The Global Ocean Data Analysis Project (version GLODAPv2.2023), a bias-corrected observational ocean biogeochemical dataset, serves as the ship-based observational data source for this study. Data from GLODAPv2.2023 collected south of 30° S are selected, including concurrent measurements of hydrographic properties, nutrients (NO3, PO4 and SiO4), and carbonate system parameters (TA, DIC, and pH), as detailed in Table 1. Before reprocessing the data, two cruises (Expocode: 316N19871123 and 318M19771204) were excluded due to noisy at depth or large quality control (QC) adjustments, as reported by Carter et al. (2021b). Subsequently, the remaining 160 cruises undergo secondary QC and adjustment check. Measurements flagged as poor quality includes TA and DIC values with adjustments exceeding ±10 µmolkg-1, pH adjustments greater than ±0.015 pH units, and nutrient data with multiplicative adjustments surpassing 10 % (Carter et al., 2018; Olsen et al., 2016). The precise adjustment values are documented in the GLODAPv2 Adjustment Table, accessible at https://glodapv2.geomar.de/ (last access: 10 April 2025). Following this QC step, five cruises are excluded for TA, three cruises (1112 measurements) for DIC, one cruise (1474 measurements) for PO4, and one cruise (940 measurements) for SiO4 (see detailed exclusions in Table A1). Although no significant offsets were identified in pH measurements, one cruise (Expocode: 49HG19950414) are excluded as noted by Carter et al. (2018). Importantly, quality control is performed independently for each variable. Subsequently, TA and DIC measurements are retained only when nutrient observations are available, following Carter et al. (2021b). The pH data in GLODAPv2 comprise a mixture of spectrophotometric- and potentiometric-derived measurements. To ensure data consistency, pH measurements are homogenized to align with pH calculated from TA and DIC, following Carter et al. (2018). Classification of pH data are conducted based on documentation available from https://cchdo.ucsd.edu/ (last access: 10 April 2025), as shown in Table A2 and Fig. A1.

The ESPER_LIR and ESPER_NN model were trained using data from GLODAPv2.2020, whereas the CANYON-B model utilized the original GLODAPv2 release. Assessment dataset for model performance comparison is identified from the GLODAPv2.2023, consisting of cruises added subsequent to the GLODAPv2.2020 release (i.e., cruise numbers ≥ 2107). Initial comparative analysis among CANYON-B, ESPER_NN, and ESPER_LIR models is conducted using this assessment data.

The Argo float data were download from the Argo Data Assembly Canters (GDACs; ftp://ftp.ifremer.fr/ifremer/argo/dac/, last access: 23 February 2025) and processed using adapted code from the SAGEO2 toolbox. This dataset comprises three types of Argo floats (Core Argo, BGC-Argo, and Deep Argo) for reconstructing carbonate system parameters and nutrients using models. Since 2000, the Core Argo network has provided high-resolution temperature and salinity profiles with broad coverage (0–2000 dbar at 10 d intervals), forming the foundation for extensive studies of oceanographic processes. Building upon this framework, the BGC-Argo extends observational capabilities by employing biogeochemical sensors to measure oxygen, pH, and nitrate. To address ongoing uncertainties regarding deep ocean, the recent deployment of Deep Argo floats enables data collection down to 6000 dbar in targeted Southern Ocean basins, providing unprecedented insights into carbon dynamics in abyssal waters.

Rigorous quality control leads to the exclusion of three categories of problematic data: (1) floats on the Argo Program's grey list identified for sensor drift or transmission errors; (2) floats with 10 or fewer operational cycles, due to insufficient calibration stability; and (3) aberrant profiles with incomplete measurements. Additionally, only adjusted data flagged as “Good” or “Probably Good” (QC flags 1 and 2, respectively) were included. The remaining floats were systematically classified based on the presence or absence of oxygen data. This classification yielded two distinct float categories, underpinning our dual-pathway analytical approach and ensuring robust estimation across diverse observational regimes. Overall, this study includes data from 4346 Argo floats, of which 525 are equipped with oxygen sensor providing 73 296 profiles, and the remaining 3821 floats without oxygen sensor providing 647 650 profiles (Table 1).

There are substantial spatial sampling gaps in the high-quality GLODAP data, particularly in the high-latitude Southern Ocean (Fig. 1). Furthermore, Fig. 1 reveals a pronounced seasonal bias toward the austral summer, with nearly four times as many measurements collected during this period compared to winter. In contrast, Argo floats provide extensive spatiotemporal coverage, owing to their flexible deployment and consistent 10 d sampling cycles. Although the number of Argo floats equipped with oxygen sensors have increased greatly in recent decades (Fig. 1f), the Argo observational network is still predominantly composed of Core-Argo floats without oxygen sensors, which constitute over 85 % of the dataset and achieve nearly complete spatial coverage across the Southern Ocean (Fig. 1a and c). The broad coverage offers an unprecedented foundation for reconstructing carbon system dynamics in the region. However, because of current limitations in data quality and correction methods (Maurer et al., 2021; Williams et al., 2017), nitrate and pH measurements from BGC-Argo floats are not used in this study; only temperature, salinity, and O2 are employed. Ongoing improvements in quality control and correction procedures may enable the incorporation of these measurements in future studies.

This study employs a dynamically adaptive framework to reconstruct carbonate system parameters and nutrients by integrating heterogeneous Argo float observations with high-quality GLODAP measurements. Figure 2 illustrates the overall workflow for generating gridded products in the Southern Ocean. Based on performance comparisons (see Sect. 4.1 for detail), the best-performing model is applied to reconstruct key biogeochemical tracers (oxygen, nitrate, phosphate, silicate) as well as carbonate system parameters (TA, DIC, pH). To accommodate differences in observational capabilities among Argo floats, particularly regarding the presence or absence of oxygen sensors, input observations are dynamically sorted into two reconstruction pathways:

Full-parameter pathway (green in Fig. 2): This pathway utilizes all available measured variables, including hydrographic properties, and dissolved oxygen concentrations from floats equipped with oxygen sensors.

The resulting dataset includes both reconstructed variables (derived indirectly from Argo profiles) and direct high-quality ship-based observations.

Typical methods for calculating Cant from DIC measurements include the ΔC* method (Gruber et al., 1996, Gruber, 1998; Sabine et al., 2004), the extended multiple linear regression (eMLR) method (Gruber et al., 2019b), and the Tracer combining Oxygen, inorganic Carbon, and total Alkalinity (TrOCA) method (Touratier et al., 2007; Touratier and Goyet, 2004) has been widely applied. Although the ΔC* method has been widely used to quantify Cant inventory, it relies on parameter customization through Optimum Multiparameter (OMP) analysis, which requires direct nutrient measurements of NO3, PO4, and SiO4 – data not available in our dataset. The eMLR method relies on repeat hydrographic measurements (Friis et al., 2005), which are not available for Argo profiles, and its output reflects temporal change in Cant rather than absolute concentration. For these reasons, this study employs the TrOCA method, which is relatively straightforward, extensively utilized in Southern Ocean studies (Metzl et al., 2024; Zhang et al., 2023), and has been demonstrated to be reliable through comparative analyses (Lo Monaco et al., 2005; Vázquez-Rodríguez et al., 2009; Mahieu et al., 2020; Zhang et al., 2023). 1TrOCA=O2+aDIC-12TA2Cant=TrOCA-TrOCA0a=O2+1.279DIC-12TA-e7.511-(1.087×10-2)θ-7.81×105TA21.279 where θ is potential temperature, °C; O2 is dissolved oxygen, µmolkg-1; DIC is total inorganic carbon, µmolkg-1; TA is total alkalinity, µmolkg-1. In the full-parameter pathway, calculated values utilized observed oxygen concentrations alongside model-derived TA and DIC estimates. Conversely, the hydrography-only pathway employed model-derived estimates for all three variables (oxygen, TA, and DIC). Finally, the Cant values are scaled to the reference year 2013 to deal with exponential increase of anthropogenic CO2 burden in the climatological products (see Sect. 3.3) (Carter et al., 2021b; Tanhua et al., 2007). A detailed description of the scaling method is given in the Appendix B1. It should be noted that the TrOCA approach is limited to waters below the euphotic layer (Lo Monaco et al., 2005), therefore the Cant estimates above 100 m are excluded from this dataset.

The aragonite saturation state (Ωar) is calculated using the CO2SYS software (v3, Sharp et al., 2023), requiring TA, DIC, temperature, salinity, and pressure as inputs to minimize uncertainty in the results (Orr et al., 2018). The following thermodynamic parameterizations are employed: carbonic acid dissociation constants from Lueker et al., 2000, hydrogen fluoride (HF) dissociation constants from Perez and Fraga, 1987, the ratio of total boron (B_T) to practical salinity (Sp) from Lee et al. (2010), and bisulfate dissociation constants (KHSO4) from Dickson et al. (1990).

Profile data for each parameter are sorted into spatial bins of 1° longitude × 1° latitude bins and 84 vertical levels to generate homogenized three-dimensional gridded products. Data derived from both float- and ship-based observations are integrated into this spatial framework, ensuring robust spatial and depth coverage. To maximize data density, we construct an “All-Data Grid” by merging all available reconstructions and observations. In addition, three specialized gridded products are generated: the “Float Grid”, comprising only float-based reconstructions; the “Non-O2-Float Grid”, limited to floats without oxygen measurements; and the “O2-Float Grid”, limited to BGC-Argo floats. The latter two grids facilitate sensitivity analyses of oxygen's influence on carbonate system parameter reconstructions. All these gridded datasets serve as the basis for subsequent analyses.

Figure 3 demonstrates the vertical sampling spacing of CTD and dissolved oxygen from Argo floats. Typically, floats sample at intervals of 10 m or finer from the surface down to 200 m and at intervals of 50 m or finer between 500 to 2000 m. Floats equipped with oxygen sensors sample dissolved oxygen at higher resolution. Measured CTD profiles are prioritized, but interpolated profiles are used when concurrent oxygen data are unavailable. To align with the float sampling scheme and maximize data utilization, the water column (0–5600 m) is divided into 84 vertical depth levels (highlighted in yellow in Fig. 3): 0–100 m at 5 m intervals (20 levels), 100–500 m at 25 m intervals (15 levels), 500–2000 m at 50 m intervals (30 levels), and 2000–5600 m at 200 m intervals (19 levels). The deepest level (5600 m) corresponds to the maximum float measurement depth.

Each Argo float profile is interpolated to these predefined vertical levels using the Piecewise Cubic Hermite Interpolating Polynomial routine (“intprofile.m” in the 2nd QC toolbox, Lauvset and Tanhua, 2015). After interpolation and prior to gridding, extreme outliers are identified and removed. For each pressure level and Longhurst Biogeographical Province (available at http://comlmaps.org/how-to/layers-and-resources/boundaries/longhurst-biogeographical-provinces/, last access: 26 May 2025), the interquartile range (IQR) is calculated for each parameter, and values exceeding 1.5×IQR are flagged as outliers (Johnson and Purkey, 2024). When an outlier is detected for any parameter at a given vertical level within a bin, all variables from that profile and level are discarded. Following Gruber et al. (1998), negative Cant estimates are preserved as negative in the averaging process. Finally, for each grid cell, all valid measurements are averaged to obtain representative values, while cells without observations are left empty. This bin-averaging approach ensures that the gridded products are entirely observation-based and preserve the genuine spatial structure of the compiled dataset.

To ensure consistency and reliability between float-based reconstructions (including both full-parameter and hydrography-only pathways) and ship-based observations, three comparative analysis of reconstruction-derived variables (Cant and Ωar) are implemented.

First, methodological discrepancies between variables from each pathway are assessed using GLODAPv2 data. Specifically, values of Cant and Ωar calculated directly from ship-based observations are compared with those estimated from the full-parameter pathway (C_{ant_ship_f}, Ωar_ship_f) and the hydrography-only pathway (C_{ant_ship_h}, Ωar_ship_h), respectively. This analysis quantifies the biases inherent in each reconstruction pathway and is detailed in Sect. 4.2. Second, within regions exhibiting spatial overlap between float-derived reconstructions and independent ship-based estimates, comparisons of float-based Cant and Ωar from each pathway are performed following established cross-over quality control procedure (results shown in Sect. 4.2). Comparisons are restricted to cases where differences are within ±0.005 kgm-3 in potential density (σθ), ±0.005 in neutral density (τ), and ±100 dbar between 1400 and 2100 dbar depth range (Bushinsky et al., 2025). This targeted analysis serves to assess potential discrepancies arising from differences between oxygen-equipped floats and floats without oxygen sensors. These two analysis are essential to evaluate potential biases introduced by differences between oxygen-equipped and CTD-only Argo floats, which is particularly important as oxygen-equipped floats comprise approximately 11 % of total Argo float deployments in the Southern Ocean – potentially leading to disproportionate representation and biases in gridded products.

Finally, a detailed zonal analysis among the gridded products is conducted to evaluate how observational differences impact the spatial consistency and reliability of the final products (results show in Sect. 4.4).

Model performance is evaluated for three widely used approaches for estimating oceanic biogeochemical properties: ESPER_LIR, ESPER_NN, and the CANYON-B. The models are trained on a different version of the GLODAPv2 dataset (CANYON-B on GLODAPv2.2016; ESPER_LIR and ESPER_NN on GLODAPv2.2020). Independent assessment data not included in model training are used for comparison (Table 1). Both the full-parameter and hydrography-only reconstruction pathways are assessed.

Under the full-parameter pathway, ESPER_NN achieves the lowest RMSE for most reconstructed variables (Table 2), including TA (4.37 µmolkg-1), DIC (6.09 µmolkg-1), PO4 (0.07 µmolkg-1), and SiO4 (2.59 µmolkg-1). ESPER_LIR performs slightly better for pH and NO3. Generally, ESPER_NN's RMSE values for TA and DIC are several percent lower than those of ESPER_LIR and CANYON-B, demonstrating superior accuracy relative to the observations.

Under the hydrography-only pathway, omission of oxygen leads to a notable increase in RMSE for DIC (8.78 µmolkg-1), particularly in deep and abyssal waters (Table 2). This highlights the critical role of oxygen measurements as predictors for deep DIC. Relative to CANYON-B, both TA and DIC exhibit systematic underestimation, with mean full-column mean bias of -0.6 and -1.7 µmolkg-1, respectively. These biases are more pronounced than those reported in earlier evaluation (Carter et al., 2021b), who found biases of -0.4 and -0.8 µmolkg-1 using 2019–2020 GLODAPv2 data. The increasing discrepancies suggest that prediction errors in CANYON-B may be accumulating over time. Similar trends are observed for ESPER_LIR and ESPER_NN, indicating that periodic updates to model training datasets are necessary to mitigate future underestimation of DIC.

Uncertainty magnitudes vary among models, reflecting both differences in model structure and in uncertainty estimation methodologies. CANYON-B, which directly incorporates measurement uncertainties from input variables, produces larger uncertainty. Vertical performance analysis shows that the lowest RMSE values for most variables are found in deep and abyssal layers (comprising 24 % of full-column data), while the largest errors are found in surface waters (25 %), likely due to greater variability in surface carbonate chemistry.

Overall, the ESPER_NN demonstrates the highest accuracy and lowest uncertainty under both reconstruction pathways (Table 2), supporting its selection as the primary model for reconstructing carbonate system parameters and nutrients in the Southern Ocean throughout this work. Based on the estimated RMSE of ESPER_NN, float-derived estimates are expected to fall within twice the model's estimated uncertainty range, serving as a criterion for the quality control applied to the Argo float dataset.

We first analyze the methodological discrepancies between the two reconstruction pathways using high-quality shipboard measurements. Biases in Cant and Ωar are quantified by using ship-based observations with concurrent O2, TA, and DIC measurements (N=93667). Parameter values derived directly from measured shipboard data are compared with those calculated from reconstructed DIC, TA for two pathways (Fig. 5).

Under the full-parameter pathway, Cant exhibits a slight negative bias relative to shipboard derived values (Cant_ship_M), with a median of -0.08 µmolkg-1 and a mean of -0.16 ± 4.74 µmolkg-1 (95 % CI: [-0.19,-0.12] µmolkg-1 ). When oxygen is omitted, the Cant bias distribution broadens and shifts slightly positive, with a median of 0.07 µmolkg-1 and a mean of 0.02 ± 6.60 µmolkg-1 (95 % CI: [-0.02,0.07] µmolkg-1). Ωar biases remain small in both pathways but show a similar pattern: for the full-parameter pathway, Ωar bias is centered near zero (median=0.0006; mean=0.0006 ± 0.0433; 95 % CI: [0.0005, 0.0011]), whereas the hydrography-only pathway yields a slightly larger mean bias and spread (median=0.0021; mean=0.0021 ± 0.0618; 95 % CI: [0.0010,0.0018]). The hydrography-only pathway results in a median difference (Cant_ship_H-Cant_ship_F; Ωar_ship_H-Ωar_ship_F) of +0.1 µmolkg-1 and 0.001 units with an added methodological uncertainty of ±2.4 µmolkg-1 and ±0.02 units (Fig. A3) compared to the full-parameter pathway.

Subsequently, biases in float-based estimates from the two pathways are further assessed. The comparison results, restricted to cases within the 1400–2100 dbar depth range and specific seawater property differences (see Sect. 3.4), are shown in Fig. 6a–d. Oxygen concentrations estimated by both pathways exhibit insignificant systematic offset, indicating robust performance of float-based oxygen reconstructions. Biases in float-based Cant estimates exhibit a moderate positive correlation with oxygen biases, especially at higher latitudes, while Ωar biases exhibit a slight positive association with O2 biases, particularly in the mid-latitudes. The difference in float-based Cant and Ωar between the two reconstructed pathways are within ±10 µmolkg-1 and ±0.075 unit, respectively. These ranges are roughly half those observed for the bias distributions between float-based and ship-based values.

Overall, the two reconstructed pathways for Argo floats have a narrow bias distribution for reliable Southern Ocean analyses. Considering both the methodological uncertainty and random uncertainty (estimated following Sect. 3.5.1), the uncertainties of Cant in both pathways are about ±4–6 µmolkg-1, which remain acceptably low for large-scale biogeochemical reconstructions (e.g., compared to Pardo et al., 2014, of ±6 µmolkg-1 and Asselot et al., 2024, of ±5.2 µmolkg-1).

Climatological spatial distributions of interior carbonate system parameters are obtained by averaging measured and reconstructed values from both ship-based observations and Argo float-based reconstructions, as well as their calculated values of Cant, and Ωar. Shipboard measurements span 1972–2020, while float–based observations span 2000–2025, with oxygen-equipped floats contributing data since 2003. Figures 7 and 8 illustrate the climatological spatial distribution of the interior carbon system parameters of the Float Grid. The spatial patterns of the gridded products are generally consistent, except for Cant and Ωar in the southwestern Atlantic (see Fig. B2). To illustrate this discrepancy, the corresponding O2-Float Grid distributions are also provided. Figure 9 presents the distributions in the abyssal layer. Because observations below 4000 m are extremely scarce, the All-Data Grid is used to provide the most comprehensive depiction of deep-ocean conditions. Section 4.4 further provides a detailed comparison among these products and shipboard estimates.

The absence of continental barriers across much of the Southern Ocean and the transport of the Antarctic Circumpolar Current (ACC) result in pronounced meridional gradients dominating the spatial patterns of interior biogeochemical properties. These meridional gradients are closely linked to the spatial distribution of the circumpolar hydrographic fronts, including the Subtropical Front (STF), the Subantarctic Front (SAF), the Polar Front (PF), and the Southern Antarctic Circumpolar Current Front (SACCF), which are indicated by black lines in the climatological distribution maps. The climatological distributions further reflect inter-basin variability driven by ocean basin geometry, bathymetry, and ocean circulation differences among the Pacific, Atlantic, and Indian Oceans. The averaged profile distributions in the Pacific, Atlantic, and Indian sectors of the Southern Ocean ae shown in Figs. 7d, h and 8d, h.

The distribution of DIC and Cant exhibit strong spatial relationships (Fig. 7). In the subsurface layer, high DIC and low Cant concentrations are found south of the PF, the southern boundary of the ACC, due to upwelling of older, DIC-rich and Cant-poor deep waters (Marshall and Speer, 2012). Conversely, the northern portion of the Southern Ocean, in north of the SAF, display low DIC and high Cant concentrations attributed to the transport of Subantarctic Mode Water (SAMW) (Talley, 2013). As depth increases into the intermediate layer (1400–2000 m), Cant concentrations decline significantly, accompanied by increases in DIC. Cant demonstrates pronounced basin-scale variability, with notably low concentrations (0–5 µmolkg-1) in mid-to-high latitudes of the southeastern Pacific Ocean, particularly south of the SACCF between 120 and 180° W. Conversely, higher Cant concentrations (>20 µmolkg-1) are observed in the Pacific sectors and areas south of the PF in the eastern Antarctic region. Regions with elevated DIC typically show lower Cant concentrations, and vice versa. In the deep and abyssal layer (2000–5600 m, Fig. 9c), the spatial patterns of DIC and Cant remain unchanged, and their vertical profiles flatten. Both DIC and Cant exhibit relatively high concentrations in the eastern Antarctic region, where Antarctic Bottom Waters (AABW) forms (Morrison et al., 2020). This enrichment is consistent with AABW-driven transport of anthropogenic carbon into the deep ocean.

The accumulated Cant uptake and increased DIC concentrations intensify OA, leading to declines in both pH and Ωar. Spatial distributions of pH and Ωar (Fig. 8) closely resemble those of DIC (Fig. 7a–d). In the subsurface layer, pH exhibits a spatial distribution pattern nearly identical to DIC. However, Fig. 8e demonstrates distinctly lower Ωar value south of the STF. Both pH and Ωar values decrease markedly from the surface to approximately 1000 m, with more gradual declines at depths below 1000 m. In the intermediate and deep layer, the Pacific Sector shows the lowest pH and Ωar values, followed by the Indian Sector and Atlantic Sector, respectively.

The Float Grid including the Non-O2-Float Grid and the O2-Float Grid represents observation-based climatological products with strong capability to resolve fine-scale horizontal and vertical distributions of interior ocean carbonate system parameters. Figure 10 presents the latitudinal distributions of Cant and Ωar across the Southern Ocean for four gridded products in this study as well as GLODAP-derived data. Additionally, we apply the TRACE method (Carter et al., 2025) to estimate Cant, generating a gridded dataset as described in Sect. 3.3, which serves as an additional comparison (Fig. 10b). TRACEv1 adopts a hybrid conceptual framework, with surface-ocean estimates that are observation-based, whereas deep-ocean fields are model-based and tuned against observations.

The four gridded products show latitudinal variations that are closely aligned with the GLODAP-derived data (green lines). Notably, in the intermediate layers characterized by variety in Cant distributions across oceanic basins (Fig. 8f), the GLODAP-derived dataset north of 45° S exhibit pronounced zonal gradients, likely due to sparse longitudinal sampling. In contrast, our products, benefiting from enhanced spatial coverage, better capture the integrated regional variations. The TRACE-derived dataset (purple lines) yields lower Cant concentration than the TrOCA-derived values, particularly in intermediate waters. This difference may partly arise because the model-based approach is difficult to constrain accurately in regions with sparse transient tracer observations and complex vertical structures associated with Southern Ocean upwelling (as suggested by Carter et al., 2025), leading to deviations from the observation-based TrOCA estimates. In deep and abyssal layers, Cant concentrations show an increasing trend from lower latitudes toward higher latitudes (60–70° S). This pattern may be linked to the formation of AABW, which drives transport of Cant into the deep oceans (Zhang et al., 2023).

A direct comparison between the O2-Float Grid (red lines) and the Non-O2-Float Grid (black lines) elucidate differences attributable to Core Argo versus BGC Argo observations. Reconstructions of Cant and Ωar are broadly consistent across subsurface and intermediate layers for both float types. Significant discrepancies and steep gradients among gridded products are evident in all water layers south of 65° S. Figure A7 illustrates the geographical coverage of float and ship-based observations at high latitudes. In the abyssal layer, float observations are restricted to the eastern Weddell Sea near 4100 m, as they may not be representative of abyssal-layer conditions. Notably, a hotspot of high Cant and vertically confined low Ωar is identified in the southwestern Atlantic Ocean near the SAF and PF in the Non-O2-Float Grid (Fig. 10a). These noticeable differences likely arise from the limited number of shipboard training samples in this region, leading to reduced performance of the machine-learning reconstructions. For regional studies in the southwestern Atlantic, we therefore recommend using the O2-Float Grid provided in this study. The corresponding vertical profiles are presented in Appendix B3. Although float-based measurements introduce additional uncertainties, their extensive spatial and temporal coverage enables our products to offer unprecedented insight into the previously under-sampled Southern Ocean, particularly in the deep ocean. Overall, the All-Data Grid offers a comprehensive representation of the Southern Ocean interior, while the O2-Float Grid and Non-O2-Float Grid demonstrate the potential and limitations of Argo-based reconstructions for studying carbon dynamics.

The uncertainty of gridded products arise from both the uncertainties in the parameter estimates and mapping (sampling) errors. Considering the nonnegligible trends of accumulated Cant, the evaluation of uncertainty in this section mainly focuses on the anthropogenic CO2. Both random errors and potential biases contribute to the uncertainties in the Cant estimates. In Sect. 4.1, the random errors for individual measurements have been estimated to be about ±4–6 µmolkg-1 in the full-parameter pathway and the hydrography-only pathway. The potential bias including uncertainty in inversion algorithm assumptions and formula parameters are more difficult to assess quantitatively, but have little effect on the climatological distribution. The mapping errors, reflecting uncertainties introduced during spatial interpolation, are also challenging to evaluate precisely.

Traditionally, distribution maps of carbonate system parameters, including Cant, were constructed using limited GLODAPv2 cruises data (sampling stations are shown in Fig. 1b), with spatial coverage extended via regression and interpolation methods (Gruber et al., 2019b; Sabine et al., 2004; Barth et al., 2014). In these earlier products, mapping errors strongly depended on the vertical and horizontal data distribution and were assumed to be less than 15 % (Sabine et al., 2004). In contrast, our gridded products leverage the statistical advantage of aggregating multiple independent observations, resulting in gridded uncertainties that are smaller than individual observation uncertainties, and mapping discrepancies reduced to below 7.5 % (Fig. A9). Although our approach may underestimate uncertainty due to potential representativity error, our dataset offers a significant improvement in both accuracy and spatial representativeness over previous gap-filling approaches, and it further extends coverage to the ocean bottom, whereas earlier analyses were largely restricted to the upper 0–3000 m (Gruber et al., 2019a). This enhancement is especially valuable for robustly assessing variability and climatological trends in the historically data-sparse Southern Ocean.