Mapping Homogeneous Response Areas for Forest Fuel Management Using Geospatial Data, K-Means, and Random Forest Classification

Round 1

Reviewer 1 Report (Previous Reviewer 1)

The revised paper has covered my last concerns. I have no other questions.

Author Response

Thank you very much!

Reviewer 2 Report (New Reviewer)

Although the manuscript is mostly well-written, a few sections including the abstract, methods, and discussion could be improved.

Abstract:

The abstract is well-written, and a theoretical background is provided. The research gap is presented and so is the consequent research question.

However, applying accuracy assessment techniques to determine/conclude the robustness or performance of the approach is not enough. A high-level summary of results from the other statistical approaches should be reported after some interpretation. [i.e. Report a summary of results from the Kruskal Wallis test and the Wilcoxon-Manny Whitney test as well].

Introduction:

The introduction is well-presented. The knowledge gap is well-motivated and supported by the literature.  However, more support or additional motivation should be provided in supporting a shift into consideration of an ecological context when characterizing forest fuel. 

Furthermore, there is a need to provide additional motivation for why the advocating of HRAs for this purpose especially when in lines 90-96 authors provide a comprehensive account of the shortfalls of this HRAs approach.

A good motivation for Machine learning is provided. 

 

Material and methods:

The study area section should be moved upwards in the method section (line 114) before the current account and figure 1.

Some information on the Kruskal Wallis test and the Wilcoxon-Manny Whitney test must be included in this section.

Results:

Improve the linkages between the presentation of the results from the Kruskal Wallis test and the Wilcoxon-Manny Whitney test to the study objectives.

Provide details on why the median was used in line 285 (or in the method section).

Author Response

Please see the attachment.

Author Response File: Author Response.docx

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

Forest fuels are one of core component for forest fire management, thus accurate description of forest fuels is necessary for fire management strategies. The present study provided a hybrid method that combine the benefits of unsupervised cluster and supervised classification techniques to estimate Homogeneous Response Areas. The methods were applied to three different regions, and yield higher overall accuracy. The methods and results may be interesting to the wide readers of Forests. Some suggestions and comments were provided to authors for improving the manuscript further.

General

       The amount and distribution of forest fuels is a quite complex problem in forest management practice. The authors only selected the variables of precipitation, temperature, altitude, canopy height and canopy cover to quantify their amount, however some important information were not considered, e.g., the inflammability of different species, site (slope, aspect, position; influence the growth and fuels); in addition, the paper provided some results of HARs, while the correctness and rationality were not clear. Thus, some filed measured fuels datasets were highly recommended to evaluate it.

 

Specific

1, in “2.1 Study area”, the information of altitudinal should be added;

2, L140-144, except for altitudinal, other terrain information e.g., slope, aspect, position are also important for forest growth, as well as for forest fuels, which should also be considered in the analysis;

3, in “Canopy characteristics”, except for canopy height and canopy cover, the species composition are also important for forest fuels, since some species are easy to be fired, while other are difficult. Thus, information on species or forest types were highly recommended.

4, figure 3, the results of ANOVA test should be labeled on the figure;

5, some filed measured fuels datasets were highly recommended to verify the results;

Author Response

"Please see the attachment."

Author Response File: Author Response.docx

Reviewer 2 Report

The authors present Homogeneous Response Area prediction as a method for informing efficient fire fuel sampling. The problem and their solution are well defined in the abstract and introduction. This method would be beneficial for other related field-based studies such as forest inventory and ecological monitoring.

For all of their efforts developing this work, they do not appear to have engaged deeply with fire science side of the work. For example, there is little discussion of the fire regime in their study area, and almost no discussion about fire and fuel dynamics in their study area. This left me wondering if the selection of variables were suitable for characterising the types of fuels influencing fire activity in the study area. Consequently, they have produced maps of areas that I have no doubt are homogeneous in terms of their climate, altitude and canopy characteristics, but it is not at all clear if these areas have homogenous fuel. I think there are three solutions to improve this paper:

1. Clarify the links between the input data and fuels that are relevant to the fire regiems in each study area.

2. If that process indicates that the present variables are not entirely suitable, run the analysis with suitable variables.

3. If either of the above are not possible, then the paper should be re-presented as a general demonstration of methods that can be used for stratifying sampling in landscape surveys.

Finally, a little more detail about decision making in the analysis is needed for transparency (see comments below).

Overall this has potential to be a quality work, it just requires a bit of thought about the application.

 

Specific comments:

Line 46-47 Different regions of the world describe fuels differently. If those are the fuel types used in your study area, make that explicit, as other regions use different fuel classifications. For example, in Australia, we often refer to surface, near-surface, elevated and bark fuels.

Line 116 What is meant by superficial fires?

Line 106-118 Is fire an issue or a risk in this area? It sounds like it is pretty manageable and perhaps not the best case study. Further, if the purpose is to understand fuel characteristics across the landscape, it would be helpful to know a little about fuel and fire dynamics here, and in the introduction. For example, is fire fuel or moisture limited? Which fuel layers are of most concern? And given information about fuel characteristics, what would land managers do?

Line 119-128 You talk about the geology/soils in these study areas but not in the first study area. Try to be consistent in the characterisation. Also, I think it would help if you explicitly stated what the contrast was between these two additional study areas and why that is beneficial for the study. And, what are the fire regimes in these two additional areas?

Line 132 (Data description) I assume that fire history, and possibly other disturbances) will be the main cause of temporal variation in fuel. So why have you not included fire history in you data? If that data were not available that’s fine, but you should acknowledge that as a considerable omission. Also, the three variables you have chosen are going to be very strongly correlated. Did you considered using the DEM to produce any aspect/slope/relief variables that might explain environmental variation at a finer scale than the climate data? And what about vegetation type mapping, is that available/useful?

As I understand it, you are choosing variables that are expected to be associated with variation in fuel. So you really need to explain what effect each variables is expected to have on fuel in your study area.

Line 147 Here you explain that canopy characteristics influence crown fire, but in the study area section you state that the study area only has low severity (i.e. not burning the canopy) fires. So why include variables associated with canopy fire in the analysis?

Line 177-179 What were the outcomes of these preliminary checks? I would imagine there is correlation between the climate variables and elevation and canopy cover. If there was, when was it considered too strong?

Line 186-191 Again, you have told us about the methods you used, but not about how you interpreted them to make decisions.

Line 233 How is altitude a more meaningful variable than temperature? Temperature has direct effects on metabolic processes and moisture availability. Altitude is just a proxy for various environmental variables as you state in the following sentence. I think altitude is justified given it is a proxy and has finer resolution, but just be careful with your wording.

Line 229-236 This information should be in the methods, not the results.

Line 277-279 What evidence do you have that canopy cover influenced understorey and ground material? I don’t recall any use of sub-canopy data in your analysis. Unless I am wrong, this interpretation needs to be in the methods with a reference, or a clear indication that is speculation.

Line 336-337 But you didn’t assess in fuel conditions. I assume the ‘fuel conditions’ you refer to here are the canopy characteristics. But in the methods section you state that fires are only of low severity, so the canopy fuel is irrelevant as it is not being burnt. Unless there are high severity fires too?

Line 388-390 You have not provided sufficient evidence to convince me that the HRAs you produced will represent homogenous fuel conditions. I think the absence of disturbance history as in input is significant, and you haven’t really explained how the input variables relate to fuel characteristics that are relevant to fire activity in your study area. Without that information, the reader can’t be confident that what you produced has anything to do with fuel.

Author Response

"Please see the attachment."

Author Response File: Author Response.docx

Reviewer 3 Report

This paper presents a method, which allows eliminating subjectivity in estimating HRAs spatial distribution, using Artificial Intelligence Machine Learning techniques. However, some problems should be addressed:

1)  the title is not appropriate. In this paper, the authors only used some of too many machine learning techniques.

2) The authors claimed “Once the samples were classified, they were used as training samples to develop a model to classify the locations into HRAs using Machine Learning techniques”. However, the authors only determine the number of classes by using K-means clustering. Then how to determine the class of each sample?

3) In machine learning, there are many algorithms for classification, such as representative SVMs. The authors should compare more classifiers to show the effectiveness of machine learning algorithms on the application.

Author Response

"Please see the attachment."

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The authors have responsed my concerns positively, thus I have no other questions.

Reviewer 3 Report

I think that the authors' responses are not satisfactory, e.g. the response to the last problem. As claimed in title, the authors focus more on machine learning approaches. For example, the authors identified each cluster using Random Forests algorithm. Then, it is guaranteed to be good when using Random Forests? If you do not compare them, there is no theoretical evidence for supporting your viewpoint.

In the response for the second problem, the authors claimed” Each of these groups represents a class, which were subsequently identified across the study areas using the Random Forests algorithm. We could label each group based on their characteristics; however, our goal was not to label them, but to classified them based on fuel characteristics measured in the field.”

“Your goal is not to label the samples?”

 Then, you do not label each of the samples, how to determine the class of each sample Random Forests algorithm? The authors claimed in the response to the third problem that “once this is achieved, practically any machine learning supervised classification algorithm could be applied.” Do you know that supervised learning needs class labels?