TLDR: A new deep learning model, CASPIAN-v2, provides fast and accurate predictions of coastal flooding under various sea-level rise and shoreline protection scenarios. It significantly outperforms traditional methods and other deep learning models, reducing prediction errors by nearly 20%, and can generalize across different geographic regions like Abu Dhabi and San Francisco. This makes it a practical tool for urban planners to quickly assess flood risks and design effective adaptation strategies, drastically cutting down the computational time from months to seconds compared to traditional physics-based simulations.
Coastal cities around the globe face an increasing threat from climate change and rising sea levels, making accurate and efficient flood prediction more critical than ever. Traditional methods, while precise, are often too slow and computationally expensive for large-scale urban planning. A new deep learning model, CASPIAN-v2, offers a promising solution, providing rapid and accurate coastal flood predictions under various sea-level rise (SLR) and shoreline adaptation scenarios.
Developed by researchers Bilal Hassan, Areg Karapetyan, Aaron Chung Hin Chow, and Samer Madanat, this novel, lightweight Convolutional Neural Network (CNN)-based model significantly outperforms existing state-of-the-art methods. It has been shown to reduce the mean absolute error (MAE) in predicted flood depth maps by nearly 20% on average. Crucially, the model demonstrates the ability to generalize across diverse geographical contexts, having been successfully tested using datasets from two distinct regions: Abu Dhabi and San Francisco.
The CASPIAN-v2 model addresses key challenges in deep learning for flood prediction, such as data scarcity and the need for high-dimensional outputs. It leverages a vision-based framework, treating flood prediction as a computer vision task. This involves transforming discrete shoreline protection scenarios into 2D spatial input maps, allowing the CNN to learn the complex geometric relationships between protected areas and resulting flood patterns. The framework also employs data augmentation techniques, like random cutouts, to expand limited training datasets, a vital advantage in data-scarce environments.
The architecture of CASPIAN-v2 is designed for robustness and efficiency. It features an encoder to extract hierarchical features, a bottleneck stage with novel multi-attention ResNeXt (MARX) blocks to refine these features, and a decoder that reconstructs high-resolution flood maps. A key innovation is the integration of SLR data directly into the decoder through specialized SLR-Enhanced Encoding (SEE) blocks, allowing the model to condition its predictions on different climate scenarios. Furthermore, a custom hybrid loss function, combining Huber, Log-Cosh, and Quantile losses, helps the model handle outliers and the imbalanced nature of flood data, where non-inundated areas vastly outnumber flooded ones.
The model’s performance was rigorously evaluated against a suite of traditional machine learning and deep learning models. While traditional methods were faster to train, they lacked the ability to capture complex spatial patterns. CASPIAN-v2, however, demonstrated superior accuracy across various metrics, including average mean absolute error (AMAE), average root mean square error (ARMSE), and the Dice Similarity Coefficient (DSC), which measures the spatial overlap between predicted and true flood extents. For instance, it achieved a DSC of 0.8437, a 31.05% improvement over the best traditional machine learning model.
One of the most significant advantages of CASPIAN-v2 is its computational efficiency. Generating a single flood scenario using traditional physics-based hydrodynamic simulators like Delft3D can take anywhere from 3.5 to 73 hours, depending on the region and complexity. In contrast, CASPIAN-v2 can predict outcomes for 72 scenarios in just under 16 seconds. This monumental reduction in computational time transforms a months-long endeavor into a near-instantaneous task, making it a practical and scalable tool for real-world coastal planning.
Beyond its quantitative performance, the model also offers valuable qualitative insights. Visual comparisons show that CASPIAN-v2’s predictions closely align with ground truth inundation maps, accurately capturing localized flooding effects and sharp inundation boundaries even in complex, unseen scenarios. Explainable AI techniques, such as Grad-CAM analysis, reveal that the model focuses its attention on vulnerable, unprotected shoreline segments, empirically validating its decision-making process and fostering trust among decision-makers.
The researchers also implemented a deep ensemble method to quantify predictive uncertainty, showing that the model effectively identifies regions where its predictions are less reliable. This self-awareness is invaluable for coastal planners, allowing them to trust high-certainty predictions for general assessments while flagging high-uncertainty zones for further, more detailed study.
Also Read:
- Improving Temperature Downscaling with a Physics-Informed Diffusion Model
- Enhancing Spatiotemporal Prediction with Historical Data Guidance
In conclusion, CASPIAN-v2 represents a significant advancement in flood prediction, offering a robust, adaptable, and interpretable approach to address the complexities of diverse geographical regions, protection scenarios, and climate variability. It is poised to become an essential tool for coastal resilience planning, empowering decision-makers, engineers, and legislators to develop effective mitigation strategies in response to the growing impacts of climate change and sea-level rise. For more details, you can refer to the full research paper here.
