A Knowledge-Based Strategy for Interpretation of SWIR Hyperspectral Images of Rocks

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Review of “A Knowledge-Based Interpretation Strategy for SWIR Hyperspectral Images of Rocks”

This manuscript presents a workflow for extracting specific absorption features from shortwave infrared hyperspectral data using minimum wavelength mapping techniques, and deriving mineral maps (classifications) using an expert-based decision tree. This approach will be of interest to the readership of Remote Sensing and has many strengths, as explained nicely in the publication, however I suggest that:

1. a more detailed literature review is needed to differentiate the proposed workflow from previous approaches including Tetracorder and PRISM from the USGS, “The Spectral Geologist” from CSIRO and decision-tree based workflows presented by previous authors (e.g., Qasim et al (2022) and Thiele et al. (2021), to name a few; though I’m sure there are many more).
2. quantitative or qualitative validation of the mineral mapping results, ideally using e.g., XRD or automated mineralogy data. Without this validation it is not possible for the reader to assess if the proposed method is producing accurate and meaningful results.

My other more detailed comments are listed below:

Line 10: “invariant between scenes” — and sensors? Do the authors expect their decision tree to be transferable to other sensors, even of the same type and wavelength range, given different calibrations often cause (small?) shifts in wavelength position?

Line 24: What is the difference between “spatial patterns” and “textures”? Reword or be more specific.

Line 58: Agreed! Consider also emphasizing that SAM is dominated by large trends/features, rather than the “sharp” absorption features that are typically considered to be diagnostic.

Line 62 - 68: This should be expanded to several paragraphs reviewing similar approaches published in previous works (e.g., Qasim et al., 2022 and Thiele et al., 2021, to name a few), and established workflows used in e.g., Tetracorder and PRISM MICA (USGS) and TSG (CSIRO).

Line 68: A brief paragraph mentioning supervised machine learning approaches to mineral mapping is also needed here, as well as a (brief) discussion of their limitations (e.g., need for extensive, site-specific training data).

Figure 1: I struggled to understand this figure, and would suggest that it could be removed and replaced by a paragraph or in-text list. If retained, Steps 4 and 5 must be depicted in more meaningful detail. Also note missing arrows between 1,2 2,3 and 3,4.

Line 92: Consider moving this section (on hyperspectral data acquisition and preprocessing) to the start of the methods section.

Line 97: How might pixel sampling distance (“spatial resolution”) influence the results? Does the presented approach assume each pixel is made from a pure mineral?

Line 117: Were the subset spectral regions hull corrected before quantifying the absorption feature position and depth? Which method (e.g., quadratic, polynomial or gaussian fitting) was used to extract the absorption feature depth and position? How was this method adapted to fit multiple features? …given how important this step is to the presented workflow, it should be described in significantly more detail.

Line 129: Specify that these “mathematical operations” are band-ratios (with the exception of Shannon Entropy).

Line 132: Why was Shannon Entropy used instead of simpler measure of spectral variance (e.g., the variance or standard deviation of the bands).

Line 136: Add a paragraph to this section somewhere outlining the assumptions made / trade-offs when classifying pixels comprising mineral-mixtures. E.g., what would happen if a pixel contained kaolinite and illite? Or chlorite and muscovite? Or carbonate and illite? Interference between the various absorption features can make interpretation of minimum wavelength maps very challenging…

Figure 2. I would consider using the functional group notation (e.g., from Laukamp et al., 2021) rather than misleading “spectral mineralogy” names (e.g., pixels classified as illite using this scheme could in reality be any of a whole suite of different clay minerals?).

Line 201: A quantitative (or qualitative / visual) comparison of this validation data and the classification results should form a core part of this manuscript. Otherwise it is impossible for the reader to assess if the results are accurate or meaningful.

Line 376: How can “incorporation of expert knowledge” be both a strength and a weakness? I think I understand what is meant, but some clarification/justification/rewording is required.

Line 381: This section should be extended to discuss the previously mentioned challenges associated with mixed (multi-mineral) pixels, especially for coarser (e.g., mm) spatial resolution data.

Line 403: This is an entirely unsupported claim, given (1) the small size and limited geological variability of the test data set, and (2) the lack of presented validation data.

Kind regards,

Sam Thiele

 

References

Kokaly, R. F., King, T. V. V., & Hoefen, T. M. (2011). Mapping the distribution of materials in hyperspectral data using the USGS Material Identification and Characterization Algorithm (MICA). International Geoscience and Remote Sensing Symposium (IGARSS).

Thiele, S. T., Lorenz, S., Kirsch, M., Acosta, I. C. C., Tusa, L., Herrmann, E., ... & Gloaguen, R. (2021). Multi-scale, multi-sensor data integration for automated 3-D geological mapping. Ore Geology Reviews, 136, 104252.

Qasim, M., & Khan, S. D. (2022). Detection and relative quantification of Neodymium in Sillai Patti Carbonatite using decision tree classification of the Hyperspectral Data. Sensors, 22(19), 7537.

Comments on the Quality of English Language

While the writing was relatively concise and easy to follow, several areas could be significantly improved (e.g, I found paragraph 1 to be unnecessarily convoluted and repetitive). Also check the spelling of Illite on line 158.

Author Response

We thank Sam Thiele for the careful review that helped to improve the manuscript significantly. Please find our replies to the issues raised and the implemented changes in the attached document.

The quality of the English language in the manuscript was checked and improved by a native English speaker.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

A brief summary - The work offers an alternative approach for generation of mineral maps, not based in statistical assumptions but on expert knowledge.

General concept comments - The approach developed considers expert knowledge in several steps of processing for generation of mineral maps.

This methodology can overcome the drawback of the most common used methods of mineral mapping, but not all, using hyperspectral images are based in statistical values |thresholds, that cannot be reproduced in different images. Another advantage is the focus on the wavelength position of the absorption features.

Despite the validity of the proposed approach, it is worth considering that this method also introduces some subjectivity in many steps.

References - More recent references could be done, in parallel with the original developments of hyperspectral tools that are old. The number of self-citations of the authors are nearly half. Even if required, others should be added.

Specific comments -

Line 65-69 steps i) to iii) are subjective. That does not seem an advantage versus statistical values even if considering selection of thresholds.

Line 373 - The requirement of an expert has limitations and also should be highlighted be highlighted that in a different geological environment another expert will introduce more variability. What would be the impact of setting the system with a different expert knowledge for developing the processing chain? Its worth explore this theme.

Improvements for next research: the final maps results could be improved with more quantitative results for validation.

Author Response

We thank reviewer 2 for the careful review that helped to improve the manuscript significantly. Please find our replies to the issues raised and the implemented changes in the attached document.

The quality of the English language in the manuscript was checked and improved by a native English speaker.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

1. The classification thresholds in Table 2 (e.g., 0.25/0.38/0.50 for dt_albedo) rely on expert experience but lack a quantified empirical basis. Supplement this with threshold sensitivity analysis or supporting evidence from spectral libraries (e.g., USGS).
2. The mineral mapping results (Fig. 5d) lack quantitative accuracy metrics (e.g., confusion matrix, overall accuracy/Kappa coefficient). Validation against point-counting data from thin-section analysis is recommended to assess classification accuracy.
3. Processing time for the automated workflow (HypPy scripts) is unreported. An efficiency analysis using larger datasetsshould be supplemented.
4. Details in the scatterplots of Fig. 7 are difficult to discern. Provide high-resolution inset panels or supplementary high-resolution versions online.
5. The literature review appears limited. Expanding references to include recent advances in hyperspectral mineral mapping would enrich the scholarly context of this work.

Author Response

We thank reviewer 3 for the careful review that helped to improve the manuscript significantly. Please find our replies to the issues raised and the implemented changes in the attached document.

The quality of the English language in the manuscript was checked and improved by a native English speaker.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

It is interesting that the manuscript focuses on a Knowledge-Based Interpretation Strategy for SWIR Hyperspectral Images of Rocks. 

1.The arrows in Figure 2 indicate unclear meanings and errors. It is suggested that "T" or "F" be marked above each arrow.

2. In the L353, " in the" is repeated.

Comments on the Quality of English Language

Some of the syntax needs further improvement.

Author Response

We thank reviewer 4 for the careful review that helped to improve the manuscript significantly. Please find our replies to the issues raised and the implemented changes in the attached document.

The quality of the English language in the manuscript was checked and improved by a native English speaker.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Dear editor, authors,

Many of the comments from my previous review have been sufficiently addressed, however I still have several reservations regarding the lack of presented validation information/results. Hence, I suggest that the paper requires moderate revisions to address the following points before it can be accepted.

Line 44: While this now provides a (minimally) brief overview of the available mineral mapping methods, it should be extended to also mention emerging machine learning based approaches ("mineral upscaling"), for which many recent publications are available. 

Line 133: The minima extraction and interpolation procedure is now well explained, but please clarify how secondary minima (i.e. the second and third deepest absorption features) were identified: this cannot be achieved by just identifying the lowest hull-corrected reflectance bands (as implied), as these will be associated to the deepest absorption feature only. 

Line 140: Please clarify which mineral groups and wavelength range the "SiOH" absorption is referring to; this is much less discussed in the literature than the AlOH, FeOH and MgOH absorptions? (I could also find no mention of it in the cited publication [Laukamp et al.]).

Line 349: While this is a good place to start validating the resulting mineral maps, it is not sufficient for a reader that is not intimately familiar with the authors' sample set, preventing independent assessment of the presented approach and results. I suggest that it is necessary to include:

(1) at least one figure showing representative photomicrographs of the "ground-truth" petrography/thin-sections/microprobe analyses from several samples, alongside corresponding mineral maps derived from the hyperspectral workflow.

(2) a table showing the spectrally dominant "modal" mineralogy of each sample (as predicted by the hyperspectral decision tree), and corresponding brief thin-section description (as a qualitative validation of the correspondence between predicted and observed minerals).

Without these independent data it is impossible to independently assess the hyperspectral results.

I suggest that these figures/tables and accompanying text should be added as a new subsection to the results, titled e.g., "3.3. Comparison to petrographic observations".

Kind regards, 

Sam Thiele

Author Response

We thank reviewer 1 again for taking the time to review our manuscript. Our reply to the issues raised and implemented changes are in the attached document.

Author Response File: Author Response.pdf