Mine Pit Wall Geological Mapping Using UAV-Based RGB Imaging and Unsupervised Learning

Round 1

Reviewer 1 Report

The manuscript entitled “Mine Pit Wall Geological Mapping Using UAV-Based RGB Imaging and Unsupervised Learning” presents results from analysing RGB photos for clustering analysis of pit walls. The motivation for using RGB photos instead of hyper-spectral data is provided, and the approach and attempt to use more simple data is fully justified. The performance of the clustering algorithms is discussed and evaluated. The manuscript of well written and I only spotted as few places with mistyping.

 

I enclose an annotated version of the manuscript where I highlighted descriptions which were not fully clear to me. The authors should carefully looks at these in order to improve the manuscript.

 

My main concern with the methodology is the large dimension of input vectors in the classification. Each tile contains a number of pixels and they are treated individually despite the goal is to classify each tile as belong to only one class. The approach essentially imply that each input vector is the size in pixels of each tile multiplied by three. The dimension become huge and furthermore position in a tile starts to matter in the clustering. A much simpler approach would be to calculate and average RGB vector (dimension 3x256) for each tile and use. this as inout. If you want more descriptive input, you could add estimated standard deviations of the mean etc. Such an approach would result in much shorter computation time, eliminate spatial dependency in tiles, and you may even not need to use the auto-encoder. Averaging is typically providing robust data sets by eliminating scatter and noise. A similar averaging approach could also be applied to hyperspectral data and you would then have smaller dimensions compared to the one used in this manuscript.

 

The authors need to provide more information about computational time. This is important in order for this to be of practical use during mining operations where decisions is needed to be taken fast.

Comments for author File: Comments.pdf

Author Response

A point-by-point response file is attached.

Author Response File: Author Response.pdf

Reviewer 2 Report

The use of RGB cameras in the case of image analysis in a complex lithostratigraphic sequence, where the interpretation and understanding of microtracks in the process of mining is the main indicator of further predictions of mining works, cannot in any case be based only on the improvement of algorithms, it is simply the work of a mining geologist.

Unfortunately, in this paper I do not see any added value of using UAV methods for the interpretation of different "mining situations", so I am of the opinion that the work should be rejected.

Author Response

A point-by-point response file is attached.

Author Response File: Author Response.pdf

Reviewer 3 Report

The authors present the article entitled “Mine Pit Wall Geological Mapping Using UAV-Based RGB Imaging and Unsupervised Learning” (Reference number remotesensing-2235360).

This research aims to utilize aerial RGB images captured by unmanned aerial vehicles (UAVs) and unsupervised learning algorithms to identify various geological formations in limited sections of pit walls at two mining locations. The authors use several algorithms to explore their ability to create geological maps in the absence of accurate ground truth information.

The manuscript is well-written, with a clear and well-organized structure, and deals with an interesting topic related to the current use of remote sensing technology with unmanned aerial vehicles. It addresses a key issue in UAV remote sensing, as RGB cameras are cheaper than LiDAR or other techniques. Indeed, more research is needed in this regard, and this article achieves this by going deeper into this topic.

However, the main flaw of the article is the lack of ground truth information and GCPs. The authors pointed this out, but it should be discussed properly. The authors could have employed LiDAR as “ground truth” or validation, for example. In fact, the authors state in the introduction "Terrestrial-based remote sensing methods such as tripod-mounted LiDAR or hyperspectral (HS) sensors", but later they omit aerial methods, such as UAV LiDAR (e.g. Zenmuse L1), which is only slightly more expensive than the equipment they have employed. I understand it is difficult to gather LiDAR data and maybe authors do not have this kind of equipment, nevertheless, they should add information about LiDAR vs photogrammetrically-generated 3D point clouds, and create a proper discussion. Moreover, the authors could have employed an RTK UAV or they could have created a larger ROI (Region Of Interest) to accurately locate Ground Control Points (GCPs) and improve the precision of this work. Therefore, the authors should elaborate further on these points.

Consequently, I recommend a minor revision of the manuscript.

-              Introduction

Authors should include information and references about UAV LiDAR.

-              Material and methods

In “2.4. Photogrammetry”: Authors should include information about the photogrammetric workflow. i.e. information about the options selected for the steps “Align photos”, “Build Dense Cloud”, “Build Mesh”, “Build Texture”, “Build DEM”…

The figures are legible, and the quality is good. Except Figures 6 and 7, which have the same caption: “An illustration of the cluster map generation process using K-Means clustering only.”. I suppose the Figure 7 caption is wrong.

Authors should consider adding the manufacturer's information after each instrument and the developer's software.

-              Results

The results are good.

Line 563 indicates that “The red clay is, however, not identified as its own group. Despite the finer resolution and noisier appearance, segmentation does yield some useful structural information that was not captured prominently by the classification approaches.”. It could be improved by adding an explanation, like the authors did in Line 459 “This is likely due to the presence of…”

-              Discussion

The authors should elaborate further on the absence of ground truth information and GCPs and how this lack of information can impact the model. The authors pointed this out, but it should be discussed properly. Moreover, they should add at least a paragraph about the impact using of LiDAR vs photogrammetrically-generated 3D point clouds.

-              Conclusions

The conclusions are fine.

-              References

Some references are incomplete. Examples in line 744 (“37. Pix4D Pix4Dmapper V4.1. User Manual.”) and line 753 “42. Kingma, D.P.; Ba, J. Adam: A Method for Stochastic Optimization. 2014.”. Please, check the format of the references.

 

Specific comments:

Line 38

“Terrestrial-based remote sensing methods such as tripod-mounted LiDAR or hyperspectral (HS) sensors”

And what about aerial LiDAR? Or aerial HS cameras?

Line 124

“Kinross Gold’s Bald Mountain”

Please add coordinates and crs.

Line 147

“DJI Inspire 2”; “Zenmuse X5S”…

Please add the manufacturer's information after each instrument.

Line 165

Expand the caption: explain what the points are and their coordinates and crs.

Line 221

“QGIS”

Add the developer’s information. The same for line 228 (DJI GS Pro) and line 253 (Metashape).

Line 368

Figures 6 and 7 have the same caption. I understand that figure 7 caption is wrong.

Line 601

“but GCPs are difficult to place on pit walls due to accessibility issues”

 

However, they could be placed on the perimeter of the Region Of Interest.

Author Response

A point-by-point response file is attached.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Acceptable as is. I an satisfied with the response to reviews comments