Towards Automated Target Picking in Scalar Magnetic Unexploded Ordnance Surveys: An Unsupervised Machine Learning Approach for Defining Inversion Priors
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe topic of this paper is interesting, proposing an automated target picking method for UXO detection. The content of the paper is complete. so it is recommended to accept it after revision. Some of my comments are as follows.
1） How long does clustering take? Because the author has a process of selecting the number of clusters.
2） There are overlapping areas between different ellipses. How to deal with abnormal fields in overlapping areas？
3） How to explain why the number of subegions increases after adding Gaussian noise?
4） In discussion part, authors motioned that “Our method is robust to multiple targets being close together?” How to determine whether a subegion is a single target or multiple targets？
Author Response
Thank you very much for taking the time to review our manuscript, and for your insightful and thought provoking questions! We've attached the PDF of the revised manuscript, which now has the following additions in response to your questions.
1. We added the following discussion of run time to the end of Section 2, Materials and Methods:
"Our code in github logs the time taken for each significant part of the pipeline. In particular, the clustering algorithm is O(n^2) in the number of input points, and the run time naturally increases with larger surveys, complex geology, and more targets. In order to manage this computation time when many points are to be clustered, the algorithm divides the survey into smaller subproblems for the clustering and then stitches them back together. Recorded times for the computation of the linkage matrices for our surveys ranged from 17 to 113 seconds. End to end, the entire process took seconds for smaller surveys (around 5000 m^2), and up to several minutes for more complex cases."
2. We added the following to Section 3, Results:
"In the figures below, every ellipse represents a proposed subregion with an unknown number of targets. The number and location of targets for the entire area of the ellipse will be determined downstream in the inversion step independently of any other regions. In the case of overlapping ellipses, the overlapping region will be considered multiple times in different inversion computations. In the event of one inversion resolving a target in the overlapping area but not the other, we recommend taking the more conservative result. If both inversions place a target in the same position this can be resolved as a single target."
3. We added the following to Section 4, Discussion:
"As seen in Figure 3, when we add Gaussian noise, more inversions are proposed, and the area of proposed regions tends to increase, with most of these new proposals covering irrelevant areas. This is an interesting phenomenon which deserves some extra explanation. Whether we add noise or not, when we winsorize and normalize the data, we produce an approximate normal distribution. Then the expected number of "points of interest" can be written in terms of Φ, the cumulative distribution function of a standard normal, and should remain unchanged. However, the added noise does affect the distribution of these points of interest throughout the survey, as they are now spread further apart, rather than being concentrated around, e.g., UXO or geographic features. When we perform our clustering step, we then find more numerous, and less targeted clusters, and thus more and larger subregions proposed for inversion. However, all targets signatures within the magnetic data set are still captured."
4. We added the following to Section 4, Discussion:
"Because we use a conservative, globally-determined threshold, our method is robust to the possibility of multiple closely clustered targets. That is, even if multiple targets produce a locally homogeneous subregion of high readings, we still correctly identify this subregion as an area of interest to be subjected to additional scrutiny through inversion. Each subregion is taken to contain an unknown number of targets and the method relies on the downstream inversion process to resolve the true number of targets present. This is achieved in the inversion step by attempting to fit increasing numbers of dipoles and evaluating which value provides the best fit, whether that is none, one, or multiple."
Thank you once again for your input!
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsI believe that this article presents a fairly high-quality approach to automating inversion for simple local objects in a simple containing environment, and in this regard it can be useful and interesting. The positive point is that the authors adequately assess the capabilities of the proposed approach, clearly indicate the limits of its applicability, and honestly show the results in the presence of geological interference. Of course, this work cannot be considered as 'Article' covering a scientific problem, but as a 'Communication' it may well be published. I hope that in the future the authors will also consider those problems that they have now removed from the work and transferred to the inversion stage (distinguishing between magnetic fields created by ammunition and geology), especially since approaches that allow solving this problem are already known (for example, vector cascade inversion).
Below are some minimal comments to make the article easier to read:
1. Move figure 1 from section 1 to section 2.2.
2. 49-56 either remove or reformulate as the research hypothesis, and not as a result of its proof.
3. I propose to divide Figure 4 (two real cases) into two different figures and place them next to the text in which they are discussed.
I wish the team of authors success in the development of this research.
Author Response
Thank you very much for taking the time to review our manuscript, and for your helpful suggestions for improving clarity. We've incorporated each of your suggestions, and believe that our manuscript now flows more naturally. We've attached the PDF of the revised manuscript, which includes the changes below:
1. We moved figure 1 to section 2.2.
2. We moved the substance of lines 49-56 to the Discussion section.
3. We divided the previous figure 4 into the new figures 4 and 5, which are now adjacent to the relevant sections of text.
Thank you once again for your input!
Author Response File: Author Response.pdf
