This manuscript presents a practical procedure for using free data of Sentinel-2 and Google Earth Engine to create an inventory of shallow landslides. The procedure was applied to two landslide areas and the results were compared with the manual landslides. Overall, the methodology is well presented and the data analysis is performed in detail. The methodology shows the potential to create inventories of landslides that are fundamental to landslide research. I have some minor comments for the authors to consider:

(1) Figs. 1 and 2 included the rock formation of the studied sites. What is the role of geological information in the creation of landslide inventories?

(2) Line 190, Eq.1, could the authors elaborate how the eq. 1 is applied to determine the NDVI value of the studied area? What is the 'pixel' size ( the calculation area) for each of single NDVI?

(3) The authors should detail how a polygon of PLs is formed?

(4) Reference to Eq. 2 and Eq. 3 should be provided. The definitions of Eq. 2 and Eq. 3 seem to be different from other studies.

(5) Figs. 4 and 8 C, could the authors remove the PL data to clearly show the color map of NDVI?

(6) Fig. 13, there is no 'C' in the caption. Also, comment on how C is created. From A and B, it seems that the kernel density points/locations are slightly different.

(7) The methodology still needs to be improved to capture the landslide inventory accurately. Could the authors comment on the limitations and the area for improvement? 

(8) Instead of having operators change the mapping parameters and visually compare the ML and PL, are there any qualitative ways (e.g. machine learning, optimization methods...)to get the most appropriate parameters to have the highest degree of match between ML and PL?

The study of Notti et al. deals with the semi-automatic recognition of shallow landslides for the compilation of event-based inventories over two study areas in NW Italy. In general, complete landslide inventories are very important products for the study of susceptibility, hazard and risk. Therefore, I retain this study a very interesting contribution for NHESS. Before publication, I would like the authors to answer a couple of general comments and a series of detailed comments, all listed below.

General comments

• In my understanding, the manual delineation of landslides for the calibration and in this case the validation of the automatic procedure is the most time-requiring activity related to the proposed procedure. For repeatability in other areas/regions, would it be possible to structure the study so to give an indication of how large the training areas should be, compared to the total investigation domain, to allow for quality (high performance) results? Also, if a training area that is just a part of the entire domain is selected, what characteristics should it have in terms of morphology, land-use and other properties?
  
  In line 251-255 authors explain that for a study area small training areas are selected while for the other case study the whole domain is used. If for the same study area, small training areas and then the whole domain are used, does the performance of the procedure change? By a sort of iteration, is it possible to define an ideal dimension of the training area? Is it possible to confirm this ideal dimension for the second area or it changes due to the different morphological and land-use conditions?
• In the study area Section, I suggest to add a climatological setting description particularly (but not only) focused on extremes frequency and its possible variation in recent years.
• Results should be discussed against previous literature with major detail.
• I suggest a general revision of English and a thorough formal proofreading as well.

Detailed comments

1. L12: it is the first step towards […] --> unclear
2. L17: hatted --> hit?
3. L19: well match --> match well.
4. L21: Keywords --> I suggest to use keywords that are not included in the title
5. L25: during flash floods […] shallow landslides --> what is the typical depth range of what you define shallow landslide? Also, flash floods and shallow landslides are two different phenomena both related to extreme rainfall events, they can happen simoultaneously but I would not connect them directly.
6. L41: satellite images resolution nowadays is not so different […] --> which is the typical resolution?
7. L43: what do you mean for dedicated acquisition planning? Aren’t satellite orbits sort of fixed and so the revisiting time defined?
8. L54-61: I suggest to rewrite this paragraph making explicit the general and the specific objectives. In addition, emphasize the novelty in comparison to previous literature.
9. L94: more steep slope in Serravalle Formation --> more steep slope than what?
10. L102: NW Alps have been affected --> always? Recently? Has the frequency changed with time? Some of this should be included in the climatological setting to be added in the study area section.
11. L105: Especially on a short time interval --> unclear, please quantify the short time interval.
12. L109: 650-700 mm --> in what time? Five days?
13. L112: almost accurate --> what is it meant for almost? Which method was used?
14. L129: intensity --> hourly intensity or instantaneous intensity?
15. L149-150: the difference between a map of areas most affected by landslides and a map of landslides is the difference between generally unstable areas and single landslides?
16. L154-157: please specify the resolution difference between satellite images and high-resolution images. Also, which is the source of high-res images?
17. L161: using slope and other geomorphological parameters --> In this phase I would say terrain and geomorphological properties. Details in the following sections.
18. L175: is the same period of the year of this point ii more restrictive than the period June-September?
19. L181: averaged NDVI --> spatially averaged?
20. L182: filtered by cloud cover --> 5%?
21. L184: these constraints --> it is not very clear how the Novak algorithm takes into consideration the constraints listed in the bullet list I, ii, iii.
22. L194: manually select NDVIvar threshold --> in a single image (raster) is the threshold the same or it can vary from area to area?
23. L195-196: the whole sentence --> It is unclear, can you please provide an example (or a couple)?
24. L202: the value is empirically based[…] --> in this case the threshold is unique, right?
25. L203-204: additional filters maybe introduced --> such as? In addition, maybe based on what? It’s a bit obscure.
26. L209-210: The parameters used to […] --> was a formula developed?
27. L222: The iteration step aims […] --> when is the iteration stopped? Which is the threshold used to exit it?
28. L239: The Boolean raster […] --> how is this obtained? With a single 10 m x 10 m cell there are four 5 m x 5 m cells. Did you use an average value? Majority? Else?
29. L247: Table 1 --> I believe it could help the reader to have the vent date in the Table.
30. L271-272: algorithm based on Novak et al. (2021) --> is it exactly the same algorithm described in Section 3.1 or there are some differences?
31. L273: TS --> please define the acronym.
32. L274: estimate the recovery of vegetation --> if this is a specific objective of your study, please clarify it in the introduction.
33. L288: The characteristics of FP --> what characteristics?
34. L294: landslide dimension or land use --> why not using terrain properties such as aspect, slope, flow accumulation?
35. L296: Table 3 --> can you please clarify the difference between PP and PD? How did you distinguish between the two? Does it mean that in PP, Pl is larger than ML while for PD is the opposite?
36. L305-308: The comparison […] characteristics --> methods not results.
37. L330: not filtered because the hydrographic network has no precise geocoding --> unclear.
38. L346-347: The intersection […] parameters --> methods not results.
39. L355: the methodology detects a landslide in 60% of the cases --> These TP-FP analysis results mean that the methods returns a over-representation of landslides. What about FN? Please discuss Fig 6C and 6D for this aspect as well.
40. L377: Fig. 7 --> what about false negative?
41. L386-387: we obtained the NDVIvar […] --> which was?
42. L387: Fig. 4D --> 8D?
43. L400-401: It is possible […] manual mapping --> unclear.
44. L404-405: the trigger points […] not detected --> This would not be a problem for an inventory intended for landslide susceptibility analysis, it maybe a problem for risk assessments. I suggest a brief discussion.
45. L419: Fig 6A --> 10A?
46. L427-428: The better performances […] 2019 event --> Are training areas absolutely comparable?
47. L454: it is possible to note […] small landslides --> unclear.
48. L476-477: The ML and PL […] case studies --> unclear, agree with each other?
49. L479-489: whole paragraph --> These results are sort of calibration results (training and application areas are the same), correct? What if you apply the method outside the training area? This relates to my first general comment.
50. L578: their density matches […] rainfall events --> unclear.
51. L582: landslides manually --> landslides manually detected?
52. L583: good agreement --> I believe it is worth mentioning that some parameters of the automatic recognition needed calibration.
53. L584-585: whole sentence --> unclear, >60% of landslides with areas larger than 67 m2 (2/3 of 100 m2) but <20% for landslides smaller than 100 m2?
54. L595: for middle latitudes, the best comparison is with summer images --> This was an assumption in selecting the images (only June-September), I suggest not to present it as a conclusion.