Skilful decadal-scale prediction of fish habitat and distribution shifts

The study addresses whether decadal-scale climate predictions can be converted into actionable forecasts of marine fish habitat and distribution shifts on timescales relevant to management and stakeholders. While climate impact assessments often focus on 50–100 year horizons, many decisions occur on seasonal to decadal scales. Recent advances show greater predictability in the ocean on annual-to-decadal timescales, particularly in the North Atlantic. Concurrently, climate-driven species redistribution is widespread in the ocean and has already triggered socio-economic impacts and international conflicts (e.g., the North Atlantic mackerel dispute). The purpose is to demonstrate that initialized decadal climate predictions can skillfully forecast fish habitat and infer distribution shifts several years in advance, thus providing anticipatory information to support adaptation, fisheries management, and conflict avoidance.

Prior work has established notable decadal predictability in the North Atlantic for sea surface temperature, ocean heat content, AMOC, CO2 uptake, and subpolar gyre dynamics. Marine organisms have shown pronounced climate-driven range shifts, often faster than on land, with implications for fisheries and governance. Conflicts over transboundary stocks have been documented and are projected to intensify. Existing marine ecological forecast efforts indicate skill at sub-seasonal to seasonal scales and suggest potential at multi-annual scales in many large marine ecosystems. For the target species, previous studies identified key environmental thresholds: mackerel distribution limited by an ~8.5°C August SST isotherm near Greenland; bluefin tuna foraging requiring access to surface waters ≥10–11°C; and blue whiting spawning distributions shaped by subsurface salinity (250–600 m) with a dome-shaped response between ~35.3–35.5 psu. These established habitat models provide a foundation to couple with climate predictions.

Study region: Northern North Atlantic, a region with strong multi-annual to decadal predictability. Species and habitat models: Three case studies—Northeast Atlantic mackerel (Scomber scombrus), Atlantic bluefin tuna (Thunnus thynnus), and blue whiting (Micromesistius poutassou). Habitat definitions: mackerel habitat defined by the August 8.5°C SST isotherm in waters around Greenland (south of 70°N); bluefin tuna thermally suitable feeding habitat defined by the August 11°C SST isotherm south of Iceland; blue whiting spawning habitat defined using a published probabilistic habitat model driven by subsurface (250–600 m) March salinity and other covariates, with calibration to match observed adult distribution areas. Climate predictions: Five CMIP6 DCPP-compliant decadal prediction systems with annual initializations, producing retrospective forecasts up to 10 years ahead (historical forcing to 2014, SSP2-4.5 from 2015), yielding an 85-member grand ensemble. Uninitialized CMIP6 historical and SSP5-8.5 simulations were used as a reference forecast set. Observations: HadISST v1.1 for SST and EN4 v4.2.1 for subsurface salinity (with corrections), monthly fields on 1° grids. Distribution/abundance: International Blue Whiting Spawning Stock Survey (standardized since 2004) for distribution area; ICES stock assessment for spawning stock biomass. Comparable survey datasets for mackerel and bluefin tuna distributions were not available; shifts were documented in literature. Data processing: For models and observations, subsurface salinity averaged over 250–600 m (layer-thickness-weighted). Variables of interest (August SST; March subsurface salinity) extracted and bilinearly regridded to a common 0.5° grid over species-specific regions. Observational climatologies computed for 1985–2014. Bias correction applied via full-field approach per model and lead time: anomalies relative to model climatology (same 1985–2014 period) added to observed climatology to produce bias-adjusted full-field forecasts. Ensemble means computed per model and across all 85 realizations (grand ensemble). Habitat metrics and verification: Habitat models applied to bias-corrected forecasts to compute annual habitat areas (km²) within predefined regions; applied likewise to observations to obtain observed habitat metrics. For blue whiting, the probabilistic model threshold was calibrated to align upper quartile of annual adult distribution areas with habitat estimates. Persistence forecasts constructed by carrying forward a given year’s habitat area up to 10 years. Multi-year averages (3-, 5-, and 9-year centered windows) computed for observations, forecasts, persistence, and uninitialized projections; lead time defined to the center of the averaging window. Skill assessment: Comparison periods—1961–2018 for SST-based metrics (mackerel, bluefin tuna); 1985–2018 for salinity-based metrics (blue whiting) due to pre-1985 observational inconsistencies. Metrics: Pearson correlation coefficient (r), Mean Squared Error Skill Score (MSESS), and Continuous Ranked Probability Skill Score (CRPSS, across all 85 members). Uncertainty estimated by bootstrapping (1000 resamples) to provide 90% confidence intervals and significance (one-tailed tests); pairwise comparison of metrics (e.g., forecast vs persistence).

- Initialized climate prediction systems skillfully forecast key physical drivers at five-year lead times in most of the study domain (SST in August; subsurface salinity in March), providing a solid basis for habitat predictions. - Annual habitat area forecasts show high skill: grand-ensemble Pearson correlations up to ~0.75 across species, with statistically significant skill (p < 0.05) at all leads (1–10 years) for all three species. Skill remains robust at decadal leads; grand ensemble typically outperforms individual models. - Forecasts outperform simpler baselines at multi-annual leads: while 1–2 year persistence performs comparably due to ocean inertia, for lead times ≥3 years all species’ forecasts significantly exceed persistence (p < 0.05 or better) in correlation, and show significant improvements in MSESS and CRPSS as well. - Initialized predictions outperform uninitialized CMIP6 projections used as forecasts, especially in probabilistic performance: CRPSS is significantly higher for initialized systems for all lead times and species, reflecting greater accuracy and precision due to initialization. - Multi-year averaging markedly increases skill: for 9-year averages, r reaches 0.95 (mackerel), 0.94 (bluefin tuna), and 0.74 (blue whiting), and forecasts remain significantly better than persistence (p < 0.01). Both mean-state (r, MSESS) and distributional skill (CRPSS) are significant at decadal scales. - Case insights: Forecast systems could have foreseen habitat changes linked to the North Atlantic mackerel distribution shift into Greenlandic waters and anticipated the decline and partial recovery of blue whiting spawning habitat, consistent with observed distribution contraction/expansion modulated by stock biomass and fishing pressure. - Source of predictability: Skill arises primarily from low-frequency (decadal) variability (e.g., subpolar gyre, AMV, anthropogenic warming) well captured by initialized models; multi-year averaging filters high-frequency atmospheric-driven noise, further enhancing predictability. - Large ensembles add value: The 85-member grand ensemble is consistently among the top performers, underscoring the benefits of ensemble size and multi-model information.

The findings demonstrate that initialized decadal climate predictions can be translated into skillful forecasts of fish habitat areas and, by extension, constraints on distribution shifts, on time horizons directly relevant to fisheries management and international negotiations. The strong performance at multi-annual to decadal leads indicates that predictable low-frequency ocean variability is effectively captured and propagated through biological habitat models. While single-year forecasts already capture substantial variability, multi-year averages substantially boost skill by suppressing unpredictable interannual atmospheric noise. These results address the core question by showing statistically significant improvements over persistence and over uninitialized projections, including in probabilistic terms essential for decision-making. They highlight the potential for foresight to support adaptation strategies, reduce conflict risk for transboundary stocks (e.g., mackerel), and inform flexible, resilient governance arrangements. The asymmetry between habitat and realized distribution underscores that habitat forecasts are necessary but not sufficient for predicting presence; nevertheless, they provide actionable bounds for spatial planning, survey design, and allocation decisions.

This work provides a clear demonstration that initialized decadal climate predictions can yield skillful forecasts of marine fish habitat and inform likely distribution shifts several years to a decade in advance. Forecasts outperform persistence and uninitialized projections and achieve very high skill when averaged over multi-year windows. These capabilities offer tangible value for fisheries management, international quota negotiations, and climate adaptation, particularly for ocean-dependent communities and nations, including SIDS and the Global South. Regularly produced, ensemble-based decadal forecasts can underpin climate services targeting living marine resources. Future directions include expanding the framework to additional species and regions, integrating more ecological drivers (e.g., prey fields, oxygen, ecosystem interactions), enhancing distribution models that link habitat to realized presence and abundance, improving observational datasets and survey coverage, and embedding forecasts within adaptive management protocols and transboundary agreements to operationalize anticipatory decision-making.

- Habitat vs distribution: Habitat models predict where conditions are suitable, not where fish will be present. Predictive skill is asymmetric—better at predicting absence than presence due to unmodeled biotic and behavioral processes (competition, predation, migration, schooling, reproduction) and other environmental factors (food quantity/quality). - Observational constraints: Lack of suitable, consistent monitoring datasets for mackerel and bluefin tuna distributions limited direct verification; blue whiting surveys enabled more thorough evaluation. - Data limitations: Inconsistencies in salinity observations prior to the mid-1980s restricted verification for salinity-based habitat (blue whiting) to 1985–2018. - Baseline comparators: At short lead times (1–2 years), persistence performs similarly due to ocean inertia, limiting relative gains. - External pressures: Stock size and fishing pressure can rapidly alter distribution independently of habitat (e.g., blue whiting mid-2000s), constraining forecast interpretability for realized distributions. - Scope: Analysis focuses on the North Atlantic; generalization to other basins requires evaluation given regional differences in predictability and model performance.