We thank the reviewer for their comments and the opportunity to clarify the scope of our work. The reviewer notes that the manuscript does not employ deep learning methods and combines data from passive optical and SAR satellites in a way they consider not novel. We would like to clarify that our study does not claim novelty in the combined use of optical and SAR data, nor does it focus on the application of deep learning techniques. Rather, the main contribution  - as explicitly outlined in the manuscript - lies in the development and adaptation of eight distinct metrics to generate, for the whole of Europe, a comprehensive dataset describing structural diversity at multiple spatial resolutions. To our knowledge, this represents the first attempt to systematically map structural complexity at a quasi-continental scale.

The review also includes DOIs referencing studies that are presented as closely related to our work. After careful examination, we found that some of the provided DOIs appear to be incorrect or point to studies that differ substantially in scope and methodology, particularly in relation to the construction of a dataset describing the structural diversity of European forests.

We hope this clarifies the scope and contribution of our work.

Below, we have pasted the DOI verification performed using the DOI Foundation’s resolver (https://www.doi.org/), along with the corresponding bibliographic information where available:

• 10.1088/1748-9326/ac5f6d
  
  DOI NOT FOUND
  
  
  
  
• 10.1016/j.foreco.2023.120987
  
  McKinney, Caleb M., Ronald E. Masters, Arjun Adhikari, Bijesh Mishra, Omkar Joshi, Chris B. Zou, and Rodney E. Will. "Forage quantity and protein concentration changes across a forest-savanna gradient with management implications for white-tailed deer." Forest Ecology and Management 538 (2023): 120987.
  
  
  
  
• 10.3390/rs14143345
  
  Shao, Z., Zhang, X., Zhang, T., Xu, X., & Zeng, T. (2022). RBFA-Net: a rotated balanced feature-aligned network for rotated SAR ship detection and classification. Remote Sensing, 14(14), 3345.
  
  
  
  
• 10.5194/bg-18-1234-2021
  
  DOI NOT FOUND
  
  
  
  
• 10.1002/eap.2567
  
  Grinde, Alexis R., Melissa B. Youngquist, Robert A. Slesak, Stephen R. Kolbe, Josh D. Bednar, Brian J. Palik, and Anthony W. D'Amato. "Potential impacts of emerald ash borer and adaptation strategies on wildlife communities in black ash wetlands." Ecological Applications 32, no. 4 (2022): e2567.
• 10.1109/TGRS.2022.3156789

             DOI NOT FOUND

We thank Referee #2 for the constructive comments.  We are pleased that the referee recognises the relevance of a wall-to-wall, spatially explicit dataset on structural diversity for European forests. Below we respond point-by-point, focusing on clarification of design choices, scope, and interpretation, as appropriate for the ESSD public discussion stage

General comments

I feel this study lacks clarity in the choices that were made: e.g. why these particular eight structure metrics? Why Sentinel-1/2 and ALOS-Palsar? I think clarification on a couple of methodological steps and decisions will help to assess whether the metrics and datasets presented here, fill the existing gap of structural diversity metrics for European forests and for which purposes and models the datasets can be used.

We agree with the referee that clearer justification of several key methodological choices would improve the transparency and interpretability of the study. While the rationale for the selection of the structural diversity metrics and the choice of Sentinel-1, Sentinel-2 and ALOS-PALSAR as predictor datasets is introduced in the manuscript, we acknowledge that these aspects would benefit from being stated more explicitly. In particular, the sensor combination was chosen to enable the production of spatially and temporally consistent, wall-to-wall estimates of forest structural diversity suitable for large-scale and long-term monitoring. These points are addressed in detail in the responses to the specific comments below and will be consolidated in a revised version of the manuscript.

Specific general comments:

I appreciate this study and the focus on structural diversity of European forests, as well as the integrated use of several openly-available EO-datasets. I would welcome a couple of clarifications on general decisions taken and processing steps. I think the preprint and the usability of the dataset would be enhanced by this. These points do not necessarily correspond to one line or particular paragraph in the preprint, hence this ‘specific general comment’-section

1. Existing structural diversity indices or datasets

While this study presents quite a complex assessment of structural diversity, I am missing a paragraph on existing structural diversity indices and datasets. Not all of the existing datasets are wall-to-wall or some only address vertical or horizontal structural diversity but I still think these have to be mentioned and their advantages and disadvantages should be addressed. This would help to identify the current knowledge gap and underline the need for other structural diversity metrics, such as those presented here.

I would expect that at least (but not limited to these): the GEDI L2B products (PAI, FHD, PAVD, …), GEDI L4C WSCI (also z, xy), Forest canopy structural complexity (CSC), are addressed.

1.5. Especially concerning the provided GEDI products (L2B and L4C): Did you assess these products as potential target variables? Why did you choose to calculate 8 new metrics?

We thank the referee for this valuable comment. We fully acknowledge that several structural complexity indicators already exist and have proven highly useful for characterising specific aspects of forest canopy structure, including GEDI L2B variables, GEDI L4C WSCI products, and other canopy structural complexity (CSC) indices described in the literature. Structural complexity is, however, a multifaceted concept that can be interpreted and quantified in different ways, depending on the ecological dimension of interest and the intended application (e.g. Coverdale & Davies 2023, ; LaRue et al. 2023).

We acknowledge that aspects of vertical profile heterogeneity are already addressed in existing GEDI-derived indices, including Foliage Height Diversity (FHD) and the GEDI Waveform Structural Complexity Index (WSCI). These indices have proven highly valuable for characterising forest structural complexity at large scales. At the same time, recent work has shown that both FHD and WSCI exhibit strong scaling relationships with canopy height, with the relative contribution of different waveform layers varying across biomes (e.g. de Conto et al., 2024).

In the context of this study, our objective was therefore not to directly use existing GEDI products as target variables, but to develop structural diversity metrics that explicitly quantify heterogeneity while reducing direct dependence on top-of-canopy height (RH98). This motivated the selection of complementary metrics based on the distributional properties of GEDI relative height profiles and canopy cover, allowing us to characterise structural heterogeneity within and among GEDI observations in a way that is interpretable and suitable for wall-to-wall prediction.

We agree that the manuscript would benefit from a clearer and more consolidated explanation of why these eight metrics were selected and what each of them represents. The metrics were chosen to explicitly span three complementary dimensions of structural diversity: vertical heterogeneity within individual canopy profiles, horizontal heterogeneity among GEDI observations within a spatial unit, and combined multivariate structural diversity. In a revised version of the manuscript, we will expand this discussion to better position the proposed metrics with respect to existing structural diversity datasets and clarify the specific knowledge gap they are intended to address.

References:

Coverdale, T. C., & Davies, A. B. (2023). Unravelling the relationship between plant diversity and vegetation structural complexity: A review and theoretical framework. Journal of Ecology, 111(7), 1378-1395.

de Conto, T., Armston, J., & Dubayah, R. (2024). Characterizing the structural complexity of the Earth’s forests with spaceborne lidar. Nature Communications, 15(1), 8116.

LaRue, E. A., Knott, J. A., Domke, G. M., Chen, H. Y. H., Guo, Q., Hisano, M., Oswalt, C., Oswalt, S., Kong, N., Potter, K. M., & Fei, S. (2023). Structural diversity as a reliable and novel predictor for ecosystem productivity. Frontiers in Ecology and the Environment, 21(1), 33–39

1. Choice of 8 metrics

I understand that these 8 metrics address vertical, horizontal and combined structural diversity and therefore depict the multidimensional structural diversity. However, I feel the manuscript misses a paragraph that explains why these metrics where chosen, and a little more detail on what the individual metric depicts. I am aware that this is briefly mentioned in 2.1.2, 2.1.3 and 2.1.4 (error in preprint, line 186 the subsection should be 2.1.4) but I think it would be good to elaborate on this to ensure that every reader can follow understand the choice made.

We agree with the referee that the manuscript would benefit from a clearer and more consolidated explanation of why these eight metrics were selected and what each of them represents. While the rationale for the metrics is briefly introduced across Sections 2.1.2–2.1.4, we acknowledge that this information should be made more accessible to readers.

The eight metrics were deliberately chosen to span three complementary dimensions of structural diversity. Metrics derived from the distributional properties of GEDI relative height profiles characterise vertical heterogeneity within individual canopy profiles, capturing differences in vertical layering and profile shape. Metrics based on variability among GEDI observations within a spatial unit describe horizontal heterogeneity, reflecting spatial variation in canopy height and cover. Finally, multivariate metrics integrate both vertical and horizontal information to represent combined structural diversity within a single framework.

This design was intended to provide an interpretable yet comprehensive representation of forest structural diversity, while avoiding redundancy with top canopy height and among metrics. In a revised version of the manuscript, we will add a dedicated paragraph synthesising this rationale and briefly describing the ecological meaning of each metric to improve clarity for non-specialist readers.

1. 10km, 5km, and 1km spatial resolution

Can you please specify how you chose these 3 spatial resolutions? Can you also address the potential application/uses of structural diversity metrics at these spatial resolutions? Do modelling studies operate at these resolutions? I am just wondering, as the all the inputs have a much higher spatial resolution and also certain potential applications that I can think of (besides modelling) would benefit from a higher spatial resolution, e.g. disturbances (small-scale disturbances are very common), edge effects, forests become more fragmented, …

We thank the referee for this important comment. We agree that the rationale for selecting the three spatial resolutions and their intended applications should be more explicitly articulated in the manuscript.

The choice of 1 km, 5 km, and 10 km resolutions reflects a trade-off between spatial detail, sampling density of GEDI observations, and the robustness of the derived structural diversity metrics. Because GEDI provides sparse footprint-level measurements, reliable estimation of structural diversity within a spatial unit requires a sufficient number of GEDI shots. Coarser spatial resolutions therefore improve metric stability, whereas finer resolutions provide greater spatial detail at the cost of higher uncertainty.

The 10 km resolution was identified as the most robust scale for continental-scale analyses and large-area monitoring, and is broadly compatible with the spatial aggregation typically used in regional ecosystem modelling and model–data integration studies. The 5 km and 1 km products were additionally provided to support applications requiring finer spatial detail, such as regional ecological analyses and biodiversity assessments.

We acknowledge that certain applications, including the analysis of small-scale disturbances, edge effects, and forest fragmentation, would benefit from even finer spatial resolution. However, at such scales the limited density of GEDI observations substantially constrains the reliable estimation of structural diversity metrics over large areas. For this reason, we opted to provide multi-resolution products that allow users to select the spatial grain most appropriate for their specific application, while explicitly recognising the associated trade-offs in uncertainty and model performance.

In a revised version of the manuscript, we will clarify this rationale and more explicitly link each spatial resolution to its potential applications and limitations.

Specific comments

1. line 74/75: give sources/citations from other studies.

We thank the referee for this suggestion. We agree that additional references should be provided to support this statement. Appropriate citations to previous studies will be added in the revised manuscript.

1. Figure 1: difference between solid edge boxes and dashed boxes is not clear. I am not sure what the difference between ‘data that were directly utilised’ and ‘raw, original data’ means pratically

We agree that the distinction between solid and dashed boxes in Figure 1 is not sufficiently clear. The intention was to differentiate between original input datasets and intermediate products derived from these inputs. We will revise the figure caption and legend to clarify this distinction and improve interpretability.

1. line 128: What is M?

We thank the referee for pointing this out. M refers to the number of valid GEDI shots within each spatial unit. This will be explicitly defined in the revised manuscript.

1. line 132: What are ‘our calculations’ referring to? The calculation of the eight individual structural metrics?

Yes, “our calculations” refers to the computation of the eight structural diversity metrics derived from GEDI observations. We will rephrase this sentence to make this explicit.

1. line 128 to 136: I do not understand this paragraph. Is this the GEDI shot selection? Is the reference to pixels or areas? What is the median value based on? Same question for the Z-score ?

We thank the referee for highlighting that this paragraph is unclear. The intention of this section is to describe how GEDI shots are selected and aggregated within spatial analysis units prior to the calculation of the structural diversity metrics. The reference is to spatial grid cells (analysis areas at the chosen resolution), rather than individual GEDI footprints or image pixels.

For each spatial unit, all valid GEDI shots overlapping that unit were collected and used to derive the structural diversity metrics. A minimum number of valid GEDI observations was required to ensure reliable metric estimation, and spatial units with insufficient sampling were excluded based on a threshold derived from the distribution of available GEDI shot counts. In addition, extreme metric values were filtered using a z-score–based criterion to remove outliers.

We agree that the current wording does not make these steps sufficiently transparent, and we will rewrite this paragraph in the revised manuscript to clearly describe the spatial units, filtering criteria, and terminology used.

1. Line 152-153: definition not clear. What is the expected CV of VP? I think ‘latter’ is not correct in this sentence anything. What is central tendency?

We agree with the referee that the definition in this sentence is currently unclear and that the wording is imprecise. In this context, the coefficient of variation (CV) of a vertical profile is defined as the ratio between the standard deviation and the mean of the relative height distribution for a given GEDI shot. The “expected value” therefore refers to the mean of the vertical profile, which represents its central tendency.

The sentence will be revised to explicitly state this definition and to remove the ambiguous use of “latter”, thereby clarifying that higher values of CV indicate greater vertical dispersion relative to the mean profile height and hence higher vertical heterogeneity.

1. Line 158, reference to Figure S2: be more specific in the main text and also in the Figure S2 caption. There are so many panels in this figure with different foci; please clarify which panel refers to what

We agree that the references to Figure S2 are currently too general. We will revise both the main text and the Figure S2 caption to explicitly indicate which panels correspond to which metrics and analyses.

1. Line 168: reference to Figure S2: same comment as above. Be more specific in Figure S2, which panel/s illustrate what (skewness, kurtosis?); this will increase understandability

We agree with the referee’s suggestion. We will specify which panels of Figure S2 illustrate skewness, kurtosis, and other distributional properties to improve readability and understanding.

1. Figure 4: PCA (panel D) displayed but not mentioned in text. Why was a PCA performed? What does the result show us?

We thank the referee for highlighting this point. The PCA shown in Figure 4D is currently referred to in the manuscript in the context of intercorrelation among the structural diversity metrics (lines 383–387). However, we agree that the purpose of the PCA and the interpretation of its results are not described explicitly.

The PCA was used to assess the degree of correlation among the predicted structural diversity metrics and to illustrate their complementarity across vertical, horizontal, and combined dimensions. In the revised manuscript, we will explicitly introduce the PCA in the main text, clarify its role, and provide guidance on how to interpret Figure 4D.

1. Table B1, B2: Model validation paragraph in main text (starting line 415). Some of the validation results are quite poor (both for the random and cross-validation method); e.g. 𝜎 of Canopy Cover (𝜏𝐶𝐶 ) 0.16 (random validation). I think this should be discussed more. Are the provided datasets on figshare those that showed these poor validation results? This is not clear. I am not sure if maps/datasets with such a low validation score should be included/used, e.g. from modellers as input to their models. Or what do you think? Maybe I am misunderstanding this. I think this is not highlighted/marked sufficiently in the text.

We agree that the validation results require clearer discussion. Some metrics, particularly those describing variability in canopy cover, are more challenging to predict from Earth observation data and therefore show lower performance. All datasets provided on Figshare correspond to the reported validation results. In the revised manuscript, we will more explicitly discuss the range of model performance, highlight metrics with lower predictive skill, and clarify appropriate use cases and limitations for downstream applications.

1. Table S1: GEDI points; I think conventionally it is GEDI shots or GEDI footprints but not points.

We agree with the referee that “GEDI shots” or “GEDI footprints” is more appropriate terminology than “points”. We will correct this throughout the manuscript and supplementary material.

Technical corrections

1. line 186: the subsection should be 2.1.4

We thank the referee for noting this error. The subsection numbering will be corrected to 2.1.4 in the revised manuscript.

We thank the referee for the positive and encouraging assessment of the dataset and the manuscript. We appreciate the referee’s recognition of the potential value of the proposed structural diversity metrics for a broad range of applications, which is the intended purpose of our dataset. Below we provide a point-by-point response, clarifying the rationale behind our design choices, the study’s scope, and the interpretation of the results, in line with the ESSD public discussion stage.

Overall, I find this work worthy of publication after minor revisions—see Specific Comments and the following note.

Specific comments

73-74 – I am not comfortable with the use of the term “predict”, which I associate to the weather forecast or climate projections realms. Perhaps “estimate”? Although I realise such a setup may be employed for future data too.

We thank the referee for this helpful suggestion. We agree that the term “predict” may be ambiguous in this context and could be interpreted as implying forecasting in a temporal sense. In this study, the modelling framework is used to derive spatially explicit estimates of forest structural diversity from Earth observation data. We will therefore replace “predict” with “estimate” (or “derive”) throughout the manuscript where appropriate, while retaining “predictive modelling” only when referring to the methodological framework.

82 – “A list of the metrics is reported in Table 1” this is a repetition with later L. 115. At this stage you could simply refer to Section 2.1 as opposed to the Table 1 directly.

87 – repetition. Remove “(Fig 1)”.

94-95 – repetition. Remove “also with tree cover exceeding 30%.”

We thank the referee for noting these repetitions. We will remove the redundant references and change the text as suggested.

92-98 – I understand that you employed more stringent criteria than those of FAO, but could you provide insight on your choices? E.g. 30% vs. 10% tree cover.

We thank the referee for this comment. We adopted a more stringent tree cover threshold (30%) than the FAO definition to focus the analysis on areas with clearly developed forest canopies and to reduce potential noise from sparsely treed or transitional land-cover types. This choice aims to improve the robustness of the derived structural diversity metrics, particularly given the reliance on GEDI observations and their sensitivity to canopy structure. We will clarify this rationale in the revised manuscript.

104 /Figure 1. “The process culminates with yellow boxes” – Make sure you refer to the right color (I don’t see yellow color in the figure).

We thank the referee for pointing this out. We will correct the color reference in the figure caption to ensure consistency with the actual figure.

126 – Why filter? To attenuate noise or other?

Filtering was applied to reduce noise and remove unreliable observations, ensuring robust estimation of structural diversity metrics. We will clarify the purpose of the filtering step in the manuscript.

158 – Correct typo: “an unimodal” to “a unimodal”.

Lines 176 / 186 – The sections here have the same number – 2.1.3! Also, if you make a distinction between vertical, horizontal, and v. and h. combined, then L. 177 should go to 2 diversity indices, and the following section with the remaining 3 combined indices, in accordance with Table 1.

189 – Watch for the top index in the capital sigma, it is missing.

We thank the referee for identifying these issues. The typos, subsection numbering, and missing index notation will be corrected in the revised manuscript.

213 – “The variables used as ML predictors were calculated from Sentinel-1, Sentinel-2, and ALOS-Palsar-2 observed data”. The predictors sources are presented very swiftly at this point, as if these data sets are the obvious choice. Can the authors provide a context as to why these observed data are fit for this purpose?

222 – Same as the comment above. At this line you report that 47 predictors were derived. If you could provide some information on how you got to this set of predictors, the trade-offs you had to deal with.

We agree that additional context is needed here. Sentinel-1, Sentinel-2, and ALOS-PALSAR-2 were selected because they provide complementary information on canopy structure, including optical properties, backscatter sensitivity to vegetation volume and moisture, and textural characteristics. The final set of predictors was derived to capture these complementary dimensions while balancing information content and redundancy. We will expand this section to better explain the rationale and trade-offs underlying the predictor selection.

239 – not sure I understand: “We selected all the valid images captured over Europe within a six-month window, centred around the day of maximum NDVI from the Sentinel-2 dataset” – so, of the entire dataset, you picked the day of maximum NDVI and took 3 months before and after it?

We thank the referee for highlighting this ambiguity. The six-month window refers to selecting observations within three months before and after the date of maximum NDVI for each pixel, rather than a single day across the entire domain. We will rephrase this sentence to clarify the temporal selection procedure.

327-341 – provided that, as the authors write, the 2nd approach is less prone to overfitting and more adequate for gridded data where spatial autocorrelation exists, why didn’t the authors just go with the 2nd approach?

We thank the referee for this insightful question. We agree that spatial cross-validation is better suited to account for spatial autocorrelation and to reduce optimistic bias in gridded datasets. However, the choice of validation strategy remains an active topic of discussion in the literature (Wadoux et al. 2021) , and different approaches serve different purposes.

In this study, we therefore included both spatial and random cross-validation. Spatial cross-validation was used to provide a more conservative assessment of model generalisation in the presence of spatial autocorrelation, while random cross-validation was retained to facilitate comparison with previous studies and to characterise overall model behaviour. In the revised manuscript, we will better justify the use of both approaches and clarify their respective roles and limitations.

Wadoux, A. M. C., Heuvelink, G. B., De Bruin, S., & Brus, D. J. (2021). Spatial cross-validation is not the right way to evaluate map accuracy. Ecological Modelling, 457, 109692.

355 – Figure 2 (but valid also for Figure 3 and annex figures). While the layout is clear and well organised, I suggest the colorbar to have: larger tick label font size, and a color palette with discrete bins, not continuous gradient. Also, possibly ticks with even values.

We thank the referee for the suggestion. We will improve the readability of the colorbars by increasing tick label size and considering a discretised colour palette where appropriate.

360 – Figure 3 is a very interesting figure unveiling insightful information on the relationship between precipitation and temperature in the different metrics. I wonder if there is a chance to point a region or two in the figure’s climate space. For instance, at L. 375, you refer to the characteristics for the Mediterranean region, it would help interpretation to find the region in Figure 3 if that is possible.

379-380 –It may seem obvious but I find it interesting that low Precip. and relatively high mean temperature is associated with the lowest levels of diversity.

384-385 – This is also interesting, that results are insensitive to the grain size.

We thank the referee for these observations. We will consider whether brief clarifying statements can be added to guide interpretation without overloading the figures.

413 – the Figure 4d graph introduces PCA. I would introduce it in the manuscript text providing insight on why it is there and what it tells us.

We agree that the PCA should be more clearly introduced in the manuscript. Although Figure 4D is currently referenced in the context of intercorrelation among the structural diversity metrics, the purpose of the PCA and the interpretation of its results are not described explicitly. The PCA was used to assess correlations among the predicted metrics and to illustrate their complementarity across vertical, horizontal, and combined dimensions. In the revised manuscript, we will explicitly introduce the PCA in the main text and provide guidance on how to interpret Figure 4D.

416-417 – So the model with Shannon index achieved highest scores, while the one with convex hull the lowest. Is there an explanation?

We thank the referee for this question. The higher performance of the Shannon index likely reflects its sensitivity to overall structural heterogeneity and its robustness to noise, whereas convex hull–based metrics may be more sensitive to outliers and sampling density. We will discuss this difference in the revised manuscript.

445-446 – This indicates well the novelty of this work! “indices based solely on optical data fail to capture crucial aspects of structural heterogeneity”.

We thank the referee for highlighting this aspect of novelty.

470 – Well crafted the “Potential applications”! At the third point, the one on Earth System Models, I would specify that these constitute the CMIP6 and soon to the CMIP7 set of simulation models contributing to the IPCC reports and their important guidance on present and future climate.

We thank the referee for the suggestion. We will consider explicitly referencing CMIP6 and forthcoming CMIP7 Earth system models in the discussion of potential applications.

Final note on clarifying the approach in the introduction.

As a reader I struggled at first to understand the use of predictors in this work. I think the authors could make an effort in the text to frame their work and clarify that combining different data sources forces to deal with gaps that need to be dealt with.

So on predictors, I would suggest adding a basic clarification, something like:

> Predictors are used to bridge the gap between sparse GEDI LiDAR observations and the need for continuous, large-scale forest structure maps. They act as observable proxies - derived from optical (Sentinel‑2), radar (Sentinel‑1, ALOS), and texture metrics - that contribute to describe canopy height, cover, and complexity. By feeding these predictors into a machine learning model, the study extrapolates GEDI-derived structural metrics across the entire domain, enabling wall-to-wall mapping and regular updates using available satellite data.

We thank the referee for this very helpful suggestion. We agree that the role of predictors could be more clearly framed in the introduction. In the revised manuscript, we will add a concise explanation clarifying that predictors are used to bridge the gap between sparse GEDI LiDAR observations and the need for continuous, wall-to-wall estimates of forest structural diversity, leveraging complementary optical and radar information through a machine learning framework.