The Hauraki Gulf is a semi-enclosed, oligotrophic shelf sea up to 150 m deep, located in the northeast of New Zealand. Its seafloor comprises diverse volcanic and alluvial sediments ranging from coarse calcareous sand and gravel to fine sands and silts rich in clay (Manighetti and Carter, 1999). Water circulation and sediment transport are dominated by the south-eastward flowing East Auckland Current and strong tidal currents (Manighetti and Carter, 1999; Sharples, 1997; Zeldis et al., 2004). Terrestrial organic C input is limited and mainly confined to the Firth of Thames, while in the inner and outer Gulf, sediment organic matter derives from both terrestrial and marine sources (Sikes et al., 2009). Anthropogenic impacts have affected the Gulf for decades, including bottom trawling, dredging, and sand mining (Hauraki Gulf Forum, 2020, 2023). Bottom trawling generally occurs at depths > 50 m (Fig. S1) wile commercial scallop dredging took place in shallower areas (< 50 m) but has been banned since 2022 (Hauraki Gulf Forum, 2020, 2023).

In February and March 2024, 57 sites were sampled aboard the R/V Te Kaihōpara (Fig. 1). Sediments were collected using a HAPS corer (KC Denmark, 14 cm diameter), with three replicate cores taken per site. Subsamples from the surface sediment (0–3 cm) were collected for characterisation using acid-cleaned (10 % HCl) polypropylene syringes with cut-off tips, while paired small sediment cores were taken for the resuspension assay using acid-cleaned acrylic core liners (inner diameter = 3.4 cm, height = 14 cm). Sediment samples for characterisation were immediately transferred to acid-cleaned polypropylene containers, kept on ice, and frozen at -20 °C at the end of each sampling day. The small intact cores were stored with open tops in dark seawater tanks at ambient bottom-water temperature (± 2 °C), and resuspension assay incubations were conducted within two hours of sampling.

To quantify aqueous CO₂ production in resuspended sediments, we conducted resuspension assays following Bartl et al. (2025). The assay incubates undisturbed sediments in small cores and resuspended sediments in glass bottles, measuring temporal changes in oxygen concentration to determine sediment oxygen demand (SOD). For both treatments, the upper 3 cm of sediment – typically disturbed by trawling and dredging (Hiddink et al., 2017) – were incubated. Optimal sediment-to-water ratios to maintain oxic conditions were 1:8 for sandy sediments (250 mL bottles) and 1:17 for muddy sediments (500 mL bottles). For the resuspension treatment, sediment was added to glass bottles pre-filled with filtered seawater and the sediment was resuspended by gentle inversion for 30 s. In both undisturbed cores and resuspension treatments, SOD (mmol m⁻² h⁻¹) was calculated from oxygen concentrations measured at the start and end of the 4–6 h incubations (OXROB10 probe and FireSting GO2 meter, Pyroscience, Germany). Values were normalised to incubation time and sediment surface area. Linear decline in oxygen, a requirement for this calculation, was validated through preliminary tests (Bartl et al., 2025). At each site, three pairs of control cores and resuspension treatments were incubated in the dark at ambient bottom-water temperature, conditions were monitored with loggers (Hobo Pendants, USA). Incubations showing > 30 % oxygen decrease between initial and final measurements, or before and after shaking, were discarded (Bartl et al., 2025). Organic C mineralisation to CO₂ was estimated from SOD using a respiratory quotient (RQ) of 0.9 for inner-shelf sites (< 50 m depth) and 0.85 for outer-shelf sites (50–200 m; Jørgensen et al., 2022). Resuspension-induced CO₂ production (RCO2P) was calculated as the difference in CO₂ production between resuspended and undisturbed samples. The factor increase in CO₂ production was obtained by dividing resuspended values by undisturbed values. RCO2P serves as a proxy for the vulnerability of sediment organic C to severe resuspension, with higher RCO2P indicating greater vulnerability.

Organic matter content was determined by loss on ignition, burning dried (60 °C for 7 d) sediments at 450 °C for four hours. For grain size analysis, sediments were digested with 10 % H2O2 for six weeks to remove organic matter, washed with deionised water, and dispersed in 5 % sodium hexametaphosphate before laser diffraction analysis using a Malvern Mastersizer-3000 (Malvern, UK). Four grain size classes were derived: coarse sand (500–2000 µm), medium sand (250–500 µm), fine sand (63–250 µm), and mud (< 63 µm). This grouping reflects the bimodal grain size distributions observed in most samples while keeping the number of grain size factors for data analysis manageable. Shell hash and gravel were quantified by wet sieving and weighing two dried fractions: > 2 mm (shell hash and gravel) and < 2 mm (remaining sediment). Phytopigments (chlorophyll a and phaeophytin) were extracted from 1 g of homogenised, freeze-dried sediment using 3 mL of 90 % aqueous acetone over 24 h (Buffan-Dubau and Carman, 2000; Sun et al., 1991). Pigment absorbances were measured before and after acidification with a spectrophotometer (Duetta, Horiba Scientific), and concentrations were calculated following Lorenzen (1967). Chlorophyll a concentration was used as an indicator of algal biomass, while the ratio of organic matter content to total phytopigment concentration (Chl a + phaeophytin) served to characterise short-term organic matter freshness (Miatta and Snelgrove, 2021). Total phytopigment concentration was used rather than Chl a because both photosynthetic Chl a and its degradation product represents labile organic matter components (Pusceddu et al., 2010). Lower ratios indicate fresher, less degraded material.

A total of 171 samples were analysed for each sediment characteristic alongside 171 resuspension assays. After quality assessment of the assay data (lost cores, > 30 % oxygen decline, or large macrofauna), two sites (6 data points) and 21 additional data points were removed. Two Chl a values were interpolated as the mean of the remaining site replicates, resulting in a complete dataset of 144 samples. To identify relationships between sediment heterogeneity and RCO2P, we applied supervised machine learning using boosted regression trees (BRT; (Friedman, 2001). BRTs are ensemble models that sequentially build decision trees, with each tree correcting errors from the previous iteration, improving predictive performance. They capture non-linear relationships and interactions, making them suitable for analysing complex ecological datasets while maintaining strong predictive power (Lucas, 2020; Rubbens et al., 2023). BRTs have been widely applied for predicting species distributions, fishing effort, ecosystem services (Cimino et al., 2020; Lohrer et al., 2020; Soykan et al., 2014) and linking biogeochemical variables to environmental factors (Panaïotis et al., 2025; Rijkenberg et al., 2011).

Our dataset included the following variables indicative of sediment heterogeneity: water depth (Depth), shell hash and gravel content (S/G), coarse sand (C-Sand), medium sand (M-Sand), fine sand (F-Sand), mud (Mud), organic matter content (OM), and the organic matter-to-total phytopigment ratio (OM : Phyto). The response variable was resuspension-induced CO₂ production (RCO2P). Mud, M-Sand, and OM were highly collinear (r>|0.8|; Fig. S2). While multicollinearity does not affect BRT predictions, it complicates interpretation of feature importance and interactions (Boulesteix et al., 2012; Dormann et al., 2013; Lucas, 2020). We therefore excluded Mud and M-Sand, retaining OM as it is the substrate for C mineralisation. Replacement tests using Mud and M-Sand instead of OM produced similar BRT results (Table S1). To perform BRT the data set was split into a training set (75 % of the samples) and a testing set (25 % of the samples). Hyperparameters were tuned via grid search and 4-fold cross-validation: number of trees = 1000, maximum depth = 3, minimum samples per leaf = 3, learning rate = 0.005, and subsampling = 0.8. To ensure robustness, 50 BRT iterations were run. Model interpretation employed SHAP (SHapley Additive exPlanations) values to assess feature importance, interactions, and feature relationships to modelled RCO2P (Li, 2022; Lundberg et al., 2018; Lundberg and Lee, 2017). Mean absolute SHAP values were used to derive overall feature and interaction importance. SHAP dependence plots visualised how modelled RCO2P increased (positive SHAP values) or decreased (negative SHAP values) across feature values. Feature importance rankings and dependence plots were then used to identify clusters with differing RCO2P. Within each cluster, Pearson correlation analysis was used to determine relationships between features and RCO2P. All analyses were conducted in Python (v3.12.7) using the packages scikit-learn (v1.6.1; Pedregosa et al., 2011) and SHAP (v 0.47.1; Lundberg et al., 2020).

The sampled sites covered a wide range of sediment types and depths, with S/G, C-Sand, M-Sand, F-Sand, and Mud contents ranging from 0 %–40 %, 0 %–62 %, 1 %–54 %, 2 %–68 %, and 1 %–84 %, respectively. Grain size was spatially variable: coarser sediments dominated channels and areas around Hauturu (Fig. 2a–c; Fig. S3), while finer sands and mud prevailed near the main coast and at deeper outer-shelf sites (Figs. 2d,  S3). OM content ranged from 0.9 %–9.6 %, with highest values at outer-shelf sites and inner-shelf areas west of Coromandel Peninsula (Fig. 2e). OM freshness varied widely irrespective of OM content or water depth (OM : Phyto = 1.2–16.1; Fig. 2f). Algal biomass, indicated by Chl a, was mainly concentrated at shallow inner-shelf sites, reaching up to 8 µg g⁻¹ dw, but Chl a concentrations of 2–4 µg g⁻¹ were also detected at several sites deeper than 50 m, suggesting either microphytobenthos presence or substantial sedimentation of algal material (Fig. 2g). CO₂ production rates were 1.4–19.5 times higher in resuspended sediments than in undisturbed sediments (Fig. S4). RCO2P showed strong spatial variability (0.1–3.5 mmol CO₂ m⁻² h⁻¹), with elevated rates observed at multiple shallow and deep sites (Fig. 2h).

The 50 BRT model iterations performed well in predicting RCO2P, with an R2 of 0.58 ± 0.11 and a root mean squared error of 0.54 ± 0.07 mmol CO₂ m⁻² h⁻¹, indicating good performance for highly variable environmental data. OM was the most important feature, showing the highest mean absolute SHAP value and ranking first across all 50 model runs (Table 1). C-Sand was the second most important feature, with a mean absolute SHAP value 3.6 times lower than that of OM. Depth and F-Sand had similar SHAP importances (∼ 0.1), consistently ranking third and fourth, while Chl a and S/G were least influential (Table 1). The most significant feature interactions were OM – C-Sand and OM – Depth, which had higher SHAP values than Chl a and S/G but remained lower than the top four individual features. Overall, RCO2P variability was primarily driven by the individual effects of OM, C-Sand, Depth, and F-Sand, with only minor contributions from interactions.

SHAP dependence plots revealed predominantly non-linear relationships between modelled RCO2P and features (Fig. 3). Across the OM gradient, SHAP values shifted from negative to positive at ∼ 3 % OM, indicating highest RCO2P in sediments containing 3 %–10 % OM (Fig. 3a). Similarly, sediments with C-Sand below 8 % and from depths of 56–100 m exhibited positive SHAP values, linking these conditions to higher RCO2P (Fig. 3b, c). Although less influential overall, F-Sand and OM : Phyto showed more linear relationships with RCO2P, with highest SHAP values at F-Sand > 27 % and OM : Phyto < 4.7, suggesting that greater CO₂ production corresponds to higher F-Sand content and fresher organic matter (Fig. 3d, e). No clear relationships were observed for Chl a or S/G (Fig. 3f, g).

To visualise which sediment types are most vulnerable to CO₂ release upon resuspension, we produced scatter heat maps from the original dataset and used BRT results to identify clusters (Fig. 4, Table 2). Lowest RCO2P rates (0.5 ± 0.3 mmol m⁻² h⁻¹) occurred in sandier sediments (> 8 % C-Sand, > 27 % F-Sand) with low organic matter content (< 3 %) and shallow depths (< 56 m), forming a distinct OM-poor cluster (Fig. 4a, b). Sediments in this cluster are from the Colville and Jellicoe Channels and around Hauturu and show a low to moderate risk of CO₂ release (Fig. 2H, Table 2). RCO2P in the OM-poor cluster correlated positively with OM and F-Sand but negatively with C-Sand, indicating stronger resuspension-driven mineralisation where OM and finer fractions are higher (Fig. 4c). At depths ≥ 100 m, both OM and RCO2P were high and relatively uniform, with RCO2P about four times higher than in the OM-poor cluster (Table 2). This Deep-cluster showed a negative correlation between RCO2P and OM, and slightly higher rates at lower OM : Phyto ratios, reflecting fresher organic matter (Fig. 4e).

Mixed sediments (> 3 % OM, < 8 % C-Sand) from intermediate depths (56–100 m,) formed a Mixed cluster with RCO2P ranging from 0.8–3.5 mmol m⁻² h⁻¹, equivalent to 3–19.5 times higher CO₂ release than in undisturbed sediments (Fig. 4b, Table 2). No clear correlations were found here (Table 2, Fig. 4d). Within the Mixed cluster the highest 10 % of RCO2P values (> 2.6 mmol m⁻² h⁻¹) occurred along interactions between OM and C-Sand (C-Sand = 270.08 × OM^−2.93, R2 = 0.51, p = 0.003) and OM and Depth (Depth =-4.97× OM + 87.45, R2 = 0.32, p = 0.029; Fig. 4a, b). Together, the Mixed and Deep cluster cover 73 % or the sampled sites and are moderately to very highly vulnerable to releasing CO₂ when disturbed (Table 2, Fig. 2h).

The stimulation of organic C mineralisation through sediment resuspension can involve multiple physical, chemical and biological mechanisms (Hulthe et al., 1998; Kleber et al., 2021; van Nugteren et al., 2009; Pusceddu et al., 2005b) which in sum contribute to the RCO2P rates that we measured. While we expected variability of RCO2P due to sampling across substantial sediment heterogeneity, the extent of variability and the location of apparent hotspots was surprising. We found three clusters of C storage vulnerability in which RCO2P appears to be regulated by differing environmental conditions. Firstly, low to moderate vulnerability in the shallow, coarse-grained sediments around Hauturu and in the channels is likely linked to strong tidal and residual currents that winnow fine organic particles and/or naturally enhance organic matter turnover resulting in a smaller resuspension impact (Boudreau et al., 2001; Manighetti and Carter, 1999). The positive relationship of RCO2P with organic matter quantity and finer grained sand content suggests that substrate availability and strong (less F-Sand) vs. calm (more F-Sand) hydrodynamic conditions were controlling the magnitude of the resuspension impact in the OM-poor cluster. At the other end of the spectrum, deep, muddy, OM-rich sediments showed high to very high C storage vulnerability likely due to the consistently high pool of OM that accumulates in the deep outer shelf area of the Hauraki Gulf. Interestingly, here RCO2P was ∼ 10 % higher when OM was fresher, suggesting that quality of the resuspended organic matter influences mineralisation when the substrate is abundant.

We could not derive mechanistic explanations of RCO2P variability in the mixed sediment cluster based on our sediment data. This suggest that there may be other environmental variables that could explain the spatial variability of resuspension impacts in these sediments. Measures of organic matter quality or bioavailability are likely to influence RCO2P and may have been underrepresented in our set of features. While our use of simple and cost-effective proxies (e.g. loss-on-ignition OM and OM:Phytopigment ratio) enables broad applicability (Bartl et al., 2025) it may overlook important compositional nuances of OM. Incorporating additional measures, such as C : N ratios, δ13C signatures, or n-alkane signatures (Sikes et al. 2009), may help resolve spatial RCO2P patterns. RCO2P variability could also be linked to the concentration and quality of dissolved organic matter that is resuspended with the sediment and can contain a considerable fraction of readily degradable compounds that could enhance microbial mineralisation (Kujawinski, 2011; Lengier et al., 2024; Reader et al., 2019). Lastly, the composition and activity of the microbial community and their response to being resuspended could influence RCO2P irrespective of the amount, composition or biochemical quality of organic matter (DesRosiers et al., 2022). On a broader scale, spatially variable OM supply from benthic or pelagic primary production or lateral transport by cross-shore bottom currents could also drive the spatial pattern of RCO2P in the Hauraki Gulf (Chang et al., 2003; Zeldis et al., 2004). Since we can only speculate about potential underlying mechanisms of resuspension impacts in the Mixed cluster, future targeted experiments on the role of organic matter bioavailability and the microbial response to resuspension in mixed sediment types will shed more light on underlying mechanisms of resuspension-induced organic C mineralisation.

Our BRT model explained ∼ 58 % of variability in RCO2P forming a strong basis for using BRT models like ours to predict C storage vulnerability from spatial patterns of sediment characteristics. Sediment grain size and organic matter data often have high spatial resolution and thus represent useful surrogates to quantify large-scale sediment C vulnerability. Improving the performance of our BRT model to capture more of the currently unexplained variability (∼ 40 %) can perhaps be achieved through integrating more nuanced features as discussed above (see Sect. 4.1). Additionally, history or intensity of trawling can influence both the long-term concentration and reactivity of OM as well as how it is mediated through benthic faunal communities (Hale et al., 2017; Pusceddu et al., 2014; Tiano et al., 2022; Zhang et al., 2024). As a result, RCO2P could be partly driven by the sampling site's trawling disturbance history. However, incorporating a measure of historical disturbance frequency may only improve assessments if the data on trawling activity is collected at the resolution necessary to link it to local sampling coordinates. Compared to first-order model estimations that rely on assumptions of degradation constants and organic C lability (Atwood et al., 2024; Luisetti et al., 2019; Muñoz et al., 2023; Sala et al., 2021), predictions from a BRT model approach are based on empirical measurements of sediment characteristics, resuspension-induced organic C mineralisation rates and the relationships and interactions that they form. This makes it a powerful tool that integrates environmental variability making predictions more realistic and detailed at the regional scale offering opportunities for meaningful regulatory actions for trawling.

The resuspension assay provides a simple measure of oxic organic C mineralisation in the top 3 cm of sediment, offering insight into the vulnerability of seafloor C storage. Porz et al. (2024) used a similar approach in their model, defining sediment organic C vulnerability in the top 10 cm as the maximum potential oxic organic C remineralization rate. Their modelled oxic organic C remineralization rates (0.1–100 mmol C m⁻² d⁻¹) align with our empirical RCO2P measurements (2–88 mmol C m⁻² d⁻¹). Other experiments investigating the immediate impact of sediment resuspension reported mineralisation rates to be 1.1–4.7 times higher in resuspension treatments compared to controls (Almroth-Rosell et al., 2012a; Niemistö et al., 2018; Ståhlberg et al., 2006). This is comparable but at the lower end of the range of our measurements where RCO2P was 1.4–19.5 times higher in resuspended sediments. The difference may be attributable to a thicker sediment layer being resuspended in our assay (3 cm) compared to the other experiments (0.3 µm to 1 cm) and the different methodological approaches (e.g. in situ chamber vs. bottle incubations). Overall, the comparability of out measurements to both model and experimental studies supports the assay's relevance for resuspension impact assessments.

Two methodological aspects of the assay need to be considered when linking the assay to trawling impacts and organic C mineralisation. Firstly, the assay determines potential CO₂ production after severe resuspension through shaking up sediment in a bottle and therefore may not reflect the true mechanical impact of individual trawling gear and trawling technique (O'Neill and Ivanović, 2016; Rijnsdorp et al., 2021). However, the high sampling frequency that is possible with the assay enhances our spatial understanding of sediment C vulnerability and thus where trawling would be most impactful. Secondly, the assay converts sediment oxygen demand to CO₂ production using respiratory quotients (see Sect. 2.2). This quantification may overestimate CO₂ production if reduced species are oxidised alongside organic C. We incubated the surface sediments from an oligotrophic system where oxygen and nitrate penetration depths (O₂: 3–6 mm, nitrate: 12 mm) and total-to-diffusive oxygen uptake ratios (TOU / DOU = 2.4) indicate strong macrofaunal influence on redox conditions and minimal accumulation of reduced species in both sandy and muddy sediments (30–128 m depth; Cheung et al., 2024). This aligns with findings of low acid volatile sulfide concentrations (AVS) in non-impacted Hauraki Gulf sediments (0–1 µmol g⁻¹ w/w) compared to higher AVS levels in sediments impacted by a mussel farm (2–12 µmol g⁻¹ w/w, Wilson and Vopel, 2015). Future use of the resuspension assay, particularly in more reduced or eutrophic sediments, should include a validation of the SOD approach by measuring changes dissolved CO₂ or DIC during the incubations.

As our understanding of sediment C storage vulnerability grows, its integration into spatial management of demersal fisheries becomes inevitable. In the Hauraki Gulf, recent discussions on confining trawling pressures focused on protecting reef-forming species and habitats, while sediment C storage vulnerability was not considered (Bennion et al., 2024). Trawling remains allowed in areas deeper than 50 m (Bevin, 2025) which, based on our results, includes nearly all sediments with high to very high C vulnerability. This leaves Hauraki Gulf sediments at risk to lose their climate-stabilizing C storage function and will contribute to the green-house gas emission of the NZ fisheries industry. By integrating sediment heterogeneity and the resulting spatial variability of resuspension impacts, highly vulnerable areas can be identified and protected. In a recent modelling study, Porz et al. (2024) compared different seafloor protection scenarios in the North Sea and found that protection based on C vulnerability was most efficient for preserving organic carbon and maintaining benthic macrofauna biomass. This highlights that seafloor carbon protection generates benefits not only for organic C storage, but also for benthic species and habitats ultimately maintaining the undisturbed functioning of seafloor ecosystems.