The stakes could hardly be higher. More than 40 percent of the global population lives in coastal regions, and sea-level rise alone is projected to erase up to half of the world’s sandy coastlines by the end of the century. The hazards now converging on those shores are not isolated events but compound extremes, in which multiple drivers strike simultaneously and amplify one another. Hurricane Harvey offered a devastating demonstration in 2017, when record-breaking rainfall, river discharge, and surface runoff combined with a prolonged storm surge to produce catastrophic compound flooding in Houston, causing approximately 152.5 billion US dollars in damage and ranking as the second-costliest natural disaster in United States history. Research cited in the review shows that precipitation is the dominant contributor to changes in compound flooding frequency across 83 percent of global coastlines.

The physical backdrop to this challenge is shifting in ways that conventional models struggle to capture. Assessments from the Intergovernmental Panel on Climate Change indicate with high confidence that both the frequency and intensity of heavy precipitation have increased across most land regions since the mid-twentieth century, and CMIP6 climate model projections point to substantial further increases in extreme precipitation under medium- to high-emission scenarios, particularly across northern Europe and adjacent seas. Storm surges and extreme sea levels, meanwhile, exhibit pronounced spatial heterogeneity: surge magnitudes are projected to increase in regions such as the Baltic Sea in response to shifting wind patterns, while other basins, including parts of the Mediterranean and Atlantic, remain uncertain or strongly modulated by large-scale climate oscillations such as the North Atlantic Oscillation. These nonlinear, non-stationary dynamics are precisely the kind of multivariate complexity that machine learning was designed to untangle.

To understand how the field has responded, researchers led by Kamran Tanwari of the University of Szczecin in Poland conducted a PRISMA-compliant systematic search of the Scopus database, screening 666 initial records down to 73 studies that applied artificial intelligence or machine learning to coastal climatic extremes. Each study was assessed across seven methodological domains adapted from Cochrane risk-of-bias guidance and the PROBAST+AI framework for prediction models. The results reveal a field in rapid but uneven growth: most studies appeared between 2021 and 2025, and the overwhelming majority originated from just two countries, the United States with 24 studies and China with 18, leaving the most vulnerable regions of the Global South strikingly underrepresented. Storm surges and extreme sea levels dominated the hazard landscape, examined in 44 studies, while compound flooding and heavy precipitation each appeared in only 11.

Deep learning was the clear methodological favorite, used in 57 of the 73 studies, with artificial neural networks, convolutional neural networks, and long short-term memory networks the most common architectures. The appeal is fundamentally computational. Process-based coastal models such as ADCIRC and MIKE 21 solve depth-integrated shallow-water equations, while spectral wave models like SWAN track wave propagation, refraction, shoaling, and dissipation, all of which demand enormous computing resources. AI surrogates, by contrast, can generate forecasts in a fraction of a second, making them highly attractive for early warning systems. Reported storm surge lead times typically range from 1 to 24 hours, but the review found a nearly universal pattern: model skill degrades significantly, sometimes dropping to random chance, once the prediction window extends beyond 48 to 72 hours, because data-driven models excel at interpolating within their training data but frequently fail to extrapolate to unprecedented extremes.

The review also uncovered a troubling pattern of contradictory findings that suggests the field lacks standardized benchmarking. In shoreline detection, one study found that a traditional object-based image analysis using random forest outperformed a convolutional neural network approach, directly contradicting earlier results. In surge prediction, some authors argue that standard LSTM networks are superior for handling temporal dependencies in water level sequences, while others found that hybrid ConvLSTM architectures, which combine spatial feature extraction with temporal regression, provide a more robust balance, and still others demonstrated that simple neural networks or even linear regression can perform comparably well at specific tide gauges with short or patchy records. This may indicate that the rush toward ever-deeper, more complex hierarchical models sometimes outpaces actual operational necessity at individual sites.

Perhaps the most consequential finding concerns what the models do not understand. When explainable AI techniques were applied to high-performing architectures, including Vision Transformers, researchers discovered a startling contradiction: despite achieving high statistical accuracy, some models primarily learned the autocorrelation of historical water levels rather than the actual physical mechanisms of wind-surge coupling. For mitigation design, where AI is increasingly used to optimize dynamic seawalls, revetments, and drainage infrastructure through reinforcement learning and Bayesian networks, this mechanistic shortcoming is a serious liability, because models that do not grasp physical drivers cannot reliably extrapolate the stress loads of unprecedented extreme events. The review also cautions that highly optimized hard defenses can promote a false sense of security, potentially leading to catastrophic disasters when climate extremes inevitably exceed design thresholds.

The gap between prediction and practice extends across the entire application spectrum. The review classified AI applications into six progressive domains: monitoring, prediction, mitigation, adaptation, resilience, and uncertainty quantification. Monitoring and prediction dominate the literature, while resilience emerged as the most critically underrepresented domain, requiring the integration of land use, elevation, soil, and post-event reconnaissance data through sophisticated architectures such as causal spatio-temporal graph neural networks. Only 11 studies explicitly addressed multi-decadal climate trajectories, and there is strong consensus that advanced adaptation models, which use reinforcement learning to navigate deep uncertainty and identify flexible policy pathways, remain largely confined to academic journals with very little uptake by actual coastal policymakers. The algorithms frequently optimize policies in a theoretical vacuum, overlooking conflicting political interests, municipal funding obstacles, and the time lags required to pass coastal zoning laws.

Uncertainty quantification, the domain the authors identify as the bridge to operational trustworthiness, is practiced in only a quarter of the reviewed studies. The most sophisticated approaches now distinguish aleatoric uncertainty, the inherent noise in meteorological inputs, from epistemic uncertainty, the internal ignorance of the neural network itself, using Bayesian hierarchical models, Gaussian process regression, and Monte Carlo simulations. But these frameworks carry a cruel irony: generating the massive ensembles of synthetic data required for robust uncertainty propagation effectively neutralizes the primary advantage of AI, its computational speed. Moreover, explainable AI tools such as attention mechanisms and permutation feature importance currently validate only statistical correlation, falling short of proving true physical consistency or causality.

The review’s authors close with four priorities for advancing coastal AI: developing multivariate frameworks capable of representing compound and cascading extremes; integrating physics-informed neural networks and hybrid hydrodynamic-AI architectures so that models obey fundamental physical laws when extrapolating beyond their training data; incorporating socioeconomic scenarios and Shared Socioeconomic Pathways to shift the focus from short-term prediction toward long-term resilience; and applying transfer learning to adapt models developed in data-rich settings to the highly vulnerable, data-scarce regions of the Global South. They also point to the rapid rise of machine-learning weather emulators such as GraphCast, Pangu-Weather, FourCastNet, and GenCast, which could supply atmospheric forcing for coastal models but require rigorous assessment of local accuracy and performance during rare extremes. For now, the verdict is clear: purely data-driven models must be considered complementary to, not a substitute for, physics-based hydrodynamic models, especially when navigating the deep uncertainty of a changing climate.

Subject of Research: Application of artificial intelligence and machine learning to modelling coastal climate extremes such as storm surges, extreme sea levels, and compound flooding

Article Title: Artificial intelligence for modelling coastal climate extremes: a global systematic review

Article References: Tanwari, K., Shi, X., Śledziowski, J., Giza, A., & Terefenko, P. (2026). Artificial intelligence for modelling coastal climate extremes: a global systematic review. Regional Environmental Change, 26(4), Article 207. https://doi.org/10.1007/s10113-026-02689-6

Image Credits: AI Generated

DOI: 10.1007/s10113-026-02689-6

Keywords: artificial intelligence, coastal hazards, storm surge, compound flooding, extreme sea levels, machine learning, deep learning, climate change, sea-level rise, uncertainty quantification, explainable AI, coastal adaptation