Global seasonal forecasts of marine heatwaves

Abstract

Marine heatwaves (MHWs)-periods of exceptionally warm ocean temperature lasting weeks to years-are now widely recognized for their capacity to disrupt marine ecosystems^1-3. The substantial ecological and socioeconomic impacts of these extreme events present significant challenges to marine resource managers^4-7, who would benefit from forewarning of MHWs to facilitate proactive decision-making^8-11. However, despite extensive research into the physical drivers of MHWs^11,12, there has been no comprehensive global assessment of our ability to predict these events. Here we use a large multimodel ensemble of global climate forecasts^13,14 to develop and assess MHW forecasts that cover the world's oceans with lead times of up to a year. Using 30 years of retrospective forecasts, we show that the onset, intensity and duration of MHWs are often predictable, with skilful forecasts possible from 1 to 12 months in advance depending on region, season and the state of large-scale climate modes, such as the El Niño/Southern Oscillation. We discuss considerations for setting decision thresholds based on the probability that a MHW will occur, empowering stakeholders to take appropriate actions based on their risk profile. These results highlight the potential for operational MHW forecasts, analogous to forecasts of extreme weather phenomena, to promote climate resilience in global marine ecosystems.

Fig. 1. Skill of global MHW forecasts.

Maps indicate MHW forecast skill, as measured using the SEDI, for the 73-member ensemble of forecasts obtained from six global climate forecast systems for the period 1991–2020. SEDI scores range from −1 (no skill) to 1 (perfect skill). Scores above (below) zero, indicated by grey contours, indicate skill better (worse) than chance, and skill that is significantly better than random forecasts at the 95% confidence level is indicated by black contours. MHW forecasts were initialized every month, with lead times up to 11.5 months; a subset of lead times is shown here. a–d, 1.5 months (a), 3.5 months (b), 6.5 months (c) and 10.5 months (d). Areas with permanent or seasonal sea ice coverage are masked in white.

Fig. 2. Predicting the onset and persistence of MHWs.

a, SEDI for 3.5-month lead forecasts (as in Fig. 1). Example locations are indicated by coloured circles and are referred to in the text as Mediterranean Sea (red), Indo-Pacific (blue-green), Eastern Equatorial Pacific (gold), California Current System (pink), Gulf Stream (green) and Brazil Current (blue). b, Forecast MHW probability leading up to the initial appearance of observed MHWs. For each 1° × 1° grid cell, forecast probabilities for each lead time preceding the first month of observed MHWs are averaged across all events from 1991 to 2020. Coloured lines correspond to individual locations in a, whereas the grey line and shading indicate the global median and the 25th–75th, 10th–90th and 0–100th percentiles. For reference, a horizontal dashed line at 10% marks the base rate of MHW occurrence; probabilities higher than 10% indicate that MHW forecasts correctly predict elevated MHW likelihood from 0.5 to 11.5 months in advance (for example, for 30% probability, forecasts are indicating that the likelihood of a MHW occurring has tripled). c, Comparison of observed and predicted mean MHW duration (that is, on average how long MHWs last once established at a given location). Each dot represents the mean duration of all events in a 1° × 1° grid cell, with coloured markers corresponding to locations in a. The strong correlation (r = 0.83) shows that the global spatial pattern of mean MHW duration is reproduced well by forecasts. For temporal correlations of observed and predicted MHW durations at individual locations, see Extended Data Fig. 4.

Fig. 3. Influence of ENSO on MHW forecast skill.

a, Difference in 3.5-month lead forecast skill (SEDI) between periods when ENSO is in an active state and when it is in a neutral state. Active states include both positive and negative phases, defined here as the upper and lower quartiles of the oceanic Niño index (ONI), respectively. b, Time series of globally averaged 3.5-month lead forecast skill, with ENSO state (as measured by the ONI) indicated by the colours. Although 3.5-month lead forecasts are shown here, the patterns of enhanced or suppressed skill also hold for other lead times.

Fig. 4. Adjusting thresholds to support decision-making based on risk tolerance.

a, Observed MHW intensity (SST anomaly) shown as a function of MHW forecast probability threshold for 3.5-month lead forecasts in the Coral Triangle (orange) and Galapagos Islands (purple) regions. For a given threshold, SST anomalies are averaged over all times when the forecast probability was at or above that threshold. b, As in a, but for rates of false positives (solid lines) and false negatives (dashed lines). Note, a and b have the same x axis.

Extended Data Fig. 1. Forecast MHW probability varies with MHW intensity.

Maps show the mean 3.5-month lead forecast MHW probability associated with periods of a, no observed MHW (<90^th percentile of SST anomalies) and observed MHWs that are b, “weak” (90^th–95^th percentile of SST anomalies) or c, “strong”(>95^th percentile). Forecast probabilities higher (lower) than 10% indicate an elevated (reduced) likelihood of MHW occurrence. A positive relationship between MHW forecast probability and observed MHW strength is indicative of forecast skill.

Extended Data Fig. 2. Observed and predicted MHWs for sample locations.

a, Mean observed MHW intensity (the average SST anomaly during MHWs), with markers corresponding to locations in Fig. 2. b–g, Time series of 3.5-month lead forecast MHW probability (blue bars) and observed SST anomalies (black, with MHWs indicated in red). Panel letters correspond to locations shown in a.

Extended Data Fig. 3. MHW forecast skill as a function of season.

Maps show 3.5-month lead forecast skill, as measured by the SEDI, for forecasts initialized in each season: a, December-February, b, March-May, c, June-August, d, September-November.

Extended Data Fig. 4. Comparison of observed and predicted MHW duration.

Maps show the correlation (Pearson correlation coefficient) between observed and predicted MHW duration at each location.

Extended Data Fig. 5. Temperature trends can influence MHW frequency and forecast skill metrics.

a, Time series show the global frequency of MHW occurrence (percentage of the ice-free global ocean in a MHW state at each monthly time step) calculated from SST anomalies with linear 1991–2020 trends removed (solid lines) and with trends retained (dashed lines). b, Time series of 3.5-month lead forecast skill metrics (Symmetrical Extremal Dependence Index, SEDI; Brier Skill Score, BSS; and Forecast Accuracy, FA). Skill metrics are calculated using globally aggregated forecasts each month (for example, forecast accuracy for a given month is the fraction of the ice-free global ocean for which the MHW state that month was corrected predicted). c, As in b, but for individual components of the 2x2 contingency table.

Extended Data Fig. 6. MHW forecast skill as a function of MHW duration for forecasts based on daily and monthly SST data.

For locations in a (which are the same as those in Fig. 2 and Extended Data Fig. 2), 3.5-month lead MHW forecast skill (SEDI) is shown as a function of mean MHW duration calculated from b, daily and c, monthly CCSM4 output.

Extended Data Fig. 7. Comparison of lead time dependent MHW forecast skill for forecasts based on daily and monthly SST data.

a–f, For locations in Extended Data Fig. 6, forecast skill (SEDI) is shown as a function of lead time calculated from daily (lines) and monthly (circles) CCSM4 output. Daily skill is smoothed with a 30-day running mean for plotting.

Extended Data Fig. 8. Comparison of MHW forecast skill metrics.

Maps show a, SEDI, b, Brier Skill Score (BSS), and c, forecast accuracy (FA) for 3.5-month lead MHW forecasts. Perfect forecasts would yield a score of one for all three metrics, while the skill expected from random forecasts is 0 for SEDI and BSS, and 0.82 for FA (indicated by gray contours).