TL;DR
This paper introduces HIDRA2, a novel deep-learning architecture for highly accurate and computationally efficient sea level and storm tide forecasting in the complex Northern Adriatic region. HIDRA2 significantly outperforms previous deep learning models and state-of-the-art numerical ocean models, particularly for extreme sea level events, by effectively capturing the influence of atmospheric conditions and crucial seiche dynamics.
What Problem Does This Paper Address?
Coastal communities worldwide are increasingly vulnerable to rising sea levels and more frequent extreme events like storm surges, exacerbated by climate change. The Northern Adriatic, a shallow, semi-enclosed basin, is particularly susceptible to coastal flooding and navigation disruptions due to its unique bathymetry and the interplay of tides, atmospheric forcing, and basin-specific oscillations called seiches. Accurately forecasting sea levels, especially during storm events, is critical for civil safety, marine traffic, and coastal flood response. While traditional numerical ocean models exist, they are computationally expensive and can struggle with the complex dynamics of this region. Previous deep learning approaches, like HIDRA1, showed promise but still fell short in predicting extreme sea levels. This paper aims to bridge this gap by developing a more advanced deep learning model that is both accurate and computationally efficient.
Methods
The core of this work is the HIDRA2 architecture, a deep neural network designed for sea surface height (SSH) prediction. Key innovations include:
- Novel Feature Encoders: HIDRA2 employs specialized encoders for atmospheric data (wind and pressure), tidal signals, and past SSH observations. These encoders are designed to extract relevant spatial and temporal patterns.
- Atmospheric Encoder: This component processes historical atmospheric data (wind and pressure) through a two-stage process. The first stage uses convolutional layers and a "prototype matching" layer to capture spatial features within short time windows. The second stage uses 1D convolutions and residual blocks to entangle temporal information across longer periods.
- Tidal and SSH Encoders: These encoders process future tidal forecasts and past SSH/tide data, respectively, using a similar convolutional and residual block structure.
- Feature Fusion and Regression: A crucial "fusion-regression block" recalibrates and combines the features from all encoders. It then injects past SSH information and uses fully connected layers with residual connections to predict the full SSH for the next 72 hours.
- Predicting Full SSH: Unlike its predecessor (HIDRA1), HIDRA2 directly predicts the full SSH, not just the residual (difference from astronomical tide). This is found to be beneficial, especially for storm events, likely due to capturing non-linear interactions between tides and storm surges.
- Ablation Studies: Extensive ablation studies were conducted to systematically remove components (encoders, input data types) to quantify their individual importance and validate design choices.
Datasets
- Training Data:
- Sea level observations from the Koper Mareographic Station (2006-2018), processed into hourly data points.
- Atmospheric input data (10m winds and mean sea level pressure) from a single member of the ECMWF Ensemble Prediction System (2006-2018), spatially cropped to the Adriatic basin and linearly interpolated to hourly time steps.
- Evaluation Data:
- A completely independent dataset from June 1, 2019, to December 31, 2020.
- This includes ECMWF atmospheric predictions and Koper sea level observations.
Key Insights and Ideas
- Seiche Dynamics are Crucial: The paper highlights that accurately capturing the Adriatic basin's seiche modes (fundamental and first excited) is vital for precise storm tide forecasting. HIDRA2 demonstrates a superior ability to represent the energy content of these seiche bands compared to other models.
- Full SSH Prediction Advantage: Predicting the complete sea surface height, rather than just the residual (deviation from astronomical tide), significantly improves performance during storm events. This suggests the model learns to implicitly handle complex non-linear interactions between tides and storm surges.
- Atmospheric Forcing Dominates: Ablation studies reveal that the atmospheric encoder provides the most critical predictive information, underscoring the importance of accurate wind and pressure forecasts for sea level prediction.
- Deep Learning for Geophysical Phenomena: HIDRA2 exemplifies how deep learning, when combined with geophysical understanding, can create highly effective and computationally efficient models that rival or surpass traditional numerical models for complex phenomena like storm tides.
Key Results
- Overall Accuracy: HIDRA2 achieved a 13% reduction in overall Mean Absolute Error (MAE) compared to HIDRA1.
- Storm Event Accuracy: For storm events, HIDRA2's MAE was reduced by a significant 25% compared to HIDRA1.
- Extreme Event Performance: HIDRA2 demonstrated superior performance in both the low sea level tail (important for marine traffic) and the high sea level tail (critical for flood response).
- Comparison to Numerical Models: HIDRA2 outperformed both the NEMO and SCHISM numerical ocean models in terms of MAE, RMSE, and accuracy across all sea level bins and forecast lead times.
- Flood Detection: HIDRA2 achieved a high F1 score (87.59%) for flood event detection, balancing precision and recall better than NEMO and HIDRA1.
- Seiche Representation: HIDRA2 most accurately represents the energy density peak of the ground state Adriatic basin seiche and is a close second to SCHISM for the first excited eigenmode.
- Computational Efficiency: HIDRA2 inference is extremely fast, taking only 4 ms for a 72h forecast on a single GPU, compared to the computationally intensive nature of numerical models.
Limitations and Future Work
- Single Location Focus: The current HIDRA2 model is trained and evaluated for a single location (Koper). The authors suggest exploring multi-point prediction setups using data from multiple tide gauges to improve spatial accuracy.
- Real-time Data Integration: Future work could incorporate real-time in situ measurements (e.g., synoptic observations, satellite scatterometer data) to further enhance model performance.
- Generalization: Investigating HIDRA2's performance in other Mediterranean locations or similar semi-enclosed basins (e.g., Baltic Sea, Red Sea) is proposed to assess its generalization capabilities.
- Understanding High-Frequency Errors: While HIDRA2 excels in seiche and tidal bands, errors in higher-frequency bands (below 10.9h) during specific storm events (like the November 2019 Venice flood) warrant further research, potentially linked to transient, localized atmospheric phenomena.
Summary updated