Review: A free, open-source method for automated mapping of quantitative mineralogy from energy-dispersive X-ray spectroscopy scans of rock thin sections

General Comments

The manuscript describes the analysis of 14 tonalite thin sections by SEM-EDS mapping and demonstrates an approach of mineral identification by the random forest classifier. The results are analyzed for the modal mineralogy and some geometric properties of the samples. The focus is on the approach, which was realized through open-source software only, giving access to any user of such data without the need of programming skills or expensive software. The need for and advantages of free software are an important topic within the community. High prices and restrictions of proprietary software are often a problem. Therefore, developing and providing the access to such kind of analysis is an important contribution to the scientific community and the readers of the journal.

In general, the results of the manuscript are convincing and the conclusions sound. The findings are well presented. The discussion might need some additions (see specific comments below). The structure, however, seemed a bit problematic, as there is some redundancy or repetition in the general approach and the used method. The reader has to move a lot in the manuscript when a topic is split into four different chapters. The reader would benefit from a condensation of the description.

I would suggest making a shorter version of the general approach (current chapter 2) with no discussion or recommendations (l. 134f for the procedure), but just the plain and neutral description of the method. Then in chapter 3 explain in more detail why certain parameters are important and were used in this study. Then the discussion should mention further explanations and limitations of the methods. Otherwise, there is in parts no real separation of results and discussion/interpretation.

As a personal preference: The style of the manuscript is in parts similar to a user manual. For a scientific paper, I would suggest to aim at a more neutral style of description in the results part.

In general, this work is useful research and should be published after the comments have been addressed.

Specific comments:

L. 28: Do you consider hyperparameter a generally known expression? If not, it would be better to use a more general word in the abstract.

Ll. 39f: I would suggest quoting the original papers regarding quantitative mineralogy, e.g.: 

•    Sutherland, D. N. and Gottlieb, P.: Application of automated quantitative mineralogy in mineral processing, Miner. Eng., 4, 753–762, 1991.

•    Sutherland, D., P. Gottlieb, R. Jackson, G. Wilkie und P. Stewart (1988). "Measurement in section of particles of known composition." Minerals Engineering 1(4): 317-326.

•    Gu, Y.: Automated scanning electron microscope based mineral liberation analysis An introduction to JKMRC/FEI mineral liberation analyser, Journal of Minerals and Materials Characterization and Engineering, 2, 33, 2003

L. 52f: I am not aware of a WDS system that can map mineral phases. I would be interested in more information.

L. 53f: I am not sure about LA-ICP-MS, here. From the quotation, you might mean Laser Induced Breakdown Spectroscopy (LIBS).

L. 77: “predicted mineralogy”. You cannot predict mineralogy. What you mean is the modal mineralogy or mineral abundance.

Ll. 88ff: The last paragraph of the introduction is normally a short summary of the hypothesis or goal of the paper and how it might be reached. You start with the goal in l. 88, but then continue to explain classification algorithms. Maybe you could move the algorithm part up and leave the goals to the last paragraph.

L. 96. “random sampling”: Do you mean a random sample?

L. 110. “In the remainder of this study”: Better: Furthermore, …

L. 113. “hope”: You could write that you intend to make it available to a broader community, which is good, but hope is the wrong word in this context. An ideal scientist would be neutral. The same applies to l.614.

L. 118: You should clarify that you mean EDS data from a SEM. There are other methods producing EDS spectra, some of which you list in the introduction, so until you mention it, the reader assumes EDS data from any source.

L. 127: How do you deal with overlapping peaks (e.g. S/As/Pb or Ti/Ba etc.)? What influence would overlaps have on the classification? Would it be possible to have several lines (e.g. K-alpha, K-beta, L-alpha) for one element?

Ll. 132ff: This raises several questions: If you need to know the minerals already, how would you do that? This would mean a lot of work done twice (manual identification + RF classification). What happens, if you miss one or more minerals? Can you analyze unknown samples?

L. 138 “electron beam”: It could also be X-rays or laser from what you mentioned in the introduction. Consider also the comment for l. 118.

L. 139: How would the smoothing help, when a peak was identified incorrectly? Please clarify.

L. 158: I believe 30% training data would be standard. How did you select the training areas? Randomly?

L. 165 expression “mineral phase”: phase is a defined technical term. The algorithm assigns a mineral name to a pixel, since EDS contain have only chemical information, which cannot always distinguish phases (polymorphs, limitations in element analysis such as light elements). 

Ll. 164ff: What about minerals that have not been trained? Are all pixels classified? Can the similarity between training pixel and the classified pixel be described?

Ll. 173f: How did you find that 10 pixels are best? I assume this would vary, based on the sample type, grain size and resolution of the image. Can you offer an explanation, why pixel sizes influence accuracy? I would assume that, if you have less spatial blur (mean filter), the results would be more accurate. Could it be a user bias, as users prefer large grains for training areas? 

L. 175: You should explain voting ties and how they are treated in the introduction.

L. 221: How can a beam step size and magnification result in a map? Please rephrase. Also, is beam step size the same as pixel size?

Chapter 3.2: If you have a mean raster with 20’000’000 pixels and measure 200 ms per pixel, how can you calculate an acquisition time of 3.5 h? Please explain.

L. 239: Why this recommendation? What if apatite is important to a study? Why would you combine minerals? One would assume that this affects accuracy negatively. Do you have any information on how many training areas are necessary and how large they should be? You should have a discussion on the selection of training areas in the discussion section.

L. 246f: Move the explanation of virtual rasters to the overview of the method or the introduction.

Figure 2: Could you plot the data from XRD and point counting as well for comparison? Then you could shorten the paragraph above.

L. 337: Why did you choose these three thin sections?

Ll. 363f “This indicates that the models correctly predicted apatite when attempted but the models often neglected to predict apatite.”: Why would the models not try to predict apatite, if they are trained for apatite? Please explain.

L. 356 and Ll. 391f: You use questions to structure your manuscript. I would suggest avoiding that and use a more descriptive style.

Table 3: You could add a line with the standard parameters for comparison. If I understood correctly, you used the standard values for what you showed in figure 2?

Ll. 535-543: How do you evaluate the effort to label training data compared to what you describe? In how far can the training data be extrapolated to unknown samples or other studies? As the training requires most of the work and time, what are the limits of this approach? – Some of the questions are answered further in the manuscript, which raises the question, if the text could be restructured.

Ll.545-558: How can you make sure that you actually have identified only one grain of e.g. biotite? With EDX, you get the chemical distribution, but if you have monomineralic aggregates or touching biotite grains, how can you separate those grains in order to calculate correct grain areas/ sizes? – After reading chapter 5.3, I realized that you addressed part of the comments.  The manuscript would benefit in structure and clarity, if you could combine the paragraphs or address the problems/limitations in the discussion parts directly where you mention the analysis.

General Discussion: Do you have information on the time invested in creating and testing the models? Can you put in relation how much time the mineralogical analysis (e.g. XRD) costs compared to your approach? 

Technical corrections:

L. 108 “requires no programming on the part of the user”: This sounds a bit odd. Do you mean no programming by the user? 

L. 177: Can a map be interrogated?

L. 219: “studied” instead of “study”?

L. 500: comma is unnecessary.

The authors should address the issue of the real added value of their work with respect to the pertinent literature. 

It is not clear the assumption at the basis of considering the results used for the training on a region as useful for a test area.