Ecological forecasts for marine resource management during climate extremes

The study addresses the need for forward-looking ecological information to support marine resource management under increasing climate variability and change. While ocean physics and ecology have seen advances in sub-seasonal to seasonal forecasting, uptake by end-users is impeded by technical debt, uncertainty communication, and perceived necessity of high-resolution regional forecasts. The central questions are whether existing management tools can be reliably reconfigured for forecasting using readily available global sea surface temperature (SST) forecasts, and how forecast resolution and ensemble size (global versus downscaled) affect ecological forecast skill. Focusing on the California Current Ecosystem (CCE), the authors test if global forecasts can provide adequate skill for management-relevant indices without regional downscaling, thereby enabling proactive decision-making during climate extremes such as marine heatwaves.

The paper situates itself within rapidly advancing atmospheric and ocean forecasting for ecological applications, noting predictability from seasonal to decadal scales driven by features like ocean memory and life-history traits. Prior studies demonstrate the utility of global forecasts for regional ecological applications and the development of flexible management tools to support climate adaptation. However, barriers include the assumption that high-resolution regional forecasts are required and the lack of operational downscaled products. The authors reference work on marine heatwave prediction, fisheries and aquaculture decision support, ecosystem indicators in upwelling systems, and prior implementations of the Habitat Compression Index (HCI) and loggerhead bycatch risk tools. They also discuss limitations in end-user uptake, emphasizing uncertainty representation and compatibility with management needs.

Study design: Two existing management tools originally based on observed SST were reconfigured for forecasting with lead times of 0.5–11.5 months using both global (∼1° resolution) and dynamically downscaled (0.1°) SST forecasts in the California Current Ecosystem. Forecast data sources: - Global forecasts: North American Multimodel Ensemble (NMME) monthly SST forecasts from six active models (1981–present, model-specific), up to 9–12 months lead, 1° resolution. Total of 73 ensemble members across models; each member represents different initial conditions to sample climate variability. - Regional downscaling: A subset of forecasts from the CanCM4 model (ensemble members 2, 8, 10) were dynamically downscaled using ROMS configured at 0.1° over the California Current region. Bias-corrected forcing was produced by computing global forecast anomalies relative to observed climatology, adding anomalies to high-resolution atmospheric and oceanic climatologies (ERA5 for atmosphere; SODA v2.1.6 for ocean), and applying these at ROMS surface and lateral boundaries. Downscaled forecasts are available daily for 1990–2010. Case study tools and computation: 1) Habitat Compression Index (HCI): - Domain: Surface waters off California (35–40°N), within 150 km of the coastline. - Definition: Number of grid cells with SST lower than a monthly SST threshold within 150 km of shore, normalized by total grid cells in the 150 km band to yield values in [0,1]. Monthly SST thresholds are the 1981–2010 mean SST from the coast to 75 km offshore. Lower HCI indicates high compression (reduced cool habitat). High compression events were defined as HCI values below the long-term mean. 2) Temperature Observations to Avoid Loggerheads (TOTAL): - Domain: Southern California Bight around 34°N, 120°W, corresponding to the Loggerhead Conservation Area. - Definition: Six-month rolling mean of SST anomalies (SSTA) over the domain. Potential closures are evaluated for June, July, and August, based on SSTA in the preceding six months. Forecast closures are recommended when SSTA exceeds a threshold corresponding to historical closure-triggering anomalies. Observed SSTA threshold of 0.41 corresponds to the 82nd percentile (1981–2010) and 74th percentile (1981–2020); thresholds were converted to percentiles to enable inter-forecast comparisons and lead-time-specific decisions. Forecast configurations compared: - Global Full Ensemble: 73 global ensemble members. - Global Reduced Ensemble: 3 global ensemble members (to match downscaled ensemble size). - Downscaled Ensemble: 3 downscaled ensemble members (CanCM4 members 2, 8, 10). Skill assessment: - Lead times: 0.5 to 11.5 months; skill assessed for each target month. - Metrics: (1) Pearson correlation between observed and forecast values (HCI for HCI tool; SSTA for TOTAL), with significance accounting for autocorrelation via effective degrees of freedom and Fisher’s Z confidence intervals; (2) Forecast accuracy (fraction correct), evaluated against random-chance baselines that depend on event frequency f (for HCI, f ≈ 0.5 so random accuracy ≈ 0.5; for TOTAL, thresholds correspond to f ≈ 0.26 or 0.18 depending on configuration); (3) Symmetric Extreme Dependence Index (SEDI), equitable and base-rate independent, with values >0 indicating skill above random. Accuracy and SEDI significance were estimated by bootstrap resampling (1000 random forecast realizations per month/lead time) to derive 95% confidence intervals. Operational examples: - HCI forecasts were evaluated across 1981–2010 (and extended where available) including the 2014–2016 marine heatwave period. - TOTAL forecasts were evaluated for closure months (June–August) including historical closures (Aug 2014; Jun–Aug 2015; Jun–Aug 2016), assessing earliest lead time with significant skill and closure prediction.

- Global SST forecasts can skillfully drive ecological forecast tools with lead times up to 11.5–12 months. - HCI (whale entanglement risk proxy): - Forecasts of high compression events were typically skillful from 0.5 to 11.5 months lead across the year, with higher skill in winter and spring. - Significant skill for high compression extended to 8.5 months for February–March forecasts. - During the 2014–2016 marine heatwave (Mar 2014–Dec 2016), 94% of forecasts (n = 13 initializations) correctly identified high compression over the 33-month period; high compression was predicted up to 11.5 months in advance, with most false negatives at the heatwave onset. - Example: May 2005 high compression was correctly forecast at 0.5, 1.5, and 3 months lead. - TOTAL (loggerhead bycatch risk/closure indicator): - Closures were skillfully forecast at 6.5 months lead, with significant skill by some metrics out to 11.5 months. - Example: The August 2014 closure was correctly forecast as early as November 2013; closures in Jan–Aug 2015 and 2016 were correctly forecast 11.5 months in advance. - Global vs downscaled forecasts: - The Global Full Ensemble (73 members) was generally more skillful than the Downscaled Ensemble (3 members), despite coarser spatial resolution, due to larger ensemble size better capturing environmental variability. - When the global ensemble was reduced to 3 members, its skill dropped markedly and fell below the downscaled forecasts, indicating that downscaling improves individual member skill but ensemble size strongly influences overall forecast performance. - Practical implication: Operational ecological forecasting for management actions (e.g., RAMP decisions for whale entanglement mitigation, loggerhead closure planning) can be supported using readily available global forecasts without mandatory regional downscaling.

The results directly address whether existing ecological management tools can operate in a forecast mode with actionable lead times. Both HCI and TOTAL achieved skillful forecasts up to 11.5 months, demonstrating the feasibility of integrating forward-looking information into proactive management. The finding that a large global ensemble outperforms a small downscaled ensemble underscores that ensemble size and characterization of variability can outweigh spatial resolution for certain applications, especially when the ecological index is driven by basin-scale SST anomalies. This lowers barriers to adoption in regions lacking downscaling capacity. For management, HCI forecasts could be incorporated into the Risk Assessment and Mitigation Program (RAMP) to inform gear restrictions, closures, or season length adjustments ahead of high-risk periods, shifting responses from reactive to proactive. TOTAL forecasts could inform preparation for monitoring and potential closures in the Southern California Bight, where rare loggerhead occurrences make sampling programs difficult to mobilize on short notice. The study also outlines pros and cons of global and downscaled approaches (e.g., access, computational cost, biogeochemical availability), offering a decision-support framework (Fig. 5) to guide investments in regional downscaling. Overall, the work demonstrates that accessible global forecast systems can deliver management-relevant ecological predictions, encourages testing forecast performance before assuming resolution limitations, and provides a pathway to broaden ecological forecasting implementation across diverse regions.

This study reconfigures two operational marine management tools (HCI and TOTAL) into forecasting systems and shows they can deliver skillful predictions 0.5–11.5 months ahead, including during unprecedented climate extremes. Key contributions include: (1) demonstrating that readily available global SST forecasts can underpin accurate ecological forecasts without mandatory regional downscaling; (2) quantifying the critical role of ensemble size in forecast skill; and (3) providing practical guidance for when to invest in regional downscaling. These advances enable more proactive, climate-ready ecosystem-based management. Future directions include co-developing an operational HCI forecast with RAMP end-users; expanding tools to incorporate additional variables (including biogeochemical fields where downscaling may be necessary); testing and deploying similar forecast configurations in other coastal systems; and exploring strategies to combine regional downscaling with larger ensembles where feasible to maximize skill.

- Geographic scope is limited to the California Current Ecosystem, which exhibits relatively high SST predictability influenced by basin-scale dynamics; results may differ in regions with weaker large-scale controls or dominant fine-scale processes. - Variables: Only SST (and anomalies) were considered. Applications needing biogeochemical fields or subsurface structure may require regional downscaling and coupled models. - Downscaled forecasts covered a shorter period (1990–2010) and used only three ensemble members, constrained by computational/storage limits; this restricts comparison breadth with the global ensemble and may underrepresent the benefits of downscaling. - Threshold definitions (e.g., TOTAL closure percentiles) and categorical event framing can influence accuracy metrics and may need tailoring to specific management contexts. - Early onset periods of extreme events (e.g., early 2014) showed reduced forecast detection, indicating sensitivity during regime transitions.