Bridging Human Expertise with Machine Learning and GIS for Mine Type Prediction and Classification

Round 1

Reviewer 1 Report

 The manuscript combines expert knowledge, machine learning algorithms and GIS technology to conduct research on demining area identification and classification, which is a good topic and of outstanding practical value. The design of the research is sound, and the results are credible. The paper is well presented and has high quality graphics. The algorithm designed in the paper obtains very high prediction accuracy in the test area and has the potential to be generalized. 

However, there are still some issues with the manuscript that are worth discussing.

(1) The algorithm proposed in the paper has been validated on only one unseen dataset. Do the results have good reliability? I think the paper should clearly state how well the model generalizes to other regions or different environmental conditions.

(2) Although the paper mentions that the model's practicality and effectiveness were affirmed by experts, it does not detail whether the model has been field-tested and validated. Moreover, it does not delve into the ethical and safety issues that the model might bring, such as the risks of false positives and false negatives.

(3) The scarcity of landmine data may limit the depth and breadth of model training. I think this issue should be addressed in the discussion section.

(4) The paper does not discuss in detail the interpretability of the model, which is an important aspect in machine learning models, i.e., how to interpret the results of the model's predictions.

Author Response

Cf. responses in the attached file.

Author Response File: Author Response.docx

Reviewer 2 Report

Dear authors,

here are some remarks to your paper:

-          Misleading title, I come from the mining industry (georesources) and expected something completely different

-          Abstract, same as the title. Your abstract is not specific enough. Demining is also sometimes used as mine abandonment/closure

-          Line 97: Obviously is you paper of essential importance to your work, therefore I would elaborate further in the work (more detail)

-          L110: What is the military experience, only of demining experts, but also the soldiers who did the mining. The last one I miss in the text as an input. I would expect that soldiers in different countries run different mining operations. Therefore, also the fact, it is not clear from your work whether your approach is valid for the given area of Lebanon, or could be deploy elsewhere.

-          L122: What are the military based variables, be more specific

-          L137: performing impact, what is it?

-          L151: What is you background model? Satellite data? Drone data?

-          L151: Elaborate more on the input data, maybe introduce a table with the parameters

-          L153: Study area not explained

-          L167: This is still the introduction but you use already your results. Please restructure!

-          L202. The run interviews: What where the questions, and how where the answers incorporated in your model? More explanation needed

-          L202: I see here also the problem that soldiers who run mining operations are not included

-          L224. The different classification in anti-personnel mines and ant-vehicle mine is not really explained

-          Fig 2: reference is in the text is missing (or the figure is at the wrong spot), ppr quality, cut north arrow, no scale, no reference to the satellite imagery, more precise legends needed – Please check all the figures on this and update them. Also, the explanation of the figures in the text is far too short! More text explanation is needed.

-          Fig 3: Fig 2: reference is in the text is missing (or the figure is at the wrong spot), where is the red spot in figure 2?

-          Fig 5: Brown color (minefields) when the satellite imagery is brownish, is completely useless

-          Fig 6: Legend says the predicted model gives polygon squares but the figure shows points, adjust

-          Fig 6: Shows predicted mines outside the actual mine fields. This is a big problem (false positves) and need explanation in the text

-          L371: this is introduction, so restructure. Are the given points all? Are they in hierarchical order? Please explain

-          Fig 7: predicted mines outside mine fields, please explain.

-          Fig 8: I do not see the three blue colors

-          Fig 9. The difficulty prediction need explanation in the introduction section of this paper.

-          Fig 11: What is the left, cut frame? Bracket?

-          Table 2: Layout need update

-          Experiments section but no results section, restructuring is needed here. I also miss the on-site validation of the mathematical results. In my understanding predicting a mine filed in a “stable” conflict zone is much easier, than predicting a mine field in a post-conflict area or in an area with changing front-lines. Nothing is said on this

-          Conclusions: Restructuring and more precision on the results.

With kind regards

Author Response

Cf. responses in the attached file.

Author Response File: Author Response.docx

Reviewer 3 Report

Dear authors,

Starting from the title and introduction of your paper you declare that with a GIS analysis, and ML will be able to provide extremely important information such as Mine Type Prediction and Classification. However, as I went through the work, I saw a superficial approach to something so serious and important for those whose destinies are transformed by the presence of objects such as mines. While I respect the effort put into your work, it is concerning that you pretend to support Mine Type Prediction and Classification with a GIS analysis that doesn’t contain data that really can give you reliable information on the presence of buried mines.

For example: Lines 148-150 “We shall also highlight new extensions and innovations introduced by us in the current  approach and the proposed model. These categories include critical evaluation, mine-type  classification, and the introduction of a new feature”.

Military expertise is indeed very important, but with the lack of data from systems that can scan for mines, prediction, and classification cannot be done. How is your GIS system able to discriminate between any type of buried object and a buried mine?

Just to make the point, I am detailing some parts of your paper:

For instance,  at lines 155- 157 you explain a quite simplistic background model. The features you mentioned, such as slope, elevation, and distances to different areas, may not be enough to accurately predict mined areas without specific data related to mine detection. The high accuracy metrics mentioned in your study raise questions about the effectiveness of your approach, especially without concrete data on mine presence.

So, by integrating the above-mentioned features, in Line 166 it is noted “The accuracy metrics were respectively: 95.5%, 97.5%, and  94.1%. The RF algorithm, which yielded the best results in predicting mined areas,…”  I'm sorry, but I don't understand the logic of how from the attributes used (slope, elevation, distances…..) you can predict where mined areas are, without actually having measured data specifically to detect mines.   Your algorithm actually “predicts” any type of object…. because you don’t have anything to differentiate the objects.

In paragraph 161 (“related to Difficulty of clearance, only slope and elevation are applied”), you focus on slope and elevation for assessing the clearance difficulty, neglecting other crucial factors. Why just these 2 factors? These oversights highlight the limitations of your methodology in addressing the complexities of mine detection and classification adequately. Actually, the major factor is related to accessibility in the work site due to human factors and the safety of the demining team until reaching that zone. Another important factor is vegetation (both for mine detection as well for clearance actions, as sometimes robotic demining systems cannot reach or work in areas with vegetation). And if you have an area with vegetation other elements (for example meteorological constraints) begin to be also important.

Considering all the shortcomings in your study's methodology and its potential implications for human safety, I recommend a more comprehensive and meticulous approach to ensure the credibility and ethical considerations necessary for any research claiming to propose methods for mine detection and classification.

Overall, I encourage you to revisit and strengthen the methodology of your study to meet rigorous standards and avoid potential risks to human lives associated with inadequate mine detection solutions. Your work has the potential to make a meaningful impact, but it requires a more thorough and ethically sound approach to contribute effectively to this critical field.  

Author Response

Cf. responses in the attached file.

Author Response File: Author Response.docx