Coastal Resilience: Forecasting Storm Surge with Spatial Intelligence

by

Author: Denis Avetisyan

A new graph neural network model, StormNet, leverages connections between coastal monitoring stations to significantly reduce biases and improve the accuracy of storm surge predictions.

The system dissects hurricane storm surge forecasting by constructing a spatio-temporal graph-nodes representing gauge stations and edges quantifying their correlations-to predict localized water level offsets <span class="katex-eq" data-katex-display="false">\hat{o}_{i}(t)</span> and refine physics-based ADCIRC models, effectively learning to correct inherent biases in surge prediction.
The system dissects hurricane storm surge forecasting by constructing a spatio-temporal graph-nodes representing gauge stations and edges quantifying their correlations-to predict localized water level offsets \hat{o}_{i}(t) and refine physics-based ADCIRC models, effectively learning to correct inherent biases in surge prediction.

StormNet utilizes graph convolutional networks to model spatiotemporal dependencies and enhance hurricane forecasting up to 72 hours, addressing persistent biases in traditional ADCIRC models.

Accurate prediction of storm surge remains a persistent challenge despite advances in numerical modeling. This study, ‘Storm Surge Modeling, Bias Correction, Graph Neural Networks, Graph Convolution Networks’, introduces StormNet, a novel graph neural network designed to correct biases in storm surge forecasts by leveraging spatial correlations between coastal water-level gauge stations. StormNet demonstrably reduces root mean square error in 48- and 72-hour forecasts by over 70% and 50% respectively, outperforming sequential LSTM baselines. Could this computationally efficient, physics-informed approach represent a paradigm shift in real-time operational forecasting for extreme weather events?

Decoding the Surge: Why Prediction Remains a Critical Challenge

The vulnerability of coastal communities hinges significantly on the ability to forecast storm surge with precision, yet achieving this proves remarkably difficult due to the intricate interplay of spatial and temporal forces at play. Storm surge – the abnormal rise in sea level during a storm – isn’t simply dictated by wind speed or central pressure; it’s a highly localized phenomenon shaped by bathymetry, coastline geometry, tidal cycles, and the storm’s track and intensity, all evolving over time. Traditional predictive models, while increasingly sophisticated, often struggle to fully capture these nuances, leading to uncertainties in forecast accuracy. The dynamic nature of the ocean and atmosphere, coupled with the complex interactions between them, creates a forecasting challenge that demands continuous refinement of existing models and exploration of new predictive techniques to safeguard lives and infrastructure.

Despite the sophistication of hydrodynamic models like ADCIRC – widely used to simulate storm surge – systematic biases often impede the delivery of truly reliable forecasts. These aren’t random errors, but consistent deviations arising from the inherent simplifications within the model’s physics and the incomplete nature of historical data used for calibration. For example, ADCIRC, like many similar tools, may consistently underpredict surge heights in areas with complex bathymetry or overestimate the impact of wind stress on shallow coastal waters. Consequently, researchers dedicate significant effort to post-processing model outputs, employing statistical and machine learning techniques to identify and correct these biases, thereby improving the accuracy and trustworthiness of vital coastal hazard warnings.

Storm surge prediction, while increasingly sophisticated, remains challenged by inherent biases arising from the complex interplay of physical processes and data limitations. Current models, despite their computational power, rely on approximations of phenomena like wave-current interactions, sediment transport, and wind drag – each introducing potential error. Critically, the historical data used to calibrate and validate these models is often incomplete or unevenly distributed, particularly for extreme events, hindering the accurate capture of surge behavior under rare but devastating conditions. This combination of imperfect representation and limited data directly impacts a community’s ability to effectively prepare for and mitigate the impacts of coastal flooding, emphasizing the ongoing need for model refinement and comprehensive data collection efforts to enhance forecast reliability and bolster disaster resilience.

StormNet consistently outperforms a sequential LSTM model in predicting Hurricane Idalia's impact, as evidenced by lower RMSE values (indicated by smaller data point radii) at most gauge stations for both 24- and 36-hour forecasts, particularly around landfall, and is visualized with station performance overlaid on the storm track.
StormNet consistently outperforms a sequential LSTM model in predicting Hurricane Idalia’s impact, as evidenced by lower RMSE values (indicated by smaller data point radii) at most gauge stations for both 24- and 36-hour forecasts, particularly around landfall, and is visualized with station performance overlaid on the storm track.

StormNet: Mapping the Chaos with Spatiotemporal Graphs

StormNet employs a spatiotemporal graph representation to model coastal storm surge dynamics. In this framework, coastal gauge stations are defined as nodes within a graph, and the edges connecting these nodes represent the spatial relationships and dependencies between stations. This allows the model to consider not only the individual measurements at each station, but also how those measurements are correlated with measurements from neighboring stations. The ‘spatiotemporal’ aspect is achieved by extending this graph structure to incorporate temporal information, effectively creating a dynamic graph that evolves over time and captures the propagation of storm surge events. This representation facilitates the application of graph neural networks to learn complex relationships within the coastal network and improve forecasting capabilities.

StormNet employs both Graph Convolutional Network (GCN) and Graph Attention Network (GAT) layers to explicitly model spatial relationships between coastal gauge stations. GCN layers aggregate feature information from neighboring stations, weighted by the graph’s adjacency matrix, to capture broad spatial correlations. GAT layers refine this process by learning attention weights that dynamically determine the importance of each neighboring station’s features during aggregation. This allows the model to prioritize information from stations with stronger influence on the target station’s storm surge, accounting for variations in geographical proximity and hydrodynamic connectivity. The combined use of GCN and GAT enables StormNet to effectively capture complex spatial dependencies within the storm surge field, improving forecast accuracy compared to methods that treat stations as independent entities.

The Long Short-Term Memory (LSTM) network within StormNet addresses the temporal dependencies inherent in storm surge data by processing sequential observations from coastal gauge stations. This recurrent neural network architecture is specifically designed to learn and retain information from past storm surge patterns, enabling the model to identify trends and anticipate future surges based on historical data. The LSTM component receives input from the graph convolutional and attention layers, and its internal memory cells allow it to weigh the importance of past observations when making predictions, ultimately improving forecast accuracy by accounting for the time-varying nature of storm surge events. This capability is crucial for capturing phenomena like lagged effects and persistent anomalies that influence surge height and timing.

StormNet addresses systematic errors present in ADCIRC storm surge forecasts through a bias correction mechanism. The model is trained on historical storm surge data and corresponding ADCIRC predictions to identify consistent discrepancies. By learning these biases – which can be spatially and temporally variable – StormNet generates a correction factor applied to the raw ADCIRC output. This correction process improves forecast accuracy by reducing both mean absolute error and root mean squared error when compared to standalone ADCIRC forecasts, and provides a more reliable prediction of storm surge heights and arrival times. The effectiveness of this bias correction is validated through performance metrics calculated on held-out testing datasets representing diverse storm events.

StormNet is a novel model architecture designed for <span class="katex-eq" data-katex-display="false">...</span> and comprises [details of architecture - replace this with details from the image].
StormNet is a novel model architecture designed for ... and comprises [details of architecture – replace this with details from the image].

Validating the Algorithm: From Historical Data to Hurricane Idalia

StormNet’s training and evaluation leveraged a comprehensive ‘Historical Storm Archive’ encompassing a wide range of meteorological conditions and storm characteristics. This archive included data from numerous historical storms, spanning varying intensities, track densities, and landfall locations within the model’s geographic domain. Utilizing this diverse dataset ensured the model was exposed to a multitude of possible storm scenarios, improving its generalization capability and robustness when predicting storm surge from previously unseen events. The archive’s breadth facilitated a statistically significant assessment of StormNet’s performance across a representative spectrum of potential storms.

StormNet’s performance was quantitatively evaluated using the Root Mean Squared Error (RMSE), a standard metric for assessing the accuracy of numerical predictions. Comparative analysis against baseline ADCIRC forecasts revealed substantial bias reduction in storm surge predictions. Specifically, the implementation of StormNet resulted in a greater than 70% reduction in RMSE for 48-hour forecasts and over 50% reduction in RMSE for 72-hour forecasts. These results demonstrate a significant improvement in the model’s ability to accurately predict storm surge height and timing compared to the established ADCIRC system.

StormNet’s predictive capabilities were validated through a case study utilizing Hurricane Idalia. Analysis of Idalia’s storm surge demonstrated the model’s ability to accurately forecast surge heights and inundation extents under actual hurricane conditions. Specifically, comparisons between StormNet’s predictions and observed water levels during Idalia showed a high degree of correlation, confirming the model’s effectiveness in real-world operational scenarios. This validation extends beyond statistical metrics, demonstrating practical utility for coastal hazard prediction and mitigation efforts.

StormNet bias correction predictions exhibit a distribution of root mean squared errors (RMSE) across stations, with the median RMSE indicated by an orange line and individual station values for NOAA-NOS (blue) and TCOON (red) detailed in Table 3.
StormNet bias correction predictions exhibit a distribution of root mean squared errors (RMSE) across stations, with the median RMSE indicated by an orange line and individual station values for NOAA-NOS (blue) and TCOON (red) detailed in Table 3.

Beyond Prediction: Empowering Resilience Through Actionable Intelligence

Coastal emergency managers now have access to a dynamic resource in the form of the CERA platform, which delivers StormNet’s complex storm surge forecasts through an easily navigable interface. This integration transcends static data delivery; CERA offers interactive visualizations, allowing officials to explore predicted inundation levels, potential impacts on critical infrastructure, and optimal evacuation routes. By translating sophisticated meteorological modeling into an accessible format, the platform empowers proactive decision-making, moving beyond reactive responses to coastal threats and facilitating targeted resource allocation before, during, and after a storm event. The system’s user-friendly design minimizes the learning curve, ensuring that vital predictive information can be rapidly understood and acted upon, even under the pressure of an impending crisis.

The convergence of StormNet forecasts within the CERA platform fundamentally shifts coastal hazard response from reactive to proactive. Accurate storm surge predictions, delivered through this integration, empower emergency managers to move beyond generalized evacuation orders and instead implement targeted strategies focused on the most vulnerable areas. This precision allows for efficient resource allocation – directing personnel, equipment, and aid precisely where it’s needed most – minimizing disruption to unaffected communities and maximizing the effectiveness of the response. Ultimately, this capability translates to reduced risk for coastal populations and a more resilient approach to managing the increasing threats posed by severe weather events.

Ongoing development of StormNet prioritizes a multi-faceted approach to data assimilation, with immediate efforts centering on the integration of real-time wave measurements. These measurements are crucial for a more nuanced understanding of storm surge dynamics, as wave action significantly influences the height and inland penetration of coastal flooding. By incorporating data from buoys, radar systems, and potentially satellite altimetry, researchers aim to improve the model’s ability to accurately predict localized variations in surge, particularly in areas with complex bathymetry or sheltered coastlines. This enhanced granularity will not only refine overall surge forecasts but also allow for more precise assessments of wave-driven hazards, such as coastal erosion and overwash, ultimately bolstering the platform’s predictive power and supporting more effective mitigation strategies.

The long-term effectiveness of StormNet, and platforms like CERA which utilize its forecasts, hinges on its adaptability to a changing world and broader applicability. Current development prioritizes extending the model’s reach to encompass a significantly larger expanse of coastline, offering protective insights to previously unmonitored communities. Crucially, researchers are integrating projections of future sea-level rise and altered storm patterns – driven by climate change – into the predictive algorithms. This proactive inclusion of climate scenarios isn’t merely about extending the forecast timeframe, but fundamentally reshaping the model’s baseline assumptions, ensuring that StormNet remains a relevant and reliable tool for coastal resilience decades into the future, even as environmental conditions shift and intensify.

StormNet accurately forecasts water levels at selected gauge stations across 12-, 48-, and 72-hour horizons (shown in green), significantly improving upon observed (blue) and uncorrected forecasts (orange) and enabling reliable assessment against NOAA's flooding thresholds (yellow, red, purple).
StormNet accurately forecasts water levels at selected gauge stations across 12-, 48-, and 72-hour horizons (shown in green), significantly improving upon observed (blue) and uncorrected forecasts (orange) and enabling reliable assessment against NOAA’s flooding thresholds (yellow, red, purple).

The study demonstrates a commitment to exposing the underlying mechanisms of prediction, challenging the ‘black box’ nature of many machine learning models. This aligns with a fundamental principle: true understanding comes from dissecting systems, not simply accepting their outputs. As Blaise Pascal observed, “The eloquence of angels is never so convincing as the evidence of facts.” StormNet, by explicitly modeling spatial correlations via graph neural networks – effectively reverse-engineering the relationships between coastal gauges – provides that evidence. It’s not enough to know a surge is coming; the model reveals why, reducing bias and extending the reliable forecast horizon by leveraging the inherent connectedness of the system, a principle mirroring Pascal’s emphasis on demonstrable truth.

Beyond the Horizon

The pursuit of accurate storm surge forecasting, as demonstrated by this work, isn’t simply about shrinking error margins-it’s a sustained challenge to the limits of predictive modeling. StormNet’s approach, leveraging graph neural networks and spatial correlations, is less a solution than a carefully constructed probe-a way to interrogate the inherent chaos of fluid dynamics and data assimilation. The model effectively transfers information between gauges, but the fundamental question of what constitutes ‘true’ surge remains stubbornly resistant to complete resolution. It invites further examination of whether the correlations exploited are merely artifacts of the measurement network itself, or genuinely reflect underlying physical processes.

Future iterations will undoubtedly focus on expanding the scope-integrating higher-resolution data, incorporating more complex physical models, and scaling to truly global forecasts. However, a more interesting line of inquiry lies in deliberately introducing controlled perturbations to the model. Can one strategically destabilize the system to reveal previously hidden vulnerabilities, or uncover emergent behaviors? The goal shouldn’t be to build an infallible predictor, but rather a system capable of quantifying its own uncertainty-a model that knows what it doesn’t know.

Ultimately, the value of any predictive model rests not in its ability to foretell the future with precision, but in its capacity to illuminate the intricate dance between order and chaos. StormNet represents a step toward that illumination, a demonstration that even the most turbulent systems can be understood-not by conquering them, but by systematically dismantling their assumptions.

Original article: https://arxiv.org/pdf/2604.20688.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2026-04-24 00:14