AI in coastal management

By Mark Ellis, Associate Director FCERM & CDM Advisor, WSP UK

Cutting through the hype to find the signal

Artificial intelligence is the most over-hyped and under-talked-about technology in flood and coastal erosion risk.

Over-promised, because the discourse on AI in infrastructure and environmental management is rife with frictionless, transformational, change-ready promises that rarely translate into the reality of coastal data, institutional frameworks, and operational realities.

Under-discussed, because behind the hype there is real, meaningful work being done - applications of machine learning that are making improvements to our forecasting, monitoring, and modelling of coastal systems - and the profession is not yet having an informed discussion about the role of this work, what it requires to be successful, and what it cannot do.

This article is an attempt to do that. Not from the point of view of a data scientist, but from the point of view of a coastal engineer: where does AI bring value to the decision-making process?

Where AI is making a difference

The most successful applications of AI and machine learning in coastal management have a common feature: they do something that the traditional way, or traditional methods, cannot do or cannot do fast enough or to a large enough scale for operational use.

A good example is surrogate modelling. Numerical models based on physics - such as Delft3D, MIKE, SWAN and XBeach - continue to be the state-of-the-art for predicting coastal hydrodynamics and morphodynamics. They are sophisticated, tested, and physics-based. They can be very slow. It can take hours or even days to run a process-based model for a complex estuary or open coast, making them unsuitable for real-time prediction or rapid exploration of scenarios.

Surrogate models - usually neural networks or Gaussian process emulators - can mimic those results in a matter of seconds. They don't substitute for physics. They learn from it. A surrogate model can be trained to replicate flood inundation extents, wave overtopping volumes, or morphological change for a set of inputs with very high fidelity, while being run at a fraction of the time. This makes them quite valuable for forecasting, for feeding near-real-time predictions to digital twins, and for running many scenarios that are too expensive to simulate in full physics.

Machine learning is also making inroads in pattern recognition in monitoring data. Automated coastline extraction from remote sensing. Landform classification from LiDAR. Identification of structural issues in drone surveys. Change detection across repeat surveys. These are things that humans can do, but not in the volumes or timeliness needed. Machine learning brings analysis to data.

Flood forecasting is also improved. Machine learning models that learn from past rainfall-runoff relationships, tidal surges, or interactions between compound flood events can support physics-based flood forecasting systems to deliver fast ensemble forecasts that support timely decision-making by emergency responders. The Environment Agency's flood forecasting system already includes machine learning elements, and the trend is towards more as data and models grow.

These are not hypothetical uses. They are being used, published and in some cases, deployed. They are a real advance.

The data quality problem

Data is an understated dependency of all those applications.

Machine learning models learn from data. If the data isn't dense, objective, consistent, or representative of the data to be predicted, the model won't be good. It doesn't matter how complex the model is. A neural network with 10 years of wave data for calm conditions will not learn how to predict stormy conditions. A surrogate model trained in one estuary won't necessarily work in another. A computer vision model for chalk cliffs will not work to detect erosion on soft clay.

Many times, England's coastal data is the type of data you don't want for machine learning. Data is often not collected continuously. Methods are non-standardised, introducing biases. They are not spatially comprehensive - some fronts are well monitored, others are not. Rare events, which are often of most interest, are just that. And the long-term changes that are most important for strategic planning - the 10-year sediment budget, progressive changes in estuary geometry - need longer data series than are typically available from digital monitoring.

This doesn't rule out AI. But it means the data cleaning, preparation and verification work that is a prerequisite to applying machine learning is at least as critical as model building. This is the aspect that receives less attention in the debate about AI in coastal management. It's "we have data" straight to "we have a model", without sufficient consideration of the "boring" question of whether we have the right data for the model to work.

This has investment implications. If we want to do AI-enabled coastal management, we need to put the same effort into the monitoring programs that provide the data. Time series that are consistent, continuous, and well-documented are not only beneficial for adaptive management but also provide the bedrock for any serious machine learning application.

The adoption gap

There is a considerable gap between a research paper describing a machine learning application in the coastal sciences and an operational tool used by a local government officer or consulting engineer for decision-making.

It isn't a technical gap. It is institutional, cultural, and practical.

The majority of existing AI applications in coastal management are research models. They are built by PhD students and research teams, written up in papers, and presented at conferences, and then - often - nothing. The models are stored on GitHub. The algorithms are published in articles that may not be read by practitioners, or they may be read but cannot be translated into practice by practitioners. The leap from idea to product doesn't always occur.

There are reasons for this. Academic incentives favour publication, not deployment and maintenance. Tools need to be supported, documented, trained, and embedded into workflow - work that is less than glamorous and not funded in most academic settings. Consultancies and local governments may not have the data science skills to assess, refine, and implement research models.

The result is a paradox. The literature is filled with examples of applications. Practices are still the same. There's a lot of talk about AI at conferences, but on Monday morning, we are back to spreadsheets and legacy modelling packages.

A concerted effort is needed from several sources to address this. Funders could focus more on the deployment of research and practitioner involvement. Consultancies could develop applied data science capabilities besides the usual modelling teams. Software vendors could add machine learning capabilities to existing software systems. And the profession could include data literacy in its education and CPD regimes, so that new generations of coastal engineers can understand and at least critique their use, even if they don't develop them.

AI and the coastal digital twin

My favourite application of AI in coastal management soon is as a supporting element of digital twins. As I've written previously, a coastal digital twin is an operational platform that brings together monitoring data, asset condition information, and predictions to help inform ongoing, adaptive decision-making. The concept is sound. But the difficulty of conducting physics-based predictions fast enough and often enough to satisfy an operational platform has been an obstacle.

Surrogate models solve that problem. A digital twin can invoke a trained emulator to provide a flood inundation forecast, a wave overtopping estimate, or a morphological change projection in seconds - rather than wait hours for a full physics-based numerical simulation - is more valuable. It can react to monitoring information. It can interactively run scenarios. It can assess trigger points in the current state of the system, rather than based on the most recent periodic update.

Machine learning also provides data assimilation, which makes a digital twin dynamic. Automatic analysis of satellite images to update the shoreline. Clustering algorithms that detect changes in equipment condition from surveys. Anomaly detection that detects when conditions are not as they should be. These are what transform an information store into a decision support system.

But crucially, the AI-based components of a digital twin must be transparent. A surrogate model can provide a flood extent, but with no information about how confident it is, how it was trained, or what it was trained on, is a curse, not a blessing. The model needs to tell the practitioner and decision-maker not only what it predicts, but also how confident it is in its prediction in a context. Uncertainty, training assumptions, and model limitations are not optional. These are essential for model use.

Getting it right

AI will not revolutionise coastal management. It will not be a silver bullet to resolve the challenges of data gaps, institutional capability, financial resources, and citizen trust. If someone claims otherwise, they're trying to sell you something.

But machine learning, when used appropriately, can help. It can speed up predictions that are too slow for real-time operations. It can automate analyses that are too inefficient for real-time monitoring. It can facilitate digital twin platforms that can integrate data, models, and decision-making in unprecedented ways.

To realise that potential, the industry must invest in three areas at once: monitoring data to train AI models, institutional capacity to implement and sustain AI-powered tools, and a professional culture of openness and scepticism to ensure that those tools are used safely.

The buzz around AI in infrastructure is noisy and increasing. The signal is there, but it is fainter. The challenge for the profession is to hear the signal, use it, and not fall for the hype.

This article appeared in issue 12 of Flood Industry magazine (May/June 2026) - you can view it here.