the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 27 Mar 2025
Climate Adaptation-Aware Flood Prediction for Coastal Cities Using Deep Learning
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 prohibitive 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
- Preprint
(44493 KB) - Metadata XML
-
Supplement
(4405 KB) - BibTeX
- EndNote
Status: open (until 08 May 2025)
-
RC1: 'Comment on egusphere-2025-838', Anonymous Referee #1, 25 Apr 2025
reply
The manuscript proposes a novel deep learning framework for predicting coastal flooding (only tidal). The framework has been applied to the urban areas of Abu Dhabi and San Francisco. The proposed model performs better in comparison to the existing DL methods. The manuscript is free from grammatical errors and is interesting to read. However, the following serious concerns need to be addressed,
- The developed model is only tuned for tidal flooding, and it does not account for the storm surge, which is a major source of coastal flooding. Especially, San Francisco is vulnerable to flooding by hurricanes. The paper doesn’t mention the idea behind the omission of flooding due to storm surge, and why modelling only tidal flooding is crucial in these example cases. The motivation for setting up this kind of prediction system is very unclear.
- Also, what is the level of damage that can occur due to tidal flooding (as the storm surge is not included here) needs to be discussed so that the readers can understand the significance or the need for this framework.
- The manuscript needs to contain the details of how the hydrodynamic model, whose results have been used for training the DL model, is validated. What kind of events were used? How closely they matched the observed flooding also needs to be included in the paper.
- The paper argues that the developed model is computationally efficient, but it fails to elucidate the details on how fast it is in comparison to the hydrodynamic model. This is very crucial to establish the significance of this work.
- There is no spatial validation of the over-/under-/correctly predicted flooded areas. For example, see Nithila Devi et al. (2024), where a spatial fitness index has been used to assess the spatial accuracy.
- In San Francisco, there is also riverine and pluvial flooding present. How only modelling the tidal flooding sufficient in this case for assessing the flood protection capabilities?
- Flood protection measures on the coast are very vaguely discussed. It is very difficult to understand what they are, how they function, and how they are incorporated into the hydrodynamic and DL models. It will be nice to discuss with figures showing their placement, functionality and model representation explicitly.
- Regarding the writing style, I find that the paper is more from a computer science background, rather than explaining clearly the application and relevance to the hydraulic modelling application. For example, the manuscript lacks details on training data used, accuracy of the training data, spatial prediction accuracy, specifics of protection measures, etc., as mentioned in the earlier comments.
- The paper is very wordy with several irrelevant information and lacks crucial information necessary for the understanding of the readers. For example, information on kind of damages due to tidal flooding, validation etc. Please discuss the important information briefly rather than just citing other works.
The following are the specific comments listed for each section,
Introduction and Related Works
- The introduction is very wordy and unorganised. A separate section of “related works” is unwarranted and has a lot of redundant information. Therefore, sections 1 and 2 (introduction and related works) need to be shortened and merged into one, ensuring that there is a proper flow of content in them.
- Lines 17-20 can be shortened.
- Lines 26 – 34 need to be shortened, retaining information relevant to the kind of study the paper deals with.
- Please use terms like computationally expensive/intensive instead of the term “computationally prohibitive” if possible.
- Mention literature related to computationally efficient modelling approaches (subgrid, parallelisation, simplified models, etc.), such as De Almeida et al. (2013), Neal et al. (2012), Li and Hodges (2019), Sanders and Schubert (2019), Nithila Devi et al. (2024), etc. Discuss such methodologies and bring out the importance of DL/ML/AI in flood forecasting and modelling.
- Highlight the need for high-resolution modelling, especially in the complex urban terrain and cite relevant literature.
- Line 56, what do you mean by high dimensionality of outputs?
- Since this paper expands on a deep-vision based framework, explain briefly about this in the introduction so that the readers from diverse backgrounds can appreciate the work. (Line 57)
- The paper lacks a dedicated and concrete objective description; rather, it mentions the conclusions from the study in the introduction. Please move the lines 61 – 81 to conclusions or discussions, if redundant, considering removing them.
- Lines 83 – 90, redundant information. Please remove them while merging sections 1 and 2.
- Lines 91 – 98, irrelevant information.
- Mention what kind of flood protection strategies exist here.
- Lines 123 – 125, not clear.
- How is the training done for different SLRs?
- Figure 1 can be moved to Methodology and please describe the overall framework and the steps involved there.
Study Area and Data Description
- Line 136, what do you mean by environmental effects?
- Briefly describe OLU discretization in a few sentences for the benefit of readers.
- Mention past flood damages caused by tides.
- Briefly mention here about the training dataset.
- How was the Delft 3D model validated? How accurate is it?
- Rather than points, use line or raster to areas susceptible to flooding. Figure 2a.
- Table 1 – What does main set, hold out set, etc., mean? Please explain when you have a table or figure in the manuscript for a general reader.
- What are Shamal winds? Describe in a sentence or two with relevant literature.
- It is unclear what you mean by three months. Period of simulation or the computational time itself? What is the significance of choosing this?
Method
- Section 4.1, please shorten this to retain relevant general information. The rest can be moved to supplementary so that the readers don’t lose interest.
- Describe a spatial fitness metric.
Results
- Tabulate and discuss the values of the spatial fitness index for different DL/ML models.
- Table 5, please describe the performance of the proposed method compared to the existing methods in the paper.
- Please use a zoomed-in figure to illustrate the effect of the representation of the protection measure in the manuscript.
- Section 6.2.2, Lines 475 – 478, lacks a clear explanation.
- Section 6.2.4, a repeated heading, please be more specific and clearer.
- Line 494, how is the ground truth information collected? What is the associated error?
- Line 506, summarize in a few sentences about data augmentation.
- Figure 9, please move it to the supplementary.
References
Bates, P. D., Horritt, M. S., & Fewtrell, T. J. (2010). A simple inertial formulation of the shallow water equations for efficient two-dimensional flood inundation modelling. Journal of Hydrology, 387(1–2), 33–45. https://doi.org/10.1016/j.jhydrol.2010.03.027
De Almeida, G. A. M., & Bates, P. (2013). Applicability of the local inertial approximation of the shallow water equations to flood modeling. Water Resources Research, 49(8), 4833–4844. https://doi.org/10.1002/wrcr.20366
Li, Z., & Hodges, B. R. (2019). Modeling subgrid-scale topographic effects on shallow marsh hydrodynamics and salinity transport. Advances in Water Resources, 129, 1–15. https://doi.org/10.1016/j.advwatres.2019.05.004
Neal, J., Schumann, G., & Bates, P. (2012). A subgrid channel model for simulating river hydraulics and floodplain inundation over large and data sparse areas. Water Resources Research, 48(11), 1–16. https://doi.org/10.1029/2012WR012514
Nithila Devi, N., & Kuiry, S. N. (2024). A novel local‐inertial formulation representing subgrid scale topographic effects for urban flood simulation. Water Resources Research, 60(5), e2023WR035334.
Sanders, B. F., & Schubert, J. E. (2019). PRIMo: Parallel raster inundation model. Advances in Water Resources, 126, 79–95. https://doi.org/10.1016/J.ADVWATRES.2019.02.007
Citation: https://doi.org/10.5194/egusphere-2025-838-RC1 -
RC2: 'Comment on egusphere-2025-838', Anonymous Referee #2, 28 Apr 2025
reply
Anonymous Referee #2
Summary
This manuscript presents CASPIAN-v2, a novel deep learning framework for predicting coastal flooding under varying sea level rise (SLR) scenarios and shoreline protection strategies. The authors test their approach on two distinct geographical regions (Abu Dhabi and San Francisco) and demonstrate superior performance compared to state-of-the-art methods. The paper makes several contributions, including a new CNN-based architecture, comprehensive datasets from vulnerable coastal areas, and validation of generalizability across different SLR scenarios.
While the work represents a significant advancement in data-driven coastal flood prediction, several aspects require substantial revision before the manuscript is suitable for publication in HESS.
General Comments
- Model Architecture Presentation and Justification
The CASPIAN-v2 architecture is sophisticated but presented in an overly complex manner that hampers understanding. Figure 4 contains too much information without sufficient explanation of design choices. The authors introduce multiple novel components (MARX blocks, SEE blocks) without adequate justification for these specific innovations over simpler alternatives.
Example: In Section 4.1.2, the rationale for integrating ResNeXt blocks with CBAM is not clearly connected to the specific challenges of flood prediction. The authors should explain why this combination addresses spatial dependencies in coastal flooding better than other attention mechanisms.
Recommendation: Provide a simplified schematic of the architecture alongside the detailed one and clearly justify each novel component in relation to the specific requirements of flood prediction tasks.
- Computational Efficiency Analysis
A primary motivation for developing surrogate models is the computational burden of physics-based simulations. However, the paper lacks a rigorous comparison of computational efficiency between the proposed model and alternatives.
Example: While Table 5 thoroughly compares prediction accuracy, it contains no information about training times, inference times, or memory requirements. This is particularly important given that lines 45-49 on page 2 emphasize computational burden as a key limitation of current approaches.
Recommendation: Include a comprehensive analysis of computational efficiency, comparing training and inference times across all evaluated models, and explicitly stating the practical time savings compared to hydrodynamic simulations.
- Uncertainty Quantification
The model provides deterministic predictions without addressing prediction uncertainties, which is crucial for risk assessment and decision support in coastal planning.
Example: The error maps in Figures 5-7 show where predictions differ from ground truth, but they don't indicate the model's confidence in its predictions, which is essential for reliable risk assessment.
Recommendation: Incorporate uncertainty quantification into the model (e.g., through ensemble methods, Bayesian techniques, or prediction intervals) or thoroughly discuss this limitation and its implications for practical use.
- Data Imbalance Handling
Figure 9 reveals severe class imbalance in the dataset, with non-inundated areas predominating. While the authors acknowledge this challenge, they don't adequately explain how their approach specifically addresses it.
Example: Section 7 mentions the imbalance issue but doesn't describe specific techniques beyond the hybrid loss function that were employed to mitigate its effects. It's unclear how the model achieves its reported high accuracy despite this challenge.
Recommendation: Elaborate on specific techniques used to address data imbalance, potentially including specialized sampling strategies, data augmentation approaches tailored to rare flood events, or custom components in the architecture designed for imbalanced spatial data.
- Real-world Application Context
The practical utility of the model for coastal planning is asserted but not demonstrated through concrete examples or integration pathways.
Example: The conclusion claims CASPIAN-v2 is "an essential tool for coastal resilience planning" (lines 539-540, page 28), but doesn't provide specific guidance on how planners might integrate this tool with existing decision-making frameworks.
Recommendation: Include a case study or conceptual workflow showing how the model could be integrated into actual coastal planning processes, identifying key stakeholders and decision points where the model adds value.
Specific Comments
- Mathematical Notation Inconsistency
The paper uses inconsistent notation, particularly in Section 4.1, making the mathematical formulations difficult to follow.
Example: In Equations 1-5, subscripts sometimes denote indices and sometimes represent different variables entirely. The relationship between tensors across equations is not always clear.
Recommendation: Standardize notation throughout the paper and provide a notation table for reference.
- Evaluation Metrics Justification
While the paper employs multiple evaluation metrics, the rationale for these specific choices and their relevance to practical flood prediction applications isn't fully explained.
Example: The threshold exceedance metric (δ > Δ) is introduced in Section 5.4, but its practical significance for flood risk assessment isn't discussed.
Recommendation: Justify the choice of each evaluation metric in terms of its relevance to practical flood prediction applications and decision-making contexts.
- Data Preprocessing Details
The data preprocessing section (3.3) lacks sufficient detail on critical aspects that could impact model performance.
Example: The method for mapping inundation coordinates onto a 1024×1024 grid (lines 182-184, page 9) is mentioned but not described in detail, despite this being a critical step that affects the spatial resolution of predictions.
Recommendation: Provide more detailed explanation of preprocessing steps, potentially with illustrative examples showing the transformation from raw data to model inputs.
- Ablation Study Presentation
The paper mentions ablation studies in the supplementary material but doesn't adequately summarize key findings in the main text.
Example: Line 341-342 on page 16 mentions "extensive ablation studies" but doesn't present the key insights derived from these experiments.
Recommendation: Include a summary table of ablation study results in the main text, highlighting the contribution of each novel component to overall performance.
- Figure Clarity and Interpretation
Several figures are complex and difficult to interpret, with insufficient explanation in captions and text.
Example: Figure 7 compares model predictions across different approaches, but the subtle differences between models are difficult to discern with the chosen color scale and presentation format.
Recommendation: Improve figure clarity through better color scales, simplified presentations, or additional explanatory elements like difference maps to highlight where each model performs better or worse.
Summary Assessment
This manuscript presents valuable research on deep learning for coastal flood prediction, with promising results that could significantly advance the field. However, major revisions are needed to address issues related to model architecture presentation, computational efficiency analysis, uncertainty quantification, data imbalance handling, and real-world application context.
With these improvements, the paper has the potential to make a significant contribution to both the technical literature on deep learning for environmental modeling and practical coastal planning applications.
Citation: https://doi.org/10.5194/egusphere-2025-838-RC2
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 112 | 24 | 6 | 142 | 12 | 6 | 5 |
- HTML: 112
- PDF: 24
- XML: 6
- Total: 142
- Supplement: 12
- BibTeX: 6
- EndNote: 5
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1