AI Predicts Coastal Floods in Seconds

Coastal cities worldwide face escalating threats from sea-level rise and flooding, endangering millions and critical infrastructure. Traditional flood prediction methods, while accurate, are slow and computationally intensive, taking days or weeks to simulate scenarios. This delay hampers timely planning for protective measures like seawalls. Researchers have now developed a deep learning model that predicts flood patterns in seconds, offering a rapid tool for city planners to assess risks and design effective coastal defenses.

The key finding is that this AI model, named CASPIAN-v2, significantly outperforms existing methods in accuracy and speed. It reduces the mean absolute error in flood depth predictions by nearly 20% compared to state-of-the-art techniques. By treating flood prediction as an image analysis task, the model learns from data to forecast how water will spread under various sea-level rise and shoreline protection scenarios, such as which areas to fortify with barriers.

Methodologically, the team used a convolutional neural network (CNN) architecture trained on hydrodynamic simulation data from two coastal regions: Abu Dhabi and San Francisco Bay. These simulations, generated with tools like Delft3D, provided ground-truth flood maps under different conditions. The model processes input data on shoreline protections and sea-level rise to output high-resolution flood inundation maps. A key innovation is the use of data augmentation, where random cutouts from the data expand the training set, helping the model generalize to unseen scenarios.

Results show that CASPIAN-v2 achieves high accuracy, with an average R-squared value of 0.9385, indicating it explains most of the variance in flood data. On test sets, it maintained low error rates, such as a mean absolute error of 0.0453 meters, and correctly identified non-flooded areas with over 99% accuracy. The model also demonstrated strong generalization, performing well on holdout datasets with complex protection schemes not seen during training. For instance, it accurately predicted floods in scenarios where only specific segments of the coastline were protected, as shown in Figure 6 of the paper.

In practical terms, this technology enables city planners to quickly evaluate hundreds of coastal defense strategies, something previously infeasible due to computational limits. For example, in Abu Dhabi, simulating one flood scenario with traditional methods takes up to 73 hours, whereas CASPIAN-v2 completes it in under a second. This speed allows for rapid iteration in designing mitigation plans, helping communities adapt to climate change impacts like rising seas and storm surges.

Limitations include the model's dependence on the quality of input simulation data; inaccuracies in terrain or environmental factors could affect predictions. The paper notes that while the model handles various sea-level rise scenarios, it does not incorporate all possible extreme events, such as riverine floods or intense rainfall, leaving room for future enhancements to cover a broader range of hazards.