Climate Adaptation-Aware Flood Prediction for Coastal Cities Using Deep Learning

TL;DR

This work develops CASPIAN-v2, a lightweight CNN surrogate for rapid coastal flood prediction under multiple sea-level-rise and shoreline-adaptation scenarios. By framing flood mapping as an image-to-image task and incorporating SLR context through SEE blocks and a MARX-based bottleneck, the model achieves high spatial accuracy and strong generalization across Abu Dhabi and San Francisco Bay while dramatically reducing computational cost relative to physics-based Delft3D simulations. The paper introduces two region-specific datasets, a hybrid loss to handle data skew, and extensive ablations demonstrating the benefits of MARX, SEE, and end-to-end training plus fine-tuning for unseen SLR levels. Additional contributions include explainable AI insights via Grad-CAM, a deep-ensembles uncertainty framework, and open-source release of data and code. Overall, CASPIAN-v2 offers a practical, scalable tool for coastal flood management that can accelerate reconnaissance of protection strategies while enabling detailed verification with high-fidelity simulators when needed.

Abstract

Climate change and sea-level rise (SLR) pose escalating threats to coastal cities, intensifying the need for efficient and accurate methods to predict potential flood hazards. Traditional physics-based hydrodynamic simulators, although precise, are computationally expensive and impractical for city-scale coastal planning applications. Deep Learning (DL) techniques offer promising alternatives, however, they are often constrained by challenges such as data scarcity and high-dimensional output requirements. Leveraging a recently proposed vision-based, low-resource DL framework, we develop a novel, lightweight Convolutional Neural Network (CNN)-based model designed to predict coastal flooding under variable SLR projections and shoreline adaptation scenarios. Furthermore, we demonstrate the ability of the model to generalize across diverse geographical contexts by utilizing datasets from two distinct regions: Abu Dhabi and San Francisco. Our findings demonstrate that the proposed model significantly outperforms state-of-the-art methods, reducing the mean absolute error (MAE) in predicted flood depth maps on average by nearly 20%. These results highlight the potential of our approach to serve as a scalable and practical tool for coastal flood management, empowering decision-makers to develop effective mitigation strategies in response to the growing impacts of climate change. Project Page: https://caspiannet.github.io/

Figures (17)

• Figure 1: AD study area shown on the map of the United Arab Emirates. (a) All areas susceptible to flooding under a 0.5 m SLR scenario without any shoreline protections (b) The 17 OLUs defined along the AD shoreline where protections are to be tested for their effectiveness.
• Figure 2: SF Bay study area shown on the map of the United States (a) All areas susceptible to flooding under a 0.5 m SLR scenario without any shoreline protections (b) The 30 OLUs defined along the SF Bay shoreline where protections are tested for their effectiveness.
• Figure 3: An overview of the proposed framework for coastal flood prediction. It begins with hydrodynamic simulations based on SLR data and coastal protection scenarios to generate raw flood data, which is then processed into spatial flood maps. The CASPIAN-v2 model, trained on these maps, predicts inundation patterns and flood extent. The framework can be fine-tuned with new data for improved adaptability. The different colored paths represent training (red), inference (green), and fine-tuning (blue) stages.
• Figure 4: A simplified schematic of the CASPIAN-v2 model architecture. The model consists of an encoder Stage that uses feature extraction (FE) blocks to create a compressed representation of the input map. At the Bottleneck, a series of multi-attention ResNeXt (MARX) blocks refine these features. The decoder stage then uses feature reconstruction (FR) blocks to generate the high-resolution output flood map. Crucially, SLR data is integrated into the decoder via the SLR-enhanced encoding (SEE) blocks and again before the final output, allowing the model to produce predictions conditioned on different climate scenarios.
• Figure 5: Evaluation of CASPIAN-v2 on the test datasets. (a) Ground truth inundation maps for representative AD and SF scenarios. (b) Predicted inundation values. (c) Absolute error distributions of predicted inundation values. Darker shades of blue indicate higher absolute errors, ranging from near 0% to greater than 25%. The magenta insets provide zoomed-in views of specific OLUs to illustrate the effect of protection measures. For instance, the inundation is shown to be minimal inland of the protected OLU-17 in AD, whereas significant flooding occurs near the unprotected OLU-20, a dynamic that the model precisely captures.