A Comparative Study of Convolutional and Recurrent Neural Networks for Storm Surge Prediction in Tampa Bay

TL;DR

This study benchmarked three deep learning architectures—CNN-LSTM, LSTM, and 3D-CNN—for surrogate storm surge prediction in the Tampa Bay region using high-resolution atmospheric forcing from CFSv2 and tide-gauge observations. By standardizing data inputs, training, and evaluation, it quantified generalization performance and extreme-event robustness, including a Hurricane Ian case with a pronounced negative surge. CNN-LSTM emerged as the most reliable model, achieving a test loss of 0.010 and an R2 of 0.84, while LSTM overfitted to training data and 3D-CNN showed stability issues but competitive overall metrics. The results suggest that hybrid spatiotemporal architectures like CNN-LSTM are promising for operational storm surge surrogates, with potential for further gains from seq2seq and expanded input strategies.

Abstract

In this paper, we compare the performance of three common deep learning architectures, CNN-LSTM, LSTM, and 3D-CNN, in the context of surrogate storm surge modeling. The study site for this paper is the Tampa Bay area in Florida. Using high-resolution atmospheric data from the reanalysis models and historical water level data from NOAA tide stations, we trained and tested these models to evaluate their performance. Our findings indicate that the CNN-LSTM model outperforms the other architectures, achieving a test loss of 0.010 and an R-squared (R2) score of 0.84. The LSTM model, although it achieved the lowest training loss of 0.007 and the highest training R2 of 0.88, exhibited poorer generalization with a test loss of 0.014 and an R2 of 0.77. The 3D-CNN model showed reasonable performance with a test loss of 0.011 and an R2 of 0.82 but displayed instability under extreme conditions. A case study on Hurricane Ian, which caused a significant negative surge of -1.5 meters in Tampa Bay indicates the CNN-LSTM model's robustness and accuracy in extreme scenarios.

Figures (7)

• Figure 1: Location and Extent of Wind data used in the study. (a) The map shows the geographical location of Tampa Bay and the extent of the atmospheric data collected for the study spanning longitudes -82 to -85 and latitudes 26 to 29. (b) The location of the NOAA 8726520 St. Petersburg, FL station within Tampa Bay.
• Figure 2: The CNN-LSTM architecture processes atmospheric data using a combination of CNN layers to capture spatial features and LSTM layers to process sequential dependencies.
• Figure 3: The LSTM Architecture utilizes LSTM layers to process both atmospheric and tidal data.
• Figure 4: The 3D-CNN Architecture employs 3D convolutional layers to capture both spatial and temporal features directly.
• Figure 5: Training and Validation Loss Curves for CNN-LSTM, LSTM, and 3D-CNN Models. The plot shows the training and validation losses over 15 epochs for each model.