Skillful multiyear prediction of marine habitat shifts jointly constrained by ocean temperature and dissolved oxygen

Marine species’ aerobic metabolism is constrained by temperature (T) and dissolved oxygen (O2), which together determine habitat viability. Ocean warming reduces oxygen solubility and ventilation while increasing metabolic oxygen demand, potentially shrinking or shifting viable habitats. Because T and O2 vary on interannual-to-decadal timescales, forecasting associated habitat shifts on multi-year horizons may be possible. A mechanistic, trait-based framework—the Metabolic Index—links O2 supply and temperature-dependent metabolic demand to define habitat viability. This study asks: how skillfully can multiyear habitat shifts, quantified by the normalized Metabolic Index across diverse ecotypes and depth habitats, be predicted using an initialized Earth system prediction system? The purpose is to assess potential predictability, identify mechanistic drivers (O2, T, S), and characterize ecotype-dependent predictability across North American Large Marine Ecosystems (LMEs), thereby informing marine resource management.

Previous work developed and validated a physiologically based Metabolic Index that better explains species distributions than temperature- or oxygen-only models and projects large global reductions in aerobic habitat under high-emissions scenarios, especially at high latitudes. Earth System Models have demonstrated skill predicting physical and biogeochemical ocean variables on seasonal to decadal scales, and have been used for ecosystem-related predictions (e.g., ocean acidification). However, the interannual-to-decadal predictability of the Metabolic Index itself and its mechanistic controls has remained unclear. Prior studies indicate oxygen variability can be a key driver of observed population and range changes in regions with strong O2 variability, suggesting oxygen may enhance predictability beyond temperature-based forecasts. This work bridges initialized decadal prediction with the metabolic framework to quantify potential predictability of habitat viability.

- Metric: Focus on the normalized Metabolic Index φ = Φ/Φcrit (Φcrit = 1), where Φ is the ratio of environmental O2 supply to temperature-dependent metabolic O2 demand. Species traits include hypoxic tolerance (A0) and temperature sensitivity of hypoxia vulnerability (E0). Habitat is viable when φ > 1. Analyses primarily use A0 = 10 atm−1 and E0 spanning −0.2 to 1.0 eV, defining ecotypes. - Domain and depths: Eleven North American LMEs and two depth habitats: 0–200 m (epipelagic) and 200–600 m (mesopelagic/thermocline). - Prediction system: Community Earth System Model Decadal Prediction Large Ensemble (CESM-DPLE; 40 members, annual initialization 1954–2017 from a Forced Ocean–Sea Ice reconstruction, FOSI). Ocean resolution ~1° with 60 vertical levels; biogeochemistry via BEC model (diagnostic, no feedback on physics). Drift correction applied via lead-time-dependent climatology subtraction. - Predictand and potential predictability: Due to sparse O2 observations, forecasts are verified against the FOSI reconstruction rather than observations, yielding potential predictability (upper bound). A simple reconstruction persistence forecast provides a low-cost baseline. - Ecotype selection for detailed analyses: Three representative E0 ecotypes: low (−0.2 eV; e.g., sea squirt), medium (0.4 eV; e.g., common littoral crab), high (1.0 eV; e.g., northern/deep-water shrimp). - Skill metrics: Anomaly Correlation Coefficient (ACC) and Normalized Mean Absolute Error (NMAE) between yearly anomalies of DPLE ensemble-mean forecasts and FOSI, for lead years 1–10. ACC significance assessed at 95% via Student’s t with effective degrees of freedom. Predictability timescale is the maximum contiguous lead year span starting at LY1 with significant ACC. Differences between DPLE and persistence ACCs tested for significance accounting for common predictand. - Mechanistic attribution: First-order Taylor linear decomposition of φ into components attributable to O2, T, and S; further decomposition of the O2 component into oxygen solubility (function of T and S) and apparent oxygen utilization (AOU). A variance budget partitions φ variance into component variances and covariances to diagnose controls across regions, depths, leads, and ecotypes. - Habitat shift characterization: Time-mean φ fields and interannual standard deviation (1954–2017) used to map spatial and vertical habitat extents (φ > 1) and their interannual expansion/contraction and shoaling/deepening.

- Overall predictability: φ is potentially predictable in the upper 600 m, with CESM-DPLE outperforming persistence across most LMEs, depths, and leads. Average predictability timescales improve from ~2 years (persistence) to as much as ~6 years (DPLE), with significant ACC improvements (ΔACC) ranging from ~0.1 (IPH, SEUS) to ~0.4 (Aleutian Islands). - Depth dependence: Subsurface layers exhibit stronger, longer skill. Examples include predictability up to ~9 years in the Gulf of Alaska (300–500 m) and up to ~7 years in the Aleutian Islands (100–300 m). Northeast LMEs show both surface (<100 m; up to ~7 years) and subsurface improvements (to ~4 years). SEUS and IPH show weak surface skill but much higher skill at 200–600 m (e.g., SEUS ACC increase ~0.3, NMAE decrease ~0.4; predictability up to ~6 years vs 1–2 years for persistence). - Regional patterns: Enhanced ACCs over mid-shelf EBS, northern Gulf of Alaska, southern and northern California Current, and Labrador shelf. Some open ocean regions (e.g., south of Aleutians) exhibit long predictability as well (up to ~9 years). - Reemergence: California Current shows reemergent ACC at LY 4–10 centered near ~100 m deep at LY4, deepening to ~300 m by LY10. A weaker reemergent pattern occurs near-surface in IPH. This likely reflects mean advection of subsurface water mass anomalies. - Mechanistic controls: Oxygen dominates predictability of φ in most regions, depths, and leads. Exceptions occur on high-latitude shelves (inner EBS, Labrador shelf) and in some Atlantic subtropical coastal regions and the Labrador Sea, where temperature contributes comparably or dominantly, especially near-surface (<100 m). Salinity contributions are generally minor, mainly in Labrador shelf/sea. Variance budgets show O2 (particularly AOU) drives φ variance in Pacific LMEs; in NE Atlantic LMEs, O2sol, AOU, and their covariance all contribute, with notable O2–T covariance. - Ecotype dependence: Predictability differs strongly across E0 traits, especially in NE LMEs. In NEUS upper layer, predictability ranges from 1–2 years for some mid-range E0 to up to ~10 years for low-E0 (e.g., E0 = 0 eV) ecotypes. NEUS deeper layer shows up to 8–10 years for low-E0 (<0.1 eV) but ~4 years for medium–high E0. NW LMEs (upper layer) and SEUS (deeper layer) show longer predictability (~6–7 years) for medium-E0 ecotypes, with limited skill (1–4 years) for low/high E0. Trait-space differences are largely driven by the O2 component in many Pacific LMEs (especially at depth), whereas T plays a stronger role in NE LMEs, with S contributing for some low-E0 cases.

Combining an initialized decadal prediction system with a mechanistic metabolic framework substantially increases multiyear forecast skill of aerobic habitat viability relative to persistence. The results directly address the research question by demonstrating that φ is predictable on interannual-to-decadal timescales across North American LMEs and depth habitats, with oxygen generally providing the dominant source of predictability. Where ventilation is strong (e.g., high-latitude shelves), temperature becomes a key determinant, particularly near the surface, reflecting relaxation of oxygen constraints. Variance and decomposition analyses emphasize the critical role of AOU in driving variability and predictability in much of the Pacific, while mixed O2sol–AOU–T contributions characterize the NE Atlantic LMEs. Reemergent skill in the California Current and IPH suggests dynamical memory via advection of subsurface anomalies. Ecotype-specific analyses reveal that predictability is not uniform across metabolic traits; some ecotypes (e.g., low-E0 in NEUS) exhibit very long predictability, underscoring the value of trait-informed forecasts for management. These findings highlight oxygen observations and forecasts as pivotal for advancing skillful, actionable habitat predictions and resource planning (e.g., port operations, gear adaptation, and anticipating predator–prey overlap shifts).

This study demonstrates that initialized Earth system decadal predictions, combined with a trait-based metabolic framework, provide skillful multiyear forecasts of aerobic habitat viability across 11 North American LMEs and two depth layers. Predictions markedly exceed persistence, especially at depth, with oxygen generally dominating predictability, and temperature playing a stronger role in ventilated, high-latitude shelves. Trait-space analysis identifies substantial ecotype-dependent predictability, enabling species-specific forecast applications. The approach offers a pathway to operationally relevant habitat forecasts to inform fisheries and ecosystem management. Future work should prioritize: verifying realized skill with expanded oxygen observations, improving representation of AOU-related processes (ventilation and biological O2 production/consumption), increasing model resolution (global high-resolution or regional downscaling), and further diagnosing reemergence mechanisms driving multiyear predictability.

- Verification uses a model reconstruction (FOSI) rather than observations for O2 due to data sparsity, so results represent potential predictability (upper bound), not realized skill. - Coarse ocean resolution (~1°) may not resolve fine-scale processes important for physical and biogeochemical variability, potentially limiting forecast fidelity in coastal/shelf regions. - Ocean biogeochemistry is diagnostic (no feedback onto physics), which may constrain coupled predictability pathways. - Model and reconstruction biases exist; oxygen observational products are only recently emerging, limiting comprehensive validation. - Skill and predictability vary by region, depth, and ecotype; near-surface predictability remains limited in some LMEs (e.g., SEUS, IPH).