Review article: Harnessing Machine Learning methods for climate multi-hazard and multi-risk assessment

Ferrario, Davide Mauro; Sanò, Marcello; Maraschini, Margherita; Critto, Andrea; Torresan, Silvia

doi:10.5194/egusphere-2025-670

Preprints

https://doi.org/10.5194/egusphere-2025-670

Preprints

03 Mar 2025

| 03 Mar 2025

Review article: Harnessing Machine Learning methods for climate multi-hazard and multi-risk assessment

Davide Mauro Ferrario, Marcello Sanò, Margherita Maraschini, Andrea Critto, and Silvia Torresan

Abstract. In recent years, interest in data-driven methods, such as machine learning and multivariate statistics for multi-hazard and multi-risk assessment has surged, due to their ability to integrate vast amounts of data in modelling complex non-linear relationships between hazard and risk factors. This review explores data-driven methods in climate multi-hazard and risk analysis, focusing on four themes: (i) data processing and collection; (ii) hazard identification, prediction and analysis; (iii) risk analysis; and (iv) future risk scenarios under climate change. Key findings highlight the extensive use of machine learning to combine Earth observations and climate data for downscaling and land use and land cover characterisation; the application of deep learning for hazard prediction; the use of ensemble methods for risk analysis; and the growing emphasis on explainable AI frameworks. Training of supervised machine learning approaches on past impacts to model future risk through climate projections also emerged as a significant area. Future research should prioritize multi-hazard interactions, particularly triggering and cascading effects, integrate dynamic vulnerability and exposure factors, and address uncertainties associated with using machine learning for extrapolation. Advancements in Earth observations and textual data integration, alongside the development of open-access disaster catalogues, will be crucial for improving multi-risk analyses and supporting AI-driven early warning systems tailored to regional needs.

Received: 13 Feb 2025 – Discussion started: 03 Mar 2025

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Preprint (PDF, 1879 KB)

Notice on discussion status
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
Preprint (1879 KB)

Download & links

The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.

Journal article(s) based on this preprint

26 Jun 2026

Review article: Harnessing data-driven methods for climate multi-hazard and multi-risk assessment

Davide Mauro Ferrario, Marcello Sanò, Margherita Maraschini, Andrea Critto, and Silvia Torresan

Nat. Hazards Earth Syst. Sci., 26, 2975–3024, https://doi.org/10.5194/nhess-26-2975-2026,https://doi.org/10.5194/nhess-26-2975-2026, 2026

Short summary

Davide Mauro Ferrario, Marcello Sanò, Margherita Maraschini, Andrea Critto, and Silvia Torresan

Interactive discussion

Status: closed

RC1:
'Comment on egusphere-2025-670', Anonymous Referee #1, 31 Mar 2025
This manuscript reviews machine learning (ML) and statistical approaches for climate-related multi-hazard and multi-risk assessment. It is organized around four themes—data processing, hazard prediction, risk analysis, and future scenarios—and incorporates auxiliary methods such as explainable AI and copula modeling. While the topic is timely and relevant, the manuscript has issues in contribution, analytical depth, structural clarity, and language quality.
Major Concerns:
Lack of Scientific Novelty or Conceptual Contribution: The paper lacks scientific novelty and conceptual contribution. It does not introduce new concepts, frameworks, or theoretical insights. Instead, it compiles existing literature without offering a critical synthesis or identifying research gaps. The review does not significantly advance the understanding of multi-hazard or multi-risk modeling compared to previous reviews, and its largely descriptive discussion limits its value as a synthesis resource.

No PRISMA flow diagram is provided, and Search strings, filtering criteria, and quality assessment processes are not disclosed. Suggest including a PRISMA diagram and methodological appendix (even in supplementary materials) to ensure transparency.

Poor Language Quality: The manuscript contains many grammatical errors, awkward phrasing, and redundant or overly long sentences. The language undermines clarity and makes the manuscript difficult to read. Given its role as a review article, this significantly reduces its accessibility and utility to the scientific community. A full professional language revision is essential.

Weak Structure and Inconsistent Framing: The manuscript shifts terminology from "machine learning methods"from the title to "data-driven methods" in the text, without clear justification. The manuscript includes a statistical method (copula), which is not a machine learning method. Poor coherence across sections: Overlap and redundancy between sections, and not clearly defined (e.g., 3.1.2 vs. 3.1.1 regarding satellite images observed soil moisture belongs to EO or climate data).

Additional Comments:
Lines 30–35: “...can advance multi-hazard and multi-risk...” → vague and informal phrasing.

Line 206: “his information…” → typo.

Line 313: “even if...” → incorrect conjunction; use “although” or “even though.”

Line 328: “images form...” → should be “images from.”

Line 335: “3.1 Multi-hazard…3.2.1 Identify...” → unclear subsection formatting.

Line 414: “With regards to” → incorrect phrase; use “With regard to.”

Line 421: “...showing popular results...”
Citation: https://doi.org/10.5194/egusphere-2025-670-RC1
- AC2:
  'Reply on RC1', Davide Mauro Ferrario, 27 May 2025
  We thank the reviewer for the constructive comments, which have helped us improve the clarity, structure, and scientific contribution of the manuscript. Below, we address each of the concerns raised:
  Scientific Novelty and Conceptual Contribution
  
  We respectfully disagree with the assessment that the manuscript lacks scientific novelty. While we do not propose a new algorithm or model, we believe the manuscript makes an original contribution by reframing the fragmented field of data-driven multi-hazard and multi-risk modeling through a structured and forward-looking synthesis. Moreover:
  We'll reframe the four core research questions, which collectively reflect a risk-modeling pipeline, from data processing to hazard prediction, risk assessment, and future scenario modeling. This framing provides a clear conceptual lens through which to interpret recent developments in the field.
  
  We'll focus the synthesis on actionable themes that cut across methods and applications, including:
  
  The integration of Earth Observation (EO) and textual data for multi-risk modeling;
  
  The challenge of model transferability and scalability across regions and hazard types;
  
  The role of model interpretability and explainable AI in machine learning application for multi-risk;
  
  The use of hybrid models (e.g., physics-informed neural networks) to bridge physical realism and data-driven inference.
  
  We'll revise the discussion and conclusion sections to more explicitly identify research gaps, including:
  
  Limited use of ML approaches that incorporate causal reasoning or address cross-sectoral risks;
  
  A need for better spatiotemporal validation techniques in multi-risk contexts;
  
  The importance of uncertainty quantification (and communication), particularly under future climate change scenarios.
  
  We believe this review is among the first to explicitly integrate machine learning and statistical approaches (including copulas) through a multi-hazard lens, and to assess their value for both scientific insight and decision support across multiple scales.
  PRISMA methodology
  
  We appreciate this comment and will add a PRISMA flow diagram and a methodological appendix (now included as supplementary material) to better describe the search strings, screening criteria, and inclusion/exclusion processes, improving the transparency and reproducibility of our literature selection.
  Poor Language Quality
  
  We fully agree that language quality is essential for a review article. In response, the manuscript will undergo a thorough professional-style language revision. We'll address issues related to grammar, structure, and clarity by:
  Simplifying overly long or technical sentences;
  
  Removing redundancy across sections;
  
  Improving flow, particularly in the methods and discussion.
  
  These changes will enhance the readability and accessibility of the manuscript for a broad interdisciplinary audience.
  Weak Structure and Inconsistent Framing
  
  We will clarify our use of the terms "machine learning" and "data-driven methods" in the introduction. The term "data-driven methods" is explicitly defined as an umbrella term that includes both machine learning techniques and other statistical approaches. This choice is motivated by the key role played by non-ML statistical tools (such as copulas) in multi-hazard modeling. We'll revised section titles, transitions, and content to reduce redundancy and improve coherence across the manuscript. In particular the overlap between Sections 3.1.1 and 3.1.2, will be clarified distinguishing between EO-based data sources and climate reanalysis/model data. We hope these revisions address the reviewer’s concerns and demonstrate our commitment to improving the clarity, rigor, and contribution of the manuscript.
  
  Citation: https://doi.org/10.5194/egusphere-2025-670-AC2
RC2:
'Comment on egusphere-2025-670', Anonymous Referee #2, 07 Apr 2025

Major:

1, The manuscript outlines four key research questions, but their link to data-driven climate multi-hazard analysis is not clear. Please clarify how these questions address existing gaps, improve risk assessments, and contribute quantitatively or qualitatively to the field.
2, Data assimilation is a critical component of modern climate analysis yet is missing from the discussion. Include recent developments that demonstrate its role in enhancing multi-hazard/risk assessments.
3, Uncertainty quantification in climate risk studies is necessary for the analysis of hazard/risk. However, they are not discussed.
4, The current focus is on pure data-driven models. Please discuss the emerging hybrid modeling that integrate physical laws, and compare their strengths and limitations.
5, Clarify the differences and complementarities between ML and copula techniques. Using practical examples if possible.
6, Reframe the discussion to emphasize how Earth observation data, combined with ML and copula techniques, leads to improved multi-hazard and risk assessments that benefit decision-makers. Try to rephrase from the perspective of assessments instead of technical comparisons.
Minor:

1, Ensure all citations (e.g., “Linkov et al., 2022” and “S. Yu & Ma, 2021”) are consistently and correctly formatted.
2, ‘understand’ is a strong word for many data-driven models as they do not really capture the underlying generative process. Replace “understand” with “modeling,” “characterization,” or “representation” to better reflect the capabilities of data-driven models.
3, Some grammatical and stylistic edits are suggested for improved readability.
4, Correct the numbering so that the “Multi-hazard” subsection is labeled 3.2 instead of 3.1.

Citation: https://doi.org/10.5194/egusphere-2025-670-RC2
- AC1:
  'Reply on RC2', Davide Mauro Ferrario, 27 May 2025
  We thank Reviewer 2 for the constructive and detailed feedback, which has helped improve the clarity, scope, and scientific depth of our manuscript. Below, we provide a point-by-point response to each of the reviewer’s comments.
  1. Link between the research questions and multi-risk
  We have revised the introduction and clarified the framing of the four research questions to better articulate how they address current gaps in the literature and contribute to advancing data-driven multi-hazard and multi-risk assessment.
  RQ1 focuses on the challenge of data preparation and preprocessing in ML-based hazard modeling. We highlight how this includes increasing hazard resolution, bias correction, and the integration of diverse sources of exposure and impact data, such as Earth observation and unstructured textual data (e.g., social media). This step is critical in enabling ML applications where data scarcity and heterogeneity persist.
  
  RQ2 addresses the modeling of multiple interacting hazards, emphasizing the use of Deep Learning techniques for forecasting compound events and copula models for representing statistical dependencies. This question targets methodological gaps in representing both simultaneous and consecutive hazard interactions.
  
  RQ3 explores how ML can support risk assessment, in addition to hazard modeling. We discuss the importance of data availability on past impacts and the need for model interpretability to ensure that outputs are actionable and trustworthy in real-world applications.
  
  RQ4 introduces the integration of future climate scenarios into multi-risk modeling. We have strengthened the revised manuscript by explicitly discussing the role of uncertainty quantification, including ensemble modeling, probabilistic approaches, and interpretability methods.
  
  These questions collectively offer a structured framework that links data engineering, compound hazard detection, multi-risk modeling, and future scenario analysis, contributing both technically and practically to the field.
  2. Role of Data Assimilation
  We fully acknowledge the importance of data assimilation in climate science and weather forecasting, particularly its role in producing high-quality reanalysis datasets that underpin many hazard models (e.g., ERA5, CHIRPS). However, we consider data assimilation a distinct field of study, more closely tied to physical modeling and numerical weather prediction.
  To maintain focus, we have not included a detailed discussion on this topic. That said, we have added a brief reference in the discussion to acknowledge the enabling role of data assimilation in the broader modeling ecosystem and to direct interested readers to relevant literature.
  3. Uncertainty Quantification
  We agree that uncertainty quantification is a key component of hazard and risk analysis. The revised manuscript now includes a specific discussion that distinguishes between input data uncertainties, such as variability across ensemble climate projections or differences between reanalysis products; model-based uncertainties, stemming from generalization limits and the structure of ML models. We highlight key methods, including probabilistic ML, Bayesian networks (Fan et al., 2021; Salvaña, 2025; Tierz et al., 2017), and Gaussian Processes (Li and Wang, 2025; Tazi et al., 2024), which provide built-in estimates of prediction confidence. We also emphasize how interpretability tools like SHAP contribute to transparency and communication of uncertainty in decision-making.
  4. Hybrid Modeling (ML + Physical Constraints)
  We appreciate this suggestion and have added a dedicated section on hybrid modeling approaches that integrate physical constraints into ML frameworks. We focus particularly on Physics-Informed Neural Networks and related hybrid architectures that embed conservation laws or differential equations directly into the learning process. These approaches are gaining traction in domains such as climate (Lütjens et al., 2021; Yang et al., 2023), hydrology and compound flooding (Feng et al., 2023; Munoz et al., 2024; Xu et al., 2024). We compare these hybrid approaches with pure ML models, noting that hybrid models often generalize better and produce more interpretable outputs, but can be harder to train and require expert-defined physical constraints (Bonfanti et al., 2024; Cuomo et al., 2022; Farea et al., 2024; Kashinath et al., 2021). We have also linked this topic to RQ1 and RQ2, where the trade-off between physical realism and data-driven flexibility is most relevant.
  5. Differences and Complementarities between ML and Copulas
  We have revised Section 4.3 to more clearly distinguish the roles of ML and copula techniques. While ML excels in detecting high-dimensional, non-linear patterns, copulas are better suited for modeling explicit dependencies, especially in the tails of distributions, a crucial feature for compound risk assessment. We now include examples where the two approaches are complementary: ML models can generate features or classify compound events, while Copulas can then model the joint probability of those features or events (Couasnon et al., 2018; Jiang et al., 2023). This combination offers interpretable, probabilistic outputs while leveraging the predictive power of ML.
  6. Role of Earth Observation (EO) in multi-risk assessments
  We have revised the discussion to better highlight the value of EO data, particularly when integrated with ML, in supporting actionable multi-hazard and multi-risk assessments. Rather than emphasizing technical differences, we now focus on operational value and decision support:
  Timely and spatially explicit indicators: EO enables detection of hazard, exposure, and impact metrics (e.g., vegetation stress, land cover change, water availability), improving early detection and event characterization.
  
  Expanded geographic coverage: EO provides access to consistent observations in remote and data-scarce areas, facilitating model scalability and application in developing countries (Chauhan et al., 2025; Kabiru et al., 2023).
  
  Recovery monitoring and disaster response: EO data such as nighttime lights and mobility proxies support tracking short-term disaster response and long-term recovery, helping planners assess evolving vulnerabilities (Qiang et al., 2020).
  
  These points are now emphasized in the revised manuscript to align the discussion with practical needs of stakeholders and decision-makers.
  
  Minor Comments
  All citations will be reviewed and updated for consistency and format.
  
  We'll revise the manuscript to avoid the term “understand” when referring to ML models, replacing it with more appropriate terms such as “model,” “characterize,” or “represent.”
  
  The manuscript will be subject to a comprehensive language revision, improving grammar, style, and readability.
  
  Section numbering will be corrected, and “Multi-hazard” is now properly labeled as Section 3.2.
  
  We thank the reviewer again for their valuable and constructive feedback, which has led to a stronger and more focused manuscript. We hope that the revisions address all points satisfactorily.
  
  REFERENCES
  
  Bonfanti, A., Santana, R., Ellero, M., Gholami, B., 2024. On the generalization of PINNs outside the training domain and the hyperparameters influencing it. Neural Comput Appl 36, 22677–22696. https://doi.org/10.1007/s00521-024-10178-2
  Chauhan, V., Gupta, L., Dixit, J., 2025. Machine learning and GIS-based multi-hazard risk modeling for Uttarakhand: Integrating seismic, landslide, and flood susceptibility with socioeconomic vulnerability. Environmental and Sustainability Indicators 26, 100664. https://doi.org/10.1016/j.indic.2025.100664
  Couasnon, A., Sebastian, A., Morales-Nápoles, O., 2018. A Copula-Based Bayesian Network for Modeling Compound Flood Hazard from Riverine and Coastal Interactions at the Catchment Scale: An Application to the Houston Ship Channel, Texas. Water (Basel) 10, 1190. https://doi.org/10.3390/w10091190
  Cuomo, S., Di Cola, V.S., Giampaolo, F., Rozza, G., Raissi, M., Piccialli, F., 2022. Scientific Machine Learning Through Physics–Informed Neural Networks: Where we are and What’s Next. J Sci Comput 92, 88. https://doi.org/10.1007/s10915-022-01939-z
  Fan, Y.R., Yu, L., Shi, X., Duan, Q.Y., 2021. Tracing Uncertainty Contributors in the Multi‐Hazard Risk Analysis for Compound Extremes. Earths Future 9. https://doi.org/10.1029/2021EF002280
  Farea, A., Yli-Harja, O., Emmert-Streib, F., 2024. Understanding Physics-Informed Neural Networks: Techniques, Applications, Trends, and Challenges. AI 5, 1534–1557. https://doi.org/10.3390/ai5030074
  Feng, D., Tan, Z., He, Q., 2023. Physics‐Informed Neural Networks of the Saint‐Venant Equations for Downscaling a Large‐Scale River Model. Water Resour Res 59. https://doi.org/10.1029/2022WR033168
  Jiang, T., Su, X., Zhang, G., Zhang, T., Wu, H., 2023. Estimating propagation probability from meteorological to ecological droughts using a hybrid machine learning copula method. Hydrol Earth Syst Sci 27, 559–576. https://doi.org/10.5194/hess-27-559-2023
  Kabiru, P., Kuffer, M., Sliuzas, R., Vanhuysse, S., 2023. The relationship between multiple hazards and deprivation using open geospatial data and machine learning. Natural Hazards 119, 907–941. https://doi.org/10.1007/s11069-023-05897-z
  Kashinath, K., Mustafa, M., Albert, A., Wu, J.-L., Jiang, C., Esmaeilzadeh, S., Azizzadenesheli, K., Wang, R., Chattopadhyay, A., Singh, A., Manepalli, A., Chirila, D., Yu, R., Walters, R., White, B., Xiao, H., Tchelepi, H.A., Marcus, P., Anandkumar, A., Hassanzadeh, P., Prabhat, 2021. Physics-informed machine learning: case studies for weather and climate modelling. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences 379, 20200093. https://doi.org/10.1098/rsta.2020.0093
  Li, J., Wang, H., 2025. Gaussian Processes Regression for Uncertainty Quantification: An Introductory Tutorial.
  Lütjens, B., Crawford, C.H., Veillette, M., Newman, D., 2021. PCE-PINNs: Physics-Informed Neural Networks for Uncertainty Propagation in Ocean Modeling. https://doi.org/arXiv:2105.02939v1
  Munoz, D.F., Daramola, S., Sakib, M.S., Moftakhari, H., Moradkhani, H., 2024. Advancing Compound Flood Modeling with Hybrid Approaches: An Integrated Statistical, Process-based, and Machine Learning Modeling Perspective, in: AGU Fall Meeting 2024; Session: Natural Hazards / Multihazard Flood Modelling: Form Inland to Coast; Poster No.1 Id. NH14C-01.
  Qiang, Y., Huang, Q., Xu, J., 2020. Observing community resilience from space: Using nighttime lights to model economic disturbance and recovery pattern in natural disaster. Sustain Cities Soc 57, 102115. https://doi.org/10.1016/j.scs.2020.102115
  Salvaña, M.L.O., 2025. Multi-Hazard Bayesian Hierarchical Model for Damage Prediction.
  Tazi, K., Orr, A., Hernandez-González, J., Hosking, S., Turner, R.E., 2024. Downscaling precipitation over High-mountain Asia using multi-fidelity Gaussian processes: improved estimates from ERA5. Hydrol Earth Syst Sci 28, 4903–4925. https://doi.org/10.5194/hess-28-4903-2024
  Tierz, P., Woodhouse, M.J., Phillips, J.C., Sandri, L., Selva, J., Marzocchi, W., Odbert, H.M., 2017. A Framework for Probabilistic Multi-Hazard Assessment of Rain-Triggered Lahars Using Bayesian Belief Networks. Front Earth Sci (Lausanne) 5. https://doi.org/10.3389/feart.2017.00073
  Xu, Q., Shi, Y., Bamber, J.L., Ouyang, C., Zhu, X.X., 2024. Large-scale flood modeling and forecasting with FloodCast. Water Res 264, 122162. https://doi.org/10.1016/j.watres.2024.122162
  Yang, Q., Hernandez-Garcia, A., Harder, P., Ramesh, V., Sattegeri, P., Szwarcman, D., Watson, C.D., Rolnick, D., 2023. Fourier Neural Operators for Arbitrary Resolution Climate Data Downscaling. https://doi.org/arXiv:2305.14452v2
  
  Citation: https://doi.org/10.5194/egusphere-2025-670-AC1

Interactive discussion

Status: closed

RC1:
'Comment on egusphere-2025-670', Anonymous Referee #1, 31 Mar 2025
This manuscript reviews machine learning (ML) and statistical approaches for climate-related multi-hazard and multi-risk assessment. It is organized around four themes—data processing, hazard prediction, risk analysis, and future scenarios—and incorporates auxiliary methods such as explainable AI and copula modeling. While the topic is timely and relevant, the manuscript has issues in contribution, analytical depth, structural clarity, and language quality.
Major Concerns:
Lack of Scientific Novelty or Conceptual Contribution: The paper lacks scientific novelty and conceptual contribution. It does not introduce new concepts, frameworks, or theoretical insights. Instead, it compiles existing literature without offering a critical synthesis or identifying research gaps. The review does not significantly advance the understanding of multi-hazard or multi-risk modeling compared to previous reviews, and its largely descriptive discussion limits its value as a synthesis resource.

No PRISMA flow diagram is provided, and Search strings, filtering criteria, and quality assessment processes are not disclosed. Suggest including a PRISMA diagram and methodological appendix (even in supplementary materials) to ensure transparency.

Poor Language Quality: The manuscript contains many grammatical errors, awkward phrasing, and redundant or overly long sentences. The language undermines clarity and makes the manuscript difficult to read. Given its role as a review article, this significantly reduces its accessibility and utility to the scientific community. A full professional language revision is essential.

Weak Structure and Inconsistent Framing: The manuscript shifts terminology from "machine learning methods"from the title to "data-driven methods" in the text, without clear justification. The manuscript includes a statistical method (copula), which is not a machine learning method. Poor coherence across sections: Overlap and redundancy between sections, and not clearly defined (e.g., 3.1.2 vs. 3.1.1 regarding satellite images observed soil moisture belongs to EO or climate data).

Additional Comments:
Lines 30–35: “...can advance multi-hazard and multi-risk...” → vague and informal phrasing.

Line 206: “his information…” → typo.

Line 313: “even if...” → incorrect conjunction; use “although” or “even though.”

Line 328: “images form...” → should be “images from.”

Line 335: “3.1 Multi-hazard…3.2.1 Identify...” → unclear subsection formatting.

Line 414: “With regards to” → incorrect phrase; use “With regard to.”

Line 421: “...showing popular results...”
Citation: https://doi.org/10.5194/egusphere-2025-670-RC1
- AC2:
  'Reply on RC1', Davide Mauro Ferrario, 27 May 2025
  We thank the reviewer for the constructive comments, which have helped us improve the clarity, structure, and scientific contribution of the manuscript. Below, we address each of the concerns raised:
  Scientific Novelty and Conceptual Contribution
  
  We respectfully disagree with the assessment that the manuscript lacks scientific novelty. While we do not propose a new algorithm or model, we believe the manuscript makes an original contribution by reframing the fragmented field of data-driven multi-hazard and multi-risk modeling through a structured and forward-looking synthesis. Moreover:
  We'll reframe the four core research questions, which collectively reflect a risk-modeling pipeline, from data processing to hazard prediction, risk assessment, and future scenario modeling. This framing provides a clear conceptual lens through which to interpret recent developments in the field.
  
  We'll focus the synthesis on actionable themes that cut across methods and applications, including:
  
  The integration of Earth Observation (EO) and textual data for multi-risk modeling;
  
  The challenge of model transferability and scalability across regions and hazard types;
  
  The role of model interpretability and explainable AI in machine learning application for multi-risk;
  
  The use of hybrid models (e.g., physics-informed neural networks) to bridge physical realism and data-driven inference.
  
  We'll revise the discussion and conclusion sections to more explicitly identify research gaps, including:
  
  Limited use of ML approaches that incorporate causal reasoning or address cross-sectoral risks;
  
  A need for better spatiotemporal validation techniques in multi-risk contexts;
  
  The importance of uncertainty quantification (and communication), particularly under future climate change scenarios.
  
  We believe this review is among the first to explicitly integrate machine learning and statistical approaches (including copulas) through a multi-hazard lens, and to assess their value for both scientific insight and decision support across multiple scales.
  PRISMA methodology
  
  We appreciate this comment and will add a PRISMA flow diagram and a methodological appendix (now included as supplementary material) to better describe the search strings, screening criteria, and inclusion/exclusion processes, improving the transparency and reproducibility of our literature selection.
  Poor Language Quality
  
  We fully agree that language quality is essential for a review article. In response, the manuscript will undergo a thorough professional-style language revision. We'll address issues related to grammar, structure, and clarity by:
  Simplifying overly long or technical sentences;
  
  Removing redundancy across sections;
  
  Improving flow, particularly in the methods and discussion.
  
  These changes will enhance the readability and accessibility of the manuscript for a broad interdisciplinary audience.
  Weak Structure and Inconsistent Framing
  
  We will clarify our use of the terms "machine learning" and "data-driven methods" in the introduction. The term "data-driven methods" is explicitly defined as an umbrella term that includes both machine learning techniques and other statistical approaches. This choice is motivated by the key role played by non-ML statistical tools (such as copulas) in multi-hazard modeling. We'll revised section titles, transitions, and content to reduce redundancy and improve coherence across the manuscript. In particular the overlap between Sections 3.1.1 and 3.1.2, will be clarified distinguishing between EO-based data sources and climate reanalysis/model data. We hope these revisions address the reviewer’s concerns and demonstrate our commitment to improving the clarity, rigor, and contribution of the manuscript.
  
  Citation: https://doi.org/10.5194/egusphere-2025-670-AC2
RC2:
'Comment on egusphere-2025-670', Anonymous Referee #2, 07 Apr 2025

Major:

1, The manuscript outlines four key research questions, but their link to data-driven climate multi-hazard analysis is not clear. Please clarify how these questions address existing gaps, improve risk assessments, and contribute quantitatively or qualitatively to the field.
2, Data assimilation is a critical component of modern climate analysis yet is missing from the discussion. Include recent developments that demonstrate its role in enhancing multi-hazard/risk assessments.
3, Uncertainty quantification in climate risk studies is necessary for the analysis of hazard/risk. However, they are not discussed.
4, The current focus is on pure data-driven models. Please discuss the emerging hybrid modeling that integrate physical laws, and compare their strengths and limitations.
5, Clarify the differences and complementarities between ML and copula techniques. Using practical examples if possible.
6, Reframe the discussion to emphasize how Earth observation data, combined with ML and copula techniques, leads to improved multi-hazard and risk assessments that benefit decision-makers. Try to rephrase from the perspective of assessments instead of technical comparisons.
Minor:

1, Ensure all citations (e.g., “Linkov et al., 2022” and “S. Yu & Ma, 2021”) are consistently and correctly formatted.
2, ‘understand’ is a strong word for many data-driven models as they do not really capture the underlying generative process. Replace “understand” with “modeling,” “characterization,” or “representation” to better reflect the capabilities of data-driven models.
3, Some grammatical and stylistic edits are suggested for improved readability.
4, Correct the numbering so that the “Multi-hazard” subsection is labeled 3.2 instead of 3.1.

Citation: https://doi.org/10.5194/egusphere-2025-670-RC2
- AC1:
  'Reply on RC2', Davide Mauro Ferrario, 27 May 2025
  We thank Reviewer 2 for the constructive and detailed feedback, which has helped improve the clarity, scope, and scientific depth of our manuscript. Below, we provide a point-by-point response to each of the reviewer’s comments.
  1. Link between the research questions and multi-risk
  We have revised the introduction and clarified the framing of the four research questions to better articulate how they address current gaps in the literature and contribute to advancing data-driven multi-hazard and multi-risk assessment.
  RQ1 focuses on the challenge of data preparation and preprocessing in ML-based hazard modeling. We highlight how this includes increasing hazard resolution, bias correction, and the integration of diverse sources of exposure and impact data, such as Earth observation and unstructured textual data (e.g., social media). This step is critical in enabling ML applications where data scarcity and heterogeneity persist.
  
  RQ2 addresses the modeling of multiple interacting hazards, emphasizing the use of Deep Learning techniques for forecasting compound events and copula models for representing statistical dependencies. This question targets methodological gaps in representing both simultaneous and consecutive hazard interactions.
  
  RQ3 explores how ML can support risk assessment, in addition to hazard modeling. We discuss the importance of data availability on past impacts and the need for model interpretability to ensure that outputs are actionable and trustworthy in real-world applications.
  
  RQ4 introduces the integration of future climate scenarios into multi-risk modeling. We have strengthened the revised manuscript by explicitly discussing the role of uncertainty quantification, including ensemble modeling, probabilistic approaches, and interpretability methods.
  
  These questions collectively offer a structured framework that links data engineering, compound hazard detection, multi-risk modeling, and future scenario analysis, contributing both technically and practically to the field.
  2. Role of Data Assimilation
  We fully acknowledge the importance of data assimilation in climate science and weather forecasting, particularly its role in producing high-quality reanalysis datasets that underpin many hazard models (e.g., ERA5, CHIRPS). However, we consider data assimilation a distinct field of study, more closely tied to physical modeling and numerical weather prediction.
  To maintain focus, we have not included a detailed discussion on this topic. That said, we have added a brief reference in the discussion to acknowledge the enabling role of data assimilation in the broader modeling ecosystem and to direct interested readers to relevant literature.
  3. Uncertainty Quantification
  We agree that uncertainty quantification is a key component of hazard and risk analysis. The revised manuscript now includes a specific discussion that distinguishes between input data uncertainties, such as variability across ensemble climate projections or differences between reanalysis products; model-based uncertainties, stemming from generalization limits and the structure of ML models. We highlight key methods, including probabilistic ML, Bayesian networks (Fan et al., 2021; Salvaña, 2025; Tierz et al., 2017), and Gaussian Processes (Li and Wang, 2025; Tazi et al., 2024), which provide built-in estimates of prediction confidence. We also emphasize how interpretability tools like SHAP contribute to transparency and communication of uncertainty in decision-making.
  4. Hybrid Modeling (ML + Physical Constraints)
  We appreciate this suggestion and have added a dedicated section on hybrid modeling approaches that integrate physical constraints into ML frameworks. We focus particularly on Physics-Informed Neural Networks and related hybrid architectures that embed conservation laws or differential equations directly into the learning process. These approaches are gaining traction in domains such as climate (Lütjens et al., 2021; Yang et al., 2023), hydrology and compound flooding (Feng et al., 2023; Munoz et al., 2024; Xu et al., 2024). We compare these hybrid approaches with pure ML models, noting that hybrid models often generalize better and produce more interpretable outputs, but can be harder to train and require expert-defined physical constraints (Bonfanti et al., 2024; Cuomo et al., 2022; Farea et al., 2024; Kashinath et al., 2021). We have also linked this topic to RQ1 and RQ2, where the trade-off between physical realism and data-driven flexibility is most relevant.
  5. Differences and Complementarities between ML and Copulas
  We have revised Section 4.3 to more clearly distinguish the roles of ML and copula techniques. While ML excels in detecting high-dimensional, non-linear patterns, copulas are better suited for modeling explicit dependencies, especially in the tails of distributions, a crucial feature for compound risk assessment. We now include examples where the two approaches are complementary: ML models can generate features or classify compound events, while Copulas can then model the joint probability of those features or events (Couasnon et al., 2018; Jiang et al., 2023). This combination offers interpretable, probabilistic outputs while leveraging the predictive power of ML.
  6. Role of Earth Observation (EO) in multi-risk assessments
  We have revised the discussion to better highlight the value of EO data, particularly when integrated with ML, in supporting actionable multi-hazard and multi-risk assessments. Rather than emphasizing technical differences, we now focus on operational value and decision support:
  Timely and spatially explicit indicators: EO enables detection of hazard, exposure, and impact metrics (e.g., vegetation stress, land cover change, water availability), improving early detection and event characterization.
  
  Expanded geographic coverage: EO provides access to consistent observations in remote and data-scarce areas, facilitating model scalability and application in developing countries (Chauhan et al., 2025; Kabiru et al., 2023).
  
  Recovery monitoring and disaster response: EO data such as nighttime lights and mobility proxies support tracking short-term disaster response and long-term recovery, helping planners assess evolving vulnerabilities (Qiang et al., 2020).
  
  These points are now emphasized in the revised manuscript to align the discussion with practical needs of stakeholders and decision-makers.
  
  Minor Comments
  All citations will be reviewed and updated for consistency and format.
  
  We'll revise the manuscript to avoid the term “understand” when referring to ML models, replacing it with more appropriate terms such as “model,” “characterize,” or “represent.”
  
  The manuscript will be subject to a comprehensive language revision, improving grammar, style, and readability.
  
  Section numbering will be corrected, and “Multi-hazard” is now properly labeled as Section 3.2.
  
  We thank the reviewer again for their valuable and constructive feedback, which has led to a stronger and more focused manuscript. We hope that the revisions address all points satisfactorily.
  
  REFERENCES
  
  Bonfanti, A., Santana, R., Ellero, M., Gholami, B., 2024. On the generalization of PINNs outside the training domain and the hyperparameters influencing it. Neural Comput Appl 36, 22677–22696. https://doi.org/10.1007/s00521-024-10178-2
  Chauhan, V., Gupta, L., Dixit, J., 2025. Machine learning and GIS-based multi-hazard risk modeling for Uttarakhand: Integrating seismic, landslide, and flood susceptibility with socioeconomic vulnerability. Environmental and Sustainability Indicators 26, 100664. https://doi.org/10.1016/j.indic.2025.100664
  Couasnon, A., Sebastian, A., Morales-Nápoles, O., 2018. A Copula-Based Bayesian Network for Modeling Compound Flood Hazard from Riverine and Coastal Interactions at the Catchment Scale: An Application to the Houston Ship Channel, Texas. Water (Basel) 10, 1190. https://doi.org/10.3390/w10091190
  Cuomo, S., Di Cola, V.S., Giampaolo, F., Rozza, G., Raissi, M., Piccialli, F., 2022. Scientific Machine Learning Through Physics–Informed Neural Networks: Where we are and What’s Next. J Sci Comput 92, 88. https://doi.org/10.1007/s10915-022-01939-z
  Fan, Y.R., Yu, L., Shi, X., Duan, Q.Y., 2021. Tracing Uncertainty Contributors in the Multi‐Hazard Risk Analysis for Compound Extremes. Earths Future 9. https://doi.org/10.1029/2021EF002280
  Farea, A., Yli-Harja, O., Emmert-Streib, F., 2024. Understanding Physics-Informed Neural Networks: Techniques, Applications, Trends, and Challenges. AI 5, 1534–1557. https://doi.org/10.3390/ai5030074
  Feng, D., Tan, Z., He, Q., 2023. Physics‐Informed Neural Networks of the Saint‐Venant Equations for Downscaling a Large‐Scale River Model. Water Resour Res 59. https://doi.org/10.1029/2022WR033168
  Jiang, T., Su, X., Zhang, G., Zhang, T., Wu, H., 2023. Estimating propagation probability from meteorological to ecological droughts using a hybrid machine learning copula method. Hydrol Earth Syst Sci 27, 559–576. https://doi.org/10.5194/hess-27-559-2023
  Kabiru, P., Kuffer, M., Sliuzas, R., Vanhuysse, S., 2023. The relationship between multiple hazards and deprivation using open geospatial data and machine learning. Natural Hazards 119, 907–941. https://doi.org/10.1007/s11069-023-05897-z
  Kashinath, K., Mustafa, M., Albert, A., Wu, J.-L., Jiang, C., Esmaeilzadeh, S., Azizzadenesheli, K., Wang, R., Chattopadhyay, A., Singh, A., Manepalli, A., Chirila, D., Yu, R., Walters, R., White, B., Xiao, H., Tchelepi, H.A., Marcus, P., Anandkumar, A., Hassanzadeh, P., Prabhat, 2021. Physics-informed machine learning: case studies for weather and climate modelling. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences 379, 20200093. https://doi.org/10.1098/rsta.2020.0093
  Li, J., Wang, H., 2025. Gaussian Processes Regression for Uncertainty Quantification: An Introductory Tutorial.
  Lütjens, B., Crawford, C.H., Veillette, M., Newman, D., 2021. PCE-PINNs: Physics-Informed Neural Networks for Uncertainty Propagation in Ocean Modeling. https://doi.org/arXiv:2105.02939v1
  Munoz, D.F., Daramola, S., Sakib, M.S., Moftakhari, H., Moradkhani, H., 2024. Advancing Compound Flood Modeling with Hybrid Approaches: An Integrated Statistical, Process-based, and Machine Learning Modeling Perspective, in: AGU Fall Meeting 2024; Session: Natural Hazards / Multihazard Flood Modelling: Form Inland to Coast; Poster No.1 Id. NH14C-01.
  Qiang, Y., Huang, Q., Xu, J., 2020. Observing community resilience from space: Using nighttime lights to model economic disturbance and recovery pattern in natural disaster. Sustain Cities Soc 57, 102115. https://doi.org/10.1016/j.scs.2020.102115
  Salvaña, M.L.O., 2025. Multi-Hazard Bayesian Hierarchical Model for Damage Prediction.
  Tazi, K., Orr, A., Hernandez-González, J., Hosking, S., Turner, R.E., 2024. Downscaling precipitation over High-mountain Asia using multi-fidelity Gaussian processes: improved estimates from ERA5. Hydrol Earth Syst Sci 28, 4903–4925. https://doi.org/10.5194/hess-28-4903-2024
  Tierz, P., Woodhouse, M.J., Phillips, J.C., Sandri, L., Selva, J., Marzocchi, W., Odbert, H.M., 2017. A Framework for Probabilistic Multi-Hazard Assessment of Rain-Triggered Lahars Using Bayesian Belief Networks. Front Earth Sci (Lausanne) 5. https://doi.org/10.3389/feart.2017.00073
  Xu, Q., Shi, Y., Bamber, J.L., Ouyang, C., Zhu, X.X., 2024. Large-scale flood modeling and forecasting with FloodCast. Water Res 264, 122162. https://doi.org/10.1016/j.watres.2024.122162
  Yang, Q., Hernandez-Garcia, A., Harder, P., Ramesh, V., Sattegeri, P., Szwarcman, D., Watson, C.D., Rolnick, D., 2023. Fourier Neural Operators for Arbitrary Resolution Climate Data Downscaling. https://doi.org/arXiv:2305.14452v2
  
  Citation: https://doi.org/10.5194/egusphere-2025-670-AC1

Peer review completion

AR – Author's response | RR – Referee report | ED – Editor decision | EF – Editorial file upload

ED: Reconsider after major revisions (further review by editor and referees) (09 Jun 2025) by Aloïs Tilloy

AR by Davide Mauro Ferrario on behalf of the Authors (02 Dec 2025) Author's response Author's tracked changes Manuscript

ED: Referee Nomination & Report Request started (15 Jan 2026) by Aloïs Tilloy

RR by Anonymous Referee #3 (01 Feb 2026)

Suggestions for revision or reasons for rejection

Thank you for the opportunity to review this manuscript. The authors have undertaken a substantial effort to review data-driven methods—especially machine learning (ML)—for multi-hazard and multi-risk assessment.
The topic is timely, interdisciplinary, and of great relevance to climate adaptation and disaster risk reduction. The paper is well-structured, covers a wide range of methods, and provides useful summary tables.
The work has a number of considerable strengths:
• Focuses on data-driven methods (especially machine learning and multivariate statistics) for climate multi-hazard and multi-risk assessment, a critical and rapidly evolving intersection of climate science and risk management.
• Organized logically into four main themes with further sub-themes providing a coherent roadmap of the field.
• Covers a wide range of methods including mainstream ML and specialized statistical tools like copulas for modeling dependencies between hazards—showcasing methodological depth.
• Repeatedly highlights Explainable AI (XAI) and Uncertainty Quantification (UQ), addressing key challenges in applying AI to scientific and decision-making contexts.
• Considers climate datasets, Earth Observation products, and textual data (social media, news), reflecting the trend toward multi-source data fusion
• Clearly outlines future priorities, such as modeling multi-hazard interactions, integrating dynamic vulnerability, and improving uncertainty propagation.
Some areas should also be considered by the authors:
• The PRISMA guidelines are cited but not rigorously followed. Key elements of a systematic review—such as a quality assessment of included studies, a risk of bias analysis, etc —are missing.
• The work uses key terms (like multi-hazard, multi-risk, compound, cascading, etc.) largely descriptively and not in accordance with any existing frameworks (e.g., IPCC, UNDRR). As this is a common feature of this diverse field, how to the authors assess potential inconsistencies in this besides sematic key-word matching?
• The work treats vulnerability and exposure largely as static in mapping exercises, with limited discussion of dynamic, social justice, governance, or human behavior.
• The review heavily biases towards ML while important statistical approaches (e.g., Bayesian networks, spatial statistics, agent-based modeling) are mentioned only briefly.
• The authors should also provide statistics on the application domains (as presently literature appears limited to “Scopus, English-only, 2010–2024”) and especially highlight if any potential regional bias in selected studies with a tendency to minimize the contribution of the Global South where risk dynamics differ.
• Although uncertainty quantification (UQ) is mentioned, the work does not systematically address how this is (or isn’t) integrated across the modeling chain.
• Presently, Figure 3 is loosely integrated into the narrative of the work.
• Finally. The work places emphasis on machine learning without adequate discussion of its limitations, overhyped applications, or failures in real-world risk management. Which should have been part of the robustness of the work.
• The review should discuss any potential ethical, political, or justice dimensions of using AI/ML in risk assessment—such as data colonialism, algorithmic bias, or the risk of reinforcing inequalities.

Potentially this can be established as systematic review article that will be highly valuable for researchers and practitioners working at the intersection of climate risk, disaster management, and machine learning. However, the work should also address the gap between technical models and actionable risk reduction and explain how complex machine learning outputs can be used by stakeholders for real-world decisions.

Hide

ED: Reconsider after major revisions (further review by editor and referees) (02 Feb 2026) by Aloïs Tilloy

AR by Davide Mauro Ferrario on behalf of the Authors (30 Mar 2026) Author's response Author's tracked changes

EF by Mario Ebel (31 Mar 2026) Manuscript

ED: Publish subject to minor revisions (review by editor) (15 May 2026) by Aloïs Tilloy

AR by Davide Mauro Ferrario on behalf of the Authors (20 May 2026) Author's response Author's tracked changes Manuscript

ED: Publish as is (28 May 2026) by Aloïs Tilloy

ED: Publish subject to technical corrections (04 Jun 2026) by Bruce D. Malamud (Executive editor)

AR by Davide Mauro Ferrario on behalf of the Authors (10 Jun 2026) Author's response Manuscript

Journal article(s) based on this preprint

26 Jun 2026

Review article: Harnessing data-driven methods for climate multi-hazard and multi-risk assessment

Davide Mauro Ferrario, Marcello Sanò, Margherita Maraschini, Andrea Critto, and Silvia Torresan

Nat. Hazards Earth Syst. Sci., 26, 2975–3024, https://doi.org/10.5194/nhess-26-2975-2026,https://doi.org/10.5194/nhess-26-2975-2026, 2026

Short summary

Davide Mauro Ferrario, Marcello Sanò, Margherita Maraschini, Andrea Critto, and Silvia Torresan

Viewed

Total article views: 24,231 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
18,401	5,542	288	24,231	417	509

HTML: 18,401
PDF: 5,542
XML: 288
Total: 24,231
BibTeX: 417
EndNote: 509

Views and downloads (calculated since 03 Mar 2025)

Month	HTML	PDF	XML	Total
Mar 2025	851	219	10	1,080
Apr 2025	665	285	20	970
May 2025	441	296	35	772
Jun 2025	613	440	25	1,078
Jul 2025	370	360	5	735
Aug 2025	2,031	351	10	2,392
Sep 2025	9,689	360	66	10,115
Oct 2025	555	185	10	750
Nov 2025	609	310	9	928
Dec 2025	504	727	25	1,256
Jan 2026	613	630	40	1,283
Feb 2026	642	505	13	1,160
Mar 2026	614	680	11	1,305
Apr 2026	105	108	1	214
May 2026	76	70	6	152
Jun 2026	23	16	2	41
Jul 2026	0

Cumulative views and downloads (calculated since 03 Mar 2025)

Month	HTML	PDF	XML	Total
Mar 2025	851	219	10	1,080
Apr 2025	665	285	20	970
May 2025	441	296	35	772
Jun 2025	613	440	25	1,078
Jul 2025	370	360	5	735
Aug 2025	2,031	351	10	2,392
Sep 2025	9,689	360	66	10,115
Oct 2025	555	185	10	750
Nov 2025	609	310	9	928
Dec 2025	504	727	25	1,256
Jan 2026	613	630	40	1,283
Feb 2026	642	505	13	1,160
Mar 2026	614	680	11	1,305
Apr 2026	105	108	1	214
May 2026	76	70	6	152
Jun 2026	23	16	2	41
Jul 2026	0

Viewed (geographical distribution)

Total article views: 24,175 (including HTML, PDF, and XML) Thereof 24,175 with geography defined and 0 with unknown origin.

Country	#	Views	%

Cited

Latest update: 03 Jul 2026

Download

The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.

Preprint (1879 KB)
Metadata XML

Short summary

This review explores how Machine Learning (ML) can advance multi-hazard and multi-risk going through four main themes: data processing, hazard prediction, risk assessment, and future climate scenarios. It shows how ML is widely used for Earth observations and climate data processing, with Deep Learning applied for hazard prediction and ensemble ML methods for risks, and how future research moving towards analysis of multi-hazard interactions, dynamic vulnerability and early warning systems.


Total:	0
HTML:	0
PDF:	0
XML:	0

Review article: Harnessing Machine Learning methods for climate multi-hazard and multi-risk assessment

Journal article(s) based on this preprint

Interactive discussion

Interactive discussion

Peer review completion

Suggestions for revision or reasons for rejection

Journal article(s) based on this preprint

Viewed

Viewed (geographical distribution)

Cited

7 citations as recorded by crossref.