TL;DR
This paper introduces a novel multi-output Gaussian Process (GP) model that can simultaneously learn from time-varying input data (like storm surge and wave height) and predict spatially varying flood impacts (like inland water levels). This approach significantly improves coastal flood hazard assessment by directly incorporating the functional nature of both inputs and outputs, leading to more accurate and faster predictions crucial for early warning systems.
What Problem Does This Paper Address?
Accurate prediction of coastal flooding is critical for disaster preparedness and mitigation. Traditional methods often rely on complex and computationally expensive hydrodynamic simulators that can take days to run, making them impractical for real-time forecasting. Existing surrogate models, which aim to speed up predictions, often simplify crucial input data (like time-series storm data) into single numbers or fail to capture the spatial variability of flood impacts. This simplification can lead to misleading predictions, especially in complex coastal systems where both the temporal evolution of environmental conditions and the spatial extent of flooding are vital. The authors aim to bridge this gap by developing a surrogate model that can handle functional (time-series) inputs and predict functional (spatial maps) outputs, offering a more faithful representation of the underlying physics.
Methods
The core of this work is a multi-output Gaussian Process (GP) model enhanced to handle functional data. The key innovation lies in the construction of a separable kernel that explicitly models correlations between:
- Functional inputs: Capturing how time-varying meteorological and oceanographic conditions influence flooding.
- Spatial locations: Accounting for how flood impacts vary across different points inland.
This separable kernel allows the model to learn from the rich temporal information in the input data and produce spatially resolved flood predictions. To manage the computational complexity, especially with large datasets, the authors leverage Kronecker-structured computations and sparse-variational approximations. They also explore Principal Component Analysis (PCA) for dimensionality reduction of functional inputs. The framework is implemented using both Python (GPflow) and R (kergp) libraries.
Datasets
The primary dataset used is a collection of hydrometeorological forcing conditions and corresponding maximal inland water height (Hmax) maps generated by the BRGM and IMT for the town of Gâvres, France. This dataset comprises 131 scenarios of functional inputs (time series of mean sea level, tide, surge, wave height, wave period, wave direction, wind speed, and wind direction) and corresponding spatial flood maps. The data was generated using the SWASH model and covers historical flood events and numerous synthetic scenarios.
Key Insights and Ideas
- Functional Inputs Matter: Treating time-series data (like storm surge) as functions, rather than scalars, significantly improves prediction accuracy by encoding more physical information into the model.
- Spatial Outputs are Crucial: Coastal flooding is inherently spatial. The model's ability to predict spatially varying flood impacts (maps of Hmax) is essential for understanding the true hazard.
- Separable Kernels Bridge Time and Space: The proposed separable kernel is the key to linking functional inputs to spatial outputs, allowing for a unified modeling approach.
- Computational Efficiency is Achievable: By employing Kronecker structures and sparse approximations, the GP model can scale to handle complex problems with many outputs and spatial points, making it practical.
- Forecasting Potential: The framework is well-suited for forecasting, as demonstrated by its ability to predict unobserved flood events with increasing accuracy as more learning scenarios become available.
Key Results
- Improved Prediction Accuracy: The functional multi-output GP (fMoGP) consistently outperformed standard multi-output GPs (MoGP) that do not explicitly handle functional inputs, showing higher Q2 values (a measure of predictive accuracy) in various synthetic and real-world coastal flooding scenarios. For instance, in the coastal flooding application, the fMoGP achieved Q2 values of approximately 95.9% compared to 93.2% for the standard MoGP.
- Efficient Computation: Predictions for a single flood map can be generated in minutes, a stark contrast to the days required by traditional hydrodynamic simulators. This speedup is critical for operational forecasting.
- Scalability: The model demonstrated good performance even with a large number of spatial design points (up to 1003) and outputs (up to 1001).
- Forecasting Performance: As the number of learning scenarios increased, the model's predictive accuracy (Q2) improved, reaching around 90% with 200 learning events for predicting unobserved flood events.
Limitations and Future Work
- Data Availability: The performance of the model is sensitive to the availability and diversity of the learning dataset. More diverse and richer datasets lead to better predictability.
- Non-negativity and Zero-Inflation: The current GP assumptions do not inherently handle the non-negative and zero-inflated nature of flood data (e.g., areas with no flooding). Alternative nonstationary GP priors could address this.
- Spatial Output Projection: The authors suggest projecting spatial outputs onto basis representations (like wavelets) as a future direction to further improve computational efficiency, especially for very large datasets.
- Incorporating Prior Information: Adapting the framework to learn critical forcing conditions that exceed defense crests, rather than just predicting water levels, is a potential avenue for more robust flood defense analysis.
Summary updated