I reviewed the paper entitled “Fatbox: the fault analysis toolbox” by Gayrin et al. The paper presents a python software using image analysis methods to automatically extract faults from DEMs, numerical and analogue models in two dimensions. In addition to reconstruct a fault network, the software can perform a time evolution analysis to characterize fault evolution in time and space. Although limited to 2D applications, as well discussed by the authors in the limitation section, the method looks good and the presented applications demonstrate well that it is promising.

Here are some comments and remarks:

The paper presents the results of applying a sequence of methods and functions to treat the data and obtain a fault network. Although the results look good, I think it would be nice to provide more details about these functionalities and illustrate the effect of the functions applied at each step from the raw data to the final fault network. While some steps are shown for the DEMs fault extraction, we do not see it for the numerical model to reconstruct a fault network from the raw plastic strain output. The jupyter notebooks illustrate this more, but I think it would be good to have these steps in the paper.

You made the choice to structure the paper using applications which leads to a lot of redundancy as some methods are the same across the applications e.g., the DEM and analogue model share the same automatic topographic analysis while the numerical and analogue models share the same approach to use strain and characterize fault evolution over time and space.

I suggest that instead of structuring the section 3 by application you could structure it by approach or functionality and present how a functionality is used for each application.

At the beginning of section 3 you could present the applications, and then describe your functionalities, how they work, what they do, when to use them etc. with some nice illustrations.

There is a problem with the references across the paper. I don’t know how this occurred but there are a lot of citations in the text that are not in the reference list. You should double check that before re-submitting the article.

Finally, because the paper is very oriented about the Python package, it is referred several times and its content is described, I had a closer look to it. I have some suggestions that I think could benefit its spread and use but I do not think that the authors need to address them for the paper to be published.

Line by line comments:

L107: “Each component of the network consists of nodes (points defined by location x- and y-coordinates) and edges (connections between nodes)”

This sentence seems appropriate to introduce here that this (nodes + edges) is called a graph, particularly because it is the title of the subsection and that you use the term graph at line 112.

L137: I am sorry, but I do not understand what “edit the network” means in this context.

L185-187: What is actually detected by the Canny algorithm? In the jupyter notebook we can see that the method is conveniently implemented in the scikit library but the actual quantity and approach utilized to detect “edges” and what it means in the physical world is not clear to me. It is a very important step of the method presented here, and even if it is not implemented by the authors, it would be nice to have an overview of the approach.

L188-189: “non-fault-related components exhibit distinct geometric signatures; they are subsequently filtered out in using dedicated functions based on curvature and length”

Interesting! However, there are no real details in the manuscript on how this important step is taken. Looking into the jupyter notebook tutorial we can see that the scikit method called “remove_small_objects” is used, with a parameter “min_size=30” pixels. This value sounds arbitrary. Is there a generic way to choose it? If the surface of the covered ground or the resolution of the DEM was different, would that value change?  L215: “a virtual cross section is drawn perpendicular to the fault axis until reaching a predefined maximum distance”

How is that predefined maximum distance chosen? In the notebook it is set to d=12 (no units indicated so I assume it is also in pixels?), same question than in the previous comment: will that value change if the surface and/or resolution of the DEM changes? How can you be confident enough that using a constant value for all faults may not lead to crossing another fault depending on where you are performing the analysis? 

L271: “In the following, we use a relative threshold optimized to highlight fault structures.”

Could you please provide a bit more details of how the optimized relative threshold is chosen?

L277: “In some cases, interpolation to a regular grid is necessary, particularly when the model dataset has variable resolution”

Why is that step necessary? If the object used in the first place is an image, the pixel grid is already regular even if the model’s mesh is not, isn’t it? Can you please provide an example for which that step is required and illustrate why?

L280: “Our library provides functions to address these issues”

This is interesting, but unfortunately, we miss details here. What are these functions? Could you be more specific about which function addresses which issue and how? It would help the readers to know what functions they should use depending on their case.

L282-284: “To optimize the network, a filtering function removes selected internal nodes (based on user-defined parameters), reducing density while preserving realistic geometric orientations.”

Could you please elaborate on that function? What does it actually do?

L290-291: “the program compares each fault at time 1 with each fault at time 2 and vice versa”

If I understand correctly, you mean that you compare two consecutive timesteps, right? If this is the case, I suggest to use “time n” and “time n+1” instead of “1” and “2” to better emphasize that it is between two consecutive steps, whatever their numbering.

L292-293: “we calculate the minimum distance between each node of a fault and each node of the neighbours and then average these distances”

I do not understand what is the “distance of each node of the neighbours”. Neighbour is not defined and thus ambiguous. Could you please be more precise about the distance between what and what is computed?

L299-300: “The article (Neuharth et al., 2022) provide a detailed study of fault system evolution in the 2D numerical models discussed earlier”

I think that it is a bit unfortunate that the illustration of the method presented in this paper is actually not in this paper but in another. You could maybe showcase some of the possibilities of the method presented here in a figure to provide the readers with examples of what they could expect from your software especially what you describe lines 302-303

L339-350: Section 3.3.2 roughly contains the same information than section 3.1.2. I suggest that you could remove this section. L355: “strain threshold to distinguish active faults”

Is strain correctly employed here? It sounds strange to use strain to identify active faults. Strain-rate would be a better quantity because strain also contains inactive faults. L357: “data are binarized”

I suggest to introduce the term binarized in section 3.2.2 where the procedure is first described. L356-360: This paragraph is also a redundant information with a previous section. I suggest to remove it.



L362-365: Is this procedure different than the one described at lines 195-197?

If not, I suggest to group them, if yes, I suggest to provide more details at lines 195-197 and then explain why here the procedure is different. 

L387-388: “The Canny edge detection is based on this topography gradient criterion.”

While the fact that the Canny edge detection uses topography gradient can be reminded in the discussion, I think it would be nice to introduce much earlier, at lines 185-187, what the Canny algorithm actually does and what type of quantity it uses. For example, the use of topography gradient criterion was not mentioned before the discussion.

Section 4.1 Fatbox options for defining a fault:

I feel like there could be more information about how to choose some parameters for each identification criterion. While the DEM analysis is slightly more detailed, the strain or strain-rate approach is summarized in a single sentence. The last paragraph is good, it states what type of choices can be made. Aren’t there more like this that you could elaborate upon to provide more insights about your toolbox and what we could do with it?

L420: “most steps can be parallelized”

What do you mean by “parallelized” in that context? Does it mean that your toolbox should run in parallel i.e., multiple mpi ranks/threads or that the steps of the procedure are independent of each other and thus you can perform each of them independently?

Because as I understood, there are steps or procedures that do not seem independent, like the time evolution and the correlation between time steps as each time step needs to be treated sequentially.



Section 4.3 limitations:

What about the limitations concerning the time correlation to identify faults over time, is it always succeeding or are there some cases for which it fails?

Although the 3 applications proposed are demonstrating several contexts in which the toolbox can be applied, they are all about rifts and extensional systems. What about convergent and strike-slip systems? Do they represent limitations? If yes which ones?

Minor comments:

L82: Mattéo et al. (2021). The parenthesis should be in front of the M as the citation is not embedded in the sentence => (Mattéo et al. 2021).

L96: parenthesis missing befor the 2)

L249: there is a parenthesis that should be remove after the “e.g.”

References :

L85: T et al., 2025. It seems that there is a problem with this citation both in text and in the reference section (L615).

L260: References are required for the wet quartzite, wet anorthite and dry olivine flow laws. (Not Neuharth et al., 2022; the actual papers that published the parameters used in the model). L261: “Beneath the lithosphere lies a weak asthenospheric layer composed of wet olivine (Neuharth et al., 2022)”

Neuharth et al., 2022 is not the publication related to the wet olivine flow law. Here, Hirth & Kohlstedt, 2003 should be cited as this is the paper describing the wet olivine flow law used and cited in Neuharth et al., 2022. 

Below are some references missing from the reference list (the one I noted, may not be exhaustive):

Panza et al., 2024

Purinton and Bookhagen, 2021

Baker and Wohlenberg, 1971

Canny, 1986

Guo and Hall, 1992

Shmela et al., 2021

Gassmöller et al., 2018

Glerum et al., 2018

Braun and Willett, 2013

Yuan et al., 2019

Saha et al., 2016

Strak and Schellart, 2016

Strak et al., 2011

Willingshofer and Sokoutis, 2009

Philippon et al., 2015

Schlagenhauf et al., 2008

Lathrop et al., 2022

Henza et al., 2010

Jourdon et al., 2025

Software/Code related remarks:

The following comments do not need a particular attention from the authors to publish the article but they could take them into account for future development to enhance code visibility, availability, reusability and collaboration.

While cloning the repository, I realized that the repo is 174 MB large, which I agree is not huge, but the code is only made of 6 Python files and a few jupyter notebooks. This size likely comes from the storage of the data used in the jupyter notebooks as tutorials/demonstration. In general, storing large data files in code repository is not recommended as it will impact users when they download, install or update the code from the repository. A better way would be to store those data in another storage place offering long time storage, and link to it to get access to the data.



In the README it is mentioned that the installation should be done using conda. I understand that it offers a simple way to install your package but it forces the users to use conda. Fortunately, your package can actually be installed without the use of conda, and except for earthpy which can also be installed without conda, there are not that much benefits of strictly restraining the installation to the use of conda. I don’t say that you should get rid of it, but I rather suggest to provide more options to the users. You could consider to add a setup.py to install using pip for instance.

The manuscript presents a new tool, FatBox for delineating faults networks either from the expression of normal faults scarps on topography or from strain maps. The paper describes the modules within FatBox and applies this tool to three different types of normal faults systems to highlight the different types insights that can arise from fault data extracted using Fatbox.  As presented, the tool has great potential for normal faults systems.  The figures are nicely constructed but the writing can benefit from greater clarity.

1. The paper has a strong focus on normal faults within extensional regimes. The paper could either 1) focus only on normal faults and accurately reflect this in the title and throughout the manuscript or 2) broaden the scope to strike-slip and reverse faults by adding examples of these faults and citations to fault detection approaches for those systems.
   1. For example, lines 45-47 talk about normal fault growth by linkage. Thes same could be said for strike slip and reverse faults.
   2. Strike-slip faults can have geomorphic expression. For example, can FatBox analyze detect strike-slip faulting along pressure ridges? Should present on this around lines 60.
   3. Figure 1 only presents normal fault scarp analysis, but FatBox could be applied to dilatational/longitudinal strain maps of reverse faulting or to vorticity/shear strain maps of strike-slip faulting.
2. The introduction explains how field mapping and numerical models investigation fault network development. The two sentences on physical experiments in the introduction do not adequately explain how these experiments are used to provide insight on fault networks. For example, the paper can explain that within the table-top experiments the material are often scaled so that stress states within the experiment represent the crust.
3. Fault connectivity is a tricky element that the process of skeletonization process may or may not do in the same way as a human mapper. Fo example, the faults of Figure 3b maybe be under-connected. How is the correct degree of connectivity assessed?
4. For the reason of c and other issues, the paper would greatly benefit from at least some validation exercise.
   1. Can FatBox map as well as human? Probably not but it certainly creates maps faster than a human. Can FatBox save time my providing a map that a human can revise more quickly than if they built from scratch? 
   2. It is possible that FatBox reduces biases introduced to fault map by human interpretation. The rigorous assessment of that is probably beyond the scope of this study but human mapping bias is well documented in the literature and could be mentioned.
5. Can the analysis work equally well for incremental strain maps and cumulative strain maps?
6. Some of the fault detection approaches within FatBox (absolute and relative strain threshold and adaptive threshold) have been used before and are not new to this study. Some discussion of the benefits and drawbacks of the different approaches along with citation to previous studies will help the reader make best use of Fatbox.

Specific comments:

• I recommend changing the term ‘analog models’ to either ‘physical experiments’ or ‘experiments with crustal analog materials’. This change in wording better distinguishes numerical models from laboratory experiments.
• The paper uses a lot of passive voice which is not very interesting to read. Try adjusting passive voice (e.g. ‘fault networks are investigated’) to active voice (“structural geologists investigate faults networks’) and you will find that the text is a lot more engaging and direct.
• When citing a paper that is just one of many examples that you could use, preface the citation with ‘e.g.’ this signifies that there are many other papers on this topic. Throughout the manuscript many citations listings need this e.g. because they do not list all of the papers on that topic.
• Rewrite the application sub heading include the type of data set analyzed. 3.2 application 1: topography from Magadi Natron basin; Application 2: strain rate maps from numerical models; Application 3: topographic and incremental strain maps from experiments

Line: 15: “Understanding complex fault networks” is a very vague and your contribution could be more effectively community by being specific.

Line 25 (and else where): the word ‘graph’ is not effectively used here and elsewhere. What do you mean? How is a fault network represented as a ‘graph’.  

Line 31-32: The definition of faults here could equally apply to opening mode joints, veins and dikes, which are certainly not faults. You are missing the key element to faults that is dip or strike slip.  

Line 59: awkward working: Maybe try “…unlike thrust faults within contractional settings where the hanging wall collapses and obscures the fault scarp”.

Lines 95-98: This listing is vague and confusing. Can be clarified by using similar set up for each example 1) uses a topography dataset from Magdi Natron basin, 2) uses a strain rate data set from a numerical model and 3) uses topography and increment strain data sets from experiments.  “Fault geometry tracking within an analogue model’ doesn’t mean anything.  

Line 124: This is only for normal fault systems so the text should be revised to state this.

Line 128: Throw and heave are displacements.  By displacement do you mean net slip?  

Line 150: Incorrect. Particle image velocimetry (PIV) is a process and not a product. What you are extracting the faults from is the strain map. This can be longitudinal strain, shear strain, dilatational strain or vorticity.  Thos strains are calculated from the displacement fields generated by PIV. But the fault maps are not extracted from PIV.

Line 153: Only in regions with low erosion and deposition rates

Line 163: Largest continental rift. The mid ocean ridges are larger.

Line 189: What does ‘dedicated functions’ mean? Are these bespoke scripts for the particular data set that the user develops?

Lines 211-212: What does this mean?

Lines 217: ‘Fault displacement and extension’  doesn’t make sense. I think you mean net slip on the faults?  By the way this is one (of many) sentences with passive voice.

Line 219: “accessible though dedicated attributes” This needs more explanation.

Line 320: incorrect PIV is a type of Digital Image Correlation. They are not synonymous. PIV calculates incremental displacement fields.  It doesn’t calculate velocity because the time step does not emerge from the PIV analysis. The analysis doesn’t consider the timing between successive photos.

Line 329: Fault extraction has been performed from experimental strain maps and that literature should be presented because it is not new to this study. The work that I’m most familiar with are for experiments of strike-slip faults.  Some studies use a globally set incremental strain threshold (e.g. Hatem et al., 2017 JSG, Visage et al., 2023 Tectonophysics). Some recent papers use a adaptive threshold for fault detection from strain maps (e.g. Chaipornkaew et al., 2022 GRL; Gabriel et al., 2025 Tektonika).  In a paper from my research group that is currently under review (under review since July!) we test the sensitivity of the detect strike-slip fault network to adaptive threshold parameters. Seems like this could interest the authors and I am happy to follow up to share the paper.

Line 371: Extension of the faults? Unclear.  Do you mean dip slip along normal faults that accommodate extension?  Once again here faults do not have displacement, they have slip as one side of the fault displaces relative to the other.

Line 441: post-extraction validation of the fault network.  This paper would be greatly strengthened with a demonstration of validation (see point d). We don’t know if the tool is useful if it is not validated.  At minimum, this part of the dsicussion can outline how such a validation could be performed and the network assessed.

Line 470-: The first statement of the conclusions implies that Fatbox provides tools for fault analysis.  The tools provided are very helpful for extracting faults, fault tracking and slip information but the analysis itself is done after that data is collected by Fatbox.  For example, if someone wanted to analyze the evolution of displacement-length data along a set of faults, Fatbox could extract the data, but the researcher would need to do the analysis of the data. Rewording the text here and in the abstract (lines 27-29) could more accurately convey the value of the tools.