A deep-learning estimate of the decadal trends in the Southern Ocean carbon storage

The study investigates how dissolved inorganic carbon (DIC) in the Southern Ocean has changed over recent decades and how these changes relate to the region’s role in absorbing anthropogenic CO2. Oceans take up roughly 25% of anthropogenic CO2 emissions, with the Southern Ocean accounting for about 40% of the oceanic sink due to cold temperatures, strong winds, and significant biological activity. Earlier analyses suggested a weakening Southern Ocean carbon sink through the 1990s, followed by a strengthening in the 2000s, likely driven by shifts in wind forcing and circulation. Yet, interior ocean carbon storage and its decadal evolution remain poorly constrained due to sparse observations and challenges in biogeochemical modeling. The authors aim to bridge this gap by developing a deep-learning model that predicts 3D DIC in the Southern Ocean from surface observations, enabling estimation of decadal DIC trends and their zonal contrasts.

Prior work has documented decadal variability in the Southern Ocean carbon sink, including a weakening in the 1990s and reinvigoration in the 2000s linked to wind-driven changes in circulation and buoyancy forcing. Biological uptake exhibits strong seasonality and interannual variability, affecting net air–sea CO2 exchange. Studies have highlighted the importance of upwelling of DIC-rich Circumpolar Deep Water, transformations into Subantarctic Mode Water (SAMW) and Antarctic Intermediate Water (AAIW), and the influence of the Southern Annular Mode. Warming and freshwater flux changes modulate stratification and upwelling strength. In the Atlantic, a weakening AMOC has implications for connectivity between deep and surface layers and DIC distribution. Observational constraints from shipboard programs (GLODAP) and biogeochemical Argo floats have expanded coverage, but disparities between ship-only and float-only products can bias estimates, motivating combined datasets and advanced methods (e.g., neural networks) that capture non-linear relationships and leverage satellite inputs.

The authors developed a deep-learning framework adapted from a U-net architecture augmented with temporal recurrence to predict Southern Ocean DIC (30–80°S, surface to 4 km). The network comprises three convolutional encoder blocks (two 3×3 conv layers with ReLU activations plus 2×2 max pooling per block), an LSTM with 1024 units to capture temporal dynamics of latent features, and three up-convolutional decoder blocks (2×2 up-convolutions with ReLU). Residual skip connections link encoder and decoder levels. The model is trained with mean squared error loss using the ADAM optimizer. To span the vertical, 22 U-nets were trained to cover 48 depth levels from the surface to 4 km (x=2 levels per model above 2 km; x=3 levels per model between 2–4 km). Inputs (predictors) include sea surface temperature (SST), sea surface height anomalies (SSH), derived surface vertical velocity w, surface ocean velocities (u, v), 10 m winds, net surface heat flux, surface chlorophyll-a, and surface pCO2. The approach uses a two-phase training strategy: Phase 1 pre-training on the Biogeochemical Southern Ocean State Estimate (B-SOSE; 1/3° spatial, 3-day temporal resolution, 2008–2012), which provides 3D DIC and associated fields and is richly sampled, especially below 2 km; Phase 2 fine-tuning on observations from GLODAPv2 shipboard measurements (1998–2019, to ≥4 km) and SOCCOM biogeochemical Argo floats (2014–2019, to 2 km) to correct potential model biases and improve realism. For Phase 1, inputs were taken from B-SOSE (SSH, u, v, w, heat flux, pCO2, chlorophyll-a) with SST and 10 m winds from ERA5 (averaged to 3-day) to aid subsequent matching to observations. Data were split into training (85% randomly sampled), in-sample validation (10% of training), and out-of-sample validation (15%). Phase 1 achieved high agreement with B-SOSE DIC (r2≈0.97; RMSE≈5.4 µmol/kg; errors mostly within ±10 µmol/kg and centered around zero; minimal vertical bias). In Phase 2, the pre-trained weights were transferred and the model was further trained to minimize RMSE against observational DIC (using only good-quality flags), with an 80/20 train/test split and 10% of training for in-sample validation. Inclusion of both shipboard and float data improved correlations and reduced RMSE relative to ship-only training (overall RMSE ≈13 µmol/kg). Chlorophyll-a gaps (e.g., under sea ice) were set to zero to align with B-SOSE treatment; for 1993–1996 when chlorophyll-a was unavailable, a climatology (1997–2019) was used. Coastal regions shallower than 1 km depth were excluded from training due to differing dynamics. After training, the model was applied to satellite/reanalysis inputs to produce DIC fields at 1° horizontal resolution and 5-day intervals for 1993–2019. Time series at each grid cell and depth were gap-filled with cubic splines under strict criteria (≥5 years per decade and ≥2/3 of a year of data at some point) to avoid over-interpolation. Seasonal cycles were removed per decade before computing linear trends. Trends not significant at the 95% confidence level (p≥0.05) were omitted. Trends in surface drivers (SST, ΔpCO2 ocean–atmosphere, 10 m zonal wind, net heat flux) were computed analogously for context. Model validation against shipboard-based local DIC trends showed strong agreement (slope≈0.95, r2≈0.98; ratios centered around 1), acknowledging sparse temporal sampling at individual stations.

- The model reproduces known spatial structures of DIC: near-surface DIC increases poleward and aligns with neutral density surfaces in the interior; the Pacific and Indian basins exhibit higher DIC than the Atlantic due to older, bottom-sourced waters. - 1993–2009: Broad decreases in interior DIC, particularly in the Pacific sector. At the surface, decreasing DIC implies lower ocean pCO2 and a strengthened carbon sink in the 2000s, consistent with prior studies. - 2010s: Reversal to increasing DIC near the surface across much of the Southern Ocean, with a notable build-up of DIC at depth (below ~1 km) in regions of CDW upwelling, indicating weakened connectivity/export from surface to interior in some sectors. - Pacific sector mechanisms: • 1990s: Predominantly positive DIC trends in the upper 1 km (50–60°S) linked to intensifying, poleward-shifting Westerlies (positive SAM), enhancing divergence and upwelling of DIC-rich CDW; deeper waters show decreasing DIC. • 2000s: Weakened near-surface DIC trends, turning negative in places, attributed to buoyancy-driven weakening of CDW upwelling (enhanced sea-ice melt and freshening) and increased contribution of low-DIC subtropical waters to SAMW/AAIW formation. • 2010s: Near-surface DIC trends turn predominantly positive despite weaker Westerlies, suggesting buoyancy forcing (surface warming) outweighs wind forcing in setting DIC. - Atlantic and Indian sectors: • 1990s–2000s: Warming and heat storage in the upper 2 km stabilize the water column and weaken wind-driven upwelling, producing decreasing DIC trends along upwelling isopycnals (45–60°S in the Atlantic; south of ~50°S in the Indian). • 2010s: Atlantic shows strengthened subsurface decreasing DIC along upwelling isopycnals with near-surface DIC increases (consistent with reduced carbon uptake potential). Indian sector exhibits negative trends south of 50°S but positive near-surface trends farther north, aligned with strong SAMW/AAIW formation enhanced by salinity fluxes and Ekman pumping. - AMOC weakening since the 1990s likely reduced connectivity between deep and surface layers in the Atlantic, consistent with progressively more negative subsurface DIC trends along upwelling density surfaces; recent subtropical AMOC strengthening may affect the Southern Ocean with a lag. - Model performance and data products: • Phase 1 vs B-SOSE: r2≈0.97; RMSE≈5.4 µmol/kg; errors centered near zero across depths. • Phase 2 vs observations: RMSE≈13 µmol/kg; including floats plus shipboard data improves correlations relative to ship-only training. • Shipboard trend comparison: slope≈0.95; r2≈0.98 between observed and model-based local trends. • Produced 5-day 3D DIC fields at 1° resolution for 1993–2019, enabling spatially resolved decadal trend analyses.

The findings extend surface-based assessments of Southern Ocean CO2 uptake by quantifying interior DIC changes and their decadal evolution. Decreases in near-surface DIC during the 2000s are congruent with enhanced ocean uptake of atmospheric CO2 inferred from surface pCO2 analyses, while the 2010s show near-surface DIC increases and subsurface decreases in some sectors indicative of weakened export and connectivity to the deep ocean. Mechanistic interpretation using water-mass transformations (SAMW, AAIW, CDW) clarifies zonal differences: wind-driven upwelling changes dominate in the 1990s in the Pacific, while buoyancy forcing and altered water-mass formation exert greater control in the 2000s and 2010s, with distinct signatures across Pacific, Atlantic, and Indian sectors. The results align with independent observational syntheses and modeling studies that attribute DIC variability primarily to circulation changes and buoyancy forcing. By leveraging surface satellite data and limited interior observations, the deep-learning framework provides rapid 3D reconstructions and trend estimates, improving spatial coverage (including winter via floats) and reconciling ship-only vs float-only biases when combined. These insights are relevant for projecting the Southern Ocean’s role in future carbon uptake and for understanding regional sensitivities to shifts in winds, heat, and freshwater fluxes.

This work introduces a fast, computationally efficient deep-learning method to estimate 3D Southern Ocean DIC from surface variables and applies it to derive decadal trends from 1993–2019. The analysis reveals: (i) decreasing interior DIC through the 1990s–2000s with near-surface decreases that enhanced carbon uptake, followed by (ii) near-surface DIC increases and deep DIC build-up in the 2010s, with mechanisms and magnitudes varying by ocean sector due to differing wind and buoyancy forcing and water-mass transformations. The approach reconciles observational sparsity by combining shipboard and float measurements for bias correction and provides high-temporal-resolution fields suitable for monitoring and process studies. Future work should separate anthropogenic versus natural DIC components, incorporate additional constraints (e.g., alkalinity, remineralization), improve representation of coastal and shelf regions, and extend the methodology for near-real-time monitoring with incoming satellite products.

- Inability to partition changes in total DIC into anthropogenic versus natural components within the current framework. - Observational sparsity, especially below 2 km and in winter for shipboard data; floats do not sample below 2 km, limiting deep validation. - Reliance on B-SOSE (only 2008–2012) for pre-training may imprint model-specific biases despite Phase 2 corrections; large imbalance between B-SOSE and observational data volumes necessitated two-phase training. - Chlorophyll-a data gaps prior to 1997 and under sea ice were filled using climatology or zeros, potentially affecting biogeochemical signals. - Exclusion of coastal/shallow regions (<1 km depth) reduces applicability in shelf seas where distinct processes operate. - Trend estimation depends on gap-filling and significance thresholds; sparse temporal sampling at fixed stations can bias local trend comparisons. - Inputs omit some potentially influential variables (e.g., total alkalinity, remineralization rates, calcification), which could improve DIC predictions.