosols complicate radiative transfer in both GL and ND viewing geometries [12], [13]. Evaluating GL ocean, GL land, and ND separately therefore clarifies how retrieval accuracy and multi-year error drift depend on the observation regime. Most published OCO-2 machine-learning studies have focused on land nadir observations [7], [9], [10]. Land glint has also been investigated [8], whereas ocean-glint evaluation remains comparatively limited.
We investigate whether SVD-compressed OCO-2 spectra, together with prior and viewing-geometry information, enable compact MLPs to rapidly reproduce ACOS XCO retrievals across ocean glint, land glint and nadir. The models predict the correction to the operational prior using residual learning [10]. We assess performance in subsequent years and examine the role of spectral and prior information. In East Asia, our nadir model achieved lower RMSEs against ACOS than the published SpT results in the common years after aligning test-screening criteria.
Materials and Methods
OCO-2 data and observation domains
We matched calibrated OCO-2 Level-1B spectra in the -A (0.765 m), weak- (1.61 m), and strong- (2.06 m) bands to ACOS Level-2 Standard records (builds B11006r–B11213r) using sounding identifiers. The Level-2 product supplies the retrieved reference (/RetrievalResults/xco2 in ppm), the operational prior (xco2_apriori), solar and viewing geometry, and surface-pressure priors [3, 5]. The -A band constrains the dry-air column and effective optical path length, while the two bands provide complementary absorption sensitivity [4, 8]. Quality screening retained outcome_flag = 1, cloud_flag_abp = 0, and cloud_flag_idp {2, 3}, selecting clear A-band preprocessor cases and clear or probably clear IDP cases [14]. This screening reduces cloud-contaminated observations and excludes unsuccessful reference retrievals, limiting their influence on the learned spectral– relationship.
Screened soundings were assigned to three operational observation domains using viewing mode and the Level-2 wind-speed field. ND included all screened soundings with mode = ND. Within mode = GL, a finite nonnegative wind speed defined GL ocean and the fill value −999999 defined GL land; wind speed was used solely for domain assignment. Figure 1 shows the 2020 validation cohorts. Deterministic sampling by sounding identifier targeted 200,000 soundings per domain and year, followed by spectral-validity and duplicate screening and equal-domain rebalancing (Text S4). The final cohorts contained 1,000,000 training soundings per domain in 2015–2019 and 200,000 validation soundings in 2020 (Table S1). Holding out the next complete year makes model selection sensitive to post-training transfer. In the 2025 archive, GL land and ND extend through 30 September and GL ocean through 1 October (denoted 2025* in figures).

Figure 1. Spatial distributions of annual mean ACOS Level-2 in grid cells for the 2020 validation cohorts (200,000 soundings per domain): (a) GL ocean, (b) GL land, and (c) ND. Unsampled cells are left blank, and coastlines follow Natural Earth.
Spectral preprocessing and SVD compression
All preprocessing parameters were determined from the 2015–2019 training archive. Because wavelength sampling is not exactly the same across detector footprints and observation dates, spectra were linearly resampled onto a common grid before SVD, ensuring that each matrix column represented the same wavelength across soundings. The reference grid used the mean dispersion relation from 173,672 valid training granule–footprint combinations. Degraded band edges were excluded, retaining the central 860, 925, and 881 channels in -A, weak , and strong , respectively [3, 9] (Figure 2). Each spectrum was normalized by the mean radiance of valid channels within a fixed 15-channel continuum window, reducing broadband amplitude differences while preserving relative absorption structure. The exact windows and trimming rule are documented in Text S1 and Figure S4.

Figure 2. Channel-wise bad-sample frequency across the 2015–2019 training archive (173,672 granule-footprints) for the (a) O₂-A, (b) weak-CO₂, and (c) strong-CO₂ bands. Shaded gray regions mark degraded band edges trimmed during preprocessing, while isolated interior bad pixels are masked during per-sounding SVD projection.
For a target wavelength between adjacent native detector channels and , the resampled radiance is linearly interpolated, and the validity indicator equals unity only when both bracketing channels are unflagged:
Each resampled spectrum was divided by the mean radiance of its valid channels within the fixed continuum window (Text S1; Figure S4). A resampled channel was masked before coefficient estimation whenever either supporting native detector pixel was flagged.
Normalized training spectra were decomposed separately for each observation domain and band using uncentered truncated SVD on the common wavelength grid [15]. After continuum normalization removes broadband radiance amplitude, uncentered SVD retains the mean normalized absorption shape directly within the basis vectors. The first 30 right-singular vectors per band capture 98.65%–99.85% of centered training spectral variance (Figure 4):

Figure 4. Fraction of remaining centered spectral variance as a function of the number of retained SVD components (up to 30) for the (a) -A, (b) weak-, and (c) strong- bands across GL ocean, GL land, and ND on the 2015–2019 training set (1,000,000 spectra per domain).
Here is the matrix of normalized training spectra, and contains the first 30 basis vectors as columns. For sounding , denotes its valid channel indices. The restricted basis has rows and 30 columns, and is the corresponding normalized spectrum as a column vector. Removing a flagged channel removes both its spectral value and the corresponding basis row. The 30 coefficients minimize the sum of squared reconstruction errors over valid channels and are obtained using the Moore–Penrose pseudoinverse [16, 17]:
The implementation verifies that more than 30 valid channels remain and that has full column rank before computing its pseudoinverse, yielding the unique least-squares solution while accounting for the non-orthogonality of the restricted basis columns after channel removal. Applying this masked least-squares projection across training, validation, and test soundings produces 30 coefficients per band and 90 spectral features in total.
Model inputs
The 90 spectral coefficients were combined with nine pre-retrieval ancillary and prior predictors to form the 99-input vector (Table I). Solar zenith angle (SZA) and viewing zenith angle (VZA) were represented by their secants, encoding the geometric optical path scaling under a plane-parallel approximation where the air-mass factor scales as sec(SZA) + sec(VZA), reducing to sec(SZA) + 1 at nadir [18]. Relative azimuth angle (AZA) was represented by cos(AZA) and sin(AZA) to preserve angular periodicity, while Earth–Sun distance and relative solar velocity captured illumination and Doppler-shift geometry. Prior state inputs comprised the full-physics and A-band preprocessor surface-pressure priors (surface_pressure_apriori_fph and surface_pressure_apriori_abp) together with the operational prior (xco2_apriori) [5, 10]. All 99 inputs were standardized within each domain using the 2015–2019 training mean and standard deviation.
| Input category | Variables | Count |
| Spectral SVD scores | 30 projection scores per band: O₂-A (0.76 µm), weak CO₂ (1.61 µm), strong CO₂ (2.06 µm) | 90 |
| Viewing geometry | sec(SZA), sec(VZA), cos(AZA), sin(AZA) SZA: solar zenith angle; VZA: viewing zenith angle; AZA: relative azimuth angle | 4 |
| Solar ephemeris | Earth–Sun distance, relative solar velocity | 2 |
| Prior state variables | Surface-pressure prior (FPH), ABP pressure prior, XCO₂ prior (xco2_apriori) | 3 |
| Total inputs | 90 compressed spectral features + 9 physical geometry and prior predictors | 99 |
Table I. Summary of the 99 pre-retrieval input features used across all retrieval models.
Prior-referenced residual learning
For sounding , the training target is the prior residual between the Level-2 XCO₂ reference and its operational prior . Each regressor estimates from the 99 inputs, and the retrieved concentration is reconstructed by adding the predicted residual to the prior:
Following Chen et al. [10], the operational prior provides an observation-specific background and the regressor predicts the difference between the ACOS retrieval and that prior. The prior is retained as an input so that the correction can depend on its value. Adding the predicted residual back to the prior gives the XCO₂ estimate, with identical prediction errors in residual and XCO₂ space (). We compare direct-XCO₂ and residual prediction under matched inputs and training settings to evaluate the effect of this target formulation (§ 3-B).
Regression models and evaluation
Within each domain, we compared validation-selected LightGBM models [19] with a fully connected MLP trained using AdamW [20] on the 99-input vector and prior-residual target (Figure 3). With domain-specific MLP training hyperparameters fixed (Tables S7 and S12), a Tree-structured Parzen Estimator search evaluated 128 candidate architectures with one to four hidden layers and 32–640 units per layer. The top three completed architectures were trained with seeds 17, 29, and 43 and selected by mean 2020 validation RMSE (Text S2). The selected hidden-layer widths were 220 for GL ocean, 164 for GL land, and 468–248 for ND. Each MLP minimized minibatch mean squared error on standardized training residuals with gradient-norm clipping at 5 and domain-specific weight decay.

Figure 3. Shared prior-referenced retrieval workflow for MLP and LightGBM. Each normalized band is compressed by SVD and masked least-squares projection to 30 spectral scores. The 90 scores are combined with six geometry and ephemeris variables, two surface-pressure priors (FPH and ABP), and the prior, then standardized using training-set statistics. Either regressor predicts the prior residual, which is converted to ppm and added to the prior to reconstruct .
LightGBM selection screened 24 predefined hyperparameter settings with seed 17, confirmed the top three settings and the default reference setting across seeds 17, 29, and 43, and selected the configuration with the lowest mean 2020 validation RMSE (Text S2). Predicted residuals were converted to ppm and added to the operational prior, and reconstructed-XCO₂ RMSE, mean bias (prediction minus reference), and were evaluated for each year and seed.
To evaluate the effect of the prediction target, we compared direct-XCO₂ and prior-residual regression using independently selected models for each target (Figure 6). Within each model family and domain, both targets used the same 99 inputs, training and validation cohorts, random seeds, and hyperparameter search spaces described above (up to 500 training epochs with a 40-epoch minimum and 30-epoch early-stopping patience for the MLPs). All 12 selected models were frozen before evaluation on the 2021–2025 test cohorts.

Figure 6. Direct- and prior-residual prediction under independent validation selection. Annual RMSEs are evaluated against ACOS for validation in 2020 and testing in 2021–2025. Dashed lines with squares denote direct prediction, and solid lines with circles denote residual prediction; red indicates MLP and blue indicates LightGBM. Points show means across seeds 17, 29, and 43. Within each model family, both targets share identical cohorts, 99-feature inputs, and hyperparameter search spaces.
East Asian benchmark using SpT screening criteria
To match the test-data screening of Chen et al. [10], we constructed an East Asian ND test cohort from archived ACOS Level-2 records and matched OCO-2 Level-1B spectra following the public SpT preprocessing criteria: successful retrieval (outcome_flag = 1), 100% land fraction, and spatial bounds of 20°N–45°N and 110°E–145°E across footprints 1–8, without the additional ABP and IDP cloud filters used in our global cohorts. After matching by sounding identifier, removing invalid spectra and duplicates, and excluding 5,479 soundings present in the 2020 global ND validation cohort, the regional test set comprised 475,134 soundings across 2020–2025 (Text S4; Table S1).
For the annual comparison, we applied the frozen global ND residual MLP directly to this regional cohort using its 2015–2019 SVD bases, training scalers, and three checkpoints selected on the 2020 global validation set. Over the common 2020–2022 test period, we compare annual RMSEs against ACOS and their three-year mean with the published base SpT results (Figure 8a).

Figure 8. East Asian ND RMSE against ACOS. (a) Frozen global ND MLP evaluated on the SpT-screened annual cohort and published base SpT, 2020–2022. (b) Fixed and rolling global ND MLPs on the SpT-screened quarterly cohort and published SpT series, 2021 Q1–2024 Q1. Red denotes MLP, blue denotes SpT, circles denote fixed weights and squares denote rolling updates. Legends give equal-year or equal-quarter means, respectively. MLP values are three-seed means.
The quarterly comparison starts from the same global ND checkpoints. For each target quarter from 2021 Q1 to 2024 Q1, we compare frozen weights with rolling fine-tuning on earlier observations. Both variants use identical target soundings from the SpT-screened East Asian cohort (Figure 8b; Text S5).
Spectral and prior sensitivity controls and error diagnostics
Spectral component controls used the same 1,000,000 training soundings per domain and 200,000-sounding 2020 validation cohorts as the main experiment. Each configuration combined the nine ancillary and prior predictors with 0, 10, 20, or 30 coefficients per band; the zero-coefficient configuration used only the nine non-spectral inputs. Lower-dimensional configurations took prefixes of the 30-component masked least-squares scores. Domain-specific hidden-layer architectures, batch sizes, and regularization settings were held fixed across component counts. Training allowed up to 500 epochs, with early stopping after at least 40 epochs and a patience of 30 epochs for a minimum validation-RMSE improvement of ppm, evaluating the best validation checkpoint (Text S2). Checkpoints and learning rates were selected on the 2020 validation set and then evaluated across the 2021–2025 test cohorts (Figure 9; Table S10).

Figure 9. Spectral component sensitivity with prior-error baselines: (a) GL ocean, (b) GL land and (c) ND. Each domain uses 1,000,000 training and 200,000 validation soundings. All configurations retain nine ancillary and prior predictors plus 0, 10, 20 or 30 coefficients per band. Points and lines show three-seed mean RMSEs. Horizontal dotted and dashed lines show the uncorrected prior RMSE for 2020 validation and the 2021–2025 mean, respectively, on the same annual cohorts. Test RMSEs are averaged equally across years within each seed before seed aggregation. Model recipes are fixed across component counts, with the common GL-land cosine horizon specified in Text S2.
We tested sensitivity to random XCO₂-prior errors without retraining the models. Each global domain retained its 2020 validation and 2024 test cohorts of 200,000 soundings per year. Zero-mean Gaussian perturbations with standard deviations of 0.5, 1, and 2 ppm were added simultaneously to the prior input feature and the prior used to reconstruct XCO₂ from the predicted residual, with the unperturbed prior serving as the baseline. Both the perturbed prior and the resulting model retrievals were evaluated against the unperturbed ACOS reference, with all other inputs and historical scalers held fixed across both model families and all three seeds (Text S6).
For each main-model sounding, prediction error was defined as reconstructed minus the ACOS reference. Annual mean bias and its 95% confidence interval were estimated from the 2.5th and 97.5th percentiles of 1,000 within-year date-block bootstrap draws, resampling observation dates while retaining all soundings and three seeds on each drawn date (Figure 11). Latitudinal bias profiles were computed across the complete years 2021–2024 by averaging errors over the three seeds and all soundings within 10° latitude bands containing at least 30 observations (Figure 12).

Figure 11. Annual mean prediction-minus-ACOS bias of the main MLPs in 2021–2025: (a) GL ocean, (b) GL land, and (c) ND. Points show the three-seed mean bias; vertical error bars indicate 95% intervals from 1,000 within-year date-block bootstrap draws with same-day soundings and seeds paired. The dashed line marks zero bias.

Figure 12. Latitude-dependent mean bias of the main MLPs relative to ACOS during the complete years 2021–2024: (a) GL ocean, (b) GL land, and (c) ND. Each point averages prediction-minus-reference errors across three seeds and all soundings within a 10° latitude band containing at least 30 observations. The horizontal dashed line marks zero bias.
Results
Compact spectral representation
Across the nine domain–band combinations, the first 30 SVD components captured 98.65%–99.85% of centered training spectral variance, leaving 0.15%–1.35% in higher-order components (Figure 4). Residual variance declined steeply within the first 10 components for the O-A and weak-CO bands (Figure 4a, Figure 4b) and within the first 20 components for the strong-CO band (Figure 4c), flattening thereafter.
Residual variability and prediction-target comparison
Across GL ocean, GL land, and ND, annual mean ACOS increased by 24.8–25.7 ppm between 2015 and 2025, accompanied by similar growth in the operational prior (Figure 5a–Figure 5c). The annual mean reference-minus-prior residuals varied over much smaller ranges: −1.22 to −0.16 ppm for GL ocean, −1.56 to −0.43 ppm for GL land, and −2.04 to −0.62 ppm for ND (right axes in Figure 5a–Figure 5c). Mean residual offsets differed among domains. Subtracting the prior therefore removes most of the long-term background growth while retaining domain-dependent offsets and interannual variability for the regressors to capture.

Figure 5. Annual mean ACOS , operational prior and their difference in 2015–2025 for (a) GL ocean, (b) GL land and (c) ND. Reference and the prior use the left axes; purple dashed curves show reference minus prior on the right axes. Each point is an annual sample mean. Left-axis limits are shared across panels, as are right-axis limits. Pale bands mark 2020 validation, and dashed vertical boundaries separate the later test years. Sample counts are given in Table S1.
After independent validation selection, direct and residual MLPs performed similarly on the 2020 validation set, where residual prediction reduced RMSE by 0.02–0.06 ppm across the three domains (Figure 6). The difference grew across the 2021–2025 test period: mean annual MLP test RMSE decreased from 1.60 to 0.89 ppm over GL ocean, from 1.68 to 1.33 ppm over GL land, and from 1.51 to 1.43 ppm for ND. The reduction was larger for LightGBM, whose mean annual test RMSEs dropped from 6.49, 6.46, and 6.79 ppm under direct prediction to 1.42, 1.72, and 1.81 ppm under residual prediction. The RMSE advantage of residual prediction widened each year for the MLP over GL ocean and GL land and for LightGBM across all three domains, consistent with the narrower interannual range of the prior-residual target (Figure 5).
Retrieval performance across observation domains
Across all three observation domains and every test year, the residual MLP improved upon the operational prior relative to ACOS and achieved lower RMSE than the validation-selected LightGBM models (Figure 7; Table II). Mean annual 2021–2025 MLP RMSEs were 0.89, 1.35, and 1.43 ppm for GL ocean, GL land, and ND, respectively, corresponding to 48.7%, 49.4%, and 57.9% reductions relative to the uncorrected prior. Over GL ocean in 2025, LightGBM RMSE reached 1.95 ppm, exceeding the prior RMSE of 1.64 ppm, whereas the MLP maintained an RMSE of 1.07 ppm (Figure 7a). We therefore focus subsequent analyses on the residual MLP. Even with residual learning, frozen-model test errors increased gradually over 2021–2025, most notably in ND (Figure 7c).
| Metric | GL ocean | GL land | ND |
| Prior RMSE | 1.743 | 2.665 | 3.401 |
| Prior R² | 0.345 | 0.179 | 0.179 |
| MLP RMSE | 0.894 | 1.348 | 1.433 |
| MLP R² | 0.829 | 0.789 | 0.855 |
| LightGBM RMSE | 1.424 | 1.724 | 1.810 |
| LightGBM R² | 0.558 | 0.652 | 0.769 |
| RMSE reduction | 48.7% | 49.4% | 57.9% |
| 2025 MLP bias | +0.452 | +0.632 | +0.651 |
Table II. Main-cohort performance across domains.

Figure 7. Annual validation (2020) and test (2021–2025) RMSE against ACOS for (a) GL ocean, (b) GL land, and (c) global ND. Points and lines show the mean RMSE across seeds 17, 29, and 43. The gray dashed line shows the uncorrected operational prior. The pale band marks the validation year.
Over GL land, annual MLP RMSE tracked the temporal shape of the prior RMSE, rising through 2023 and declining in 2024–2025 while remaining 1.14–1.47 ppm below the prior in every year (Figure 7b). In ND, prior RMSE rose from 3.01 ppm in 2021 to 3.61 ppm in 2022 (Figure 7c), driven by increases in both the prior mean bias (1.56 to 2.04 ppm) and error standard deviation (2.57 to 2.98 ppm). Over the same interval, MLP RMSE increased by 0.28 ppm (from 1.07 to 1.35 ppm), attenuating more than half of the 2022 jump in prior error.
Note: RMSE and R² are unweighted means of annual metrics for 2021–2025; model metrics are first averaged across three random seeds. Bias refers only to the MLP in 2025. RMSE and bias are in ppm; reduction is the percentage decrease in mean annual MLP RMSE relative to the operational prior.
East Asian nadir results under SpT screening criteria
Without regional retraining, the global ND MLP achieved annual RMSEs against ACOS of 1.287, 1.351, and 1.488 ppm in 2020–2022, respectively. These were lower in all three years than the corresponding published base SpT values (1.316, 1.526, and 1.666 ppm). The mean annual RMSE was 1.376 ppm, 8.4% lower than the published SpT mean of 1.503 ppm (Figure 8a).
When the same frozen model was evaluated through 2025, the annual East Asian RMSEs for 2023–2025 were 1.537, 1.476 and 1.546 ppm (Table S6).
Across 13 quarters, rolling updates reduced the global ND MLP's mean quarterly RMSE against ACOS from 1.54 to 1.06 ppm. The published base and rolling SpT means were approximately 1.89 and 1.20 ppm, respectively (Figure 8b). The rolling MLP also showed less quarter-to-quarter variation than the published rolling SpT series (standard deviations of 0.10 and approximately 0.14 ppm). The rolling models had relatively similar mean quarterly RMSEs in 2021 (1.13 ppm for the MLP versus approximately 1.25 ppm for SpT). The gap was larger in 2023 (1.00 versus approximately 1.19 ppm) and in the first quarter of 2024 (0.96 versus approximately 1.20 ppm).
Measured retrieval computation time
We benchmarked two stages separately on an Intel Xeon Platinum 8581C CPU (four threads) and an NVIDIA RTX 5880 Ada Generation GPU. Starting from 1,024 native Level-1B spectra and associated inputs already in memory, the residual MLP pipeline took 75.6–102.9 s per sounding on the CPU across observation domains (Table S9). For GL ocean, the batch took 0.105 s (102.9 µs per sounding). The median latency for a single GL-ocean sounding was 1.33 ms. This benchmark includes wavelength registration, continuum normalization, masked least-squares projection, feature assembly and scaling, and prediction with the prior added back. It excludes disk I/O, cloud screening, prior generation, and model loading.
Starting instead from 200,000 precomputed 99-feature inputs, the validation-selected LightGBM models took 26.683, 29.364, and 28.084 s on the four-thread CPU for GL ocean, GL land, and ND, respectively. The corresponding MLP times were 0.193, 0.202, and 0.409 s (Table S8). For GL ocean, the MLP took 0.96 µs per sounding in the CPU batch and 0.137 s for the full GPU batch. Starting from native spectra in memory, LightGBM took 227.8–234.7 µs per sounding in 1,024-sounding CPU batches (Table S9). For context, Chen et al. [10] reported more than 2–3 min per sounding for full-physics retrieval and 6.2 ms per sounding for SpT on an Intel Xeon Gold 5218R CPU. Our global ND MLP took 1.20 ms for a single sounding starting from in-memory native spectra on the Intel Xeon Platinum 8581C. These published times provide context; differences in hardware and processing scope preclude a direct speedup estimate.
Spectral and prior contributions and temporal error diagnostics
Including spectral scores lowered validation and multi-year test RMSE relative to the nine ancillary and prior predictors alone across all three domains (Figure 9). For GL ocean and GL land, RMSE decreased steeply up to 20 scores per band, with smaller reductions from 20 to 30 scores. In ND, the largest reduction occurred within the first 10 scores per band, followed by a nearly flat response above 10 scores.
In global ND, increasing from 20 to 30 scores per band lowered 2020 validation RMSE while increasing mean test RMSE across 2021–2025, indicating that fitting fine-scale variations in the validation year can trade off against multi-year transfer stability. The zero-PC models also improved upon the uncorrected prior in post-training years, showing that ancillary and prior predictors contribute baseline state adjustments while the spectral scores drive the primary accuracy gains.
Adding zero-mean Gaussian noise with a standard deviation of 1 ppm to the 2024 priors increased prior RMSE from 1.580, 2.734 and 3.438 ppm to 1.871, 2.913 and 3.579 ppm for GL ocean, GL land and ND, respectively. The corresponding MLP RMSEs increased from 0.912, 1.472 and 1.593 ppm to 1.075, 1.566 and 1.670 ppm. The smaller increase in MLP RMSE than in prior RMSE across all three domains indicates moderate MLP sensitivity to the prior under the tested random perturbations (Figure 10).

Figure 10. Sensitivity to random -prior errors in the complete 2024 test cohorts: (a) GL ocean, (b) GL land and (c) ND. The horizontal axis gives the standard deviation of the added zero-mean Gaussian error. The prior input and the prior added to the predicted residual are perturbed together. Model curves show three-seed mean RMSE, and the gray dashed line shows perturbed-prior RMSE; both are evaluated against the same ACOS reference. Models share the same perturbation vector and 200,000 soundings per domain. Zero noise denotes the original baseline.
Main-model annual mean biases were positive relative to ACOS across 2021–2025 and varied by observation domain (Figure 11). Bias over GL ocean remained comparatively stable and decreased in 2024, whereas GL land and ND exhibited larger positive shifts after 2022. In the 2025 cohort, mean biases (with 95% date-block bootstrap intervals) were +0.452 ppm (+0.415 to +0.489) for GL ocean, +0.632 ppm (+0.585 to +0.682) for GL land, and +0.651 ppm (+0.564 to +0.727) for ND. Latitude profiles for 2021–2024 show comparatively consistent patterns over GL ocean (Figure 12). GL land exhibits southern high-latitude underestimation in several years and stronger northern positive biases from 2022 onward. ND profiles change more substantially with latitude and year, without a similarly persistent pattern.
Discussion
Spectral information and prior correction
The spectral component controls separate baseline prior adjustments from spectral retrieval skill (Figure 9). Without spectral inputs, the zero-score model learns a conditional correction for the prior, surface-pressure priors, and viewing geometry, reflecting systematic relationships between the ACOS retrieval and its prior state. Including the leading SVD components produces the primary accuracy gains, which persist across the 2021–2025 test period. Because SVD orders basis vectors by spectral variance rather than predictive relevance to [21], [15], higher-order components mix weak absorption signals with measurement noise and fine-scale surface or instrument variability. In global ND, increasing from 20 to 30 components per band lowered 2020 validation RMSE but slightly increased 2021–2025 test RMSE, indicating that higher-order nadir components capture validation-year variations that transfer less reliably to subsequent years.
In the prior-perturbation experiments, adding random noise increased MLP RMSE by less than the added prior error in all three domains (Figure 10). Because the operational prior enters both the MLP input vector and the final reconstruction (Equation (4)), the network uses spectral and geometric inputs to adjust the predicted residual when the prior departs from the state implied by the radiance spectrum, partially compensating for random prior errors.
The remaining interannual error drift reflects shifts in both the prior residuals and the observing system. ACOS v11 priors represent large-scale background growth and climatological profile structure derived from NOAA marine boundary-layer measurements at Mauna Loa and American Samoa [22], [23]. Subtracting the prior removes the global secular trend (Figure 5), whereas regional fluxes, seasonal cycles, atmospheric transport, instrument calibration drift, and updates to the ACOS retrieval build continue to evolve across years. Consequently, the residual distribution shifts over time, producing the year-to-year residual variations in Figure 5 and the domain-dependent bias trends in Figure 11.
Spatial and temporal bias diagnostics identify the regimes that drive these post-2022 shifts (Figure 11 and Figure 12). Over GL ocean, latitudinal bias profiles remained within ±0.7 ppm across 2021–2024 (Figure 12a). Over GL land, negative biases of −1.39 and −1.54 ppm occurred at 80°–90°S in 2021 and 2022, where the test cohorts contained only 57 and 69 soundings over high-albedo snow and ice surfaces, while positive biases strengthened across northern midlatitudes after 2022 (Figure 12b). In ND, latitudinal biases varied more strongly from year to year across the tropics and northern midlatitudes (Figure 12c).
Interpreting the ocean-glint result
Among the three global domains, GL ocean achieved the lowest absolute test error, reaching 0.89 ppm mean annual RMSE over 2021–2025 with a single-hidden-layer MLP of 220 units (22,661 parameters). Whereas previous neural OCO-2 retrieval studies have focused primarily on land nadir or mixed overland soundings [7], [8], [9], [10], our ocean-glint evaluation establishes a compact baseline for fast multi-year estimation over the global oceans.
The lower retrieval RMSE over GL ocean has a clear physical and radiometric basis. Near-glint viewing over water redirects substantially more reflected sunlight toward the sensor than diffuse nadir viewing over dark ocean surfaces [11], [3], [5]. Under unsaturated observing conditions, higher photon counts increase the signal-to-noise ratio and reduce relative measurement noise in absorption-line depths, enhancing sensitivity to small -induced spectral variations [3]. Because continuum normalization removes broadband radiance scaling without altering relative channel-to-channel noise, this signal-to-noise advantage carries over to the normalized spectra. In addition, the spatially uniform ocean surface reduces lower-boundary variability relative to land, where surface albedo, topography, and continental aerosols vary strongly [12], [13].
GL ocean also had a smaller prior-error baseline: the 2021–2025 mean annual prior RMSE was 1.74 ppm, compared with 2.67 ppm for GL land and 3.40 ppm for ND (Table II). MLP reductions of 49%, 49% and 58%, respectively, show comparable proportional gains despite different absolute errors. ACOS represents ocean-glint reflection with a wind-speed-dependent Cox–Munk wave-slope model, whereas land reflection has stronger spatial heterogeneity [11], [24], [5]. These radiative and statistical differences provide complementary explanations for the observed domain contrast.
Comparison with previous nadir and regional machine-learning retrievals
The global ND experiment spans diverse observing conditions worldwide and achieved a 2021–2025 mean annual test RMSE of 1.43 ppm. Evaluating the 99-input framework within the East Asian domain (20°N–45°N, 110°E–145°E) places our results in the geographical region examined by Xie et al. [9] and Chen et al. [10].
The global ND MLP has 164,793 trainable parameters, compared with approximately 6.5 million estimated from the published SpT architecture (Text S5). After aligning the test screening criteria, the frozen MLP produced lower annual RMSEs against ACOS than the published base SpT over the common East Asian test period (Figure 8a). The result shows that a compact model trained globally can reproduce regional ACOS retrievals without local retraining.
Starting from the same global model, rolling updates improved agreement with ACOS on identical target-quarter soundings. Even without updates, the mean quarterly RMSE gap between the fixed MLP and the published base SpT widened from approximately 0.17 ppm in 2021 to 0.62 ppm in 2023 (Figure 8b). This pattern points to stronger temporal transfer for the MLP in this East Asian comparison, with rolling updates providing a further reduction in error.
Limitations and future directions
Because our models are trained and evaluated against ACOS Level-2 Standard , the reported metrics quantify agreement with the operational retrieval rather than accuracy against independent atmospheric observations. The model estimates may inherit systematic biases in the reference product. Evaluating atmospheric accuracy for carbon-flux applications requires independent validation against ground-based Total Carbon Column Observing Network (TCCON) measurements. In addition, the 2022 increase in ND prior error coincided with a transition from build B11006r to mixed B11006r and B11008r records, and post-2022 GL-land and ND biases spanned multiple processing builds [25] (Table S11). Reprocessing matched soundings under a single ACOS build would separate algorithm-version shifts from atmospheric variability. Future work can also extend this SVD-residual framework to OCO-3 and GOSAT by adapting the wavelength grid and continuum windows to each spectrometer, and use radiative-transfer simulations to isolate how signal-to-noise ratio and higher-order spectral components affect retrieval error.
Conclusions
We evaluated multi-year OCO-2 retrieval across ocean glint, land glint, and nadir observation domains by combining bandwise SVD spectral compression with prior-residual MLP regression. Trained on 2015–2019 observations, compact MLPs achieved mean annual 2021–2025 test RMSEs against ACOS of 0.89, 1.35, and 1.43 ppm for GL ocean, GL land, and ND, respectively, while buffering perturbations to the operational prior. Under the published East Asian ND screening criteria, the frozen global MLP achieved lower annual RMSEs against ACOS than the base SpT in 2020–2022 without regional retraining, and quarterly rolling updates further reduced the 2021–2024 mean quarterly RMSE from 1.54 to 1.06 ppm. Together, spectral compression and compact residual networks reproduced ACOS retrievals across the three observation domains. With native spectra and ancillary inputs already in memory, CPU batches required 75.6–102.9 μs per sounding for spectral preprocessing and prediction.
Data and code availability
OCO-2 Level-1B radiances and ACOS Level-2 Standard products are publicly available from NASA GES DISC (https://disc.gsfc.nasa.gov/). Analysis code, configurations and compact result tables for the 99-input residual framework will be made publicly available at https://github.com/Sanshi2024/oco2-svd-residual-retrieval.
Acknowledgements
We thank the NASA OCO-2 mission and the ACOS science team for providing the calibrated Level-1B spectra and Level-2 retrieval products and acknowledge financial support from the Hong Kong Jockey Club Charities Trust (Grant No. FA123), the Innovation and Technology Fund (Grant No. P0491), the General Research Fund (Grant No. 16303725), and the State Key Laboratory of Climate Resilience for Coastal Cities (Grant No. ITC-SKLCRCC26EG01).
Credit authorship contribution statement
Xianlei San conducted the experiments and wrote the original draft. Chengxing Zhai supervised the research. Hui Su, Kuai Dai, and Shuangchen Du provided substantive feedback on the manuscript and contributed to its revision.
References
- [1]P. Friedlingstein, M. O'Sullivan, M.W. Jones, R.M. Andrew, J. Hauck, P. Landschützer, et al., “Global Carbon Budget 2024,” Earth System Science Data, vol. 17, pp. 965–1039, 2025, doi: 10.5194/essd-17-965-2025.DOI
- [2]A. Eldering, C.W. O'Dell, P.O. Wennberg, D. Crisp, M.R. Gunson, C. Viate, et al., “The Orbiting Carbon Observatory-2: first 18 months of science data products,” Atmospheric Measurement Techniques, vol. 10, pp. 549–563, 2017, doi: 10.5194/amt-10-549-2017.DOI
- [3]D. Crisp, H.R. Pollock, R. Rosenberg, L. Chapsky, R.A.M. Lee, F.A. Oyafuso, et al., “The on-orbit performance of the Orbiting Carbon Observatory-2 (OCO-2) instrument and its radiometrically calibrated products,” Atmospheric Measurement Techniques, vol. 10, pp. 59–81, 2017, doi: 10.5194/amt-10-59-2017.DOI
- [4]C.W. O'Dell, B. Connor, H. Bösch, D. O'Brien, C. Frankenberg, R. Castano, et al., “The ACOS CO₂ retrieval algorithm – Part 1: Description and validation against synthetic observations,” Atmospheric Measurement Techniques, vol. 5, pp. 99–121, 2012, doi: 10.5194/amt-5-99-2012.DOI
- [5]C.W. O'Dell, A. Eldering, P.O. Wennberg, D. Crisp, M.R. Gunson, B. Fisher, et al., “Improved retrievals of carbon dioxide from Orbiting Carbon Observatory-2 with the version 8 ACOS algorithm,” Atmospheric Measurement Techniques, vol. 11, pp. 6539–6576, 2018, doi: 10.5194/amt-11-6539-2018.DOI
- [6]A. Chédin, S. Serrat, N.A. Scott, C. Crevoisier, and R. Armante, “First global measurement of midtropospheric CO₂ from NOAA polar satellites: Tropical zone,” Journal of Geophysical Research: Atmospheres, vol. 108, no. D18, Art. no. 4581, 2003, doi: 10.1029/2003JD003439.DOI
- [7]L. David, F.-M. Bréon, and F. Chevallier, “XCO₂ estimates from the OCO-2 measurements using a neural network approach,” Atmospheric Measurement Techniques, vol. 14, pp. 117–132, 2021, doi: 10.5194/amt-14-117-2021.DOI
- [8]F.-M. Bréon, L. David, P. Chatelainaz, and F. Chevallier, “On the potential of a neural-network-based approach for estimating XCO₂ from OCO-2 measurements,” Atmospheric Measurement Techniques, vol. 15, pp. 5219–5234, 2022, doi: 10.5194/amt-15-5219-2022.DOI
- [9]F. Xie, T. Ren, C. Zhao, Y. W. Gu, M. Zhou, P. Wang, K. Shiomi, and I. Morino, “Fast retrieval of XCO₂ over east Asia based on Orbiting Carbon Observatory-2 (OCO-2) spectral measurements,” Atmospheric Measurement Techniques, vol. 17, pp. 3949–3967, 2024, doi: 10.5194/amt-17-3949-2024.DOI
- [10]W. Chen, T. Ren, C. Zhao, Y. Wen, Y. Gu, M. Zhou, and P. Wang, “Transformer-Based Fast Mole Fraction of CO₂ Retrievals from Satellite-Measured Spectra,” Journal of Remote Sensing, vol. 5, Art. no. 0470, 2025, doi: 10.34133/remotesensing.0470.DOI
- [11]C. Cox and W. Munk, “Measurement of the roughness of the sea surface from photographs of the sun's glitter,” Journal of the Optical Society of America, vol. 44, pp. 838–850, 1954, doi: 10.1364/JOSA.44.000838.
- [12]H. Boesch, D. Baker, B. Connor, D. Crisp, and C. Miller, “Global characterization of CO₂ column retrievals from shortwave-infrared satellite observations of the Orbiting Carbon Observatory-2 mission,” Remote Sensing, vol. 3, pp. 270–304, 2011, doi: 10.3390/rs3020270.DOI
- [13]S.T. Massie, H. Cronk, A. Merrelli, C. O'Dell, K.S. Schmidt, H. Chen, and D. Baker, “Analysis of 3D cloud effects in OCO-2 XCO₂ retrievals,” Atmospheric Measurement Techniques, vol. 14, pp. 1475–1499, 2021, doi: 10.5194/amt-14-1475-2021.DOI
- [14]T.E. Taylor, C.W. O'Dell, C. Frankenberg, P.T. Partain, H.Q. Cronk, A. Savtchenko, et al., “Orbiting Carbon Observatory-2 (OCO-2) cloud screening algorithms: validation against collocated MODIS and CALIOP data,” Atmospheric Measurement Techniques, vol. 9, pp. 973–989, 2016, doi: 10.5194/amt-9-973-2016.DOI
- [15]N. Halko, P. G. Martinsson, and J. A. Tropp, “Finding Structure with Randomness: Probabilistic Algorithms for Constructing Approximate Matrix Decompositions,” SIAM Review, vol. 53, pp. 217–288, 2011, doi: 10.1137/090771806.DOI
- [16]R. Penrose, “A generalized inverse for matrices,” Mathematical Proceedings of the Cambridge Philosophical Society, vol. 51, no. 3, pp. 406–413, 1955, doi: 10.1017/S0305004100030401.DOI
- [17]R. Penrose, “On the best approximate solutions of linear matrix equations,” Mathematical Proceedings of the Cambridge Philosophical Society, vol. 52, no. 1, pp. 17–19, 1956, doi: 10.1017/S0305004100030929.DOI
- [18]C. Frankenberg, R. Pollock, R.A.M. Lee, R. Rosenberg, J.-F. Blavier, D. Crisp, et al., “The Orbiting Carbon Observatory (OCO-2): spectrometer performance evaluation using one-launch direct sun measurements,” Atmospheric Measurement Techniques, vol. 8, pp. 301–313, 2015, doi: 10.5194/amt-8-301-2015.DOI
- [19]G. Ke, Q. Meng, T. Finley, T. Wang, W. Chen, W. Ma, Q. Ye, and T.-Y. Liu, “LightGBM: A Highly Efficient Gradient Boosting Decision Tree,” Advances in Neural Information Processing Systems, vol. 30, pp. 3146–3154, 2017.
- [20]I. Loshchilov and F. Hutter, “Decoupled Weight Decay Regularization,” International Conference on Learning Representations (ICLR), 2019.
- [21]I.T. Jolliffe, “A Note on the Use of Principal Components in Regression,” Journal of the Royal Statistical Society: Series C (Applied Statistics), vol. 31, no. 3, pp. 300–303, 1982, doi: 10.2307/2348005.DOI
- [22]J.L. Laughner, S. Roche, M. Kiel, G.C. Toon, D. Wunch, B.C. Baier, et al., “A new algorithm to generate a priori trace gas profiles for the GGG2020 retrieval algorithm,” Atmospheric Measurement Techniques, vol. 16, pp. 1121–1146, 2023, doi: 10.5194/amt-16-1121-2023.DOI
- [23]N. Jacobs, C.W. O'Dell, T.E. Taylor, T.L. Logan, B. Byrne, M. Kiel, et al., “The importance of digital elevation model accuracy in XCO₂ retrievals: improving the Orbiting Carbon Observatory 2 Atmospheric Carbon Observations from Space version 11 retrieval product,” Atmospheric Measurement Techniques, vol. 17, pp. 1375–1401, 2024, doi: 10.5194/amt-17-1375-2024.DOI
- [24]R.R. Nelson, A. Eldering, D. Crisp, A.J. Merrelli, and C.W. O'Dell, “Retrieved wind speed from the Orbiting Carbon Observatory-2,” Atmospheric Measurement Techniques, vol. 13, pp. 6889–6899, 2020, doi: 10.5194/amt-13-6889-2020.DOI
- [25]G. Osterman and the OCO-2/OCO-3 Science Team, “OCO-2/OCO-3 Data Product User’s Guide,” Version 3, Revision B, 2 April 2024. [Online]. Available: https://docserver.gesdisc.eosdis.nasa.gov/public/project/OCO/OCO2_v11.2_OCO3_v10_L2_Data_Users_Guide_V3_RevB.pdf