Bayesian correction of satellite-derived pCO₂ reveals the biogeochemical impact of hurricanes on ocean–atmosphere CO₂ fluxes

GabrielaY.Cervantes-Díaz1✉Emailcervantes.gabriela@uabc.edu.mx

Coronado-ÁlvarezLuz

Lourdes Aurora1

T.Espinosa-Carreón1

Leticia2

Instituto de Investigaciones OceanológicasUniversidad Autónoma de Baja CaliforniaCarretera Transpeninsular Tijuana-Ensenada 3917, Fraccionamiento Playitas, CP22860Ensenada, Baja CaliforniaMéxico

2Instituto Politécnico Nacional-Centro Interdisciplinario de Investigación para el Desarrollo Integral RegionalUnidad Sinaloa81101Guasave, SinaloaMéxico

Cervantes-Díaz Gabriela Y.¹, Coronado-Álvarez Luz de Lourdes Aurora¹, Espinosa-Carreón T. Leticia²

¹ Instituto de Investigaciones Oceanológicas, Universidad Autónoma de Baja California, Carretera Transpeninsular Tijuana-Ensenada 3917, Fraccionamiento Playitas, CP, 22860, Ensenada, Baja California, México

² Instituto Politécnico Nacional-Centro Interdisciplinario de Investigación para el Desarrollo Integral Regional, Unidad Sinaloa, 81101 Guasave, Sinaloa, México

Correspondence to: cervantes.gabriela@uabc.edu.mx

Abstract

The exchange of CO₂ between the ocean and atmosphere is a key process in the global carbon cycle, highly sensitive to disturbances causes by hurricanes and tropical storms. In this study, we applied a Bayesian Recurrent Neural Network (BNNR) to correct satellite-derived surface pCO₂ estimates from the PISCES-NEMO model in the eastern tropical Pacific, off the coast of Guerrero, Mexico. The model was calibrated using in situ data collected during Tropical Storm Lidia (2017) and supplemented with synthetic cases representing extreme wind conditions. Under fair-weather conditions, the BNNR achieved low errors (MAE = 0.76 µatm; RMSE = 0.99 µatm). Although errors increased under simulated extreme winds, the model successfully captured abrupt increases in pCO₂. During Hurricane John (2024), corrected values exceeded 850 µatm, revealing strong CO₂ degassing pulses that contributed up to one-third of the seasonal air-sea CO₂ flux, a magnitude largely underestimated by deterministic products. By explicitly incorporating uncertainty through credibility intervals, the BNNR provided robust predictions in highly variable conditions. This methodological framework offers a novel tool for improving the monitoring of marine carbon cycle dynamics and for quantifying the biogeochemical impact of extreme events in coastal regions.

Keywords:

probabilistic deep learning

Bayesian neural networks

cyclonic events

satellite oceanography

carbon dioxide flux

1 Introduction

The ocean is an essential regulator of the global climate system, and at the same time, a dynamic component of the carbon cycle. Among the most relevant processes is the exchange of carbon dioxide (CO₂) between the atmosphere and the sea surface, a phenomenon strongly conditioned by physical and biogeochemical factors that exhibit marked spatial and temporal variability. Despite advances in coupled ocean models and satellite observations, estimating the partial pressure of CO₂ (pCO₂) at the sea surface remains a challenge, particularly in tropical regions affected by hurricanes and tropical storms. These areas, characterized by heterogeneity in temperature, salinity, surface winds, and primary productivity, require innovative methodologies that integrate multiple sources of information to more realistically represent air-sea carbon fluxes (FCO₂).

Satellite monitoring and coupled biogeochemical models, such as PICES-NEMO, have expanded our capacity to study ocean-atmosphere interactions at high spatial and temporal resolution. These systems incorporate key variables — including sea surface temperature (SST), surface salinity (SSS), chlorophyll concentration (Chl-a), absolute dynamic topography (ADT), and wind components (u, v) — which serve as essential inputs for characterizing ocean conditions and driving predictive models. However, these products often underestimate transient phenomena driven by intense physical forcing, with reported errors exceeding 90 µatm in dynamic coastal environments. Improving their accuracy requires methods that integrate multiple data sources and account for uncertainty.

Neural networks have established themselves as key tools for modeling nonlinear relationships in complex systems with multiple variables. In particular, Bayesian recurrent neural networks (BNNR) combine the capacity to represent temporal dynamics with explicit uncertainty quantification, an advantage over deterministic approaches. Their application is especially valuable in oceanography, where the availability of in situ data is limited and natural variability is high, as they enable the dynamic representation of physical and biogeochemical processes while generating credible intervals. Recent studies demonstrate that BNNRs can correct systematic biases in satellite-derived products and more effectively capture anomalies associated with cyclonic forcing, such as abrupt increases in pCO₂ induced by tropical cyclones. Overall, neural models—whether feed-forward, recurrent, or Bayesian—have consistently recovered pCO₂ in oceanic and coastal environments, with root mean square errors (RMSE) below 25 µatm. Previous neural network applications, including feed-forward and ensemble models, have reconstructed surface pCO₂ patterns at global and regional scales with reasonable skill (e.g. ANN and XGBoost in the Bay Bengal archived r ~ 0.75 and ± 12 µatm error; Joshi et al., 2022). However, strictly Bayesian implementations remain rare, and their capacity to capture abrupt anomalies during extreme events such as hurricanes has yet to be systematically evaluated.

In this regard, Bayesian neural networks provide crucial added value by allowing explicit estimation of predictive uncertainty and rigorous incorporation of prior knowledge. These characteristics are essential in applications that require high levels of reliability, such as in areas with low sampling or in contexts dominated by high-energy meteorological events where data extrapolation involves significant risks (Clare et al., 2022).

This study applies a BNNR to correct surface pCO₂ estimates from the PISCES-NEMO model in the eastern tropical Pacific, near Guerrero, Mexico (10°–18° N, 90°–110° W). The model was calibrated using high-frequency buoy data from Tropical Storm Lidia (2017) and tested for three contrasting meteorological systems events that occurred in 2024: Hurricane Ileana (category 2), Hurricane John (category 5), and Tropical Depression Eleven-E. Our objectives were to (1) reduce biases in satellite-derived pCO₂, (2) incorporate predictive uncertainty, and (3) evaluate the contribution of tropical cyclones to regional air–sea CO₂ fluxes. This approach enabled the bias correction of systematic biases in modeled pCO₂, the explicit estimation of uncertainty, and the calculation of corrected air–sea CO₂ fluxes. Collectively, these advances contribute to understanding marine carbon dynamics during high-energy meteorological events and have implications for climate oceanography, fisheries management, and coastal resilience (IPCC, 2021).

2 Methodology

2.1 Study area and period

The analysis focused on the eastern tropical Pacific, off the coast of Guerrero, Mexico (10° – 18° N, 90° – 110° W), a region characterized by high physical-biogeochemical variability and frequent tropical cyclone activity. The study period extended from May to November 2024 and included three contrasting events: Hurricane Ileana (September 11–17, category 2; Berg, 2024), Hurricane John (September 22–25, category 5; Pasch, 2025), and Tropical Storm Eleven-E (October 1–3; Blake, 2025) (Fig. 1). The classification was made according to the Saffir–Simpson scale (CNH, 2017). These atmospheric systems generated winds exceeding 70 m s⁻¹, abrupt decreases in sea surface temperature (SST), and notable increases in surface pCO₂. Their inclusion allowed for a comprehensive assessment of the impact of extreme forcings on air–sea CO₂ dynamics at subseasonal scales.

Figure 1. Study area showing the trajectories of tropical cyclones that impacted the eastern tropical Pacific during 2024: Hurricane Ileana (September 11–17; green line), Hurricane John (September 22–27; blue line), and Tropical Storm Eleven-E (October 1–4; pink line). For reference, the trajectory of Hurricane Lidia (August 30–September 3, 2017; yellow line) is also included. The red line within the Gulf of California denotes the path of a drifting buoy deployed during Tropical Storm Lidia, which provided high-frequency in situ measurements of pCO₂, SST, and SSS for model calibration.

2.2 In situ data for model calibration

High-frequency in situ data on pCO₂ (partial pressure of carbon dioxide), sea surface temperature (SST), and sea surface salinity (SSS) were collected during the passage of Tropical Storm Lidia (August 30–September 3, 2017) using an oceanographic buoy located in the Gulf of California (Fig. 1). These data were used to calibrate and validate the bias-correction model. Because satellite pCO₂ products were not available for 2017, predictor variables derived from the Copernicus Marine Environment Monitoring Service (CMEMS) and the Global Carbon Project (10.48670/moi-00015) were used to achieve temporal alignment, thus allowing the construction of a proxy training set under similar physical-biogeochemical conditions. This step was essential for calibrating the model under real forcing scenarios.

2.3 Satellite and modeled datasets

Daily and hourly variables from satellite products and CMEMS L4 reanalyses, together with data from the Global Carbon Project, were integrated. Key variables included: SST (°C), SSS (CMEMS, 10.48670/moi-00016), zonal and meridional wind components (u, v; Hersbach et al., 2023), chlorophyll a concentration (Chl-a; mg m⁻³; 10.48670/moi-00015), and absolute dynamic topography (ADT, cm; https://doi.org/10.48670/moi-00149). Simulated surface pCO₂ fields were obtained from the PISCES-NEMO biogeochemical model of the CMEMS L4 satellite product (10.48670/moi-00015). All datasets were interpolated to a standard regular grid with a spatial resolution of 0.083° and coupled at a daily frequency.

2.4 Data fusion and preprocessing

All variables were processed in Python (Python Software Foundation, 2023), utilizing the NumPy, Xarray, and Pandas libraries. Wind speed and daily maximum wind (m s⁻¹) were calculated from the u and v components. The final set included seven predictors: SST, SSS, Chl-a, ADT, and wind speed. Missing values were imputed conservatively. All data were normalized using MinMax scaling and interpolated spatially and temporally to fit a regular grid. The wind components were converted to daily magnitude (speed), and the maximum speed per day was also calculated.

2.5 Generation of synthetic data

To increase the robustness of the model under extreme conditions, synthetic data were generated to represent hurricane-like forcing. Multivariate Gaussian distributions with controlled noise (“jitter”) were derived from observed conditions during Lidia to simulate cases characterized by wind speeds up to 70 m·s⁻¹, low SST and SSS, and high pCO₂ values (800–850 µatm). These synthetic data were added exclusively for September 23, 2024 (the day Hurricane John struck), to prevent overfitting and maintain the model’s generalization capacity.

2.6 Bayesian neural network model

A Bayesian recurrent neural network (BNNR) was implemented in TensorFlow Probability (Abadi et al., 2016; Dillon et al., 2017), and LSTMFlipout and DenseFlipout layers (Wen et al., 2018) were used to estimate posterior distributions of weights and quantify predictive uncertainty. The model was trained exclusively on the Lidia event dataset, which includes both real and synthetic data. Training was performed using variational inference, employing the negative log-likelihood loss function and 1000 Monte Carlo samples. The architecture included an LSTM layer with 32 neurons, followed by a Bayesian dense layer, a ReLU activation (Rectified Linear Unit), and a linear output. The output consisted of corrected pCO₂ values referenced to the oceanographic buoy.

2.7 Application of the model to the 2024 dataset

The trained BNNR model was applied to the pCO₂ dataset from the PISCES-NEMO output (Aumont et al., 2015) off the coast of Guerrero for the period from May to November 2024. Daily maps of corrected pCO₂ were generated along with their predictive uncertainty. On September 23, a hybrid bias-correction was implemented that combined model predictions with physical spatial anomalies (based on wind and chlorophyll a), resulting in increased pCO₂ of up to + 300 µatm in areas with strong forcing. This strategy aimed to replicate patterns similar to those observed during Lidia, thereby ensuring spatial consistency.

2.8 Estimation of sea–atmosphere CO₂ flux

The CO₂ flux between the ocean and the atmosphere (FCO₂) was calculated using the standard formulation described by Liss and Merlivat (1986) and the transfer coefficient from Wanninkhof (2014):

$\:\:{FCO}_{2}={k*\:k}^{{\prime\:}}*{\varDelta\:pCO}_{2}$

Where k is the gas transfer rate (quadratic function of wind speed) and k′ is the solubility coefficient of CO₂, calculated based on temperature and salinity according to Weiss (1974), and ΔpCO₂ = pCO₂ ocean - pCO₂ air. A constant average pCO₂ atmospheric value of ~ 422 µatm was assumed for the study period (Lan et al., 2024). Flows were calculated in mmol C m⁻² d⁻¹ per grid cell and integrated spatially and temporally.

2.9 Model evaluation and event attribution

Model performance was evaluated by comparing corrected and uncorrected pCO₂ fields using root mean square error (RMSE), mean absolute error (MAE), and satellite bias reduction. Event-specific analyses quantified the relative contribution of each event (Ileana, John, Eleven-E) to the seasonal total of air–sea CO₂ flux (FCO₂). This methodology quantifies carbon flux variability induced by high-energy meteorological events in a dynamic region, despite limited information on biogeochemical changes resulting from meteorological events.

3. Results

3.1 Physical characteristics during meteorological events

Between May and November 2024, three tropical systems were recorded off the coast of Guerrero: tropical storms Ileana and Eleven-E, and category 5 hurricane John (CNH, 2024). Among them, John produced the strongest effects, with wind speeds exceeding 70 m·s⁻¹, a rapid surface cooling of approximately 1.5°C between September 22 and 27 (Fig. 2a), and a simultaneous increase in surface chlorophyll a (Fig. 3). The SSS exhibited short-lived positive increases during both John and Ileana, whereas Eleven-E generated disturbances of lesser magnitude and spatial extent (Fig. 2b). The T–S diagram revealed the presence of Tropical Surface Water (TSW) and a weak signal of Gulf of California Water (GCW) (Fig. 2c), consistent with the classification proposed by Fiedler and Lavín (2017).

Figure 2. Temporal evolution of sea surface temperature (SST; panel a) and sea surface salinity (SSS; panel b), showing spatial and temporal variability (mean ± standard deviation) between May and November 2024. A temperature–salinity (T–S) diagram (panel c) illustrates the surface distribution of water masses, identifying Tropical Surface Water (TSW) and Gulf of California Water (GCW), classified according to the boundaries defined by Fiedler and Lavín (2017).

3.2 Calibration of the BNNR model

The BNNR was trained with pCO₂ data collected in the Gulf of California during Storm Lidia (2017), when the oceanographic buoy drifted through the storm`s influence from August 30 to September 3 (trajectories in Fig. 1). Although Lidia's track remained outside the Gulf, its influence was strong enough to cause the buoy, and drive it northward (Fig. 1, S1). This event provided a reference series of pCO₂, SST, and SSS under extreme forcing conditions, which enabled the calibration of the Bayesian model using high-frequency time series.

After aligning buoy data with 2023 satellite predictors, the model was trained using variational inference and a negative log-likelihood loss. Calibration showed strong agreement with observed pCO₂ (MAE: 0.76 µatm, RMSE: 0.99 µatm) under normal conditions. When synthetic extreme-wind data were incorporated, errors increased (MAE: 23.4 µatm, RMSE: 28.7 µatm), reflecting higher variability during intense events but without degrading the model’s ability to capture rapid pCO₂ increases. The Monte Carlo-derived uncertainty bands encompassed observed values within 95% credibility intervals, confirming the model’s robustness and generalizability to similar events.

3.3 Integration of Satellite Data and Predictor Variables

Satellite and reanalyzed variables were integrated to construct a spatially coherent prediction framework.

Predictors included SST, SSS, chlorophyll a, Absolute Dynamic Topography (ADT), and wind components (u, v), from which wind speed and daily maximum wind speed were derived. All data were interpolated to a regular grid of 0.083° and daily frequency.

Preprocessing was performed using Python (Xarray, Pandas, and Scipy libraries), and null values were removed. Missing data were conservatively imputed. Predictor variables were scaled with MinMaxScaler. This process generated a regular grid (0.083°) of more than 200 days, providing a consistent framework for model application. The SST and SSS distributions confirmed the coexistence of tropical and local water masses, with high temporal variability.

3.4 Correction of pCO₂ modeled by PISCES-NEMO

After calibration, the BNNR model was applied to the Guerrero 2024 dataset, which showed that the model systematically corrected the PISCES-NEMO bias (Fig. 3b). The network generated daily predictions of corrected pCO₂ and associated uncertainty. Compared to the original PISCES-NEMO model values (Fig. 3a, c), the corrected predictions (Fig. 3b) showed significant increases during meteorological events, especially on September 23 (Hurricane John), with local increases of up to ~ 300 µatm in coastal areas, reaching values above 850 µatm (Fig. 3d). On days without high-energy meteorological events, the differences between corrected and original pCO₂ were less than ± 50 µatm, demonstrating the model’s specific sensitivity to physical forcing without inducing systematic artifacts (Fig. 3).

Figure 3. Surface distribution maps of partial pressure of CO₂ (pCO₂): satellite-estimated values (panel a), corrected estimates produced by the Bayesian neural network (panel b), and the difference between both products (satellite minus corrected; panel c). Maps correspond to the average of the days with the greatest variation (September 22–25). Panel d shows a temporal comparison between satellite-estimated (blue line) and corrected (red line) pCO₂, with mean ± standard deviation across the study area

3.5 Physical adjustment during Hurricane John

To represent the extreme intensity of Hurricane John more accurately, an additional physical adjustment was applied to the BNNR output for September 23. This adjustment incorporated observed satellite anomalies in wind speed (greater than 50 m·s⁻¹) and chlorophyll a (Fig. 4), resulting in spatially increased corrected pCO₂ in affected regions (Fig. 3b, d). The strategy replicated the spatial pattern observed during Lidia (2017) and improved the physical consistency of predictions, particularly in areas with elevated physical forcing and increased chlorophyll a (Fig. 4).

Figure 4. Time series of chlorophyll a (Chl-a) concentration, showing the mean and standard deviation of spatial and temporal variability between May and November 2024.

3.6 Spatial and Temporal Patterns of Corrected pCO₂

The model generated daily maps of corrected pCO₂ along with their respective uncertainty bands. The areas most affected by cyclones showed increases of more than 200 µatm compared with uncorrected values, with the extreme values were concentrated in the coastal strip and upwelling regions. Predictive uncertainty was lower on regular days (± 30 µatm) and increased during events (up to ± 80 µatm), reflecting greater variability and physical complexity during extreme conditions.

3.7 Comparison of pCO₂ between events

The magnitude of the impact was heterogeneous among the three events. Hurricane John presented the highest corrected pCO₂ values, with local maxima above 850 µatm and differences of up to + 300 µatm relative to the uncorrected product. At the same time, the associated CO₂ fluxes (FCO₂) reached their highest values, with a notable intensification from September 22 to 25 (Fig. 3b). Hurricane Ileana showed an intermediate increase, with pCO₂ differences of around 150–200 µatm, reaching maximum values close to 700 µatm (Fig. S2). The increase in fluxes was also noticeable, although of lesser magnitude and spatial extent compared to John.

Tropical Depression Eleven-E generated the most minor variations, with pCO₂ increases of less than 100 µatm (Figs. S4). Its effects were limited to a narrower coastal strip and were short-lived (October 1–4). The differences in spatial extent were also evident: John affected virtually the entire coastal strip off Guerrero, Ileana showed a more localized pattern, and Eleven-E had a restricted impact.

3.8 Variability of air–sea CO₂ fluxes (FCO₂)

Corrected CO₂ fluxes using the BNNR (Fig. 5b, d) revealed large degassing pulses, particularly during Hurricane John, with values up to twice those of the original PISCES-NEMO estimates (Fig. 5a). In contrast, the satellite product showed more uniform and moderate fluxes (Fig. 5a), indicating that deterministic models may underestimate CO₂ exchange during cyclonic forcing. The integrated time series confirmed that John produced the largest degassing pulse, Ileana an intermediate event, and Eleven-E a minimal impact (Figs. 5, S3, S5, respectively). Overall, tropical cyclones contributed up to one-third of the seasonal atmospheric carbon flux, underscoring their significant role in regional carbon variability.

Figure 5. Surface distribution maps of air–sea CO₂ fluxes (FCO₂): satellite-estimated (panel a), corrected (panel b), and the difference between both products (panel c). Maps represent the average of the most variable days (September 22–25). Panel d compares satellite-estimated (blue line) and corrected (red line) FCO₂ time series, with mean ± standard deviation for May–November 2024

3.9 Physical-biogeochemical conditions during cyclones

The environmental variables associated with the three events confirmed the strong interaction between atmospheric forcing and oceanographic structure. Sea surface temperature (SST) decreased by more than 4°C during Hurricane John (Fig. 6b), consistent with the vertical mixing processes induced by intense winds. In parallel, surface salinity (SSS; Fig. 6a) showed localized positive deviations, possibly linked to subsurface intrusion or lateral transport. Chlorophyll a (Chl-a; Fig. 6c) recorded notable increases in coastal areas, indicating a rapid biological response to upwelling and nutrient redistribution. The overlap of these three variables reinforces that the intensification of pCO₂ (Fig. 3b) and CO₂ fluxes (Fig. 5b) during John was coupled to a physical-biogeochemical feedback system, in which cooling, saline intrusion, and primary productivity combined to amplify the air–sea exchange signal.

Figure 6. Maps of environmental variables averaged for September 22–25, 2024, during Hurricane John: (a) sea surface temperature (SST, °C), (b) sea surface salinity (SSS, units), and (c) chlorophyll a concentration (Chl-a, mg m⁻³). These maps highlight the physical-biogeochemical feedbacks driving the observed anomalies

3.10 Model sensitivity to predictors

Sensitivity analysis of the BNNR identified the relative importance of each predictor in correcting pCO₂. Maximum daily wind speed accounted for approximately 35–40% of the explained variance, followed by sea surface temperature (SST) at 25–30%, sea surface salinity (SSS) at 15–20%, and, to a lesser extent, chlorophyll a and absolute dynamic topography (ADT) at less than 10%. This hierarchy indicates that atmospheric forcing is the primary modulator of pCO₂ anomalies during cyclones, while temperature and salinity primarily influence the air–sea gradient under baseline conditions.

We used SHAP (Shapley Additive exPlanations) values to quantify the contribution of each predictor (wind speed, SST, SSS, and chlorophyll a) to the BNNR pCO₂ predictions. The SHAP analysis confirmed that increases in wind speed above 40 m·s⁻¹ are associated with sudden increases in pCO₂ (> 200 µatm whereas decreases in SST greater than 2°C explain between 20–30% of the variability in the days following the passage of cyclones. Sensitivity to SSS was more limited, indicating that its role is concentrated in localized subsurface intrusion events. Taken together, these results validate that the BNNR architecture explicitly captures the nonlinear relationship between physical forcings and surface biogeochemical dynamics.

4. Discussion

A Bayesian recurrent neural network (BNNR) was implemented to correct satellite-derived pCO₂ estimates from the PISCES-NEMO model in the tropical Pacific off Mexico, focusing on extreme meteorological events, such as hurricanes and tropical storms. The bias-correction utilized in situ reference data from Tropical Storm Lidia (2017) and was applied to satellite and modeled data from May to November 2024, encompassing three tropical cyclones. This approach reduced systematic bias in the PISCES-NEMO biogeochemical model and, to our knowledge, for the first time in this region, incorporated explicit uncertainty into pCO₂ estimates. The integration of real and synthetic data facilitated the representation of extreme conditions, enhancing model performance during hurricanes. The following section presents the main findings on model performance, corrected pCO₂ dynamics, air–sea CO₂ flux (FCO₂) variability, and the biogeochemical implications of these intense events in this tropical region.

4.1 Performance of the Bayesian correction model

The BNNR effectively corrected biases in the PISCES-NEMO pCO₂ estimates, achieving very low errors under normal conditions and capturing sharp increases during cyclones. The BNNR’s ability to reduce systematic errors in pCO₂ estimates is comparable to other machine learning approaches applied to the ocean (e.g., optimized Random Forest, RMSE ≈ 15 µatm; Wu et al., 2024). Unlike deterministic methods, it provides credibility intervals, a critical feature in the context of high variability. Recent studies in simplified carbon and climate modeling have highlighted the value of Bayesian frameworks for representing uncertainty and assimilation of new observations (Bossy et al., 2022). The results suggest that the Bayesian model can complement global products by better adapting to local conditions and extreme events.

This probabilistic framework is consistent with recent advances in climate and biogeochemical modeling, where representing uncertainty is key for decision-making. The results presented here contrast with previous applications of deterministic neural networks, such as those by Bittig et al. (2018) and Denvil-Sommer et al. (2019), as they explicitly incorporate uncertainty into daily pCO₂ predictions in a highly dynamic tropical region.

4.2 Model sensitivity analysis

Sensitivity analysis identified maximum wind speed as the dominant predictor in pCO₂ correction, consistent with previous studies highlighting the role of mechanical forcing in subsurface ventilation and the intensification of air–sea fluxes (Yu et al., 2020; Wang et al., 2020). The relatively high contribution of SST corroborates that surface cooling after cyclonic events acts as a secondary modulator of the air–sea gradient, consistent with observations in the eastern Pacific and South China Sea (Sun et al., 2020; Wu et al., 2025).

The lower sensitivity attributed to SSS and chlorophyll a reflects their role as contextual modulators rather than primary triggers of pCO₂ increase. However, their incursion ensured physical-biogeochemical consistency and avoid spurious predictions, which is consistent with the findings of Somavilla et al. (2016) on the importance of including physical redundancies in ocean prediction models.

Taken together, these results reaffirm that the BNNR not only reduces the systematic bias of PISCES-NEMO but also provides an explicit disaggregation of the relative weight of each forcing factor. This level of interpretability constitutes an advance over deterministic or “black box” models and provides a useful framework for assessing the differential impact of future cyclones in climate change scenarios (IPCC, 2021).

4.3 Impact of extreme events on surface pCO₂

Cyclones generated short-lived but intense pCO₂ pulses, exceeding 300 µatm in less than 24 hours, which coincides with observations in other tropical basins (Landschützer et al., 2016; Wang et al., 2020). These increases were linked to surface cooling, vertical mixing, and the ventilation of carbon-rich subsurface waters. Similar phenomena have been reported in other tropical basins, where typhoons and hurricanes trigger abrupt increases in pCO₂ and air–sea CO₂ fluxes (Yu et al., 2020).

The heterogeneity between events was explained by the combination of intensity, trajectory, translation speed, and previous state of the system, which modulated the magnitude and spatial extent of the anomalies (Yu et al., 2020). In particular, John (cat. 5) produced the most pronounced and extensive changes, Ileana produced an intermediate response, and Eleven-E had a limited impact in magnitude and duration, in line with the dynamic control of mechanical forcing on local ventilation/upwelling.

Comparable responses of pCO₂ to cyclonic forcing have been reported in other tropical basins. For example, in the Bay of Bengal, neural network and gradient boosting models driven by SST and SSS captured anomalies of ± 12 µatm associated with monsoon-driven variability (Joshi et al., 2022). Similarly, in the northern South China Sea, typhoons produced abrupt increases in surface pCO₂ and degassing fluxes exceeding background variability, consistent with the role of extreme wind forcing (Yu et al., 2020). Positioning our results within this broader context highlights that the Bayesian bias-correction approach developed here is not limited to Guerrero, but represents a transferable strategy for other cyclone-prone tropical regions.

4.4 Biogeochemical implications in FCO₂

Using uncertainty-corrected pCO₂ and FCO₂ revealed short-lived but large degassing pulses, which contributed up to ~ 30% of the seasonal flux. This pattern supports the role of cyclones as intensifiers of air–sea CO₂ exchange in tropical regions (Resplandy et al., 2019; Riebesell et al., 2021), highlighting the importance of capturing cyclonic forcing to close seasonal balances.

The colocation of FCO₂ maxima with SST minima and chlorophyll peaks suggests a physical-biological coupling: vertical mixing and subsurface intrusion raise pCO₂ and the air–sea gradient. At the same time, nutrient injection triggers a rapid trophic response that can modulate the return to background conditions. The combination of satellite diagnostics and Bayesian bias-correction enabled the quantification of this coupling with daily resolution, a feat difficult to achieve with deterministic products or sporadic in situ sampling (Liss & Merlivat, 1986; Wanninkhof, 2014; Weiss, 1974).

4.5 Contribution of the Bayesian approach to oceanographic monitoring

The added value of the Bayesian approach can be summarized in three key aspects:

Incorporation of uncertainty: Credibility intervals enable the robustness of predictions to be assessed, a crucial property in decision-making for climate and fisheries management. This type of probabilistic representation has been highlighted in recent ocean and climate modeling studies, where uncertainty estimation is crucial for assessing carbon fluxes and their large-scale implications (Bossy et al., 2022).

Temporal flexibility: the LSTM architecture ensures consistency in time series, avoiding spurious oscillations during periods without events. The usefulness of recurrent architectures for capturing the temporal dynamics of marine processes has been validated in reconstructions of pCO₂ and other biogeochemical variables, showing clear improvements over conventional deterministic networks (Wu et al., 2025).

Increased sensitivity: the inclusion of redundant predictors, such as mean wind and maximum wind, allowed the model to respond differentially to atmospheric forcing, a feature that “black box” algorithms, like SVM or XGBoost, tend to lose by not preserving physical consistency. Recent studies confirm that integrating redundant forcings enhances predictive capacity in dynamic regions by more accurately capturing the nonlinear responses of complex ocean processes (Somavilla et al., 2016; Wu et al., 2024).

A distinctive strength of the Bayesian framework is its capacity to generate credibility intervals that support both scientific interpretation and management decisions. In practice, these intervals provide quantitative bounds on the reliability of pCO₂ and FCO₂ predictions, enabling decision-makers to distinguish between robust signals—such as degassing pulses clearly outside the uncertainty range—and conditions where the overlap of intervals advises caution. This probabilistic framing enhances the usability of bias-corrected estimates in operational contexts, for instance, early warning systems or carbon accounting schemes, where confidence in anomaly detection is as important as the mean prediction itself.

4.6 Limitations and future directions

This study has two main limitations: calibration relied on a single reference event (Lidia 2017), which may not capture the full range of conditions in the Mexican Pacific. Additionally, the synthetic data rely on statistical assumptions that require further validation with additional in situ observations from oceanographic campaigns or buoys.

Another limitation is that the input data resolution of 0.083° may miss submesoscale processes relevant during cyclones.

Future research should:

Incorporate multiple calibration events to strengthen training.

Evaluate hybrid architectures that explicitly integrate physical processes, for example, coupling with simplified biogeochemical models (Pathfinder, PISCES-NEMO; Aumont et al., 2015).

Extend the methodology to climate change scenarios, considering projections of more intense cyclones in the eastern Pacific (IPCC, 2021).

5. Conclusions

A Bayesian recurrent neural network was successfully applied to correct satellite-derived pCO₂ in the tropical Pacific off Guerrero, Mexico, during extreme meteorological events. The model reduced systematic biases, incorporated predictive uncertainty, and revealed short-lived but significant CO₂ degassing pulses, particularly during Hurricane John (2024). These pulses accounted for nearly one-third of the seasonal flux, underscoring the importance of cyclones in regional carbon dynamics.

This study demonstrates that Bayesian approaches provide a powerful and reliable framework for monitoring air–sea CO₂ exchange in highly variable environments. Beyond improving the accuracy of satellite products, they enhance our capacity to evaluate the biogeochemical impacts of extreme events.

Acknowledgment

We thank the Secretariat of Science, Humanities, Innovation, and Technology (SECIHTI) for awarding a postdoctoral fellowship to the first author (EPM-2022(3)). We also acknowledge Dr. T.L.E.C. for providing the Gulf of California oceanographic buoy data, which was essential for model calibration.

Electronic Supplementary Material

Below is the link to the electronic supplementary material

Supplementary Material 1

References

Abadi M, Barham P, Chen J, Chen Z, Davis A, Dean J, Devin M, Ghemawat S, Irving G, Isard M, Kudlur M, Levenberg J, Monga R, Moore S, Murray DG, Steiner B, Tucker P, Vasudevan V, Warden P, Wicke M, Yu Y, Zheng X (2016) TensorFlow: A system for large-scale machine learning. Proceedings of the 12th USENIX Symposium on Operating Systems Design and Implementation (OSDI 2016), 265–283

Aumont O, Ethé C, Tagliabue A, Bopp L, Gehlen M (2015) PISCES-v2: An ocean biogeochemical model for carbon and ecosystem studies. Geosci Model Dev 8(8):2465–2513. https://doi.org/10.5194/gmd-8-2465-2015

Avila LA (2017), December 20 Tropical Cyclone Report: Tropical Storm Lidia (EP142017), 30 August – 3 September 2017. National Hurricane Center. https://www.nhc.noaa.gov/data/tcr/EP142017_Lidia.pdf

Berg R (2024) diciembre 17). Tropical Storm Ileana (EP092024): 12–15 September 2024. Tropical Cyclone Report. National Hurricane Center. https://www.nhc.noaa.gov/data/tcr/EP092024_Ileana.pdf

Bittig HC, Steinhoff T, Claustre H, Fiedler B, Williams NL, Sauzède R, Körtzinger A, Gattuso JP (2018) An alternative to static climatologies: Robust estimation of open ocean CO₂ variables and nutrient concentrations from T, S, and O₂ data using Bayesian neural networks. Front Mar Sci 5:328. https://doi.org/10.3389/fmars.2018.00328

Blake E (2025) 10 de febrero). Tropical Cyclone Report: Unnamed Tropical Storm (formerly Tropical Depression Eleven-E) (EP112024), 1–3 October 2024. National Hurricane Center. Recuperado de https://www.nhc.noaa.gov/data/tcr/EP112024_Unnamed.pdf

Bossy A, Guivarch C, Gauthey J, Dumas P, Séférian R, Boucher O (2022) Simplified climate–carbon cycle models for uncertainty assessment of carbon budgets. Geosci Model Dev 15(23):8831–8855. https://doi.org/10.5194/gmd-15-8831-2022

Chen L, Gao S, Han G, Yu J, Shen Z (2019) Reconstruction of ocean surface pCO₂ using machine learning approaches. J Geophys Research: Oceans 124(7):5038–5054. https://doi.org/10.1029/2018JC014655

Clare M, Yu J, Thompson J, Wang H, Liu Y (2022) Incorporating Bayesian machine learning into environmental predictions. npj Clim Atmospheric Sci 5(1):63. https://doi.org/10.1038/s41612-022-00278-5

Denvil-Sommer A, Gehlen M, Vrac M, Mejia C (2019) LSCE-FFNN-v1: A two-step neural network model for the reconstruction of surface ocean pCO₂ over the global ocean. Geosci Model Dev 12(5):2091–2105. https://doi.org/10.5194/gmd-12-2091-2019

Dillon JV, Langmore I, Tran D, Brevdo E, Vasudevan S, Moore D, Patton B, Alemi A, Hoffman MD, Saurous RA (2017) TensorFlow Distributions. arXiv preprint. https://arxiv.org/abs/1711.10604

Hersbach H, Bell B, Berrisford P, Biavati G, Horányi A, Muñoz-Sabater J, Nicolas J, Peubey C, Radu R, Rozum I, Schepers D, Simmons (2023) A., Soci, C., Dee, D., & Thépaut, J. N. ERA5 hourly data on single levels from 1940 to the present. Copernicus Climate Change Service (C3S) Climate Data Store. https://doi.org/10.24381/cds.adbb2d47

Intergovernmental Panel on Climate Change (IPCC) (2021) Climate Change 2021: The Physical Science Basis. Contribution of Working Group I to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change. Cambridge University Press. https://doi.org/10.1017/9781009157896

Joshi R, Kumar A, Chakraborty K, Ghosh S (2022) Machine learning approaches for reconstructing the ocean surface pCO₂ in the Bay of Bengal. Sci Total Environ 806:150631. https://doi.org/10.1016/j.scitotenv.2021.150631

Fiedler PC, Lavín MF (2017) Oceanographic conditions of the eastern tropical Pacific. In P. W. Glynn, D. P. Manzello, & I. C. Enochs (Eds.), Coral reefs of the eastern tropical Pacific: Persistence and loss in a dynamic environment (pp. 59–83). Springer. https://doi.org/10.1007/978-94-017-7499-4_3

Lan X, Tans P, Thoning KW (2024) Trends in globally averaged CO₂ determined from NOAA Global Monitoring Laboratory measurements. NOAA GML Dataset. https://doi.org/10.15138/9N0H-ZH07

Landschützer P, Gruber N, Bakker DCE (2016) Decadal variations and trends of the global ocean carbon sink. Glob Biogeochem Cycles 30(10):1396–1417. https://doi.org/10.1002/2015GB005359

Li W, Zeng J, Chen L, Gao S, Han G, Yu J (2023) Advances in machine learning for air–sea CO₂ fluxes: Opportunities and challenges. Sci Total Environ 872:162066. https://doi.org/10.1016/j.scitotenv.2023.162066

Liss P, Merlivat L (1986) Air-sea exchange rates: Introduction and synthesis. In: Buat-Ménard P (ed) The role of air–sea exchange in geochemical cycling. Reidel Publishing Company, pp 113–127

Pasch RJ (2025) febrero 10). Hurricane John (EP102024): 22–27 September 2024. Tropical Cyclone Report. National Hurricane Center. https://www.nhc.noaa.gov/data/tcr/EP102024_John.pdf

Python Software Foundation (2023) Python Language Reference, version 3.x. https://www.python.org

Resplandy L, Séférian R, Bopp L (2019) Marine carbon cycle feedbacks to climate change. Nat Clim Change 9(12):919–929. https://doi.org/10.1038/s41558-019-0646-7

Riebesell U, Körtzinger A, Oschlies A (2021) The ocean carbon sink: Processes, variability and change. Front Mar Sci 8:662295. https://doi.org/10.3389/fmars.2021.662295

Somavilla R, González-Pola C, Fernández-Díaz J, Rodríguez C, Josey SA, Sánchez RF (2016) The warmer the ocean surface, the shallower the mixed layer. Biogeosciences 13(14):4359–4373. https://doi.org/10.5194/bg-13-4359-2016

Wang B, Xu Y, Cai W, Wu L, Wang G, Wang Y, Santoso A (2020) The ocean’s carbon uptake under extreme storm forcing. Sci Total Environ 749:142334. https://doi.org/10.1016/j.scitotenv.2020.142334

Wanninkhof R (2014) Relationship between wind speed and gas exchange over the ocean revisited. Limnol Oceanography: Methods 12(6):351–362. https://doi.org/10.4319/lom.2014.12.351

Weiss RF (1974) Carbon dioxide in water and seawater: The solubility of a non-ideal gas. Mar Chem 2(3):203–215. https://doi.org/10.1016/0304-4203(74)90015-2

Wen Y, Vicol P, Ba J, Tran D, Grosse R (2018) Flipout: Efficient pseudo-independent weight perturbations on mini-batches. International Conference on Learning Representations (ICLR 2018). https://openreview.net/forum?id=rJOVjyo0-

Wu C, Xu Y, Zhang J, Wang H, Wang B, Chen H, Li Y, Liu J (2024) A machine learning reconstruction of ocean surface CO₂ reveals variability and drivers. Sci Total Environ 922:170735. https://doi.org/10.1016/j.scitotenv.2024.170735

Yu P, Wang ZA, Churchill J, Zheng M, Pan J, Bai Y, Liang C (2020) Effects of typhoons on surface seawater pCO₂ and air–sea CO₂ fluxes in the northern South China Sea. Journal of Geophysical Research: Oceans, 125(8), e2020JC016258. https://doi.org/10.1029/2020JC016258

Statements and Declarations

Funding

This work was supported by multiple institutions.

Dr. T. Leticia Espinosa Carreón received institutional support from CIDIIR-IPN, which provided financial, administrative, and technical assistance, including buoy deployment, supplies, and travel for equipment maintenance.

Dr. Luz de Lourdes Aurora Coronado Álvarez was supported by the Secretaría de Ciencia, Humanidades, Tecnología e Innovación (SECIHTI) through a postdoctoral fellowship awarded to the first author (CVU 353474).

Dr. Gabriela Y. Cervantes Díaz received institutional support from the Universidad Autónoma de Baja California (UABC), which provided financial and administrative assistance, as well as coverage of publication expenses through an institutional agreement with the journal.

Competing Interests

The authors declare that they have no known competing financial interests or personal relationships that could have influenced the work reported in this paper.

The authors further declare that they have no relevant financial or non-financial interests to disclose.

Author Contributions

All authors contributed to the study’s conception and design. Material preparation, data collection, and analysis were performed by Luz de Lourdes Aurora Coronado Álvarez, Gabriela Y. Cervantes Díaz, and T. Leticia Espinosa Carreón.

The first draft of the manuscript was written by Luz de Lourdes Aurora Coronado Álvarez, and all authors commented on and revised previous versions. All authors read and approved the final manuscript.

Data Availability

The datasets generated and/or analyzed during the current study are not publicly available because they form part of an ongoing research project, but they are available from the corresponding author upon reasonable request.