The authors present a very coherent and complete literature review on the various applications for detecting IRHs in radagrams, a very challenging and important problem. This is a very useful paper for anyone who wants to study IRHs and develop novel methodologies for automatically identifying them. I really enjoyed reading this review and I believe that the paper contributes to the existing literature and should be published subject to minor corrections. 

My comments are given below: 

The authors should consider mentioning at the introduction that radio-echo sounding is also often referred to as Ice Penetrating Radar (IPR) or radioglaciology.

Line~ 40: Matlab is referenced as a software for processing radar data and automated mapping of IRHs. Although I am sure there are some Matlab scripts that can do that, Matlab in general is a high-level programming language and should not be referenced as a radar processing tool. The authors should also reference Geolitix, which is a recent commercial software that is used quite extensively by the GPR community nowadays. The unique thing of geolitix is that it is web based. Regarding Matlab, there are some packages for example “GPRlab: A ground penetrating radar data processing and analysis software based on MATLAB”. But I have never used them, so I cannot comment on the quality of these tools.

Line~ 60. The authors mention echograms are also known as radagrams. I would also add the term B-Scan, which is also widely used by the community.

Line ~ 80. I would suggest the reviewers to add the well-known book of David Daniels “Ground Penetrating Radar” as a reference.

Line ~ 85. The authors refer to radioglaciology as a remote sensing methodology. This term might exclude the ground-based in-situ systems.

Line ~ 85: The authors mention that the waves are reflected by changes in the "complex permittivity" of ice. I believe it would be better to use the term “dielectric properties” instead of “complex permittivity”. Dielectric properties  are more inclusive and hold as special cases the complex permittivity, conductivity and magnetic permeability. It would be better also to explain the term “dielectric properties” because in paragraph 95 it is said that the main source of reflections of IRH are changes in conductivity while in previous paragraph you mention complex permittivity as the main source of reflections.

Line ~ 90. The authors refer to similar references as they refer them on paragraph 80 for the same context. I think there is a repetition here that can be avoided.  

Line ~ 95. The abbreviation IRH is previously already defined.

Line ~ 95. The authors mention that the main source of reflections of IRS is the change in conductivity, while at Line~85 they mention that reflections occur due to complex permittivity.

Line ~ 100. The authors mention that crystal orientation fabric can also result in reflections. How does this translate to dielectric properties? In the reference given by the authors (Eisen 2003) it is stated that there are evidence that suggests that changes in permittivity and not conductivity give rise to reflections in IRHs.

Line ~ 105. I would suggest the authors to use the term B-Scan as well as echogram and radagram.

Line ~ 105. Along A-scope I would also suggest the term A-Scan.

Line ~ 115. The authors discuss Figure 1, where an example of old traces is illustrated, where traces needed to be differentiated. I think it would be better to show more recent radagrams that don't need to be differentiated.

Line 127. Typo “…. a simplified schematic of a …”

Line ~ 130. The authors state “The red lines in the radargram indicate IRHs. In an ice sheet, these represent the interfaces between the neighbouring ice layers of different properties, such as layers with different density or crystal orientation fabric, or they can be caused by thin individual horizons with higher conductivity, thus forming IRHs”.  From the above I understand that interfaces caused by changes in conductivity (and not permittivity) are IRHs?

Line ~ 170. Reference is needed there to support this statement.  

Line ~ 135. This statement has been repeated many times in the text so far, and I believe that could be omitted.

Line ~ 190. It would be good to outline the limitations and advantages for each type of methods.

Line ~ 210. “..avoidance of having discontinuities”, I would add “..avoidance of having discontinuities between splines”.

Line ~ 225. Both Level Set Function and Snake are part of the Active Contour methodologies. The authors should consider adding them as subsections in the same section “Active Contour”.

Line ~ 250. The Radon and Hough transform are very similar. To my understanding the Hough transform is a discreet version of the Radon transform.

Line ~ 290. This is a very generic definition of support vector machines (SVM). The original support vector machines fit hyper-planes, but with the Kernel trick SVM can deal with non-linear boundaries. Also SVM can deal with multi-class problems using strategies such as one vs one, or one vs all.

Many of the sections that define the methodologies 3.1-3.12 are flagged as AI-generated text, which often results to read very generic and definitional.

Section 3.12. Flagged as AI generated. This is a very detailed and out of scope explanation of deep learning. I believe that this paragraph is not necessary. Otherwise another paragraph is needed to define what is machine learning prior to describing Support Vector Machines.

Line ~ 345. The authors state that U-net is a type of autoencoder. An autoencoder outputs the same inputs i.e. is an un-supervised learning method that has the same inputs as outputs. U-net can be an auto-encoder if the inputs and outputs are the same, but U-nets can also use different inputs and outputs. The paper cited by the authors (Ronneberger et al., 2015) is the first introduction to U-net, used as a segmentation tool, therefore not as an autoencoder.

Line ~ 395. The authors state “Over the last decade, various studies have put forth the use of pattern recognition methodologies in the examination of ground-penetrating radar (GPR) signals (e.g. Delbo et al., 2000).” The reference is from 2000 so it cannot support this statement.

Line ~ 435. Typo “….a semi-automatic picking routine…”

Line ~ 436. Typo “ …the maximum amplitude of each…”

Chapter ~ 4.1. It is not clear why some references are in bold font while others are not.

Line ~ 785. This is a very big paragraph that needs to be divided into smaller ones.

In Figure 7, would it be easy to include machine learning (not deep learning) related papers as well?

The authors should use some figures from the most important cited papers to illustrate the results of the methods discussed in chapter 4. It would be good to showcase with some visual examples how these methods work. 

see attached file