This manuscript entitled “Wood density variation in European forest species: drivers and implications for multiscale biomass and carbon assessment in France” presents an analysis of wood density variation across temperate forest species in France. The study makes a valuable contribution to understanding wood density variation and its implications for forest biomass estimation. The authors constructed four linear models to identify the biotic and abiotic factors controlling wood density variability. They also applied the NFI-based and GIS-based models to generate a spatial distribution map of wood density across France. Finally, they evaluated the influence of wood density on biomass estimation at multiple scales, such as the plot, subregion, and country levels. A particularly interesting conclusion is that the choice of method for inferring wood density depends on the spatial scale of interest.

I have several comments regarding the methods and overall storyline of the study

Specific comments:

First, the results from the taxonomic and NFI-based models highlight the importance of tree species in explaining wood density variation. However, this may be partly due to the limited set of climate variables included. The authors only included two variables, mean annual temperature and precipitation. I think only temperature and precipitation may not fully reflect spatial differences in specific environmental conditions. It would be helpful to clarify why only these two climatic variables were selected. Would the inclusion of additional climatic and soil variables (such as soil nutrient availability, soil pH, or atmospheric humidity) alter the identified drivers?

Second, I think the inclusion of four distinct models seems somewhat redundant. The environmental model, in particular, appears less critical, as it overlaps significantly with the GIS-based model in terms of predictive capacity using abiotic variables. The GIS-based model, which incorporates both biotic and abiotic factors, could sufficiently illustrate the predictive power of remotely accessible variables (including environmental variables) without the need for a separate environmental model.

Third, given the hierarchical nature of the data (e.g., trees nested within species and plots), the use of ANOVA is not optimal. A generalized linear mixed-effects model (GLMM) would be more appropriate, as it allows variance to be partitioned into random effects (e.g., tree, species, plot) rather than attributing it entirely to fixed effects (e.g., genus, DBH, height). A GLMM would enable all density samples to be analyzed collectively, with nested random effects (e.g., trees within plot and species), thereby improving degrees of freedom and interpretive power.

Furthermore, I like this part of “Difference in forest aboveground biomass (AGB) stock depending on the method used to predict wood density”, and I additionally suggest extending this analysis to more explicitly quantify the total carbon stock (e.g., in PgC of aboveground carbon) and assess potential over- or underestimation at broad scales under different wood density inference methods.

Finally, it remains unclear how the NFI-based model, which relies on species identity, can be upscaled to generate a large-scale wood density map. Species-level information is often unavailable at broad scales, and the model's dependency on such data may limit its practical application. The authors should clarify how this limitation is addressed or propose alternative approaches for spatial extrapolation.

A more detailed description of the candidate predictor variables is needed. For instance, what is the spatial resolution of the remote sensing data used in the GIS-based model? Which specific spectral indices were selected, and from which satellite products were they derived?

Wood density variation in European forest species: drivers and implications for multiscale biomass and carbon assessment in France

Authors: Cuny, H. et al.

General assessment

The manuscript by Cuny et al. offers a valuable contribution by quantifying interspecific and intraspecific variation in wood density across France using the XyloDensMap dataset (≈110,000 increment cores). The authors use a clear modelling hierarchy: Taxonomic, Environmental, NFI-based, and GIS-based modelling framework, to disentangle ecological and structural drivers of wood density and to evaluate implications for biomass estimation across multiple spatial scales. The manuscript is well-structured, well-written, and analytically thorough. The figures are clean and intuitive, and the work is highly relevant for both forest ecology and the remote-sensing biomass community.

The core findings are robust and important: species identity dominates total variance (≈78.5%), intraspecific variation exists but remains difficult to model, and neglecting wood-density variation has minimal effect at the national scale but produces substantial biases (up to 30%) at subregional and plot scales. For these reasons, the study constitutes a significant step forward in improving estimates of biomass and carbon stocks.

There are, however, a couple of things that could be addressed to strengthen the MS. Below, I provide some comments for revision:

1. Cross-validation strategy requires clarification and extension

It is currently unclear whether cross-validation was performed through random sampling of individuals or whether spatial autocorrelation was accounted for. Because both wood density and environmental variables exhibit strong spatial structure, random train-test splits tend to inflate performance metrics.

I strongly recommend explicitly assessing spatial generalisation capabilities. This could include:

- Spatial block cross-validation (training and testing in different regions, e.g., biogeographical subregions or k-means clusters in geographic space), and/or

- Feature-space cross-validation (e.g., K-nearest-neighbour splits in predictor space), which helps assess model robustness across underrepresented combinations of climate, stand structure, and species.

2. Section 4.4 (“Why intraspecific variation…”) underexplores climate effects

The authors use mean annual climate variables, but intra-annual or interannual climate anomalies often show stronger relationships with wood formation. Given the dendrochronological nature of ring-width measurements used in the predictors, it would be interesting to explore:

- seasonal climatic anomalies,

- drought severity indices (e.g. SPEI),

- lagged climate effects, and/or

- extreme-year metrics.

Even if such variables are unavailable for the entire dataset, performing a sensitivity analysis on a subset (and reporting the result) would help clarify whether the weak climate signal is due to model formulation or intrinsic biological limits. This is especially important because the current discussion suggests ecological constraints, whereas part of the limitation may arise from the coarse temporal structure of the climate variables used.

3. Benchmarking against external biomass datasets (e.g. ESA-CCI biomass)

The manuscript demonstrates how choices in wood-density modelling affect biomass at different scales. However, it would be extremely valuable to compare the resulting biomass maps (from either the NFI- or GIS-based model) with ESA-CCI Biomass products. This is particularly relevant because the GIS-based model shows species-dependent biases (Fig. 5) and ESA-CCI biomass is known to have systematic regional biases, especially in high-biomass areas. A comparison of spatial residuals, biome-stratified biases, or scatterplots at NFI plot scale could directly link the wood-density-driven biases identified here with well-documented remote-sensing limitations.

4. GIS-based model: Could species-level information be used via the ForestPaths European tree genus map?

Recent remote-sensing products, such as the ForestPaths European tree genus map (https://zenodo.org/records/13341104), provide spatially explicit genus-level classifications across Europe. Given that the genus explains ~30% of total variance (Section 3.1.1), integrating this map into the GIS-based model could improve predictive performance and reduce the regression-to-the-mean bias documented in Fig. 5. I encourage the authors to test whether adding ForestPaths genus information improves GIS-based predictions.

Concluding remarks

This is an excellent, well-written manuscript. The dataset and modelling framework represent a significant contribution to European forest research. The key findings are robust:

- Species identity dominates wood-density variation.

- Intraspecific variation exists but is structurally difficult to predict.

- The effect of wood density on biomass is negligible at the national scale but substantial at the subregional and local scales.

- GIS-based predictors recover a significant portion of variation even without species identity.

- Addressing the points above, particularly the treatment of cross-validation, the more in-depth analysis of intraspecific drivers, and benchmarking against external datasets, will significantly enhance the robustness and impact of the study.