Communicating Uncertainty in Coastal Retreat Forecasting

coastal erosion
sea level rise
uncertainty quantification
applied AI
climate adaptation
scientific ethics
Author
Published

October 1, 2026

Presentation
Video
Extra

Abstract

Coastal retreat forecasting has become a legal requirement in France. Communes listed under the Loi Climat et Resilience must produce erosion exposure maps at thirty years and at thirty to one hundred years, horizons at which scientific uncertainty about shoreline evolution is at its greatest. The regulatory demand is for maps; the scientific reality is a set of ranges conditional on emission scenarios, on model class and on the accuracy of the underlying satellite observations. This paper argues that the resolution to that tension lies in treating uncertainty as a primary model output rather than as a caveat attached to a point estimate. We describe the design of an applied AI forecasting pipeline built on that principle, from satellite-derived shoreline observations through scenario-conditioned outputs reported as ranges with stated confidence, and we report our own exploratory results plainly, including their failure to outperform simple baselines. Drawing on the case of Le Signal at Soulac-sur-Mer, we discuss what an AI system should and should not claim when its outputs inform public decisions about coastal risk.

1 Introduction

Today, 680 million people live in low-elevation coastal zones worldwide. This number is expected to exceed one billion by 2050 (Oppenheimer et al. 2019).

France has already begun legislating on this subject. The Loi Climat et Resilience, in force since August 22nd, 2021, established a dedicated legal regime for shoreline retreat (Loi no 2021-1104 2021). Under the French urbanism code, listed commune must produce local erosion exposure maps at two mandatory horizons: 30 years, and 30 to 100 years. These regulatory horizons are the ones any coastal retreat forecasting tool must serve (Articles l. 121-22-1 à l. 121-22-12 Du Code de l’urbanisme 2026, Article l.321-15 Du Code de l’environnement 2022). Inclusion on this list is voluntary: communes deliberate to request it, and mapping becomes mandatory once they join. The list has grown accordingly from 121 communes in April 2022 (Décret no 2022-750 2022) to 371 in February 2026 (Décret no 2026-95 2026), reflecting rising municipal engagement towards adaptation to the impacts of climate change on coastal areas.

In France, the stakes have been quantified within two horizons by the Cerema (Hedou et al. 2024). By 2050, under a trend scenario assuming current defense structures are maintained, 5,200 dwellings and 1,400 business premises are projected to be at risk, at an estimated cost of €1.1 to 1.2 billion. Beyond 2050, the same study includes a deliberately unfavorable 2100 scenario in which all existing coastal defenses are assumed lost. Under this inaction hypothesis, 450,000 dwellings could be affected.

French law requires forecasts precisely at the horizons, decades out, where scientific uncertainty is the greatest. Legal obligations and epistemic limits pull in opposite directions. The question this paper asks is therefore not whether artificial intelligence can forecast shoreline retreat, but what an AI system should be able to claim when it does.

The need, moreover, is not purely predictive but also adaptive. Forecasting outputs must support decisions including evaluation, planning and financing of coastal risk, rather than simply producing numbers.

Section 2 traces the physical chain from sea level rise to shoreline retreat; Section 3 describes our data-to-output forecasting pipeline; Section 4 grounds the discussion in the case of Le Signal; Section 5 develops our central argument for treating uncertainty as a first-class output; Section 6 reports work still in progress; and Section 7 discusses the implications for AI-assisted climate decision-making.

2 Scientific background

Sea level and shorelines have been measurable globally by satellite since 1984. That allowed scientists to discover that about 24% of sandy beaches are eroding (Luijendijk et al. 2018). It is due to sea level rise. On a geological level, shorelines are the result of two phenomena: erosion and accretion (Mentaschi et al. 2018). Erosion consists in the natural process of degradation, alteration and transportation of rocks or soil at the surface of water by water, wind or gravity; while accretion is a body’s volume growth via the progressive addition of matter on its surface. In meteorology and geography, accretion is the addition of sediments alongside littorals or by rivers. Sea level rise is the underlying forcing impacting coastal retreat. Altimeters allowed scientists to detect the acceleration of sea level rise. The seas rise at a rate of approximately 3 ± 0.4 mm/y since 1993. The rising phenomenon itself is accelerating at a rate of 0.084 ± 0.025 mm/y ² (Nerem et al. 2018). Although it is possible to estimate sea level rise, the projections are conditional on IPCC SSPs, which produce medians and likelihood ranges, not point predictions (Fox-Kemper et al. 2021). At a local scale, sea level rise varies considerably. For instance, relative and absolute sea level rise can differ due to land motion and subsidence shift local exposure. This justifies local scale differences (Nicholls et al. 2021). It is possible to consider sea level rise to be relative at a local scale and absolute at a global scale. Land motion can be horizontal or vertical. In the case of subsidence, the movement is vertical towards the center of the Earth. According to the Bruun Rule, emitted in 1962, there is a historic empirical relationship between sea level rise and coastal retreat. It is a foundational two-dimensional coastal model that predicts rising sea levels cause sandy beaches to retreat landward and upward while maintaining an equilibrium profile. Material eroded from the upper beach moves offshore and raises the nearshore bottom by an equal volume (Bruun 1962). This theory faced immediate critique as the assumptions were known to be false and were never adequately validated. It was finally disproved in field studies (Cooper and Pilkey 2004). State-of-the-art research identified sea-level rise and changing wave patterns as primary drivers of shoreline retreat on open sandy coasts. It is important to acknowledge its limitations, as it is reminded that local-scale, multi-model assessments are required to manage uncertainty in predicting future coastal changes (Ranasinghe 2016).

By 2100, it is projected that up to 50% of world’s sandy beaches could be gone (Vousdoukas et al. 2020). These results are criticized, the researchers arguing that the projection rests on flawed assumptions about coastal response (Cooper et al. 2020). Two teams worked based on the same literature but still got wildly divergent headlines. That is not a failure of science, but a normal state for a young field. As the physical link itself is contested, data-driven models inherit the uncertainty without resolving it.

3 Methodological approach

Our forecasting approach can be described as three layers: a data layer, a forecasting layer, and an output layer. Each is disclosed to a different level of detail: the data sources are named precisely, the forecasting methods are described only by model class, and the output layer, where scenario-conditioned ranges replace single values, is this paper’s central design commitment.

Satellite-derived shorelines are the field’s standard observational input, with CoastSat the reference open-source toolkit operating on the public Landsat and Sentinel archive (Vos, Splinter, et al. 2019). This observational layer has a known accuracy floor: satellite-derived shorelines carry an RMSE of roughly 7 to 14 meters against in-situ surveys, and everything downstream is bound by it (Vos, Harley, et al. 2019). The extraction algorithm itself is a further source of uncertainty: benchmarking of CoastSat, SHOREX, ShorelineMonitor, CASSIE and HighTide-SDS across four sites, including Truc Vert in France, documents systematic biases between landward and seaward algorithms (Vos et al. 2023). In other words, extraction is the result of a modelling decision, not of neutral preprocessing, and it should be treated as such when uncertainty is propagated through the pipeline. Climate forcing enters the same layer through IPCC SSP scenarios, integrated as conditioning inputs rather than as a single assumed future (Fox-Kemper et al. 2021). A third input family, coastal physics (waves, sediment dynamics, bathymetry), is described here only at a generic level, consistent with the level of disclosure adopted throughout this section.

The forecasting layer spans a range of model families, from empirical relationships such as the Bruun rule, through data-driven statistical approaches, to machine-learning architectures. Rather than describing any single architecture, we position our approach against the field’s own benchmark. ShoreShop 2.0 was a blind benchmark in which 34 groups submitted predictions at short (five-year) and medium (fifty-year) horizons against withheld observations at an anonymized site; across these model families, the best submissions reached an accuracy of about 10 meters, comparable to the accuracy of the satellite data itself (Mao et al. 2025). Dating from ShoreShop 1.0, blind benchmarking has become the field’s established methodological norm (Montaño et al. 2020). Separately, research has shown that satellite-calibrated models can match in-situ-calibrated ones, which justifies a satellite-first data strategy for regions, such as ours, where dense in-situ monitoring is unavailable (Vitousek et al. 2023).

The output layer is where this paper’s design commitment is made explicit: our outputs are reported as scenario-conditioned ranges with stated confidence, rather than as single values. This is not an exotic requirement. The French institutional methodology for characterizing the shoreline retreat hazard already includes an explicit step to determine uncertainty on future shoreline positions, alongside transect analysis performed at 100-meter spacing using the Digital Shoreline Analysis System (DSAS), a tool that computes rates of shoreline change from historical shoreline positions (Bernon et al. 2016). Uncertainty characterization is already mandatory practice for the public expert body, so an artificial intelligence tool that dropped it would be a regression rather than an innovation.

4 Illustrative anchor: Le Signal, Soulac-sur-Mer

Soulac-sur-Mer illustrates the problem this paper is concerned with, but it is not a model output, and it should not be read as one. It is, rather, a case in which the physical process was already well understood and well documented, and yet the institutional response failed regardless. Built in 1967, it is made of four storeys, roughly 200 meters from the coast at construction. A second building followed in 1970 (Simian 2018). Together they comprised 78 dwellings, held by 75 co-owners (Simian et al. 2018). The winter 2013-14 storms resulted in strong erosion of the littoral; the building then being situated about 10 m from the shore. On January 24th, 2014, the building was evacuated by municipal order (Simian 2018). It remained empty for nearly a decade before being demolished in 2023.

On Gironde’s sandy coasts, the regional rate of erosion is worth -2.50 m/y. It is -1.70 m/y in the Landes. It was then revised upward from an earlier estimate of 1 to 3 m per year (Bernon et al. 2016). The local rate in Soulac-sur-Mer is worth 4-8 m/y. This sector is described as of the most eroded in Europe (Ministère de la Transition écologique).

None of this reflects a failure of forecasting in any meaningful sense. The retreat at Soulac was slow, physically continuous, and directly observable; it has been present in satellite records since 1984. What failed was not epistemic but institutional. The compensation framework set in place by the French Government excluded the marine erosion hazard. It was not cited among perils in Art. L.561-1 of the Code de l’environnement. Unlike Criel-sur-Mer cliff case, it was therefore not included as it was not assimilable to ground movement (Assemblée nationale 2014). During the State’s Counsil of August 16th, 2018, the ineligibility for the Fonds Barnier was confirmed. A budget of EUR 7 million was later admitted via the 2018 budget amendment (Simian 2018).

The building was evacuated by public authority; no expropriation procedure was ever triggered (Assemblée nationale). The Observatoire de la Côte Aquitaine published its retreat projections in 2011 for horizons 2020 and 2040 (Aubie et al. 2011). Winter 2013-14 alone exceeded the 2040 projection. In 2016, an update was commissioned as this winter’s storms invalidate the work made earlier (Bernon et al. 2016). The lesson Le Signal offers is not that the 2011 model was wrong in some simple sense; slow coastal retreat is, by its nature, difficult to bound precisely decades in advance. The deeper failure was one of format: the projection was published as a fixed date on a fixed horizon, with no accompanying language for how confidently that horizon should be trusted or how quickly it might need revision. Better forecasts, in this light, only matter if the institutions receiving them are structured to act on probabilistic information rather than on single dates.

5 Uncertainty as a first-class output

There are two kinds of uncertainties. One is aleatoric, which implies an irreducible data noise. The second is epistemic, which can be represented as a reducible lack of knowledge (Hüllermeier and Waegeman 2021).

Uncertainty does not, however, originate solely inside the model. It enters at data collection and persists through deployment, so a complete account has to follow it across the entire pipeline rather than confine it to a single estimation step (Gawlikowski et al. 2023). The methods available for quantifying it are by now extensive, spanning Bayesian approximations and ensemble techniques (Abdar et al. 2021). What matters for our purposes is that the distinction between the two families is operational rather than academic: an epistemic gap can in principle be narrowed with more or better data, whereas aleatoric noise cannot. Telling a commune which of the two dominates its exposure estimate changes what it should do about it: commission further survey work in the first case, plan around an irreducible margin in the second.

The specific danger for an AI system is miscalibration. Modern neural networks are systematically overconfident: the confidence they report exceeds the accuracy they achieve (Guo et al. 2017). Later work nuances this picture across architectures, finding some more recent model families better calibrated than the original result suggested, but it does not dissolve the underlying concern (Minderer et al. 2021). The danger, put plainly, is not that the model is wrong; models are wrong routinely and that is expected. The danger is that a wrong model looks right. Fluency and numerical precision are read as confidence by non-specialists, and the audience for shoreline retreat forecasts, whether municipal councils, planning departments or insurers, is non-specialist by construction. The concern is not particular to our application: reviews of artificial intelligence in climate science identify interpretability, generalisation and uncertainty quantification as the conditions of trustworthiness, and note the difficulty non-experts face in interpreting complex model outputs (Camps-Valls et al. 2025).

The technical means are available and need only brief mention here. Deep ensembles provide a practical and scalable baseline for predictive uncertainty (Lakshminarayanan et al. 2017). Conformal prediction offers distribution-free intervals with finite-sample coverage guarantees, applicable to an already-trained model (Angelopoulos and Bates 2023). Neither is exotic. The binding constraint on publishing uncertainty is therefore not technical availability; it is willingness to publish a range where the recipient expects a number.

Nor does the vocabulary need to be invented. The IPCC has maintained a calibrated language of confidence and likelihood, applied consistently across successive assessment reports (Mastrandrea et al. 2010). Our recommendation is accordingly conservative: AI climate tools should adopt an established calibrated vocabulary rather than devise per-product confidence scales that no user can compare across systems. How that vocabulary is deployed matters as well, and here the communication literature distinguishes direct uncertainty about a quantity from indirect uncertainty about the quality of the evidence behind it, and catalogues nine distinct ways of expressing the former (Bles et al. 2019).

The standard objection to all of this is that publishing ranges costs credibility. The evidence does not support it. In a survey experiment with more than ten thousand participants, presenting a statistic as a numerical range rather than as a point value produced at most a small reduction in the perceived trustworthiness of the number itself, and no measurable effect on trust in the source providing it (Kerr et al. 2023). The objection to publishing ranges is, in the end, commercial rather than scientific, and the evidence does not sustain it even on commercial terms.

6 Work in progress and open challenges

Our own work on this problem is exploratory, and we report it as such. We have applied classical statistical time-series approaches to satellite-derived shoreline series at two French test sectors. The results are not yet conclusive, and we make no claim that they are.

This outcome is expected rather than anomalous, and the reason is worth stating precisely. The observational input carries an RMSE of roughly 7 to 14 meters against in-situ survey (Vos, Harley, et al. 2019). At the ShoreShop 2.0 blind benchmark, the best submissions across all model families reached an accuracy of about 10 meters (Mao et al. 2025), comparable to the noise in the observations against which they were scored. A forecasting claim below 10 meters is therefore not meaningfully verifiable with currently available satellite data. Declining to make one is principled caution, not a failure to report.

The broader methodological literature supports that caution. A systematic survey of machine-learning-based science documented data leakage across seventeen disciplines and hundreds of affected papers, frequently producing conclusions far more optimistic than the underlying evidence warranted (Kapoor and Narayanan 2023). On long-horizon forecasting benchmarks specifically, simple linear baselines have been shown to match or outperform Transformer architectures (Zeng et al. 2023), and large-scale forecasting competitions continue to find statistical and hybrid methods competitive with machine learning across a wide range of tasks (Makridakis et al. 2020). Reporting non-outperformance is consistent with what this literature would lead one to expect. It is also, more to the point, the epistemic practice this paper argues for: a pipeline designed to report uncertainty honestly cannot exempt its own results from that standard.

Three challenges remain open. The first is data quality: satellite-derived shoreline series carry both label noise and temporal sparsity, and the extraction algorithm contributes systematic bias of its own (Vos, Splinter, et al. 2019). The second is non-stationarity. Chronic retreat and storm-driven episodes are not the same signal: a single winter can add a retreat of up to 20 meters on top of the chronic trend (Bernon et al. 2016), which means a model fitted on mean annual rates will misrepresent precisely the events that trigger institutional response. The third is transferability. A model calibrated on open sandy coasts cannot be assumed to transfer to cliff or estuarine settings without revalidation, and the limits of that transfer are not currently well characterized (Ranasinghe 2016).

7 Discussion

The preceding sections argue for a change in what a forecasting system outputs. This one argues for the corresponding change in what a decision-maker asks of it. Under deep uncertainty, where the parties to a decision cannot agree on the external context of the system, on how it works, or on which outcomes matter, the established response is not to optimise against a single best-estimate forecast, but to seek strategies that perform acceptably across many plausible futures (Marchau et al. 2019). Robust Decision Making is one well-developed instantiation, treating models as generators of cases for exploring assumptions rather than as predictors (Lempert 2019). The reframing matters because it relocates confidence: confidence attaches to a decision, not to a prediction. This, in our view, is what makes probabilistic outputs usable rather than merely honest. A range is an obstacle to a decision procedure that demands a single number; it is an input to one designed to test robustness.

Dynamic Adaptive Policy Pathways develops the same logic over time, sequencing actions around the tipping points at which a given measure ceases to be effective rather than committing to one forecast at the outset (Haasnoot et al. 2013). A review of pathways approaches across applied cases finds them consistently framed as sequences of actions to be implemented progressively as conditions develop (Werners et al. 2021). The link to our output design is direct. Scenario-conditioned ranges combined with stated thresholds map onto a pathways structure, because both are defined by conditions rather than by dates. A single-date output does not map onto it at all; it can only be complied with or ignored.

The cost of false confidence is not hypothetical. The maladaptation literature documents that measures taken in the name of adaptation can lock in detrimental long-term effects on the territories they were meant to protect, and argues that anticipating this risk belongs at the top of the planning agenda (Magnan et al. 2016). Acting on a confident but wrong projection can therefore be worse than acting on an acknowledged uncertain one. Le Signal returns here: the 2011 projection to horizons 2020 and 2040 was overtaken by a single winter, and its format offered no way to register that it had been.

A regulatory horizon is worth flagging as well, though cautiously. Regulation (EU) 2024/1689 establishes transparency and risk-management obligations for artificial intelligence systems (Regulation (EU) 2024/1689 Laying down Harmonised Rules on Artificial Intelligence (Artificial Intelligence Act) 2024). How those obligations apply to decision support supplied to public authorities for territorial planning is, to our knowledge, not yet settled, and we raise it as an open question rather than as a classification claim. What can be said is that a system reporting calibrated ranges is better positioned to meet a transparency expectation than one reporting a bare number.

8 Conclusion

The tension this paper began with does not resolve. French law requires exposure maps at thirty years and at thirty to one hundred years, which is to say precisely at the horizons where scientific uncertainty about shoreline evolution is largest. No improvement in modelling removes that.

What can change is what the model claims. Our position is that uncertainty should be treated as a primary output rather than as a caveat appended to a number: scenario-conditioning in place of deterministic projection and, wherever a forecast is less reliable, an explicit statement to that effect, quantified rather than qualified. The empirical case for doing so is stronger than is generally assumed, since communicating ranges rather than point values does not cost credibility (Kerr et al. 2023).

The title of this paper asks what an AI should say. The answer we arrive at is: less than it could. A system capable of emitting a precise date for a precise parcel should decline to do so when the data cannot support the claim. That restraint is a design requirement, not a limitation on the design.

9 Declarations

Data availability. All input data are publicly available. Satellite-derived shoreline positions were extracted using CoastSat, an open-source toolkit, from the public Landsat and Sentinel archives accessed through Google Earth Engine. Wave and atmospheric forcing variables were obtained from the ERA5 reanalysis produced by ECMWF, retrieved via the Copernicus Climate Data Store API. The derived shoreline and forcing series generated for this study are available from the corresponding author on reasonable request.

Code availability. The code developed for this study is confidential and is not publicly available.

Funding. This work received no external funding.

Conflicts of interest. The authors declare no competing interests.

Use of AI tools. The authors used Anthropic’s Claude to assist with drafting and language editing. The authors reviewed and edited all resulting text and take full responsibility for the content of this publication.

Acknowledgements. The authors thank Ayoub Shaït, Simon Chauvel and Redouane El Badaoui for their contributions to the data work underpinning this study.

References

Abdar, Moloud, Farzad Pourpanah, Sadiq Hussain, et al. 2021. “A Review of Uncertainty Quantification in Deep Learning: Techniques, Applications and Challenges.” Information Fusion 76: 243–97. https://doi.org/10.1016/j.inffus.2021.05.008.
Angelopoulos, Anastasios N., and Stephen Bates. 2023. “Conformal Prediction: A Gentle Introduction.” Foundations and Trends in Machine Learning 16 (4): 494–591. https://doi.org/10.1561/2200000101.
Article l.321-15 Du Code de l’environnement (2022).
Articles l. 121-22-1 à l. 121-22-12 Du Code de l’urbanisme (2026).
Assemblée nationale. 2014. Question Écrite no 29717.
Aubie, S., C. Mallet, and J. Favennec. 2011. Caractérisation de l’aléa Érosion (2020–2040) de La Côte Aquitaine Dans Le Cadre de l’étude Stratégique de Gestion Du Trait de Côte : Observatoire de La Côte Aquitaine : Rapport Final. BRGM.
Bernon, Nicolas, Christophe Mallet, and R. Belon. 2016. Caractérisation de l’aléa Recul Du Trait de Côte Sur Le Littoral de La Côte Aquitaine Aux Horizons 2025 Et 2050. BRGM.
Bles, Anne van der, Sander van der Linden, Alexandra Freeman, et al. 2019. “Communicating Uncertainty about Facts, Numbers and Science.” Royal Society Open Science 6 (5): 181870. https://doi.org/10.1098/rsos.181870.
Bruun, Per. 1962. “Sea-Level Rise as a Cause of Shore Erosion.” Journal of the Waterways and Harbors Division 88: 117–30.
Camps-Valls, Gustau, Miguel Ángel Fernández-Torres, K.-H. Cohrs, et al. 2025. “Artificial Intelligence for Modeling and Understanding Extreme Weather and Climate Events.” Nature Communications 16 (1): 1919. https://doi.org/10.1038/s41467-025-56573-8.
Cooper, Andrew, Gerd Masselink, Giovanni Coco, and Andrew Short. 2020. Sandy Beaches Can Survive Sea-Level Rise. https://doi.org/10.31223/osf.io/4md6e.
Cooper, J. G., and Orrin H. Pilkey. 2004. “Sea-Level Rise and Shoreline Retreat: Time to Abandon the Bruun Rule.” Global and Planetary Change, ahead of print. https://doi.org/10.1016/j.gloplacha.2004.07.001.
Décret no 2022-750 (2022).
Décret no 2026-95 (2026).
Fox-Kemper, Baylor, Helene T. Hewitt, Guðfinna Aðalgeirsdóttir, et al. 2021. Ocean, Cryosphere and Sea Level Change Supplementary Material.
Gawlikowski, Jakub, Cedric Njieutcheu Tassi, Mohsen Ali, et al. 2023. “A Survey of Uncertainty in Deep Neural Networks.” Artificial Intelligence Review 56: 1513–89. https://doi.org/10.1007/s10462-023-10562-9.
Guo, Chuan, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. 2017. “On Calibration of Modern Neural Networks.” Proceedings of the 34th International Conference on Machine Learning, Proceedings of machine learning research, vol. 70: 1321–30.
Haasnoot, Marjolijn, Jan H. Kwakkel, Warren E. Walker, and J. ter Maat. 2013. “Dynamic Adaptive Policy Pathways: A Method for Crafting Robust Decisions for a Deeply Uncertain World.” Global Environmental Change 23 (2): 485–98. https://doi.org/10.1016/j.gloenvcha.2012.12.006.
Hedou, F., A. Herman, and H. De Luca. 2024. Projection Du Trait de Côte Et Analyse Des Enjeux Au Niveau National – Horizons 2050 Et 2100. Cerema.
Hüllermeier, Eyke, and Willem Waegeman. 2021. “Aleatoric and Epistemic Uncertainty in Machine Learning: An Introduction to Concepts and Methods.” Machine Learning 110 (3): 457–506. https://doi.org/10.1007/s10994-021-05946-3.
Kapoor, Sayash, and Arvind Narayanan. 2023. “Leakage and the Reproducibility Crisis in Machine-Learning-Based Science.” Patterns 4 (9): 100804. https://doi.org/10.1016/j.patter.2023.100804.
Kerr, John, Anne van der Bles, Sarah Dryhurst, et al. 2023. “The Effects of Communicating Uncertainty Around Statistics, on Public Trust.” Royal Society Open Science 10 (11): 230604. https://doi.org/10.1098/rsos.230604.
Lakshminarayanan, Balaji, Alexander Pritzel, and Charles Blundell. 2017. “Simple and Scalable Predictive Uncertainty Estimation Using Deep Ensembles.” Advances in Neural Information Processing Systems 30, 6402–13.
Lempert, Robert J. 2019. “Robust Decision Making (RDM).” In Decision Making Under Deep Uncertainty: From Theory to Practice. Springer. https://doi.org/10.1007/978-3-030-05252-2_2.
Loi no 2021-1104, Journal officiel (2021).
Luijendijk, Arjen, G. Hagenaars, and R. Ranasinghe. 2018. “The State of the World’s Beaches.” Scientific Reports, ahead of print. https://doi.org/10.1038/s41598-018-24630-6.
Magnan, Alexandre, E. Lisa F. Schipper, Maxine Burkett, et al. 2016. “Addressing the Risk of Maladaptation to Climate Change.” WIREs Climate Change 7 (5): 646–65. https://doi.org/10.1002/wcc.409.
Makridakis, Spyros, Evangelos Spiliotis, and Vassilios Assimakopoulos. 2020. “The M4 Competition: 100,000 Time Series and 61 Forecasting Methods.” International Journal of Forecasting 36 (1): 54–74. https://doi.org/10.1016/j.ijforecast.2019.04.014.
Mao, Y., Giovanni Coco, Sean Vitousek, et al. 2025. “Benchmarking Shoreline Prediction Models over Multi-Decadal Timescales.” Communications Earth & Environment, ahead of print. https://doi.org/10.1038/s43247-025-02550-4.
Marchau, Vincent A. W. J., Warren E. Walker, Pieter J. T. M. Bloemen, and Steven W. Popper, eds. 2019. Decision Making Under Deep Uncertainty: From Theory to Practice. Springer. https://doi.org/10.1007/978-3-030-05252-2.
Mastrandrea, Michael D., Christopher B. Field, Thomas F. Stocker, et al. 2010. Guidance Note for Lead Authors of the IPCC Fifth Assessment Report on Consistent Treatment of Uncertainties. Intergovernmental Panel on Climate Change.
Mentaschi, Lorenzo, Michalis I. Vousdoukas, and Jean-François Pekel. 2018. “Global Long-Term Observations of Coastal Erosion and Accretion.” Scientific Reports, ahead of print. https://doi.org/10.1038/s41598-018-30904-w.
Minderer, Matthias, Josip Djolonga, R. Romijnders, et al. 2021. “Revisiting the Calibration of Modern Neural Networks.” Advances in Neural Information Processing Systems 34, 15682–94.
Montaño, Juan, Giovanni Coco, Jose A. Antolínez, et al. 2020. “Blind Testing of Shoreline Evolution Models.” Scientific Reports, ahead of print. https://doi.org/10.1038/s41598-020-59018-y.
Nerem, Robert S., Brian D. Beckley, John T. Fasullo, and Gary T. Mitchum. 2018. “Climate-Change–Driven Accelerated Sea-Level Rise Detected in the Altimeter Era.” Proceedings of the National Academy of Sciences, ahead of print. https://doi.org/10.1073/pnas.171731211.
Nicholls, Robert J., Daniel Lincke, Jochen Hinkel, and Sally Brown. 2021. “A Global Analysis of Subsidence, Relative Sea-Level Change and Coastal Flood Exposure.” Nature Climate Change, ahead of print. https://doi.org/10.1038/s41558-021-00993-z.
Oppenheimer, Michael, B. C. Glavovic, Jochen Hinkel, et al. 2019. Sea Level Rise and Implications for Low-Lying Islands, Coasts and Communities. https://doi.org/10.1017/9781009157964.006.
Ranasinghe, Roshanka. 2016. “Assessing Climate Change Impacts on Open Sandy Coasts: A Review.” Earth-Science Reviews, ahead of print. https://doi.org/10.1016/j.earscirev.2016.07.011.
Regulation (EU) 2024/1689 Laying down Harmonised Rules on Artificial Intelligence (Artificial Intelligence Act), Official Journal of the European Union (2024).
Simian, B. 2018. Situation de l’immeuble Le Signal à Soulac-Sur-Mer.
Simian, Panonacle, Hammerer, et al. 2018. Amendement no 589.
Vitousek, Sean, Kilian Vos, Kristen D. Splinter, Li Erikson, and Patrick L. Barnard. 2023. “A Model Integrating Satellite-Derived Shoreline Observations for Predicting Fine-Scale Shoreline Response to Waves and Sea-Level Rise Across Large Coastal Regions.” Journal of Geophysical Research: Earth Surface, ahead of print. https://doi.org/10.1029/2022JF006936.
Vos, Kilian, Mitchell D. Harley, Kristen D. Splinter, Jason A. Simmons, and Ian L. Turner. 2019. “Sub-Annual to Multi-Decadal Shoreline Variability from Publicly Available Satellite Imagery.” Coastal Engineering, ahead of print. https://doi.org/10.1016/j.coastaleng.2019.04.004.
Vos, Kilian, Kristen D. Splinter, Mitchell D. Harley, Jason A. Simmons, and Ian L. Turner. 2019. “CoastSat: A Google Earth Engine-Enabled Python Toolkit to Extract Shorelines from Publicly Available Satellite Imagery.” Environmental Modelling & Software, ahead of print. https://doi.org/10.1016/j.envsoft.2019.104528.
Vos, Kilian, Kristen D. Splinter, J. Palomar-Vázquez, et al. 2023. “Benchmarking Satellite-Derived Shoreline Mapping Algorithms.” Communications Earth & Environment, ahead of print. https://doi.org/10.1038/s43247-023-01001-2.
Vousdoukas, Michalis I., Roshanka Ranasinghe, Lorenzo Mentaschi, et al. 2020. “Sandy Coastlines Under Threat of Erosion.” Nature Climate Change, ahead of print. https://doi.org/10.1038/s41558-020-0697-0.
Werners, Saskia, Rachel Wise, James Butler, Edouard Totin, and Katharine Vincent. 2021. “Adaptation Pathways: A Review of Approaches and a Learning Framework.” Environmental Science & Policy 116: 266–75. https://doi.org/10.1016/j.envsci.2020.11.003.
Zeng, Ailing, Muxi Chen, Lei Zhang, and Qiang Xu. 2023. “Are Transformers Effective for Time Series Forecasting?” Proceedings of the AAAI Conference on Artificial Intelligence 37: 11121–28. https://doi.org/10.1609/aaai.v37i9.26317.

Citation

BibTeX citation:
@inproceedings{capucine_pineau2026,
  author = {Capucine Pineau, Lili},
  title = {Communicating {Uncertainty} in {Coastal} {Retreat}
    {Forecasting}},
  booktitle = {Proceedings of the 1st Conference Cycles of Cytopia},
  date = {2026-10-01},
  url = {https://cytopia.fr/pineau2026},
  isbn = {TBA},
  langid = {en}
}
For attribution, please cite this work as:
Capucine Pineau, Lili. 2026. “Communicating Uncertainty in Coastal Retreat Forecasting.” Proceedings of the 1st Conference Cycles of Cytopia, accepted, October 1. https://cytopia.fr/pineau2026.