TL;DR
This paper introduces a novel deep learning approach, the Convolutional Temporal Attention Network (U-TAE), for predicting pixel-wise flood water depth maps with a 2-hour lead time. By leveraging a new large-scale dataset called SwissFlood, the model demonstrates strong generalization capabilities and achieves a mean absolute error as low as 27 cm for water depths exceeding 1 meter, outperforming baseline models.
What Problem Does This Paper Address?
Accurate and timely flood forecasting is crucial for disaster management and early warning systems. Traditional physically-based hydrodynamic models, while effective, are computationally intensive, making real-time predictions challenging. Existing data-driven approaches often focus on predicting maximum water depth rather than detailed spatial maps, or they require extensive manual calibration for each specific scenario. This paper aims to bridge this gap by developing a data-driven model that can rapidly predict high-resolution, pixel-wise water depth maps for entire catchments with a useful lead time, while also generalizing to unseen catchments and rainfall events.
Methods
The core of this research is a deep learning model called the U-TAE (U-net with Temporal Attention Encoder). This architecture is an adaptation of a spatio-temporal encoder-decoder network originally designed for image sequence segmentation. It combines multi-scale spatial convolutions with temporal attention mechanisms. The temporal attention allows the model to focus on salient spatio-temporal features across different resolutions, enabling it to capture complex patterns in the data.
The U-TAE model takes three primary inputs:
- A digital elevation model (DEM) of the catchment.
- The current water depth at each pixel (time $t$).
- Rainfall intensity values for the next two hours, discretized into 10-minute intervals.
The model outputs a pixel-wise water depth map for the catchment at a 2-hour lead time ($t + 120$). A key innovation is the use of the Nash-Sutcliffe efficiency (NSE) metric, a standard hydrological performance indicator, as the loss function during training. This is a departure from typical regression loss functions like L1 or L2, aiming to better align the model's optimization with hydrological performance.
Datasets
The study introduces a new, large-scale dataset called SwissFlood. This dataset comprises:
- 100 catchments in Switzerland, representing urban, rural, and mixed types.
- Digital Elevation Models (DEMs) with 2m spatial resolution for these catchments.
- 1500 simulated rainfall events, derived from 380 real rainfall observation events collected by MeteoSwiss between 1981 and 2021. These events were augmented with small random perturbations to ensure robustness.
- Reference water depth maps generated by simulating these rainfall events using the CADDIES 2D cellular automata flood model.
The SwissFlood dataset is publicly available. The catchments are split into training (72), validation (18), and testing (10) sets to evaluate the model's generalization capabilities.
Key Insights and Ideas
- NSE as a Loss Function: Adapting the Nash-Sutcliffe Efficiency (NSE) as the loss function for training a deep learning model for flood prediction is a significant contribution. This metric is directly relevant to hydrological model evaluation and appears to lead to better performance, especially for deeper water levels, compared to standard L1 and L2 losses.
- Generalization is Key: The paper strongly emphasizes and evaluates the model's ability to generalize to unseen catchments and unseen rainfall events. This is a critical requirement for operational flood forecasting systems, where models need to perform reliably in novel situations.
- Spatio-Temporal Attention: The use of temporal attention mechanisms within a U-Net-like architecture allows the model to effectively learn complex interactions between spatial features (DEM, current water depth) and temporal dynamics (rainfall over time), leading to more accurate predictions.
- Data-Driven Simplicity: The model requires only three fundamental inputs (DEM, current water depth, rainfall forecast), significantly simplifying the data requirements compared to complex physically-based models that need detailed hydrological and land-use information.
Key Results
- Outperforms Baselines: The proposed U-TAE model consistently outperforms three baseline convolutional neural network (CNN) models (trained with L1, L2, and NSE losses) across various water depth ranges and for both seen and unseen catchments.
- Low Absolute Error: The model achieves a mean absolute error (MAE) as low as 0.75 cm for water depths up to 5 cm in seen catchments and 1.18 cm for unseen catchments. Crucially, for water depths exceeding 1 meter, the MAE is 26.86 cm for seen catchments and 67.76 cm for unseen catchments, with an overall absolute error of 27 cm for severe flooding.
- Qualitative Accuracy: Visualizations show that the model can accurately reconstruct flood inundation patterns, although some underestimation of high water depth areas and smoothing effects at sharp depth changes were observed.
- Impact of Loss Function: Ablation studies confirm that the NSE loss function leads to better performance than L1 and L2 losses, particularly for higher water depth ranges.
Limitations and Future Work
- Lead Time Dependency: The current model is trained for a specific 2-hour lead time. For different lead time requirements, the model would need to be retrained. The authors suggest future work could involve designing networks that output predictions for multiple lead times simultaneously.
- Deep Water Accuracy: High absolute errors are observed for very deep water. This is attributed to a potential lack of sufficient training samples with extreme depths. Future work could involve designing loss functions that give more weight to deep water predictions or collecting more data for these scenarios.
- Spatially Uniform Rainfall Assumption: The model assumes spatially uniform rainfall, which is a simplification. Integrating spatially variable rainfall data could improve accuracy.
- Simulator Dependency: The model's performance is inherently bounded by the accuracy of the CADDIES flood model used for generating the training data. Independent validation of CADDIES with in-situ data was not feasible for the large dataset.
- Extrapolation Risks: Like all data-driven models, the U-TAE may produce physically inconsistent or implausible predictions when encountering inputs that significantly diverge from the training data distribution, a limitation not present in physics-based models.
- Additional Factors: Future improvements could incorporate other factors influencing flood dynamics, such as land use, soil characteristics, backwater effects, and storm water systems, by integrating them as new input layers.
Summary updated