MV-GPRNet: Multi-View Subsurface Defect Detection Network for Airport Runway Inspection Based on GPR

Round 1

Reviewer 1 Report

The main concern with this paper is the lack of comparison with ground truth. In essence none of the claimed performance parameters can accurately be relied on unless comparison with the results of excavation can be shown. It is appreciated that this cannot easily be done on runways, but the claims will only be convincing when test sites with known and established features are used to rank the processing techniques. The paper essentially compares the techniques against each other. The authors need to address this. The use of the term noise is incorrect. GPR data is cluttered but none of the plots shown have much evidence of noise. This is concerning as it implies a lack of understanding by the authors as to the essence of the problem which is extracting targets from a cluttered environment. White noise by its nature is random in amplitude whereas the GPR data shows evidence of artefacts associated with the soil and subsoil.

Fig 14 shows multiple returns on the right-hand side labelled pipes. They look remarkably similar to the returns from rebars or cables rather than pipes which are usually more deeply buried. Do the authors know what they are? Plastic pipes are hollow and give a different signature.

There is no mention of the propagation parameters of the test sites and how these may affect the results. Why is the crack detection low? Is this a consequences of antenna design and element spacing and limited bandwidth rather than the processing? Section 5.4 is misleading - if a site has been excavated the defect is recorded and clear - the error bounds on human performance need to be stated. Generalisations are not acceptable. 

In summary the work may be more an assessment of the particular radar the Raptor 14 channel system than a truly independent method of classifying targets. The English needs attention by a native English speaker as some words like “novelly” are nonstandard.

The paper would benefit from a reduction in length, more explanation of terms such as ablation which while common in machine learning many are not to GPR based readers and improving the clarity of reading flow.

Author Response

Dear Reviewer,

Thanks for your valuable suggestion. We have responded to your comments point by point and revised our manuscript accordingly. Please see the attachment.

Best regards,

Nansha Li, et al.

Author Response File: Author Response.pdf

Reviewer 2 Report

The paper concerns the application of the GPR method for the rapid mapping of the subsurface defects in the airport runway. The topic and the experimental tests are quite interesting and of great interest for a wide community of civil engineers and applied geophysics. To date, there is a growing interest about the novel applications of the no-invasive electromagnetic sensing technologies for monitoring civil infrastructures.

The work sounds with the current state-of-the-art, I greatly appreciated the use of machine learning agorithms for GPR data analysis. I have only some minor comments.

The authors mainly focused the attention to the application of deep learning algorithms, but they completely omitted to report basic informations about the physics of the GPR method. No discussions about the velocity of the em signals and the criteria underlying the selection of the dielectric constants are reported. I strongly suggest to include more details about this aspect.

Finally, the English form could be improved. Furthermore there are many typewriting errors (see lines 379, 409 etcc) to be removed.

On the basis of these considerations, I suggest the pubblication after an accurate minor revision.

 

 

 

Author Response

Dear Reviewer,

Thanks for your valuable suggestion. We have responded to your comments point by point and revised our manuscript accordingly. Please see the attachment.

Best regards,

Nansha Li, et al.

Author Response File: Author Response.pdf

Reviewer 3 Report

Dear Authors,

The presented research describes a CNN-based algorithm for the automated detection and localisation of a number of subsurface defections. Though not novel in its essence, the article is interesting and well developed, particularly due to the inclusion of experimental and field data (not so common within the domain of AI/ML applied to GPR data) and for the provided comparison with alternative solutions.

On an overall basis, the article is properly contextualised and developed, hence easy to follow. As well, language is fine. I appreciated the inclusion of Section 2 which gives the reader a detailed analysis of the current status of the research within the field.

Here are some minor comments that could be useful to improve the overall quality of the manuscript.

·         Section 3: it would be nice to describe the adopted GPR equipment, its specifications and characteristics.

·         Section 5.1: a more detailed justification of the adopted parameters would be better, as otherwise one could wonder why you selected those specific values for resolution, thresholds and learning constraints.

·         Section 5.3: the built model description lacks information on the source and excitation characteristics.

·         Section 5.5.1: as before, some details on the survey and acquisition must be included. It would be also interesting to have more details on the robotic platforms.

·         Section 5.5.2: comments on Table 3 – which is essentially the core of the proposed work – can be in my opinion improved to better highlight strengths and limitations of the algorithm, potential indications for such results and the reason for such performance.

Finally, there are some figures that could be improved, particularly in terms of clarity and information (labels, axis, etc – see fig. 15 and 16 for example).

I hope these comments might help.

Best Regards.

Author Response

Dear Reviewer,

Thanks for your valuable suggestion. We have responded to your comments point by point and revised our manuscript accordingly. Please see the attachment.

Best regards,

Nansha Li, et al.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The paper is much improved but there are still some issues that need to be corrected

 

The authors use the terms noise and clutter incorrectly and the following clarifies the accepted understanding of clutter and noise as far as radar engineering is concerned

 

The IEEE “Standard for Radar definitions” defines clutter as unwanted echoes.

 

Noise in a radar system is considered as primarily

 

Thermal emissions of the target scene

Random currents in any and all components, including semiconductor shot noise

Data quantization effects

Purposeful random dithering

Atmospheric phenomena

Cosmic sources including cosmic background radiation

Random Interference both intentional and unintentional, external and internal

Shot noise (Poisson noise) due to the discrete nature of electric charge

Flicker noise (1/f noise, pink noise) in active devices

Plasma Noise due to random motion of charges in an ionized gas

Quantum Noise due to random currents in conjunction with motion of discrete charges

 

The introduction contains the author’s version of these terms which is incorrect. The data shown in the figures shows little evidence of noise signals due to say the receiver kTB noise but is actually data of some form received by the antenna. The structure of these signals as evidenced in Fig 2 a b and d shows signal returns that are caused by other features within the ground. While this information may be unwanted it is clutter rather than noise. Fig 2 c looks to be contaminated with RF interference. Therefore, the paper needs to be revised accordingly.

 

The paper states the Raptor radar sweep speed is 1,800 scans per second. Later on the paper states the horizontal sampling rate of the GPR is 390 A-scans per meter and the: vertical sampling rate is 1024 sample per A-scan. Why are values of 1800 scans per second and 390 scans per second being used This needs to be explained.

 

The spatial resolution based on the channel spacing is 87.5mm which means that cracks of less than 40mm wide between the channels are unlikely to be detected well. The authors need to explain why the crack detection performance is low and whether it is a fundamental limitation of the array.

 

A final spell check for grammar and punctuation would be beneficial

Author Response

Dear Reviewer,
Thanks for your suggestion and professional instruction. We have responded to your comments point by point and revised our manuscript accordingly. Please see the attachment.
Best regards,
Nansha Li, et al.

Author Response File: Author Response.docx