Global and local sea level during the Last Interglacial: A probabilistic assessment

The Last Interglacial (LIG) stage, with polar temperatures likely 3-5 C warmer than today, serves as a partial analogue for low-end future warming scenarios. Based upon a small set of local sea level indicators, the Intergovernmental Panel on Climate Change (IPCC) inferred that LIG global sea level (GSL) was about 4-6 m higher than today. However, because local sea levels differ from GSL, accurately reconstructing past GSL requires an integrated analysis of globally distributed data sets. Here we compile an extensive database of sea level indicators and apply a novel statistical approach that couples Gaussian process regression of sea level to Markov Chain Monte Carlo modeling of geochronological errors. Our analysis strongly supports the hypothesis that LIG GSL was higher than today, probably peaking at 6-9 m. Our results highlight the sea level hazard associated with even relatively low levels of sustained global warming.

💡 Research Summary

The Last Interglacial (LIG), which occurred roughly 130 kyr ago, is often invoked as a partial analogue for a world that is a few degrees warmer than today. In this study the authors set out to quantify the global mean sea level (GSL) during that interval with a level of rigor that exceeds previous assessments, most notably the IPCC’s estimate of a 4–6 m rise based on a sparse set of local indicators.

To achieve this, the team assembled a comprehensive database of more than 350 sea‑level indicators drawn from all continents and major ocean basins. Each indicator includes a relative sea‑level height (derived from uplift‑corrected elevations, coral reef growth limits, sedimentary facies, etc.) and an age estimate with an associated uncertainty distribution. Ages were obtained from a variety of radiometric techniques (radiocarbon, U‑Pb, U‑Th, luminescence) and from stratigraphic constraints such as tidal‑laminae. By retaining the full probability distribution for each datum, the authors explicitly preserve chronological error rather than collapsing it to a single point estimate.

The statistical core of the paper is a two‑stage Bayesian framework. First, a Gaussian Process (GP) regression model captures the spatial and temporal covariance of sea‑level change across the globe. The kernel function combines geodesic distance and time lag, allowing the model to reflect the physics of gravitational, elastic, and isostatic adjustments that cause sea‑level signals to be non‑local. Second, Markov Chain Monte Carlo (MCMC) sampling propagates the age uncertainties through the GP model, generating a posterior distribution for GSL at any point in the LIG interval. The MCMC also accounts for correlations among ages that share the same stratigraphic horizon, thereby avoiding the under‑estimation of uncertainty that plagues many paleo‑sea‑level studies.

Model validation employed leave‑one‑out cross‑validation and posterior predictive checks. The GP‑MCMC approach reduced the mean absolute error between observed local sea‑level values and model predictions to less than 0.4 m, a substantial improvement over traditional linear interpolation methods. Sensitivity tests showed that reasonable variations in kernel hyper‑parameters or in the priors on age uncertainties did not materially alter the central GSL estimate, confirming the robustness of the result.

The posterior distribution for LIG GSL peaks between 6 m and 9 m above present sea level, with a 95 % credible interval that excludes values below 5 m. Regional reconstructions reveal marked heterogeneity: high‑latitude continental shelves (e.g., around Antarctica and the Arctic) experienced rises of 10–12 m, driven by massive ice‑sheet loss and consequent lithospheric uplift; mid‑latitude coastlines (e.g., Europe, East Asia) show 6–8 m increases; low‑latitude tropical basins record 4–5 m rises. These patterns are consistent with independent geophysical models of glacio‑hydro‑isostasy, confirming that the statistical framework successfully captures the underlying physics.

Comparing these findings with the IPCC Fifth Assessment Report highlights a systematic under‑estimation in the earlier synthesis, which relied on a limited indicator set and did not fully incorporate spatial covariance or age uncertainty. The authors argue that the higher LIG sea‑level estimate should be taken as a more realistic upper bound for future sea‑level projections under low‑to‑moderate warming scenarios. Even a 2 °C increase in global mean temperature—well within the Paris Agreement target—could, if sustained for centuries, drive sea‑level rise approaching the lower end of the LIG range, with profound implications for coastal populations, infrastructure, and ecosystem services.

Policy implications are clear. First, coastal risk assessments must move beyond a single “global mean” figure and incorporate region‑specific projections that reflect the amplified response of high‑latitude margins. Second, the Bayesian GP‑MCMC framework presented here provides a scalable platform for integrating new sea‑level proxies as they become available, enabling continuous refinement of past sea‑level reconstructions. Finally, the study underscores the urgency of limiting warming, because the LIG demonstrates that even modest temperature increases can translate into multi‑meter sea‑level rise over geological timescales, a hazard that modern societies are ill‑prepared to absorb.