Compliments. Very interesting article and subject. In the following only few hints for improving the presentation of the work.

Abstract

Well written and rather clear. Perhaps shold be stressed better the differences among three machine learning methods.

It could be useful to follow the same order of presentation that has been illustrated in the introduction (RDF,CNN,TDA).

1) Introduction

Very good the stress about the difference among earth-triggered and rainfall-triggered events. It is a common problem also today since there is not a standard way to classify landslide information

It should be better stressed that the whole work has been included in a new python library and that it is one of the unique in this field.

2) Data

Ok, all information required and necessary for the case study are listed correctly!

3) Methods

3.1 ok well clear. Perhaps, a brief explanation on the differences between earthquake and 

rainfall triggered shapes should be included even though the reader miss the key information

about the classificator.

3.2 the method is presented clearly but in my opinion is missed a point: why we need

all these infromation extracted by the DEM? It should be remarked at least to highlight you are

presenting this second approach

3.3 here is clear that the output is a probability to belong to one or to the other class. And the

presentation of the CNN method in more readable !

4) Landsifier model evaluation

from 198-202 it is not clear how the simulation was carried out. Please revise it and may be a short table could

be helpful in this sense

4.1 OK well presented and clear

4.2 the same of 4.1

4.3 the same of 4.1

5) landsifier library

Ok for the riminder but since it represents the TITLE of the work probably few words

should be spent here. For example: which functions are embedded in? Settings and options?

Is it fast or slow in computation? Just some characteristics that can involve the reader to download and test it.

6) Discussion

The discussion are well written but in my opinion should be rather organized in order

to highlight better the outcomes of the study. Which is the best tecnique adopted? Cost e benefit of each tachnique?

Computational demand? Accuracy? All of the questions are expected by the reader aftere the presentation of the new python library.

 

from 267-282: this statement is OK but should be more integrated in the discussion of the current work: without good quality landslide data the performance of classificaction techniques may be not sufficient.

from 282-293: ok so should be stressed better the TDA peculiarities with respect to the other

from 294-310: ok but future outcomes and expected improvemente should be better highlighted!

7) Conclusions

They are well reassumed but in my opinion a the Landsifier novelties and key new element should better shown. In particular, are there any software application such as for landslide census or analysis at catchment scale?

Is it a tool useful for susceptibility mapping or also for Civil protection purpuose??

Appendix

Well organizaed and rathe clear. Perhaps the scheme B1 of appendix B should be moved to the chapter 5 and then described briefly since represents the core of your work (Landsifier library).

Dear authors, 

I would like to express my appreciation on your work. I have read it with great interest and found the overall manuscript of particular scientific relevance for the geomorphology community. 

Below I will summarize the manuscript content and later provide my feedback and suggestions, which I would like to stress here from the very beginning are very minor.

The manuscript deals with a very important topic, as it proposes a protocol to estimate the likely trigger of landslides from their shape and size characteristics. This is done by using a CNN architecture, which adds a numerical and methodological flavor to an article that adresses an important research question. In fact, as also stated by the authors, any attempt to predict landslides relies on previous information, this being usually expressed in polygonal inventories. However, often these inventories only provide the location of landlside occurrence and extent lacking to report the date. These inventories are usually geomorphological inventories (sensu Guzzetti et al. 2012) and they still represent the vast majority of the available inventories. This means that not knowing the date we cannot know the trigger responsible for the landslide occurrence, which is a fundamental requirement to then better understand the slope response  over the whole affected landscapes. 

Therefore, the protocol proposed by Rana and co-authors brings a very relevant tool for geomorphologists and for any other pratitioner, especially because of the way the authors opted to share their work through a python script. This is particularly important for repeatability and reproducibility. 

Aside from this general overview of why I think this manuscript deserve to be published, specific elements support the same conclusion. In fact, the text is extremely elegant and it flows nicely while reading it. I actually read it all in one go, which is something that not always occurs. In addition to the style and readability of it, I would like to stress the originality of the manuscript because to my knowledge at least, no other work has addressed the same issue, specifically with an open source solution to the problem. Also, the quality of scientific illustrations is very high. I am usually quite picky and yet I have no real comment to add, other that complimenting the authors.

In terms of feedback I can provide to improve the text, I have comments almost exclusively related to the literature review and what could be potentially added. One is that the authors refer to Taylor et al. (2018) in their text and rightfully so. But, at least for me a very similar if not better article has been recently published along the same lines that Taylor and co-authors introduced for the first time. The work I am mentioning is authored by Amato et al. (2021), where they also use a neural network architecture to explore landslide shape characteristics and infer on the landslide type at hand. This is even closer to your work because of the method they chose to use and I feel should be mentioned in your text. Another potential missing reference could be Lombardo et al. (2019), a paper where the trigger pattern has been derived from the inventory itself, although using a latent effect featured in a statistical model. Of course what you propose here is different but the question is basically the same: "How do I retrieve the trigger from the landslide themselves?". It is worth mentioning that I am the first author of that paper, therefore if you feel like I am imposing a reference, feel free to avoid my comment. I swear it is a genuine one, without a second interest to it.

Other than this minuscule details, your work is impecable to me and I would definitely be happy once I see it published in NHESS.

Again, congratulations. Kind regards, 

Luigi Lombardo

References

Amato, G., Palombi, L. and Raimondi, V., 2021. Data–driven classification of landslide types at a national scale by using Artificial Neural Networks. International Journal of Applied Earth Observation and Geoinformation, 104, p.102549.

Guzzetti, F., Mondini, A.C., Cardinali, M., Fiorucci, F., Santangelo, M. and Chang, K.T., 2012. Landslide inventory maps: New tools for an old problem. Earth-Science Reviews, 112(1-2), pp.42-66.

Lombardo, L., Bakka, H., Tanyas, H., van Westen, C., Mai, P.M. and Huser, R., 2019. Geostatistical modeling to capture seismic‐shaking patterns from earthquake‐induced landslides. Journal of Geophysical Research: Earth Surface, 124(7), pp.1958-1980.