Characterizing Uncertainty in Forest Remote Sensing Studies

Round 1

Reviewer 1 Report

As a general comment, there are many repetitions of the same concept (inadequate common metrics, importance of accounting for errors in reference data) in the paper. The readability of the paper would gain from some well-aimed “cuts”.

389 maybe I would rephrase as “biomass per unit area” or something like…

The increase of RMSE is about half of the increase in sigma_delta; though the trend looks nonlinear, it doesn’t seem to be strongly so for a “fair” range of sigma_delta values. In other words, the

I think that the discussion section should more explicitly state that your model only address some of the errors in reference data. Some that you mention in the introduction, e.g. positioning error at stand level, may have systematic effects at stand level that are not captured from the model, as they differ from stand to stand.

In addition, it would be worth discussing whether your error model would be ok also with single tree methods.

Although I made only a few checks in literature, I find the RMSE of ALS larger than that of SAR in Krycklan a bit surprising.  At row 268-269 you mention that the q ratio for ALS might be 1 to 2, so… Please comment.

Table 2. The results in Table 2 to me look strange. ALS has no scale error, small noise and a large displacement (and so a large RMSE) while SAR data have significant scale error, much larger noise and a comparatively small displacement. I would expect ALS to be better, and of course resorting to RMSE doesn’t help. So which model best fits the data based on your error model estimated parameters? If you want to substitute for RMSE, an objective answer is needed. The end user still will want to know the error amount… Of course Figure 6 helps interpretation, however a clear-cut evaluation is needed. Please comment.

Figure 5: Irrespective of the above comment, in Figure 5 one stand value for SAR and five for ALS predicted negative biomass values. Please explain why and the physical meaning of this result, if any.  Is it because of a systematic underestimation of tree height in ALS data? Another negative value (-10.1) can be found in Appendix B, Table B.2 at row 27.

Figure 7. the corrected RMSE* with 10 ton/ha noise has a small however clear downward trend, i.e. improves with increasing noise in reference data. Why? please comment.

Just to be sure: the corrected RMSE* is computed as square root of (14)?

The discussion and conclusion section is a bit weak. It looks too much a summary of the paper, repeating too many details. Moreover, your model is identical to Tien et al. model (that you fairly reference in the paper), so I would rephrase expressions like “our model” as in fact you (rightly in my opinion) adopt a good proposal and push it further. Discussion from row 498 to end is ok, however, as you decided to focus the paper on forest parameter estimation, you should be more specific about the relative importance of errors in reference data (what sort of randomness is the most important? Positioning errors in stands? Gross errors?...  My comment on the user standpoint on overall error should be perhaps also be addressed in this section…

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The value of this study for characterizing uncertainty in forest remote sensing studies lies in two aspects: (1) to reduce the ambiguous metrics for the quality assessment by a proposed linear parametric error model to take the bias, scaling factor, and random error into account; (2) to extend the use of the parametric error model to cover cases where the field references used for evaluation contain random errors to avoid severely misjudged quality due to the inclusion of errors from the field reference data. The proposed models seem to help in achieving the above gains. Yet, several questions listed below need to be further clarified before the merit of this work can be fairly identified.

 

1. Equations (1),(6),and (7) do not agree with each other.

2. The formula used for Figure 4 should be specified.

3. The proposed linear model plus considering random errors also in field reference does slightly reduce the RMSE estimates for the two cases. Yet, what prevents the study from further improvement? Is it because of too simplified model to fit the reality? Or the quality assessment needs to be explained by extra metrics?  

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

In this paper the authors provide an interesting approach to capture error structures when ground-truth (GT) and remote sensing (RS) datasets are compared, and to enhance error characterization using a parametric model. Importantly, they emphasize the fact that enhanced RS estimates lead to the ratio of errors RS/GT being usually lower than 2, when errors in the ground data have stronger impact. They present also an empirical example to show an application of their methodology. The paper is well structured and very well written. I recommend its publication after three specific minor improvements:

(1) Please, specify how the references can be modelled using unobserved true values, and what do you mean with the term "unobserved". Also, how would you obtain objective inventories in real case applications? 

(2) Some text pieces within lines 475-497 are redundant with previous sections. Please improve this issue.

(3) Last sentence in the paper could be more conclusive. However, l. 520-521 are redundant and do not offer a final conclusion. I would focus more in the general (and powerful) idea you aim to highlight during the paper: that improved RS techniques make the ground errors more important in relative terms (when compared to the RS errors) and that your paper provides a new method to quantify and enhance this fact.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

The responses from the authors have clarified most of my given comments and questions. Two further suggestions are added for the authors' reference in revising the manuscript.

1.Since the linear error model had been proposed in previous work by other authors, it is strongly advised to use the word "adopt" instead of "propose" to soften the originality of the model.

2.Some suggested corrections are marked in the attached PDF file.

Comments for author File: Comments.pdf

Author Response

Thank you for your feedback. We have implemented all your suggestions for this round.