Review Reports - Predicting Mineralogy with Hyperspectral Data: A Benchmark Dataset and Machine Learning Framework to Enable Hyperspectral Geometallurgy

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors This is a well-structured and technically rigorous manuscript that addresses a significant challenge in applied hyperspectral geology: scaling detailed mineralogical information from thin sections to entire drill cores. The authors present a comprehensive benchmark dataset, a novel open-source Python package (hklearn), and a systematic comparison of multiple machine learning approaches. The work is timely, relevant, and likely to have high impact in geometallurgy and mineral exploration. Below are my detailed comments, organized by major themes and specific sections, intended to strengthen the manuscript prior to publication.

The term "upscaling" is used throughout, but the core task is fundamentally supervised regression (predicting mineral abundances from spectra) followed by spatial extrapolation. The distinction between (a) learning a spectral–mineralogy relationship from sparse samples and (b) applying it to unsampled areas is sometimes blurred.
The authors correctly note that rare and accessory minerals are poorly predicted. However, the paper would benefit from a more quantitative discussion of how far (in spectral or geological space) predictions can be trusted. For example, what is the distribution of spectral angles between training pixels and the most extrapolated test pixels?
Figure 5: Does "tuned" mean initializing weights from the general model and then continuing training on local data? If so, what learning rate and number of epochs were used?
The acknowledgment that MLA has its own uncertainties (e.g., illite vs muscovite, epidote vs garnet) is important and honest. However, this point is raised only late in the discussion. Consider moving part of this caveat earlier (e.g., at the end of Section 2.4 or in the results) and quantifying its potential impact where possible.
The use of SMAPE for MLP/RMLP and inverse-weighted loss for the CNN is sensible. However, the paper does not report whether any threshold-moving or post-hoc calibration was applied to the predicted abundances.
Figure 2: The caption is very dense. Consider moving the detailed algorithmic references (NFINDR, SLIC) to the main text and simplifying the caption to the conceptual flow. The figure itself is clear, but the caption currently tries to do too much.

Author Response

We would like to thank the reviewer for their constructive comments and rigorous review. These are each addressed in the attached PDF, along with a detailed description of the corresponding changes to the manuscript.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Dear Authors,

Thank you for the opportunity to review your manuscript. While your work on the benchmarking dataset is valuable, a few clarifications and improvements would strengthen its clarity and impact. The title and use of "upscaling" could be more precise, as the manuscript primarily enables algorithm testing and application to larger datasets rather than directly advancing hyperspectral geometallurgy. Consistency in terminology (e.g., "thin-section," "thick section," "sample block") is needed. The abstract and introduction should also clearly define "upscaling" and the scope of your contributions. Your literature review would benefit from including foundational works in spectral geology and mineral classification (e.g., Swayze, Clark, Cudahy) to provide broader context.

The sample descriptions should be standardized and detailed early on, including dimensions, preparation methods, preprcoessing , and spatial resolution for all sensors. This is essential for reproducibility and understanding the data’s limitations. The data workflow, calibration, and preprocessing steps (e.g., masking, co-registration) should be thoroughly described in the metadata or alongside the published dataset. The sampling strategy and rationale for varying sample counts per drillhole should also be clarified. If superpixels are used, explain how this affects the signal-to-noise ratio and robustness of predictions. For mineral classification, provide more detail on how classes were defined and validated, particularly for spectrally similar phases (e.g., white mica vs. illite). All abbreviations and links should be explained in captions or supplementary material.

Lastly, the discussion on operational applicability could acknowledge practical challenges, such as dust or illumination variability, to avoid overpromising.

With these refinements, your manuscript will better align with its goals and enhance its clarity and reproducibility.

Detailed comments:

Title: From the content of the manuscript I do not see how there is a direct link or discussion to hyperspectral geometallurgy. What is being enabled by this benchmarking dataset is the testing of different algorithms and the application of said models to larger drillhole data sets. I don’t see how this directly enables hyperspectral geometallurgy. While aspirational, the current title is slightly misleading. This is also true for the use of “upscaling”. ML models are trained on a smaller dataset (thick sections with labels) and applied to a larger dataset. I am not 100% sure how that defines as upscaling and not just extrapolation?

Authors: there is a type in “Richard”s name

General: While many of the publications that are mentioned are relevant, it’s hard to ignore that some of the publications for mineral classification, unmixing, different applications etc. are dominated with citations from the first publishing working group. Relevant publications establishing fields of spectral geology are missing. If the authors are not familiar with them, I recommend getting familiar with them to enable citing more diverse sources.

Abstract:

“upscaling” used in abstract, acknowledging that the term can be understood differently. But this is then not described or clarified further later on.

“thin-section scale mineral maps” – slightly misleading, since you did not use thin section scanning as a technique, but used 2.5x4cm flat subsamples from the core for SEM-based training. If “thin section” and “HSI” is searched for and this publication appeared, the person who searched for thin section HSI would possibly be surprised by the actual content. I mention this further below but a consistent description of the samples would be benefitial (instead of thick section, thin-section blocks, block samples, …). Thin section scale could also apply that the spatial resolution of the mapping is similar to microscopy resolutions and scale. The actual pixel GSD is not described in the manuscript (is it 1x1mm?). Claim in l. 38 – 41 is not substantiated in the document. Other than a forward looking comment that unforeseeable insights may aid applications such as geometallurgy. Mineral mapping and labelling for geomet purposes was not discussed. E.g. accessory minerals such as talc would have a huge effect but with the current model, accessories are not predicted with high certainty.

Introduction

General comment: I believe there are other relevant articles published in the past that go beyond the publication cited from the co-authors for the applications that are listed in ll. 47-48

e.g. Swayze, Clark, Cudahy, Meyer, Rogge, Kruse, Hunt, Murphy, …

l.64 this wavelength range does not account for all carbonates, could be more precise

l.90 “range of locations” – is the benchmark dataset that is being presented purely from drill core or mixtures of sample type and platforms to collect data? Would be helpful to clarify.

l.96, the term “upscaling” especially in the context of extending predictions to larger areas can be ambiguous, my first impression was the spatial upscaling, but it seems like you refer to extrapolating models/predictions to larger datasets? Is this correct? Rephrasing may help with clarity.

l.113 I am missing some key publications that deployed classification approached for mineralogy, e.g. the Tetracorder as a dominance classifier comes to mind.

l.150/151: a brief mention of the samples in the benchmark dataset earlier on would useful to understand what you mean by “thin-section block” here.

Thin section block is often referred to as “billet”, if that’s what is talked of here.

Methods

Figure 1: Thin section blocks are now referred to as “sample blocks”. A scale would help understand that these are likely a 2cmx1cm of size and the core box is 100cmx40cm or similar.

Table 1: I believe the actual length (in meters) for the samples drillhole would be more meaningful, as drillhole depth is variable. Sample count refers to samples taken for thin section creation? Location list could be more consistent, listing countries of origin for all of them.

l.175 abbreviation MLA not explained

l.175 now referred to as “MLA thick section”. Ideally, I would find a consistent label here.

For the sample locations here, seeing that you want to upscale mineralogy, not lithology, it could be beneficial to list the spectrally active minerals expected in each sample location instead of deposit specific lithologies. This is done e.g. in 2.1.4 but not the others.

l.243-257: I think keeping the description of the instruments used here to a minimum with reference to another publication is good, however, for the benchmarking dataset, I hope the data workflow and how the data was calibrated etc. is described in detail for people to clearly understand what data they are working with and how it was manipulated, including manual adjustments for masking etc.

While keeping it short, understanding the difference in spatial resolution (if any) between the data collected on the thick section vs. the drill core for the different sensors would be very helpful. What’s the pixel ground sampling distance for the different sensors being used, including the SEM data that is used for labelling?

l.269 “thin-section sized sample block”

l.271 the superpixel approach seems interesting, however, I’m wondering, does that change the SNR that is being trained on vs. the SNR of individual pixels that are being predicted later on? Does that change the robustness of the prediction?

l.277 what are class-center spectra?

l.277 on “optimally located sam- ple blocks were identified using a sliding window approximately the same size as a thin- section block to find locations”. Could benefit from rephrasing for clarity. SAM as a matching algorithm is now introduced for the first time, but it is not mentioned what endmember spectra are being used for the SAM analysis to find matching pixels.

l.287: what makes the sampling at the other sites sporadic? Where those thick sections and samples that already existed or were all samples taken for this study and prepared into thick sections. If so, why not apply the same approach to all drillholes and locations?

Figure 2: Is the PCA calculated per box or for the whole drillhole?SLIC method abbreviation not explained. Same for NFINDR. The workflow shown in figure 2 is not fully explained in the text under 2.3

Unclear how the thick section location is found. Is that for an area with the most diverse endmember spectra in the wanted window size per box or per drill core. How did you decide how many samples to take per drill hole, seeing it varies quite a bit between the locations. Did you find an optimal number of samples per km of core/ lithology/ ..?

l.301 do you see transmission through the 3mm thick section and reflectance from the underlying superglue and glass plate? Was a 3mm thick section and glueing onto sample holder necessary for SEM or could a thicker sample of the same size be scanned?

l.311 SEM and MLA abbreviation being introduced too late

what spatial resolution was the SEM data acquired in?

l.321, in the introduction it is mentioned that XRD and XRF bulk rock analysis is not available. How could it be checked against here?

l.327 spatial resolution of RGB data was not established, neither is the SEM resolution

section 2.5 MLA coregistration: are there any insights in the accuracy of the co-registration? Seeing that this is crucial to allow for precise, correct labelling of the HSI data (at different spatial resolutions?) it’s important to touch on the detection of key points across the image if it allows for precise co-registration or if it’s e.g. 1px precise t the center but 2.5 px towards the edges? I’ve seen automated key point detection and co-registration see a relative distortion of one image to match other, I am assuming this is avoided here or monitored?

Was the data resample spatially or how are you matching data with a factor 6 pixel size difference?

If there are mixed minerals in the SEM how does that translate to the HSI?

l.348-349: could benefit from mentioning this as continuum removal and establishing if division or subtraction was used, just to allow the reader to follow the data processing instead of guessing.

l.351 what do you mean by “contain more mineral species than can feasibly be upscaled from hyperspectral data”. Is that 50 mineral species or 300? What does the “feasibility” linked to? Processing power of hardware? Interpretability of data? ..

mentioning the accessory phases, are those predominately not spectrally active? Or not relevant to describe the deposit/ lithology? The <30 pixel threshold per image refers to VNIR, SWIR, MWIR, LWIR and individual boxes or the whole drill hole, all drillholes together (you mention “We filtered the dataset”)?

l.367-372: Slightly unclear how was the HSI data combined with the MLA mineral abdundance maps? How are spectra concatenated? The ML workflow is not taking spatial correlation into account and trains only on the spectra extracted from pixel areas labelled using MLA?

l.372: link not working for me.

l.395-398: I now realize that it’s unclear to me if you are analysing and training for individual sensors or for co-registered, stacked data from all sensors. The same goes for what data is published in the benchmark data set? Individual datasets? Have you considered that applying the same savitzky golay filter will affect the different sensors differently, seeing that all sensors have different band numbers and spectral sampling? Maybe significantly so?

Table 2: since white mica is a group of minerals spectrally similar but does not necessarily match the understanding of “white mica” in a mineralogical sense it would be good to include what you refer by it here. E.g. Illite shows up as an individual group in the table but is often counted among “white mica minerals” spectrally.

Figure 3: accuracies for lower abundance minerals / smaller angle between phases(ankerite to FeOx) are very hard to read. Is it possible to split this visualization up between dominant abundance (calcite, dolomite,muscovite, plagioclase) and the others for clarity? Same for Figure 4.

TinyURL links of the upscaling results are working, however labels are not clear. MWL is not described in the publication, the link for each images top caption leads to 404 page. Unclear what upscaling model is shown of the three different tuned RMLP models. Unclear what AB stands for? You mention that XRD is missing for mineralogy validation. What is the expected accuracy for the output from the SEM MLA? Does it need to be independently verified as well?

Figure 6: F2350 and F2200 referring to the wavelength position of the feature mapping. Not explained in the caption. MWL not explained. Often considered for both position and depth of the feature. Did you threshold what pixels are shown here depending on a minimum depth? or are all features considered valuable, regardless of SNR, noise, shadowing, etc.? Not the focus of this paper but worth a mentioned. “Smoothing” etc was applied but no intensity based thresholding or fall shadow removal between core pieces because the yolo segmentation is accurately removing those pixels?

e.g. for Fig. 7 MWL F2250 and F2200 looks pretty noise for the 4-6^th sample from the right. Interesting to see the quartz show up with a strong F2200nm feature (piece of core with the X in it). Mixture? Or a low depth absorption?

l.660 EDS?

l.731 while I agree with the overall optimistic forward looking aspect of applying mineral prediction models to operational conditions (geometallurgical or mineral processing parameters), I assume you don’t mean upscaling from e.g. thick section trained models to a dusty conveyor where samples are presented at different surfaces (unlikely to find somewhat flat surfaces as for drill core), dusty, different grain sizes and particle sizes, different illumination conditions etc.?

Author Response

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The authors present a study that is both relevant to the scope of the journal and timely, considering the growing interest in mineral exploration through increasingly automated approaches based on machine learning and user-trained datasets. The manuscript represents a meaningful contribution and a step toward the development of more refined methods for this type of preliminary mineral mapping. There are some aspects of the manuscript that require revision by the authors, as outlined below:

The iThenticate similarity report indicates an overall similarity of 9% between the authors’ manuscript and other texts available online. This value is considered acceptable, particularly given that no single source accounts for more than 2% similarity.

The authors should avoid repeating in the keywords expressions that have already been used in the manuscript title, as this creates unnecessary redundancy in scientific indexing databases. Such repetition reduces the opportunity to include a broader range of terms and, consequently, to enhance the visibility of the article within the scientific community. Accordingly, the keywords “hyperspectral,” “mineralogy,” “machine learning,” “benchmark,” and “geometallurgy” are inappropriate.

The authors’ citations do not conform to the formatting standards of the publisher MDPI and should be revised accordingly. MDPI employs a sequential numerical citation style using brackets (e.g., “[1], [2], [3], [...]”), which must be consistently followed throughout the manuscript.

It is not clear how the authors establish the correlation between mineralogical data derived from BSE images obtained via SEM and the hyperspectral data in the near-infrared range. This aspect requires further elaboration, as, in its current form, the manuscript does not demonstrate a sufficiently robust or well-defined correlation between these datasets.

The authors should avoid creating subsections, such as “1.1 Approaches to hyperspectral upscaling,” unless there are at least two subdivisions at the same hierarchical level (e.g., “1.2”, “1.3”, etc.). As it stands, this subdivision is inappropriate. This also implies that the subdivisions within the already subdivided levels must be reorganized accordingly.

I understand the authors’ intention to perform an upscaling of data from SEM to multispectral datasets; however, the choice of this type of microscopy raises some concerns. It is not a directly compositional technique, despite allowing for indirect correlations. This introduces an additional level of complexity in the data integration that was not clearly explained in the manuscript. The rationale behind this methodological choice should be better justified and clarified to ensure that the proposed correlations are properly understood.

I would like the authors to further clarify how they modified the methodology of Gairola et al. (2023). Although they state that an adaptation of the method was performed, they do not explicitly describe what those modifications entail.

There is no need to present the reference “https://doi.org/10.14278/rodare.4582” as a hyperlink; it would be more appropriate to include it in a standard citation format consistent with the manuscript’s referencing style.

It would be pertinent for the authors to further elaborate on the MLA technique, particularly regarding how it was effectively employed as a method for analysis and as a basis for machine learning training in the studies presented.

Figure 8 does not exhibit the same level of quality as the other figures and should be graphically reconsidered to improve its presentation.

Author Response

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

Please find the attached.

Comments for author File: Comments.pdf

Author Response

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The previous concerns have been well addressed, so the paper can be accepted for publication.

Author Response

Comments 1: The previous concerns have been well addressed, so the paper can be accepted for publication.

Response 1: We would like to thank the reviewer for their constructive comments during the review process, and are pleased we have been able to address them appropriately.

Reviewer 3 Report

Comments and Suggestions for Authors

The authors have shown a constructive and professional approach to the revision process. They responded individually to all comments raised during the previous review round and provided detailed explanations supporting the technical decisions made in the manuscript. Their responses demonstrate a willingness to engage with reviewer feedback and to justify the methodological framework and interpretations presented in the study. Such efforts have contributed positively to the overall quality and transparency of the manuscript.

While the authors have made a commendable effort to revise the manuscript and respond to the reviewers' comments, some concerns have not yet been fully resolved and warrant further consideration. The main outstanding issues are outlined below:

The citation format adopted in the manuscript is not consistent with the standard style used by MDPI journals. References should be cited using sequential numbers in square brackets, rather than by author names in the text. Likewise, the reference list should be organized according to the order of citation in the manuscript rather than alphabetically. The authors should carefully verify the journal's formatting requirements and revise both the in-text citations and the reference list to comply with MDPI editorial standards. Consultation of recently published articles in the journal is recommended as a guide for the required formatting.

The authors should also revisit the hierarchical organization of sections and subsections throughout the manuscript. Several instances of isolated subdivisions are present, where a section is split into only a single subsection. This is structurally incorrect, as subdivisions should only be introduced when at least two subordinate sections exist at the same level. This is a matter of proper document organization rather than editorial preference. The authors are therefore requested to review the manuscript structure and eliminate any unnecessary subdivisions that are not accompanied by at least one parallel subsection.

Apart from the points discussed above, the remaining issues are predominantly editorial and formatting-related. These matters may be addressed during the subsequent stages of the editorial process under the guidance of the handling editor and the journal's editorial staff.

Author Response

Comments 1: The citation format adopted in the manuscript is not consistent with the standard style used by MDPI journals. References should be cited using sequential numbers in square brackets, rather than by author names in the text. Likewise, the reference list should be organized according to the order of citation in the manuscript rather than alphabetically. The authors should carefully verify the journal's formatting requirements and revise both the in-text citations and the reference list to comply with MDPI editorial standards. Consultation of recently published articles in the journal is recommended as a guide for the required formatting.

Response 1: We have updated the citations to match the consecutive numbering format of MDPI.

Comment 2: The authors should also revisit the hierarchical organization of sections and subsections throughout the manuscript. Several instances of isolated subdivisions are present, where a section is split into only a single subsection. This is structurally incorrect, as subdivisions should only be introduced when at least two subordinate sections exist at the same level. This is a matter of proper document organization rather than editorial preference. The authors are therefore requested to review the manuscript structure and eliminate any unnecessary subdivisions that are not accompanied by at least one parallel subsection.

Response 2: We have checked the organisation of the sections and subsections and removed isolated subdivisions as suggested.