Do AI Models Improve Taper Estimation? A Comparative Approach for Teak

Round 1

Reviewer 1 Report

This manuscript concerns the comparison of different modeling approaches for predicting stem taper for Tectona grandis in Mexico. It is a potentially interesting study, which, however, requires major revision before it might be accepted for publication.

My comments are:

1) The manuscript needs English editing.

2) The structure of the manuscript could be improved. As it stands now the manuscript is dominated by general (and difficult to follow) descriptions of the different AI methods and regression models applied. I would recommend that they are (slightly) shortened and carefully checked (further details are provided below). The Results and Discussion section partly contains methods, and it fails to focus on a core table or figure that compares the different approaches (which is now presented at the end of the Chapter, whereas it logically would be placed first.

3) It is unclear how the data were divided between estimation and validation data? Logically the validation data would be fully independent trees (not diameters at heights within trees for trees that were used also for the estimation). In case the validation was not based on fully independent trees the study has limited value.

4) Linked to the previous comment, the total number of trees included in the study was about 300. This appears to be a very small number for successful application of AI techniques, and this should be discussed in the article. (You could expect model-based approaches to be perform better than the AI techniques when the number of observations is small.)

5) The evaluations involve many different criteria and it is unclear how the ranking of the different approaches was made based on the output. Still the authors make conclusions about what methods performed well and which did not. But it is unclear how this ranking was made.

6) Lines 36-37: Clarify that this concerns diameters at different heights, i.e. taper.

7) Lines 39-40: It is unclear what the authors imply by "geometrically" here.

8) Lines 46-47: Also, large datasets for machine learning are often available, which, however, does not appear to be the case in this study.

9) Lines 49-52: But this requires large datasets

10) Lines 81-83: Are the abbreviations related to climate needed? It is unclear what information they provide.

11) Section 2.2: Clarify how the distinction between training and validation data was made. Was it based on trees or individual observations within trees?

12) Section 2.3: Very detailed descriptions of the AI methods are given, but these descriptions still are slightly difficult to follow. I would recommend that these sections are slightly shortened and that the authors carefully revise the descriptions, especially with regard to technical details, which sometimes are not really helpful for the reader, especially since they are sometimes confusing. Some examples are:

- Table 2: Describes several features that have not been described in the text

- Lines 147-152: X seems to be used to denote both numbers (or vectors) and sets?

- Equation 2: An expected value is a theoretical concept which cannot be applied in estimators. 

- Equation 3: The covariance argument contains only one term (f(x')) whereas it concerns how to random variables are probabilistically related.

- Lines 159-160: Give references?

- Equations 5-7: Difficult to follow (for me at least). Does this follow from Bayes' theorem?

- Line 190: What does this abbreviation imply?

- Line 196: Strange description of D and Xi

- Equation 8: This estimator comprises an expected value, which appears as very strange

- Lines 219-220: What is means by Mi model, xi sample, and Xi sample

- Lines 229-230: I guess a more correct description is that it mimics the way it is believed that the brain works. 

- Equation 9: Why summation from 0?

- Line 240: Change from "income of the" to "input"

- Equations 9-11: Seems to mix u, x and y? Are the three equations consistent with each other?

- Lines 271-272: What input data were used for the other AI approaches?

- Table 3: uses the same parameter symbols for both models, but mentions only the regression coefficients in one of the models.

- Line 284: What is FIML

- Line 285: With maximum likelihood the likelihood of the observed data is maximised, but maybe FIML is not based on maximum likelihood?

- Equation 14: It is good that the authors address the autocorrelation structure of the data. But were there other dependencies in the data due to the fact that a large number of observations were taken from the same tree? Would a mixed modelling approach be needed?

- Lines 297-298: If the Durbin-Watson test statistic was computed it should be presented.

13) Section 2.5: In what way was a "composite" ranking of the different methods obtained from the individual criteria?

14) Equation 15: What was the number of variables (K) in the different AI methods? Is AIC really a relevant measure for the different AI methods?

15) Equation 18: The term "mean bias error" is new to me, since the terms bias has a strict definition. Is eq. 18 something else than an estimate of the bias?

16) Equation 18: Logically I think that you should not compute "mean bias error" but the bias at different heights. As it stands now, a negative bias at low heights appears to be compensated by a positive bias at high heights. This can be observed in the (informative!) figures in Figure 4, where there seems to be a trend for many methods that the bias is negative at low heights and positive at high heights.

17) I would suggest restructuring of Results and Discussion. Move some pieces in the Methods. Focus on the comparison of the methods (Table 5 and Figure 4) and put the discussion at the end of the Chapter, once the results have been presented.

I think that the authors can revise the manuscript so that it becomes good enough for publication. However, a thorough revision is needed!

 

 

 

 

 

Author Response

The responses to the reviewers' comments can be found in the file "Response Reviewer Comments reviewer.docx".

Author Response File: Author Response.pdf

Reviewer 2 Report

This manuscript aims at comparing methods to estimate Teak taper  using four AI models and two well established non-linear regression models, concluding that the ANN was comparable to the variable exponent Kozak model and better than the compatible taper-volume equation from Fang.

It is the view of this reviewer that the authors fail to provide convincing evidence as to why the ANN should be selected as a valid method when the variable exponent method is available.  The ANN carries not only a much larger number of parameters (no parsimony there), but also, no clear indication is given how it could be used in a practical way.  

Taper equations are used not only to calculate diameters at different heights, but also to calculate commercial volumes.  Fang's equation has the virtue to sacrifice taper fit to improve volume fits (something that the Kozak model fails to do when integrated and compared to observed volumes).   The authors should address that as part of their discussion, as well as the fact that parsimony is sacrificed at the mercy of more complexity.

Finally, an example should be give how to estimate volumes using the ANN, and indicate how bad they are as compared to the compatible Fang method, resulting in substantial bias, negatively affecting the price transaction (the ultimate goal of a taper equation).

 

1. More complex method to find parameters

and 

2. Less accurate method to evaluate volumes.

 

Also please see the following comments:

Line 1 :  Correctly estimating stem diameter (remove the).

Line 21: remove “The” and just leave “Artificial intelligence methods”.

Line 49, change “the AI models” to just “AI models”.  

Line 50  least-squares fitted expressions is just one form of fitting taper models, there are also maximum likelihood and Bayesian methods.  I suggest changing “traditional least-squares” to “traditional regression models”. 

Line 50 Multicollinearity and auto-correlation are never irrelevant for AI tools, this is a false statement.   There are ways to account for those, like L1 or L2 penalizations (that are regression methods!).  Sakici’s paper did propagate the notion that these issues are accounted automatically by ANN, but they are not. They can obscure any analysis.  

 

Line 54 Random Forest suffers from pour identifiability when variable reduction methods are not utilized.  Of course many variables can used, but the risk of over parametrization is present there (under the assumption of pour correlation due to the random nature of the algorithm).

 

Line 66 change “have been shown” to “have shown”.

 

Line 70 comparison is probably not a good objective, rather model improvement in accuracy and bias sound more like a challenging objective, one that is more related to the end goal of improvement.

  

All throughout the text, use of pronouns should be reviewed.

 

 

Author Response

The responses to the reviewers' comments can be found in the file "Response Reviewer Comments reviewer.docx".

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

I am ok with publication of this manuscript in its final form