Skillful multiyear to decadal predictions of sea level in the North Atlantic Ocean and U.S. East Coast

Sea level change, driven by a complex interplay of factors, presents a significant threat to coastal communities worldwide. The consequences of rising sea levels, including increased flooding and erosion, are particularly severe along densely populated coastlines like the U.S. East Coast. While short-term sea level fluctuations are influenced by factors such as tides, waves, and storms, longer-term variations (seasonal to decadal) are primarily linked to large-scale ocean dynamics and climate variability. Anthropogenic warming contributes to a secular rise in global sea levels through thermal expansion and melting land ice. In the North Atlantic, the wind, Gulf Stream, and Atlantic Meridional Overturning Circulation (AMOC) significantly influence sea levels on interannual and longer timescales. The U.S. East Coast is particularly vulnerable, experiencing sea level rise rates exceeding the global average due to a combination of global mean sea level rise and dynamic changes associated with AMOC weakening. This study addresses the critical need for accurate seasonal to decadal sea level forecasts to aid in coastal risk mitigation and adaptation strategies. Previous research has explored seasonal sea level prediction, finding higher skill along the U.S. West Coast due to the strong influence of El Niño-Southern Oscillation (ENSO). However, the predictability of sea level beyond seasonal timescales remains under-investigated, despite its crucial role in long-term coastal planning. This study uses diagnostic analysis and initialized decadal hindcasts to investigate multiyear to decadal sea level predictability and prediction in the North Atlantic and along the U.S. East Coast, leveraging the established skill in predicting AMOC and North Atlantic upper ocean heat content on decadal timescales.

Existing literature highlights the significant societal impact of sea level rise on coastal communities (Fitzgerald et al., 2008; Kirwan & Megonigal, 2013; Kirwan et al., 2016; Woodruff et al., 2013). Studies have explored the link between storm surge flooding and rising sea levels (Rahmstorf, 2017; Yin et al., 2020), emphasizing the need for improved forecasting capabilities. Research on the U.S. East Coast specifically has demonstrated the disproportionately high rates of sea level rise compared to global averages (Ezer & Atkinson, 2014; Ezer, 2015; Sallenger et al., 2012), and attributed this to the interplay of global warming and dynamic processes linked to the AMOC and Gulf Stream (Little et al., 2017; Piecuch et al., 2019; Yin & Goddard, 2013; Yin et al., 2009; Ezer et al., 2013). Previous work has examined seasonal sea level prediction (Long et al., 2021; Long et al., 2023), but multiyear to decadal prediction remains largely unaddressed. The demonstrated skill in decadal prediction of the AMOC and related ocean heat content (Robson et al., 2012; Smith et al., 2007; Yang et al., 2013; Yeager et al., 2012) suggests the potential for skillful multiyear to decadal sea level forecasts.

This study employs a multi-faceted approach combining observational data, climate model simulations, and statistical analysis. First, a diagnostic analysis using the Average Predictability Time (APT) method (DelSole & Tippett, 2009a, 2009b) was applied to a preindustrial control simulation from the Geophysical Fluid Dynamics Laboratory's (GFDL) SPEAR_LO model (Delworth et al., 2020). The APT method identifies the most predictable components of sea level by integrating predictability over all lead times. This analysis, performed on a perfect model simulation, reveals the inherent predictability of North Atlantic sea level components, providing a benchmark for comparison with real-world prediction skills. The identified components were examined for their relationship with AMOC variability through lagged regression analysis. Second, the study evaluates the predictability of these identified components using initialized decadal hindcasts from the SPEAR_LO model. These hindcasts were initialized from an observationally constrained reanalysis (Yang et al., 2021), which utilizes observations to constrain the atmospheric and sea surface temperature components, resulting in a more realistic representation of AMOC evolution. The APT method was again applied to the hindcasts, with prediction skill verified against satellite observations of sea surface height. The externally forced sea level variability was derived from SPEAR_LO large ensemble simulations (Zhang et al., 2022) using the signal-to-noise maximizing EOF analysis (Ting et al., 2009, 2011). This allowed the separation of forced and internal variability in sea level predictions. Finally, the study focused on the U.S. East Coast, using coastal tide gauge (TG) observations to verify the initialized decadal hindcasts. The U.S. East Coast was divided into three regimes (Northeast, Mid-Atlantic, and Southeast) based on the spatial coherence of sea level variability. Composite sea level time series were created for each regime, allowing for a regional assessment of prediction skill. The linear trend was removed from the time series before analyzing multidecadal variability to isolate the contribution of internal climate variability. The impact of external forcing was examined by comparing the total sea level, the detrended sea level (after removing the linear trend) and the internal sea level component (after removing both the linear trend and the forced signal).

The APT analysis of the SPEAR_LO control simulation identified three significant, predictable sea level components in the North Atlantic. The most predictable component (APT1) exhibited a basin-wide upward trend, primarily attributable to anthropogenic warming and predictable for up to a decade. This finding is supported by the high correlation (up to 0.75) between the APT1 spatial pattern and the externally forced sea level pattern derived from SPEAR_LO large ensemble simulations. The second (APT2) and third (APT3) components were associated with AMOC variability. APT2, characterized by sea level anomalies over the western subpolar gyre extending to the U.S. East Coast, showed predictability up to 7 years in the control simulation and around 5 years in the initialized hindcasts. This component is linked to the mature phase of the dominant AMOC mode of internal variability. APT3, characterized by a tripole pattern, had a predictability of approximately 4 years in the control simulation and about 3 years in the initialized hindcasts, linked to the AMOC transition between phases. In initialized decadal hindcasts, the skill of predicting APT1 remained high (up to 10 years), while APT2 and APT3 showed slightly reduced skill (around 5 and 3 years respectively), likely due to model biases and initialization uncertainties. Analysis of U.S. East Coast tide gauge data revealed a strong correlation between observed sea level and the predicted components. The trend-like sea level component, primarily driven by external radiative forcing, showed increasing predictability from south to north along the coast. The Northeast regime exhibited high skill in predicting the detrended sea level (around 4 years), largely attributed to the influence of the AMOC mature phase. The Mid-Atlantic regime displayed reduced skill due to noisier, high-frequency variability. The Southeast regime showed a predictability of around 3 years for the detrended sea level, linked to the AMOC transition state. Ten-year sea level predictions, initialized between 1995-2003, accurately captured the transition to multiannual high sea level events after 2005, highlighting the role of AMOC in sea level prediction.

This study's findings demonstrate the potential for skillful multiyear to decadal sea level prediction in the North Atlantic and along the U.S. East Coast. The identification of both a long-term trend (driven by external forcing) and shorter-term fluctuations (driven by AMOC variability) provides a comprehensive understanding of sea level predictability. The skill in predicting the long-term trend emphasizes the importance of incorporating external radiative forcing into sea level prediction models. The ability to predict AMOC-related variability suggests that improved understanding and modeling of AMOC dynamics are crucial for enhancing sea level forecast accuracy. The regional variations in prediction skill along the U.S. East Coast underscore the importance of considering geographical factors when developing and applying prediction models. The successful prediction of the transition to multiannual high sea level events after 2005, based on initialization during years with a mature positive phase of AMOC, further emphasizes the importance of AMOC for prediction skill. These results have significant implications for coastal risk management, enabling more informed decision-making regarding infrastructure planning, adaptation strategies, and emergency preparedness. Future research should focus on improving model representation of relevant processes (land ice melt, river discharges, and coastal processes), incorporating higher-resolution models, and evaluating predictability under future climate scenarios.

This study demonstrates the skillful prediction of sea level along the U.S. East Coast on multiyear to decadal timescales, driven by both external radiative forcing and AMOC variability. The most predictable component is the basin-wide upward trend caused by anthropogenic warming, while the AMOC's mature and transition states influence shorter-term predictability. Model biases and initialization uncertainties reduce prediction skill compared to perfect model simulations. Regional variations in skill exist, with the Northeast regime showing the strongest skill in predicting the detrended component due to AMOC influence. These findings highlight the importance of AMOC for sea level prediction and have significant societal implications for coastal risk management. Future work should address model limitations, explore the impact of future climate change on predictability, and improve model resolution for better prediction accuracy.

The study acknowledges limitations in the SPEAR_LO model, including the absence of a land ice component, underestimation of river discharges, and coarse resolution affecting the representation of the Gulf Stream and coastal processes. The model does not incorporate factors like land subsidence and uplift, which also influence relative sea level change. These limitations might affect the prediction skill, particularly at interannual timescales. Additionally, while the study identifies key drivers of sea level predictability, the interaction between different factors and non-linear processes may not be fully captured. Future research should address these limitations by using higher-resolution models with more comprehensive representations of the physical processes.