the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 15 Jan 2025
Improving Forest Damage Detection and Risk Assessment from Winter Storms Using High-Resolution Satellite Data and Environmental Drivers
Abstract. Accurate assessment of forest losses and evaluation of future damage risks are crucial for effective forest management and conservation, particularly as global warming intensifies natural disturbance agents. This study introduces a novel approach combining convolutional neural networks (CNNs) with Random Forest (RF) machine learning classifiers to enhance the precision of forest disturbance detection and risk evaluation. We tested this approach on a large-scale dataset (1490 km2) with diverse forest types and environmental conditions of cool temperate-south boreal forests on Kunashir Island (Northwest Pacific). Using the U-Net deep learning architecture, we precisely identified windthrow patches from (VHR) Pléiades-1 optical satellite imagery. Resulted windthrow map was integrated with an RF classifier that utilized environmental predictors, including elevation, slope aspect, slope inclination, slope curvature, forest canopy closure, landform type, and forest vegetation type, to assess forest damage risk. Our analysis revealed approximately 21.73 km2 of the forested area as significantly disturbed, predominantly within dark coniferous forests. Elevation emerged as the most critical predictor of disturbance risk, with complex interactions observed among predictors such as canopy closure and slope steepness. This integrated approach allowed for highly accurate forest loss detection and provided valuable insights into the risk of future damage events. By combining advanced deep learning techniques with RF and detailed environmental predictors, this approach offers a robust framework for evaluating forest disturbance risk. This method pushes forward the frontiers in the precision of forest loss detection but also aids in developing effective strategies for managing and mitigating risks from future disturbance events.
- Preprint
(1600 KB) - Metadata XML
-
Supplement
(181 KB) - BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on nhess-2024-217', Anonymous Referee #1, 03 Mar 2025
General comments
The title of the preprint reads: “Improving Forest Damage Detection and Risk Assessment from Winter Storms Using High-Resolution Satellite Data and Environmental Drivers.”
Although the authors use the innovative U-Net CNN method, this approach has already been used by the same researchers for automated windthrow detection and disturbance segmentation (Kislov et al., 2021; Kislov and Korznikov, 2020). The manuscript does not introduce any notable improvements or advancements to this method. Furthermore, the so-called “risk assessment” presented in the study is, in fact, a vulnerability assessment, as risk is typically defined as a combination of vulnerability, exposure, and hazard assessments. However, the latter two components (exposure and hazard) are absent from the analysis. Throughout the manuscript, the terms risk, vulnerability, and susceptibility are often used as synonyms, despite their distinct meanings, which leads to conceptual imprecision.
One of the most significant issues that I find here, is that this analysis is not transferable to different environments (L274–276), not it is relevant for forest management, since the active forest management is no longer practiced on the island (L105).
The manuscript reads quite well, but the English language could be improved for greater clarity and precision. Additionally, the methods section would benefit from a clearer structure, and the use of terminology should be consistent throughout the text.
Specific comments
Methods:
Please stay consistent in describing well the forest stands and their species composition. E.g. on L99-101 you describe mixed forests with different species composition than later in the text.
Figures should avoid using red-green colour combinations to make them accessible to all.
How did you deal with imbalances in the data in disturbed and undisturbed areas?
Discussion:
Overall, the discussion is not sufficient for me. You only discuss three predictors and in the end of the discussion you state that the individual predictors are not universal factors for windthrow risk. You should better discuss, how is your method transferable or any of your results useful for forest management or further research.
The most important predictor in your case was elevation, but you don’t discuss this in detail and you just cite one paper that found the opposite results. You do not really discuss the effect of canopy closure – how is the fact of naturally developed open canopy or open canopy created after logging related to your results?
From your analysis the elevation was highly correlated with the vegetation type. I think you should address this in your analysis as you state in L180 and L277-279.
What about the other aspects of forest type than interception?
You don’t discuss the rest of your predictors like aspect or slope.
With your results, are you able to predict windthrow disturbance? What is your main contribution? You say on L294 that you offer critical insights into forest loss and recurrence risks, but you also say that your results are not universal (L274).
Technical corrections
L34, L35, L36 and L39 – pixel-1 – superscript
L60-61 To test the new approaches for disturbance detection, it is necessary to have a diverse study area – one that encompasses a range of geomorphological features and forest types. Furthermore, the availability of cloud-free VHR satellite images is crucial for accurate analysis.
L64 delete risk
L64 incorporating -> analysing
L65 specify where exactly are Kuril Islands, e.g.: eastern Russia
L65 In addition, it was affected
L66 ms-1 superscript
L69 disturbances – why plural? Which kind of disturbances other than windthrow?
L74 why biotic damage agents?
L80 vulnerability of forest cover -> forest vulnerability?
L88 which come -> coming
L93 Figure 1: Kuril Islands or Kurils; Figure caption: … windthrow impact on forest cover … where is it in the figure? Please add to a legend if you want to draw the attention to the black dots as the extent of the windthrow damage.
L98 which kind of spruce? Can you add the Latin name as you do for other species?
L100 please add the oak species plus the Latin name
L101 same here for fir
L103 same here: Latin name for dwarf pine, what kind of alder and Latin names of alder and willow (Salix sp.)
L104 what does it mean single or sparse trees?
L105 delete “Even”
L109 Figure 2: It is not necessary to show the location of the island anymore as you did it in the previous figure; Legend: please use the same names as in the text, e.g.: is the mixedwood forests the same as mixed boreal forests?
L135 Figure 3: legend and north arrow are missing
L142 program -> software
L145 there is no Fig. S4, do you mean the lower part of figure 4?
L147 risk assessment: it is not risk assessment – maybe vulnerability assessment? Risk is composed by the vulnerability and exposure of the forest and the windthrow hazard. What you assess in my understanding is only the forest’s vulnerability or susceptibility to the windthrow.
L149 aspect – I think aspect should be treated as categorical variable as one can divide between the main four cardinal directions or eight directions. Also, the values 0 and 360 are indicating the same direction.
L150 curvature – also the curvature should be treated as categorical variable as you explain in the text, you have basically three categories: convex, concave and flat.
L151 Figure caption: please add citation from where is this figure taken.
L155 peat -> pit
L160 why did you select the threshold of 20% of canopy closure?
L164 units -> types
L164 six -> you name seven classes, in the map you have 11 classes – clarify. Please, be consistent with the names of the different forest types.
L180 delete “Therefore”
L181 incorrect conclusions: please explain what did you do to consider this in your analysis.
L221 before you talk about disturbance risk, in the previous sentence about susceptibility and now severity of the disturbance. These terms are not interchangeable and should be clear and used in a correct way throughout the manuscript.
L226 Figure caption: Post-ETC disturbance impact of uprooted trees from satellite Pleiades-1 (a) and from the ground (b).
L236 Figure 10: Vegetation -> Vegetation type
L244 only the dark coniferous forests are dominated by spruce and fir, right? Rephrase
L249 delete the hyphen
L256 mountain birch – is it stone birch?
L265 sensitivity?
L265 why in contrast? You also observed higher windthrow damage in spruce and fir forests compared to mixed forests.
L268 why disturbances? Do you mean windthrow disturbance?
L268-L270 do you have some data to support the statement that the windthrow occurred in patches of coniferous trees or is it a field observation?
Citation: https://doi.org/10.5194/nhess-2024-217-RC1 -
AC1: 'Reply on RC1', Kirill Korznikov, 14 Mar 2025
We thank the referee for the highly informative and constructive review. We particularly appreciate the detailed attention and the effort invested in compiling a section with technical comments.
(1) We acknowledge the need to improve the consistency of terminology and enhance the overall readability of the text. Specifically, we recognize the importance of refining our use of the term "risk." In some publications addressing similar topics, the term "risk" is often used synonymously with "vulnerability." However, we agree that it will be beneficial to adopt more precise and consistent terminology throughout the manuscript as you suggested.
(2) We would like to justify the primary concern that our methodology lacks novelty. The cited papers (Kislov et al., 2021; Kislov and Korznikov, 2020) describe methods for developing semantic segmentation neural networks to detect forest windthrows on the base of very high-resolution satellite images. These papers are strictly methodological, presenting a new approach for highly accurate recognition of forest disturbances.
In contrast, the presented study applies this neural network as a practical tool to detect windthrows and investigates the factors driving the observed windthrow patterns using terrain and forest characteristics. In particular, we utilized this method to precisely detect the windthrows across a 1,500 km² volcanic landscape with diverse natural forests, ranging from broadleaf to taiga, and a pronounced elevation gradient. We aimed to understand the factors that explain why windthrows occurred there. The search for explanations behind observed phenomena is a fundamental goal of science, and this was precisely our focus in this article.
That's why we believe that the integration of the recognition results of the neural networks with additional environmental data (landforms, forest types from vegetation maps, etc.) offers a valuable contribution to understanding the causes behind large-scale disturbances. Furthermore, our methodology can serve as a valuable model for other regions affected by weather hazards, as we relied solely on publicly available remote sensing data, ensuring that our approach can be widely applied without requiring unique or specialized data sources.
While numerous scientific journals frequently publish papers on formally "new" neural network architectures for object detection or segmentation purposes in satellite imagery, often emphasizing incremental improvements in computational efficiency or accuracy (frequently arise from improved training data quality rather than architectural advancements), there is a noticeable gap in research demonstrating the practical application of these tools for scientific or applied problems. Our work seeks to address this gap by the implementation of such methods for a better understanding of windthrow patterns and forest vulnerability. Hence, we believe that our work is mainly relevant from this perspective.
(3) Regarding the referee's second significant remark that "this analysis is not transferable to different environments (L274–276), it is not relevant for forest management, since active forest management is no longer practised on the island (L105)," we would like to offer clarification. No analysis conducted in one region can be fully transferable to another region in its entirety, as each site is unique. However, there are universal factors that apply across different environments, such as the overloading of evergreen tree crowns with heavy, wet snow, which significantly increases their susceptibility to uproot/break down. However, our findings can be extrapolated to regions with similar environmental conditions, including forest types, climate/weather patterns, terrain characteristics, etc. Given the study area's location and extent, similar events may occur across the broad gradient of Northeast Asia, from the latitude of central Japan and Korea to Kamchatka. Although active forest management is no longer practised on Kunashir Island, our conclusions hold practical relevance for foresters and forest holders for this extensive region, particularly in the context of forest management strategies. We aim to address that in the revised version of the Discussion.
(4) We also acknowledge the referee's constructive comments regarding the methodological section. We agree that improvements can be made, including enhancing the quality of the figures to better support our findings.
(5) For the comments about the discussion section, we would like to clarify the following. Our results indicate that environmental predictors are intercorrelated, making it less meaningful to discuss them independently. Doing so would result in repetitive explanations about the limitations of each factor's influence and its interaction with other predictors. Since these relationships are already effectively conveyed through graphical representations, we chose to keep the discussion section concise and focused on the complex interplay of factors that determine forest susceptibility to disturbances. While we can expand the discussion section, if necessary, we believe that doing so may reduce the article's readability and clarity.
(6) Finally, regarding the final comment of the referee: "With your results, are you able to predict windthrow disturbance? What is your main contribution? You say on L294 that you offer critical insights into forest loss and recurrence risks, but you also say that your results are not universal (L274)." We agree that this aspect should have been more clearly articulated in the discussion. We recognise the need to better highlight our main contribution and remove any ambiguity in the sentences.
We are grateful for the referee’s comments, and we greatly appreciate the time and effort invested in providing such detailed feedback. As indicated above, we realise that there is space for improvement, and we will be happy to revise our manuscript according to the comments if we have such an opportunity.
Citation: https://doi.org/10.5194/nhess-2024-217-AC1
-
AC1: 'Reply on RC1', Kirill Korznikov, 14 Mar 2025
-
RC2: 'Comment on nhess-2024-217', Leon Bührle, 10 Mar 2025
Improving Forest Damage Detection and Risk Assessment from Winter Storms Using High-Resolution Satellite Data and Environmental Driver
Kirill Korznikov, Dmitriy Kislov,Jiří Doležal,and Jan Altman
General comments:
The paper uses high-resolution Pleiades images to detect windthrow damages applying convolutional neural networks (CNNs) and Random Forest (RF) methods. They further derive the importance of different environmental factors to explain the stands affected. The findings regarding important factors are then used to describe how they can support windthrow predisposition assessments.
While the analysis of natural disturbances in forests using high-resolution satellite (Pleiades) imagery and machine learning is a promising research direction aligned with the scope of Natural Hazards, the study's novelty is limited.
As Reviewer#1 addressed correctly, the most interesting part of the methodology is the detection of windthrow areas using high-resolution satellite imagery and machine learning approaches, which was already published by the authors in Kislov und Korznikov (2020). The main difference is using only one Pleiades images instead of several images (Pleiades and WorldView).
Furthermore, a similar study by Korznikov et al. (2022) also detected windthrow areas using high-resolution WorldView images and analyzed predictors for different windthrow sizes with comparable data. While investigating predictors remains relevant and warrants further exploration, the current study's predictor set is limited, and the reliability of specific datasets (e.g., Global Tree Cover and Aster DEM-derived data) is questionable. The study does not adequately discuss these limitations.
Additionally, important predictors are missing, reducing the meaning of the results. Essential factors for windthrow predisposition identified in previous research such as tree/stand height, distance to forest edges, proportion of shallow-rooting species, vertical layering, and wind exposure are not considered (Stritih et al. 2021). The authors do not discuss the potential impact of missing key parameters on predictor importance ranking.
Moreover, the study does not adequately address the broader applicability of its findings. Given the challenges in transferring the methodology to other regions, the discussion does not sufficiently explore how the results could be used for improving forest management.
Overall, the study offers limited new scientific contributions.
I also agree with Reviewer #1 that the terminology is unclear and confusing. Since risk has multiple definitions, the authors should explicitly define their usage. As Natural Hazards typically defines risk as a combination of hazard, exposure, and vulnerability, I recommend using susceptibility or predisposition assessment instead, as these terms more accurately reflect the study’s approach.
The overall structure of the paper is acceptable, and the English language is generally clear, requiring only minor refinements.
However, given the limited novelty and added value of the study, I cannot recommend publication in its current form. A relevant improvement of the study could be the actual creation of a spatial continuous predisposition assessment based on the authors findings with improved and more predictors.
Specific comments
Given that the manuscript requires substantial extensions and improvements, I will focus on the main issues. Additionally, I do not mention every instance of incorrect risk terminology, as this is a systematic problem.
Abstract:
L11/12: This study does not introduce a novel approach combining convolutional neural networks (CNNs) with Random Forest (RF) machine learning classifiers, as it was already introduced in the last publication (Korznikov et al. 2022). Please clarify that the study applies this approach rather than introducing it.
L22: The limited number of predictors and their questionable reliability mean that the proposed framework is not robust.
Introduction:
The manuscript lacks references to several state-of-the-art studies on this topic.
L33: Sentinel-2 and Landsat data are freely available since many years, not only recently.
L35 Include relevant disturbance detection studies, such as Senf und Seidl (2021) and Morresi et al. (2024).
L54 Add a reference to Stritih et al. (2021)
L57: The most critical factors for windthrow susceptibility are stand height and species composition. Do not focus only on the data available in your study.
L60: The approach is not new and was already tested. Please describe the differences compared to Korznikov et al. (2022).
Study area:
Figure 1: Where can I see the windthrow impact on forest cover mentioned in the figure description?
Extratropical Cyclone Disturbance Event
L120: How can a 39 mm precipitation event result into an 80 mm snow cover? Please clarify.
L123: Please locate the weather station in the map. How represent is this weather station for the study area?
L124: Reference is missing, which wind speeds are to be expected during strong storms in the study area.
Forest Disturbances Recognition
L130: Please provide more information about the accuracy assessment.
L131: Why were these areas excluded? Small disturbances and even single broken trees are crucial for post-disturbance management in many regions. High-resolution data provide a unique opportunity to detect such areas, which are often prioritized for sanitation felling.
Environmental Predictors
L138: Justify the selection of these specific predictors.
L140: Specify that the reported 7 m accuracy is an average. Accuracy is significantly lower in steep slopes.
Table1: Given the >10 m inaccuracy in steep slopes and the 30 m resolution, such detailed classification steps are questionable.
L167: What is the total number of points? Did you consider/check spatial correlations?
Results
Fig 11. I The y-axis labelling is unclear. I suppose that the value is a range, for example in case of slope from 0 to 6° or? Please improve the labelling.
Discussion
L241-244: Discuss why elevation is an important predictor. Is it due to random storm patterns, or does it reflect species composition at different elevations?
L245/246: Add relevant references supporting the claim that denser stands are more susceptible.
L248/249: While these statements are valid, your data cannot distinguish the causes of open stands, as no management intervention or past disturbance records were included.
L270-272: Since the analysis is based on a single storm event, generalizing about forest structure development introduces significant uncertainties.
L279: Large-scale high-resolution LiDAR acquisitions, combined with satellite data, now enable more accurate predisposition assessments only based on remote sensing data.
Conclusions
L285: As the disturbance detection was already published, only parts of the predictor analysis are new.
L287: The study does not assess future predisposition. Please revise this statement.
L289: The claim that wet snow is the primary windthrow factor is based on a single event under specific conditions and is not a robust conclusion.
L292: A comprehensive windthrow predisposition model would require additional parameters. Clarify that your statement applies only to the limited predictor set used in this study.
L295: Disturbance analyses are valuable for forest management, yet the manuscript only briefly discusses their practical implications.
Final Recommendation: Due to the study’s limited novelty, methodological limitations, and insufficient discussion of key limitations, I cannot recommend publication in its current form.
Kislov, Dmitry E.; Korznikov, Kirill A. (2020): Automatic Windthrow Detection Using Very-High-Resolution Satellite Imagery and Deep Learning. In: Remote Sensing 12 (7), S. 1145. DOI: 10.3390/rs12071145.
Korznikov, Kirill; Kislov, Dmitry; Doležal, Jiří; Petrenko, Tatyana; Altman, Jan (2022): Tropical cyclones moving into boreal forests: Relationships between disturbance areas and environmental drivers. In: The Science of the total environment 844, S. 156931. DOI: 10.1016/j.scitotenv.2022.156931.
Morresi, Donato; Maeng, Hyeyoung; Marzano, Raffaella; Lingua, Emanuele; Motta, Renzo; Garbarino, Matteo (2024): High-dimensional detection of Landscape Dynamics: a Landsat time series-based algorithm for forest disturbance mapping and beyond. In: GIScience & Remote Sensing 61 (1), Artikel 2365001. DOI: 10.1080/15481603.2024.2365001.
Senf, Cornelius; Seidl, Rupert (2021): Mapping the forest disturbance regimes of Europe. In: Nat Sustain 4 (1), S. 63–70. DOI: 10.1038/s41893-020-00609-y.
Stritih, Ana; Senf, Cornelius; Seidl, Rupert; Grêt-Regamey, Adrienne; Bebi, Peter (2021): The impact of land-use legacies and recent management on natural disturbance susceptibility in mountain forests. In: Forest Ecology and Management, 484, S. 118950. DOI: 10.1016/j.foreco.2021.118950.
Citation: https://doi.org/10.5194/nhess-2024-217-RC2 -
AC2: 'Reply on RC2', Kirill Korznikov, 17 Mar 2025
We are grateful to the referee for taking the time to review our manuscript and for their careful consideration, including attention to even the smallest details that warranted comments or raised questions. We greatly appreciate the effort invested in providing detailed feedback.
However, we noticed that several points in this review align closely with those raised by the first referee. While we understand that reviewers may naturally reach similar conclusions, this overlap makes it difficult to determine whether these remarks reflect the referee's independent assessment or were influenced by the earlier review. We believe an independent evaluation is important to ensure a fair and unbiased review process, and we hope this was the intention behind the provided feedback.
As Reviewer#1 addressed correctly, the most interesting part of the methodology is the detection of windthrow areas using high-resolution satellite imagery and machine learning approaches, which was already published by the authors in Kislov und Korznikov (2020). The main difference is using only one Pleiades images instead of several images (Pleiades and WorldView).
We would like to clarify, as mentioned in our response to the first referee, our article is not intended to focus on methodological advancements. Rather, it presents a study aimed at identifying spatial patterns affected by significant forest disturbances linked to a weather hazard. While we utilized previously developed neural network algorithms, this does not diminish the originality or importance of our study. Applying established neural network techniques to practical, large-scale environmental challenges, as we have done, is a substantial and complex task. Although the algorithm itself was developed earlier, implementing it to analyze extensive territories and uncover real-world disturbance patterns requires considerable adaptation and refinement. This practical application is a key contribution to our work, underscoring its relevance to the scientific community. We fully acknowledge the existence of prior studies that introduced the methodological foundation we employed. However, we believe it is inappropriate to question the novelty of our work on the basis that it builds on established methods. The majority of scientific research relies on well-developed approaches to advance knowledge in specific domains, and our manuscript does not claim to introduce a new method. Rather, its value lies in the novel insights gained by applying these methods to assess large-scale disturbance patterns.
Furthermore, a similar study by Korznikov et al. (2022) also detected windthrow areas using high-resolution WorldView images and analyzed predictors for different windthrow sizes with comparable data. While investigating predictors remains relevant and warrants further exploration, the current study's predictor set is limited, and the reliability of specific datasets (e.g., Global Tree Cover and Aster DEM-derived data) is questionable. The study does not adequately discuss these limitations.
Our current manuscript addresses a fundamentally different research focus compared to the cited study. While Korznikov et al. (2022) investigated a post-tropical cyclone event involving strong wind gusts and rainfall, our study examines a winter extratropical cyclone, characterized by heavy wet snow precipitation combined with wind gusts. Despite similarities in the resulting windthrow patterns, the underlying disturbance mechanisms differ significantly, as do the respective findings. We are unsure why the existence of a previous study would diminish the relevance of our current work. The scientific community routinely builds upon established findings to refine methods, expand knowledge, and explore broader applications. In our case, we utilized a previously developed technique to address a novel research question: the unknown impact of disturbance across a vast territory caused by a winter cyclone. We are also concerned by the referee’s suggestion that Global Tree Cover and ASTER DEM-derived data are “questionable.” Such a claim requires substantial supporting evidence. Both datasets are widely used in peer-reviewed studies and, while they have known limitations (as is the case with most global datasets), they remain valuable tools for large-scale environmental assessments. Dismissing these data sources outright would severely limit progress in understanding landscape-scale processes.
In our study, we deliberately relied on publicly available global datasets to ensure our approach could be applied universally, making the methodology reproducible across various regions. While we incorporated a unique vegetation cover and forest-type map derived from Sentinel-2 imagery, we designed the framework to be adaptable — alternative cartographic materials can easily replace this dataset.
The referee’s comment suggests that without highly detailed datasets (e.g., tree height, and species composition), meaningful analyses are impossible. We respectfully disagree. While such data would indeed improve precision, they are rarely available at regional or global scales. Relying solely on highly localized data risks restricting scientific insights to a few well-documented areas, overlooking the broader patterns that emerge across less-studied landscapes.
We believe that global datasets provide a reasonable and effective foundation for understanding disturbance drivers at large scales. Moreover, these insights can serve as a valuable starting point for more refined studies in regions where detailed data are available.
Lastly, we find the referee’s request for an explanation of why certain predictors were not included somewhat surprising. It is impractical to justify every possible variable that was not included, especially when data limitations are common in large-scale studies. Instead, we focused on the most relevant and accessible predictors for achieving our study’s objectives. We trust this response clarifies our rationale and addresses the referee's concerns appropriately.
Moreover, the study does not adequately address the broader applicability of its findings. Given the challenges in transferring the methodology to other regions, the discussion does not sufficiently explore how the results could be used for improving forest management.
The neural networks we employed are capable of identifying windthrow locations following forest disturbance events, provided very high-resolution satellite imagery is available. Additionally, the predictor data we used are global, publicly accessible, and suitable for replication in different regions. We believe our study has clear implications for improving forest management, particularly by providing a framework that can be adapted to various disturbance types across diverse landscapes. One of the key outcomes of our work is the demonstration. The methodology has the potential to be applied in other regions, supporting better-informed decision-making in forest monitoring and management.
Overall, the study offers limited new scientific contributions.
Although we respect the referee's opinion, we must respectfully disagree. Moreover, this statement lacks supporting evidence, making it appear subjective. Our study provides valuable insights by demonstrating the contribution of various predictors to windthrow vulnerability, particularly in the context of winter cyclonic extremes, a disturbance type that has received limited attention in the scientific literature. We believe these aspects significantly advance the understanding of forest disturbances driven by such an extreme weather event.
I also agree with Reviewer #1 that the terminology is unclear and confusing. Since risk has multiple definitions, the authors should explicitly define their usage. As Natural Hazards typically defines risk as a combination of hazard, exposure, and vulnerability, I recommend using susceptibility or predisposition assessment instead, as these terms more accurately reflect the study’s approach.
We previously addressed this concern in our response to the first referee. We fully acknowledge the need to improve the consistency of terminology and enhance the overall clarity of the text. Specifically, we recognise the importance of refining our use of the term "risk." As the referee has pointed out, "risk" can have multiple interpretations, and in some publications, it has been used synonymously with "vulnerability." The referee’s explanation convinced us to adopt more precise and consistent terminology throughout the manuscript.
However, given the limited novelty and added value of the study, I cannot recommend publication in its current form. A relevant improvement of the study could be the actual creation of a spatial continuous predisposition assessment based on the authors findings with improved and more predictors.
The criteria by which this judgment was made remain unclear, especially given that our study presents new findings and valuable insights into windthrow vulnerability. Also, the suggestion to enhance the study by creating a continuous predisposition map based on additional predictors is unfortunately not feasible for our study area. Such datasets are either unavailable or incomplete in the regions we analyzed. While detailed data may exist for select areas with intensive forest inventory systems, this is the exception rather than the rule. Limiting research to data-rich regions would risk leaving large portions of the world’s forests and disturbance phenomena, particularly in remote or underexplored areas, unstudied. We believe that exploring these regions using publicly available datasets, despite their limitations, is an important contribution to the broader understanding of forest disturbances.
L11/12: This study does not introduce a novel approach combining convolutional neural networks (CNNs) with Random Forest (RF) machine learning classifiers, as it was already introduced in the last publication (Korznikov et al. 2022). Please clarify that the study applies this approach rather than introducing it.
We appreciate the referee’s observation. However, we believe it is appropriate to describe our approach as novel, given that it has yet to be widely implemented in practice. While a similar method was previously applied in Korznikov et al. (2022), has not yet widely adopted this combined approach for practical forest disturbance detection. Novelty in research is not solely defined by introducing entirely new methods but also by demonstrating their successful application in new contexts or at broader scales. In this regard, our study presents a valuable contribution.
L22: The limited number of predictors and their questionable reliability mean that the proposed framework is not robust.
The effectiveness of a model does not necessarily improve by increasing the number of predictors, especially when those additional variables are highly correlated. Methods such as Random Forest are designed to identify the most informative predictors, and in many cases, a smaller set of carefully chosen variables with stronger explanatory power yields better results. Regarding the reliability of our predictors, the referee has not provided specific evidence or reasoning to support this concern.
L33: Sentinel-2 and Landsat data are freely available since many years, not only recently.
We apologise for the unclear presentation. What we intended to convey is that very high-resolution satellite images have become more widely accessible in recent years, as opposed to medium-resolution imagery like Landsat, which has been available for a longer period. We will revise this sentence in the manuscript to better clarify this point.
The manuscript lacks references to several state-of-the-art studies on this topic.
L35 Include relevant disturbance detection studies, such as Senf und Seidl (2021) and Morresi et al. (2024). L54 Add a reference to Stritih et al. (2021)We appreciate the referee’s suggestion and acknowledge the importance of referencing key literature. While we have already cited foundational studies such as Hansen et al. (2014), which laid the groundwork for annual Global Forest Change layers, we recognise that additional references could further strengthen the introduction. We are happy to include other relevant studies.
L57: The most critical factors for windthrow susceptibility are stand height and species composition. Do not focus only on the data available in your study.
As mentioned previously, we intentionally focused on data that were available across our extensive study area to ensure reproducibility. While canopy height may be a reasonable predictor in some contexts, many researches has shown it to be relatively weak for predicting windthrow susceptibility. For example, while canopy height is often intuitively perceived as a strong predictor, its explanatory power is limited for both individual uprooting/breaking and large-scale windthrows. Experienced forest managers understand that height alone is not the key factor. Rather, the slenderness ratio (the relationship between tree height and diameter at breast height), plays a far more important role, along with the crowns’ wind resistance. This is well-documented in specialized forestry literature. Additionally, while individual emergent trees that rise above the main canopy may indeed be more vulnerable, they are outliers relative to the overall canopy structure. In large-scale wind disturbances, a domino effect is often observed: large trees fall and subsequently, damage or break neighbouring trees, and smaller trees, due to their size and reduced stability, can also be prone to uprooting or breakage. In such a common scenario, canopy height alone becomes a weak predictor of windthrow susceptibility.
Regarding species composition, our forest-type layer, derived from Sentinel-2 spectral data, effectively captures key species-related characteristics.
L60: The approach is not new and was already tested. Please describe the differences compared to Korznikov et al. (2022).
We understand the referee’s concern and have to clarify in our manuscript how our study differs from Korznikov et al. (2022). While the cited study addressed post-tropical cyclone disturbances, our work focuses on the unique disturbance scenario caused by a winter extratropical cyclone with wet snow precipitation and wind gusts. This is a critical distinction that we believe justifies the novelty of our approach.
Figure 1: Where can I see the windthrow impact on forest cover mentioned in the figure description?
The forest loss resulting from windthrow is the impact itself, as stated. However, we understand that this may not have been conveyed, and we are willing to adjust the figure description to better highlight this.
L120: How can a 39 mm precipitation event result into an 80 mm snow cover? Please clarify.
We appreciate the attention to detail. While this relationship is a well-known phenomenon in meteorology (snow water equivalent is the amount of water contained in snowpack), we can add a brief clarification in the text to ensure clarity for readers less familiar with this.
L123: Please locate the weather station in the map. How represent is this weather station for the study area?
We did not interpolate weather parameters from this station across the entire island. Instead, we presented the recorded meteorological data during the cyclone event as factual observations rather than regional averages. Since our study does not conduct meteorological modelling, we believe further discussion of the station’s representativeness is unnecessary. However, we can add the station's location on the map if this would improve transparency.
L124: Reference is missing, which wind speeds are to be expected during strong storms in the study area.
Since our manuscript does not focus on meteorological analysis, we provided only the observed wind data directly related to the cyclone event. Nonetheless, we can include a reference that outlines typical storm intensities in the region if the editor believes this would strengthen the text.
L130: Please provide more information about the accuracy assessment.
The achieved accuracy of the neural networks is already detailed in the referenced methodological publications. Repeating this technical content would be redundant, given that these details are not the focus of our study.
L131: Why were these areas excluded? Small disturbances and even single broken trees are crucial for post-disturbance management in many regions. High-resolution data provide a unique opportunity to detect such areas, which are often prioritized for sanitation felling.
We appreciate the referee's concern. However, these small patches are typically composed of fallen branches and wooden debris rather than single trees. While we did not intentionally train the model to detect such small features, they are often embedded within larger windthrow patches and are naturally recognized by the neural network as part of the windthrows. We now see that our initial wording may have caused confusion, and we are open to rephrasing this for clarity. See also the attached file with the example of delineated single fallen trees and wood debris.
L138: Justify the selection of these specific predictors.
Our methodology is based on the reasonable assumption, as stated in the manuscript, that windthrow exposure is influenced by terrain features. Therefore, we selected predictors representing key terrain characteristics to assess their significance. While the referee suggests justifying these choices in the methods section, we note that similar justifications are also absent in the paper by Stritih et al. (2021), which the referee previously referenced. Our methodology follows standard practice in this regard.
L140: Specify that the reported 7 m accuracy is an average. Accuracy is significantly lower in steep slopes.
We will revise the text to clarify that the reported 7 m accuracy is an average.
Table1: Given the >10 m inaccuracy in steep slopes and the 30 m resolution, such detailed classification steps are questionable.
Steep slopes represent only a small proportion of the total forest area and windthrow patches (as shown in Fig. 9c). Consequently, this has minimal impact on the overall analysis. The classification steps we employed effectively illustrate patterns in forests and windthrows on more typical terrain.
L167: What is the total number of points? Did you consider/check spatial correlations?
In the referenced paragraph, we specified the number of points used in the Random Forest model and for plotting in Fig. 3. If this is unclear, we will rephrase the text to improve clarity.
Regarding spatial correlation, we recognise two important points: 1. In training classifiers like Random Forest, spatial correlation can become problematic if data points are sampled in a way that introduces bias (for example: https://doi.org/10.1038/s41586-022-04959-9). We avoided this by randomly distributing points across the island and all windthrow patches, ensuring the sample reflected predictor gradients. 2. Windthrows inherently display spatial clustering because certain landscape features favour their formation. This clustering is a natural property of the disturbance itself, not a methodological flaw. Spatial autocorrelation in this case reflects real environmental patterns rather than an artifact of the analysis.
Fig 11. I The y-axis labelling is unclear. I suppose that the value is a range, for example in case of slope from 0 to 6° or? Please improve the labelling.
We agree that the specification of quantitative categories such as “from-to” should be clearer. We will revise the labelling to explicitly indicate both the lower and upper values for each range (e.g., slope from 0° to 6°) to ensure the axes are easily interpretable.
L241-244: Discuss why elevation is an important predictor. Is it due to random storm patterns, or does it reflect species composition at different elevations?
We recognise the need to provide more detailed discussion. Elevation can influence windthrow susceptibility through complex environmental gradients such as temperature, precipitation, soil type, and species composition. We will expand this section to clarify how these factors interplay concerning windthrow vulnerability.
L245/246: Add relevant references supporting the claim that denser stands are more susceptible.
We have already referenced publications that discuss the influence of forest structure, including canopy density, on windthrow susceptibility. These studies indicate that denser stands can be more susceptible to windthrow. We stand by these references, as they provide direct evidence related to our results.
L248/249: While these statements are valid, your data cannot distinguish the causes of open stands, as no management intervention or past disturbance records were included.
We do acknowledge that some open oak forests in the southern part of the island were influenced by traditional coppicing practices. We will clarify this in the manuscript, highlighting that the open canopy patterns may have been historically shaped by the past local people's activities. The stone birch forests in the mountains have a naturally open canopy.
L270-272: Since the analysis is based on a single storm event, generalizing about forest structure development introduces significant uncertainties.
The cyclone event we analyzed was extraordinarily impactful, resulting in more than 21.73 km² of forest loss. The majority of this loss occurred in mixed and dark coniferous forests (97.1%), which is a substantial portion of the total forest area. This large-scale impact is significant, and we also observe cascading effects like a bark beetle outbreak that worsened the damage.
L279: Large-scale high-resolution LiDAR acquisitions, combined with satellite data, now enable more accurate predisposition assessments only based on remote sensing data.
We respectfully disagree with the suggestion that interpreting remote sensing data without local expertise leads to correct conclusions. We acknowledge the need for expert knowledge in interpreting any high-resolution remote sensing data (such as LiDAR) provides a powerful tool for assessing forest disturbance susceptibility when used with appropriate methodology and context taken from the ground survey data. For example, the remote sensing data will not be able to determine the domino effect of fallen trees and to reconstruct the mechanism of windthrow events.
L285: As the disturbance detection was already published, only parts of the predictor analysis are new.
We must clarify that none of the results presented in our manuscript have been published previously. The analysis, including the predictors we used and the methodology applied, is entirely original. Our approach and findings contribute novel insights to the understanding of windthrow susceptibility, and we are confident in the originality of this work.
L287: The study does not assess future predisposition. Please revise this statement.
We respectfully disagree with the assertion that the study does not assess future predisposition. While we do not attempt to model future events, our findings suggest that similar windthrow events may occur in neighbouring regions under comparable meteorological conditions. This provides useful insights into potential future risks.
L289: The claim that wet snow is the primary windthrow factor is based on a single event under specific conditions and is not a robust conclusion.
We recognise the concern that this finding is based on a single event, but it is important to note that this storm event caused forest loss comparable to the total annual loss from all disturbances in a country like Switzerland (https://gfw.global/4i9GH3a). Additionally, winter extratropical cyclones, especially those accompanied by wet snow, are not isolated events in Northeast Asia. Increased cyclonic activity, likely driven by climate change, is a significant disturbance risk for many regions with similar atmospheric conditions and forest types.
L292: A comprehensive windthrow predisposition model would require additional parameters. Clarify that your statement applies only to the
We acknowledge that a more sophisticated predisposition model could benefit from additional parameters. However, our study focuses on predictors that are most relevant to the specific event and study area. Regarding the reliability of our chosen predictors, the referee has not provided specific evidence or justification for questioning their validity. We remain confident that our predictor selection effectively captures the key factors influencing windthrow susceptibility in our study area.
L295: Disturbance analyses are valuable for forest management, yet the manuscript only briefly discusses their practical implications.
We agree with the referee that adding practical recommendations for forest management could enhance the manuscript. However, we are also mindful of the journal’s focus, and while we have provided a valuable contribution to understanding forest disturbances, we are unsure if extensive management recommendations fit the scope of the article. Nevertheless, we recognise that the study’s results can aid forest managers in making informed decisions, and we will expand on these practical applications in the revised manuscript.
We appreciate the time and effort the referee has dedicated to reviewing our manuscript and providing constructive feedback. While we respectfully disagree with some of the points raised, we believe that the suggested improvements have helped clarify several aspects of our study. We have carefully considered all the comments. We trust that the updated version addresses the concerns raised and enhances the overall quality of the manuscript.
Citation: https://doi.org/10.5194/nhess-2024-217-AC2
-
AC2: 'Reply on RC2', Kirill Korznikov, 17 Mar 2025
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 210 | 44 | 10 | 264 | 22 | 7 | 6 |
- HTML: 210
- PDF: 44
- XML: 10
- Total: 264
- Supplement: 22
- BibTeX: 7
- EndNote: 6
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1