Machine Learning Methods for Quantifying Uncertainty in Prospectivity Mapping of Magmatic-Hydrothermal Gold Deposits: A Case Study from Juruena Mineral Province, Northern Mato Grosso, Brazil

Round 1

Reviewer 1 Report

This paper addresses a methodological approach using machine learning for mineral prospectivity mapping. It is well structured and written, using an appropriate research design and sound references.

Author Response

Dear reviewer,

We thank you for all your comments,

Best regards.

Reviewer 2 Report

This article address a difficult problem, how to generate a dataset consisting of so-called negatives in the absence of actual negative data. The particulars of the problem, involving ore deposits, are less important than the soundness of the methods. This is a class of statistical problem is called a presence-only distribution. The paper follows the usual approach of attempting to find some predictor variables, then testing those to determine if the predictions lead to true positives on an independent dataset. I'm sorry to be a skeptic, but there is no reason to think that a complex machine-learning algorithm with black-box characteristics is going to be the most appropriate way to handle this class of problem. Nevertheless, this was an interesting article to read although perhaps not the most fundamentally sound approach to the problem. Overall, I think the results support the conclusions and that there is so much focus on machine-learning methods that it would be interesting to readers; however, I think that the method should be itself more critically viewed. In other words, is it worthwhile applying machine-learning to this problem at all?

Such philosophical considerations aside, the study is structured in such a way as to add more constraints, in this case simply assumptions because the negative model is just absence of evidence of a deposit, and the finding is that more constraints lead to a more certain model. This is pretty much guaranteed by the way that the algorithms (RF, SVM, KNC) determine the outcomes. A much more interesting approach would be to compare potential maps created by only buffering and by only spatial association. Perhaps other ways to generate simple datasets could be considered.

Some more specific edits as follows:

line 18 and 132: shuttle radar topography mission

line 33: Consequently, [this] lead to lack of [reproduceability]...

line 43: an "uncertain class" might be misleading, the class is actually high variation, or maybe an undefined class

line 82: covert the capital I to (i)

line 84-85: setting a minimum distance to a large value doesn't ensure the selection of true negative samples. in fact, the authors cannot really test this assertion against true negative data due to the nature of the problem being one of undefined data.

Section 3, refer to figure 2 and summarily describe all of the datasets at the start of this section. This would help me as a reader to see how the data are being used.

Line 168: this choice doesn't look arbitrary, it looks like the sill value on Figure 3. Another possible choice is approximately 5 meters, where there's another natural break.

Figure 5: What does the shading show in lower panels of Figure 5. It looks like there are supposed to be two shades of gray, but this figure is confusing as presented. Perhaps the lithologic units would be better depicted with a pattern instead of shading?

Line 272: is the increase in performance actually an increase in accuracy of positives?

Figure 9: check that axes labels are consistent with the figure caption

 

 

Author Response

Dear reviewer, please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Dear authors,

 

Congratulations on such a piece of work. I think the future of ore deposits prospection goes through matching learning development. I still see the location rates low. They are all below 25% of accuracy on location of prospects.

Is it possible to improve the results by including geochemical data or lithological affinity? Please clarify this point because it is vital. More detailed analysis may improve the results by including geochemical and lithological information in the modelling.

Author Response

Dear reviewer, please see the attachment.

Author Response File: Author Response.pdf