Re-Estimating GEDI Ground Elevation Using Deep Learning: Impacts on Canopy Height and Aboveground Biomass

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This article is based on applying a known deep learning method to estimate the ground elevation from the GEDI waveform. The authors further on evaluate how their improved detection of the ground elevation influences the canopy height and the AGB estimations. A particular attention is given to the effect of transfer learning.

I found the presented work relevant and important, but would have to read the feedback from the authors on the following points before recommending the paper for publication:

·         Title: You should probably somehow indicate in the title that you actually re-estimate the ground elevation, and check for the impacts of the novel estimation on the canopy height and the aboveground biomass.

·         Line 35: I would say that forests absorb and sequester atmospheric CO2, by transforming it into the biomass (and not only the above-ground one!).

·         You should increase the legend font size in Fig. 2 and 3. and explain slightly better the items.

·         Line 180: Can you provide few more details concerning the waveform simulations.

·         In my experience we often use ResNet18 in the 2D image classification, object detection, face recognition, but I’ve never come across a study using it for the “time series” analysis. You should maybe add a brief explanation of how this was done in [46], and how did you adapt the convolution filters to tackle the waveform at the input? Did you do in the same manner as [46] in dealing with the ECG heartbeat signal, knowing that they did for the classification purpose? More details is needed on this, I reckon …

·         Line 221: So, the ResNet18 was originally parametrized on Area 1? You state it later on, but this should be noted here as well.

·         I like Fig. 6 and the accompanying analysis! The result demonstrating the weaker increase of the RMSE with the increasing slope with respect to the GEDI L2A product is particularly interesting, and should probably be more accentuated, knowing the problems GEDI experience across the slope (due to the footprint size).

·         The same (positive) comment applies on Fig. 9 and the accompanying analysis, even though the tendencies are weaker as commented in the discussion section.

·         On the other side, I’m not really sure I grasp the pertinence of Fig. 10 …

·         Line 398: You should probably rephrase – “than” -> “with respect to”?

·         In Fig. 12 why don’t we observe the same effect as in Fig. 6 and partly in Fig. 9?

·         Also, how did you estimate the AGB with the ALS data? Another option was to redo this analysis (and potentially the canopy height one) by “translating” the GEDI profile to the inventory plot, as suggested by: N. Besic, S. Durrieu, A. Schleich and C. Vega, "Using Structural Class Pairing to Address the Spatial Mismatch Between GEDI Measurements and NFI Plots," in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 17, pp. 12854-12867, 2024, doi: 10.1109/JSTARS.2024.3425431. and  P. M. May, R. O. Dubayah, J. M. Bruening and G. C. Gaines, "Connecting spaceborne lidar with NFI networks: A method for improved estimation of forest structure and biomass", Int. J. Appl. Earth Observ. Geoinf., vol. 129, 2024. It could be worth commenting this in the discussion or at least in the perspectives section, cause the inventory plot can provide more reliably all the information you derive here from the ALS.

·         Again, I’m not really sure I grasp the pertinence of Fig. 13 …

·         Line 556: I agree with this statement!

Author Response

Thank you for your comment. Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

I believe that this paper has a great value for scientists and analysts. More studies every year rely on GEDI footprints to use it as training data for AGB and carbon wall-to-wall models based on satellite data. However, errors in estimating ground elevation (and subsequent errors in deriving relative height metrics) are often not addressed. Authors of this research attempted to cover this issue and I am certain it should be published.

However, some minor changes to text/figures must be made to clarify/simplify the representation of this study before publication.

My specific comments:

Figure 1: indeed, often the lowest rh data for some footprints (e.g., rh0) have negative values of a few meters. Authors are encouraged to provide their own explanation why here elev_lowestmode differs from rh0 so much and rh0 has so high negative value (-10.08 m). In general, figures are not used in Introduction section. Authors may think to make a conceptual part in the beginning of Methods section and visually explain their study design with Figure 1.

Lines 71-72: this expression can be cited. Authors may simply mention that rh98 is important predictor in the model and the quality of GEDI metrics directly affect the accuracy of model

Line 139: as an individual who personally visited Shizuoka and forests nearby, I can understand how other study sites (within Japan) may differ from this in terms of species composition and local biogeography. If reader can simply remember that Area 4 represents Amazon's dataset, it is neccessary to not just refer to other study sites as 'Area 1, Area 2, Area 3' whenever authors mention it, but to recall (for a reader) what it means. I would propose to use A1, A2, A3 and then occassionally clarify in the text where it is mentioned. E.g., they can use 'in boreal study site (A3) there was...' instead of 'in Area 3 there was...'. I use 'boreal' here as a simplified example, authors may clarify Shizuoka, Fukuoka and Hokkaido sites according to what they know about these regions (obviously better than me).

Figure 5: plots with too many multiple facets produce too small text of accuracy estimates.

Word 'precision' sometimes is used instead of 'accuracy'. I would track all these cases and correct it accordingly to which metric was used.

Author Response

Thank you for your comment. Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

Authors in this article identified an approach to improve ground elevation RH of the GEDI waveform and recalculate the canopy height metric and AGB incorporating ALS data through deep and transfer learning approaches. This study presented a significant topic of concern on the improvement of height metrics which led to the AGB estimation.  

The following is some detailed comments:

1- Suggest to re-write and re-arrange the introduction in a way that from a Global perspective to the specific study area.  

2- Page 3, line 83, Figure 1, Suggest to label the Amp. and Percent Energy Returned maps with (a) and (b) respectively, and include these labels with a description in the caption.    

3- Figure 2 on page 5, line 165, the legend abbreviation shown in the picture (a) needs to be described in the caption (a).

4- Suggest to re-design or re-arrange the study area map (Figure 2) as well as other maps along with legend labels so it will be neat, tidy, and readable.

5-  Page 9, line 268, section 2.7.3, suggest to include more about the AGB estimation method in detail and how it would be estimated using allometric equations or other method e.g. on lines 275 and 276, stated that recalculated RH value from a model and transfer learning were used to estimate AGB. 

Author Response

Thank you for your comment. Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Thank you for considering my remarks. As far as I'm concerned, the article is good for publishing.