General Comments

The manuscript entitled “Assessing the effect of forest management on above-ground carbon stock by remote sensing” evaluates a multi-sensor approach for forest above-ground biomass/carbon prediction and its usability for assessing the difference in biomass carbon content between managed and unmanaged forests. Both of these issues are highly topical. The potential of multi-sensor approaches is still underutilized and will become increasingly important as the range of different types of data grows. The requirements for monitoring biomass in managed and unmanaged forests is increasing in importance due to the new European regulations aiming to increase close-to-nature silvicultural practices with a simultaneous increase in monitoring requirements.

The study site and the field reference dataset are unfortunately rather small and limited to one specific ecosystem. This limits the usability of the results of this study. Nevertheless, I think the findings of this study would be a valuable resource for other researchers interested in similar topics in other areas. The manuscript is very concise. Although I generally like short and concise manuscripts, there are several points that need to be elaborated more to ensure replicability of the analyses and to verify some of the conclusions presented in the manuscript. I have itemized the key points in the ‘specific comments’ below. In addition, I have some minor suggestions or comments (listed in the ‘technical comments’ section) for the authors to consider.

Specific Comments

L8: For the clarity of the text, it might be good to consider keeping the two main topics (i.e. multi-sensor biomass prediction and assessment of managed vs. unmanaged forest) in the same order as they have been presented here throughout the manuscript (at least in the Results and Discussion sections). In my mind this would be the logical order and it would make it easier to follow the manuscript. Particularly Figure 4 is now a bit confusing since the reader does not yet know at that point what data/model has been used for the final remote sensing prediction.

L12: Here and throughout the manuscript, I think it would be important to make it clearer that you did not use GEDI data as such, but a product that is based on a combination of GEDI data and Sentinel-2 imagery. This is significant because the Sentinel-2 data undoubtedly induces some saturation in the height predictions, potentially with varying effects in different areas. I would say already here something like “…derived from a combination of the Global Ecosystem Dynamics Investigation (GEDI) and Sentinel-2 data were used…”.

  

L59: This sentence is confusing. Spaceborne LiDAR (e.g. GEDI with 25-30 m footprint) does not provide information on higher [spatial?] resolution than passive optical sensors. Also, I would not say that spaceborne LiDAR provides detailed 3D profile of forest canopies, knowing the noisiness of the observations and other complications in deriving the height predictions for the footprints. The sentence would be true for airborne LiDAR, but I think it is an overstatement for spaceborne LiDAR and gives a wrong impression to those readers who are not familiar with GEDI data. I would rather say something like it is possible to derive predictions of the canopy profile for the 25 m footprints.

[There is some problem with the line numbers at pages 6-7.]

Regarding the Sentinel-2 data description, it would be important to elaborate further two points:

1) What is the cloud probability product/layer used in the masking? To my understanding it is not part of the regular S2 L2A products, right? This product/layer should be somehow described, with reference if relevant.

2) The derivation of band values for plots is unclear to me. Did you take all pixels that had min 90% overlap with the plot and take their average? Or did you calculate weighted average, weighted by the area of plot overlap?

L166: Regarding the GEDI/Sentinel-2 canopy height product, it would be very important to have some understanding on the accuracy of the product in the study area. Did you do any evaluation how well the product seems to work in the study area? Would FFI plots be available for this analysis? Or you could use your own plots. This is a crucial issue for the interpretation of the results. It is possible that there are significant saturation effects in the product in your study area due to the Sentinel-2 data.

L167: The Sentinel-1 data needs to be described in much more detail for replicability of the study. Where did you get it from? Which data products you started the processing from? What kind of preprocessing steps were or had been done (by you or others)? How was the temporal compositing done? Details of the preprocessing are very important for radar datasets as they strongly affect the end product characteristics. If the text length is  an issue, I would rather have the data preprocessing details here, rather than general description of C-band characteristics.

L194: Do I understand correctly that you used only eight plots to derive the error metrics presented in the results section (e.g. Table 4)? If so, this is a very small number of plots. Why did you not use e.g. 1/3 vs. 2/3 split (i.e. 33% for validation)? How where the eight plots selected? Do they include plots from all forest size classes? Did you test how sensitive the error metrics are to each of the plots (i.e. how the metrics change if you leave out one of the plots, one by one)?

L232: It would be nice to see a bit more information on the results of the variable selection, to understand better how the variable selection progressed and how clearly the five chosen ones finally provided the best results (e.g. compared to using only spectral bands without indices).

L241: I am very surprised that the addition of canopy height hardly improves the results. Usually, knowledge of canopy height is very beneficial for biomass estimation. This makes me wonder how accurate and useful the GEDI/S2 height predictions are in the study area. It would be important to validate the height product in the study area to better understand its effects and usability. On the other hand, the addition of S1 has a very clear and significant positive effect. Therefore, I would also like to see the S2+S1 results. They may be nearly on the same level as S2+GEDI/S2+S1. This affects also the discussion section (L285). It would be important to understand the role of the canopy product better so that it could also be clarified in the discussion.

Technical Corrections

L32: Suggestion, perhaps change the last sentence into something like: ”However, accurately capturing carbon stocks over large areas presents both technical and logistical challenges. In this context, remote sensing provides cost-efficient means for large scale monitoring of above-ground carbon in forests.”

L51: Suggestion: “This information can be extrapolated using…”

L53: Comment: This is true for wall-to-wall mapping, which is why sampling has been traditionally used in forest inventories. However, the increasing requirements of spatially explicit information call for new approaches.

L56: Suggestion: Instead of -range, I would prefer to use either -area or -scale.

L58: Comment: I tend to disagree with this to some extent. NDVI (and numerous other vegetation indices) also tell about the vigorousity of the vegetation, not necessarily about AGB. Also, there is great variation between tree species on the NDVI (for similar AGB).

L70: Comment: When talking about Spaceborne LiDAR, they also suffer from “mixed observations” as the footprint is rather large. 

L75: Suggestion: “…, because they do not provide wall-to-wall data.”

Figure 1: Suggestion: Change “3 plots” to “Three plots”.

L110: Comment: This sounds like a very short time period for me. Is this sufficient time to reach characteristics of unmanaged forest in Belgium?

L113: Suggestion: IMFP needs to be written out somewhere. I did not find it anywhere in the manuscript.

L119: Comment: Why were dead trees included? Depending on the case, this may badly confuse remote sensing based biomass/carbon mapping.

L132: Comment: I wonder if this is well established use of the term "two-entry tariffs". This is a totally new term for me. Please check.

[There is some problem with the line numbers at pages 6-7.]

P6.fifth last line: Comment: And 8A is also not used. Perhaps it is not even available in the GEE data collection?

L153: Suggestion: I would remove ‘calibration’, saying ‘Once the field measurements were obtained…’

L159: This sounds a bit simplistic in my mind. There are a lot of complications deriving forest canopy height profile from the GEDI data. I would rather formulate it somehow so that "prediction of canopy height profile for the 25 m footprint area can be derived from the GEDI observations".

L214: Suggestion: Perhaps “Model application” instead of “Model extrapolation”.

Table 3: Comment: Might be more informative to use the plot size or the tree sizes, rather than A, B, C in the caption and the table itself?

L237: Comment: Why so short text here and most of the text in the appendix? If not limited by manuscript length rules, I would not mind seeing more results on the modelling here in the main text.

L244: Comment: Again, perhaps application would be better than extrapolation?

L290: Suggestion: You could expand the discussion on the benefits of longer L-band wavelength a bit e.g. by citing some more recent ESA CCI Biomass related references by Santoro et al., where they use combination of C and L band data.

L291-L304: Comment: This is a good discussion paragraph. Indeed, the remote sensing predictions typically gravitate towards the average, causing underestimation of high volume forests and overestimation of low volume forests.

L301: Suggestion: Should be overestimation, not underestimation, right? 

L314: Suggestion: “as” should be replaced with “was”.

L332: Suggestion: Table A1, not Table 5.

Review on Assessing the effect of forest management on above-ground carbon stock

by remote sensing

General Comments

The study consists in building a model of forest aboveground biomass that combines data sources from optical, LiDAR, and Synthetic Aperture Radar (SAR) in order to estimate forest biomass and carbon stocks at a 10 x 10 m resolution over a small-sized study area (a national park in Flanders, Belgium). The forest biomass map produced was used to evaluate the effects of forest management on carbon stocks by comparing matched pairs of managed and unmanaged sites. The model used data from Sentinel-1, Sentinel-2, and a canopy height product (Lang et al., 2022, 2023) derived from the Global Ecosystem Dynamics Investigation (GEDI) mission as predictors in a generalized additive model (GAM).

Field assessments revealed higher carbon stocks in unmanaged stands compared to managed ones while the model-based estimations did not result in significant differences. The signal saturation and the need for additional training data are among the possible reasons for this discrepancy between direct but point-wise inventories and the predicted carbon stocks. The main reasons for this discrepancy are: i) the potential saturation of the remote-sensed products at high biomass stocks and ii) calibration issues of the GAM, iii) mean absolute errors larger than the difference between managed and unmanaged forest stocks.

The study brings two contributions: one is the use of remote-sensed data trained by field data in order to assess the forest aerial biomass over a given region, the second is to use this information in order to detect differences among forest management types.

There is a large investment of the community into using remote-sensed data to produce forest and/or volume biomass estimations, with a wealth of publications in this topic. The methods used here are not new, although the use of a GAM model is not as common as the use of a non-parametric model approach. Thus overall, the level of novelty is not very high on this aspect.

The use of the model to estimate the forest biomass over a given area to test for management effects is more novel. It has limitations but also has the merit of highlighting current challenges : i) the use of remote-sensed data to improve local estimations of forest biomass is not an easy task, ii) the gains are strongly limited by the amount of field data available and iii) the differences in forest biomass related to the management is sufficiently nuanced to be difficult to determine. 

By making additional efforts into quantifying the impact of the reduced field-data sample size on the model development, and associated problems of extrapolation, the study would be much more conclusive.

Specific comments

The reduced sample size (39 per forest category, managed/unmanaged), probably too small for the calibration of the model, may be evoked as an overarching limitation to the study. Fitting complex models certainly requires a sufficiently large observational basis, unless the population studied is very homogeneous. Here, the population is probably not really homogeneous, as the range of carbon stock values seems quite large. While this does not invalidate the study, it certainly adds uncertainty to the results.

The assessment of predictions uncertainties would have brought some light and helped the diagnosis. In particular, splitting extrapolation issues from saturation issues would have been very helpful. Extrapolation issues highlight potential deficits in field sampling. The use of a convex hull to identify the proportion of extrapolated predictions (see ex. https://doi.org/10.1016/j.jag.2022.102939) would be one efficient way to respond to this. The question currently remains open, that is, the study is not very conclusive in this regard. That is a major concern.

The way the estimations at the scale of the patches were made was not sufficiently well explained. It seems that the estimations of the biomass at patches level represent an average over the estimations made at the pixel level inside. They would therefore be purely model-based estimations. Besides the mean estimation error, the bias had to be estimated too. It seems to be quite substantial, as indicated by figure 5. The bias relates of course to the signal saturation. Signal saturation is indeed another known problem, for which the integration of multiple sensors was hoped to bring a solution here. The analyses on the model predictions do not allow to attribute these problems to a lack of data for fitting, or for a saturation issue per say. It therefore remains largely inconclusive.

Technical comments

- This affirmation is false: L85: “The combined use of LiDAR, SAR and 85 passive optical remote sensing has, to our knowledge, not yet been investigated to assess above-ground biomass in temperate forests.”

It has been used, albeit with different modelling approaches: https://essd.copernicus.org/articles/15/4927/2023/?utm_source=chatgpt.com

- In Figure 3 it would be good to complement the “100%” Field dataset box with the actual number of data available (N = 78).

- The description of the GAM is wrong: L187: “A Generalized Additive Model (GAM) was chosen as a non-parametric extension of GLMs (generalized linear models)”

A Generalized Additive Model (GAM) is a semi-parametric, rather than a fully non-parametric extension, of Generalized Linear Models (GLMs). GAM models assume an underlying parametric structure for the response distribution.