Scholarlink

<p>Storm surge prediction is crucial for coastal zone management and flood risk reduction practices. Physics-based hydrodynamic models are often used for storm surge modeling. However, their inherent computational demands and the need for high-fidelity spatiotemporal data hinder their applicability at regional and continental scales. One alternative is the data-driven modeling of storm surges. However, most studies used data-driven techniques to model daily surges at individual tide gauges. In this study, we develop large-scale storm surge models utilizing Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and a hybrid of CNN and LSTM (CNN-LSTM) algorithms to predict hourly surges at multiple tide gauges simultaneously, trained with atmospheric predictor variables. Observed hourly surges from 28 long-record tide gauges along the U.S. Gulf and East Coasts are used as the predictand. CNN-LSTM was identified as the best model, achieving an average (over 28 TGs) correlation of 0.82 and a root mean square error of 0.08 during the testing period. Furthermore, the best model’s performance in capturing temporal variability, extreme-surge statistics, probability distributions, and surge-hydrograph characteristics (e.g., peak, duration, severity, and intensity) is investigated. Additionally, we demonstrate how the networks leverage plausible physical relationships between storm surge and predictor features during training and prediction. Overall, the CNN-LSTM model achieves up to 78% and 46% improvements in the correlation coefficient and root mean square error, respectively, compared with existing data-driven storm surge models. In contrast, the model yields an average 8% increase in correlation and a 25% reduction in RMSE relative to hydrodynamic model-derived storm surges. The study underscores that the proposed deep learning modeling framework efficiently predicts hourly storm surges simultaneously across the U.S. Gulf and East Coasts. The model can be used for historical reconstruction and for potential future extensions to produce storm-surge projections that facilitate the investigation of future trends, variability, and return levels under different climate-change scenarios.</p> Graphical Abstract <p></p> <p>The graphical abstract summarizes the research, including the study area, data, methodologies, and highlights the main results. The study trains large-scale deep learning models to simultaneously predict hourly storm surges at 28 long-record tide gauges (TG) along the U.S. Gulf and East coasts. Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and their hybrid (i.e., CNN-LSTM) architectures are utilized for model development. Instead of separate predictor domains for each TG, a single large domain is used, from which the DL algorithms automatically extract physically relevant features during network training and predict storm surges for all 28 TGs simultaneously. CNN-LSTM demonstrated the best performance, exhibiting higher correlation and lower RMSE, while also accurately reproducing the overall variability and pattern of the hourly observed surge series. Results also reveal that the CNN-LSTM model outperforms the existing hydrodynamic and data-driven storm surge models.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Large-Scale Deep Learning-Based Hourly Storm Surge Modeling: Application Across the U.S. Gulf and East Coasts

  • 期刊论文

  • 发表日期：

    2026年1月29日

查看全文
  • Md Abu Zafor,
  • Md Mamunur Rashid

摘要

Storm surge prediction is crucial for coastal zone management and flood risk reduction practices. Physics-based hydrodynamic models are often used for storm surge modeling. However, their inherent computational demands and the need for high-fidelity spatiotemporal data hinder their applicability at regional and continental scales. One alternative is the data-driven modeling of storm surges. However, most studies used data-driven techniques to model daily surges at individual tide gauges. In this study, we develop large-scale storm surge models utilizing Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and a hybrid of CNN and LSTM (CNN-LSTM) algorithms to predict hourly surges at multiple tide gauges simultaneously, trained with atmospheric predictor variables. Observed hourly surges from 28 long-record tide gauges along the U.S. Gulf and East Coasts are used as the predictand. CNN-LSTM was identified as the best model, achieving an average (over 28 TGs) correlation of 0.82 and a root mean square error of 0.08 during the testing period. Furthermore, the best model’s performance in capturing temporal variability, extreme-surge statistics, probability distributions, and surge-hydrograph characteristics (e.g., peak, duration, severity, and intensity) is investigated. Additionally, we demonstrate how the networks leverage plausible physical relationships between storm surge and predictor features during training and prediction. Overall, the CNN-LSTM model achieves up to 78% and 46% improvements in the correlation coefficient and root mean square error, respectively, compared with existing data-driven storm surge models. In contrast, the model yields an average 8% increase in correlation and a 25% reduction in RMSE relative to hydrodynamic model-derived storm surges. The study underscores that the proposed deep learning modeling framework efficiently predicts hourly storm surges simultaneously across the U.S. Gulf and East Coasts. The model can be used for historical reconstruction and for potential future extensions to produce storm-surge projections that facilitate the investigation of future trends, variability, and return levels under different climate-change scenarios.

Graphical Abstract

The graphical abstract summarizes the research, including the study area, data, methodologies, and highlights the main results. The study trains large-scale deep learning models to simultaneously predict hourly storm surges at 28 long-record tide gauges (TG) along the U.S. Gulf and East coasts. Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and their hybrid (i.e., CNN-LSTM) architectures are utilized for model development. Instead of separate predictor domains for each TG, a single large domain is used, from which the DL algorithms automatically extract physically relevant features during network training and predict storm surges for all 28 TGs simultaneously. CNN-LSTM demonstrated the best performance, exhibiting higher correlation and lower RMSE, while also accurately reproducing the overall variability and pattern of the hourly observed surge series. Results also reveal that the CNN-LSTM model outperforms the existing hydrodynamic and data-driven storm surge models.