Thank you for the opportunity to the review this manuscript.

The authors develop an event-based inventory of shallow landslides for an area that includes forested and unforested zones. They find that some landslide characteristics (e.g., depth) host significant relationships with the presence/absence of forest and distance to individual trees. The authors conclude that these findings improve our understanding of how forests affect shallow landslide processes.

As written, it is sometimes difficult to discern if the main focus of this study is meant to be semi-automated landslide mapping methods or developing a better understanding of landsliding in context of vegetation. Substantial text is dedicated to describing the setup, tweaking, and performance of the segmentation algorithm. Yet, given the vegetation-oriented nature of the Abstract and Conclusion, I found that (1) relatively little text is developed to review the existing literature to shape the study design as part of the Introduction, (2) the analyses do not attempt to control for the effects of other landslide-relevant variables (e.g., topographic slope angle) inside/outside the forest cover mask as part of the Methods, and (3) process-based interpretations of the observed relationships between vegetation and landslides are underdeveloped as part of the Discussion.

Regarding Point #2, my impression is that authors have developed an internally consistent methodology that facilitates reproducible observations of differences in landslide dimension inside/outside the forest mask, but it is unclear why is there no attempt to “normalize” these comparisons in the context other spatially variable factors. The noted presence of steeper slopes in the forested versus unforested area (33 degrees versus 23 degrees, respectively; LN 104-105), for example, suggests that some consideration of landslide occurrence in the context of topographic slope angle, for areas with forest/no forest, should be considered. Or, it may suggest that the forest/no forest mask is not well suited for evaluating landslide characteristics in the context of vegetation for the study area. In the absence of an effort to deconvolve these kinds of competing factors, which the authors also note in the text (LN 419-420), it is difficult to evaluate if the conclusion (i.e., that the study provides “a better understanding of the roles of forest and how they affect the processes behind shallow landslides”) is supported by the results.

Sincerely,

Matthew A. Thomas

Other Notes:

• LN 99-104: Is this a naturally occurring grassland? What is the land-use history of this area and how might that be reflected in the local vegetation patterns? Can the authors elaborate on the characteristics of the trees in the context of landslide-relevant metrics (e.g., canopy cover, rooting density and depth, etc.)?
• LN 100: How was the forest canopy cover mask created? Did the authors generate it or did it come from another source? If the authors created it, should it be introduced in the Methods, as opposed to the Study Area?
• Ln 127-128: Should a citation be provided for “seeded region growing” algorithms?
• LN 148-151: Is the tree location dataset attributed with any information about tree type?
• LN 228: Use of “scar width” (if that is the intended meaning) may be more intuitive than use of “scar thickness,” especially since “scar depth” is already used.
• Results: I understand that the authors employ a semi-automatic mapping routine to reduce subjectivity in landslide delineation across the study area, but how is the transition between source and runout handled? This seems important, as landslide area is one of the metrics that the authors consider and this metric would also affect landslide volume. With sufficiently high-resolution optical imagery and lidar, a human can usually distinguish the transition from the source area to the transport zone for shallow landslides. It’s unclear what the segmentation algorithm considers.
• LN 287-290: Suggest reduced use of “it is interesting” unless the authors are more explicit about why these statements are particularly noteworthy.
• LN 395-397: Can the authors be more specific about these discrepancies?
• LN 409-411: Why might landslide area and landslide volume be unrelated to forest location? Landslide depth (found to have a relationship) would ultimately factor into landslide volume, for example.
• LN 416-418: Can the authors elaborate as to why distance to a tree appears to have a stronger relationship than the number or density of trees surrounding a scar?

 General comments

Thank you for giving me the opportunity to review this paper, I enjoyed reading it.

The authors aim to increase our understanding of the effect of vegetation, specifically forest cover, on landslide processes. To achieve this, they develop a new semi-automatic algorithm that combines topographic data (point clouds and DEMs from ALS) with aerial imagery, using a machine learning approach (Random Forest) to map landslide scars after an intense rainfall event. In addition, a forest cover mask is created to distinguish between landslides triggered in forested and non-forested areas. The authors report significant differences in landslide scar characteristics between these two environments, such as greater depth, thickness, and pre-failure slope in forested areas.

As I mentioned earlier, I appreciated reading the paper. It is well-structured and pleasant to read, with clearly stated objectives. The issue of landslide under-detection in forested areas is indeed a common limitation in landslide inventories. I commend the authors’ effort to use an approach that integrates different types of data, which seems appropriate to meet the objectives of the present study. Nevertheless, I suggest that the authors make some modifications and add clarity to a few points before the paper is published.

Best regards,

Thomas Bernard

Specific comments

. Abstract: Well informative. I would just add that the landslide detection workflow is based both on Lidar and the use of a random forest algorithm to make it clear that it is not just based on topographic difference.

. Section 3.1: It would be useful to add a table with information on the lidar dataset (date of acquisition, mean point density, ground point density, vegetation points density, data availability). The points density is informative on the quality of the data you are using. It is quite important especially for the construction of the canopy cover mask, the detection of single tree position and the derivative maps (tree distance and tree density maps).

. L 163: “(original DoD downscaled to a 0.1 m resolution)”. I guess the authors choose to increase the resolution of the DoD to match the one of the orthophoto. I have two comments. First, why making this choice rather than decreasing the orthophoto to 0.5 m?  In this way the interpolation is based on real information. Second, can you explain the method used to downscale the DoD? Did you produce DTMs at 0.1m from the point clouds? Maybe add this information in the text.

. L 171-172: Are the non-landslide areas delineated automatically or manually?

. L182: There are some technical terms such as hypertune that are unclear for non-expert or non-users of this kind of approach. Maybe add a sentence to explain a bit even if the reader can access to the reference of the method used.

. L 191/265/269-270: Can you explain a bit why there is an overestimation of landslide locations? Do you have an idea on the number of false negative? Especially in forested areas.

. L220: “.. landslides with partial cover <90% were excluded from this binary analysis.”: I understand the point, but I don’t think it should. It would be interested to see the same results in this category because if they are not different than the >90% canopy areas, then it’s difficult to claim than the forest cover is the main parameter explaining the observed differences.

. L231-232: Why not using directly the point clouds to compute normal (thickness) distances to detect changes directly with M3C2. It is more appropriate than doing it from the DoD (vertical distance), especially on steep slopes (see Bernard et al., 2021).

. L 323: The general convention for significance is with a p-value < 0.05. Can you please explain why do you consider these results significant with a p-value of 0.10?

. L 384: Is this underestimation compared to the BWF inventory? It is not clear. Also, can you say a few words why there is an underestimation?

. L 385: Is the false positive rate pixel-based or polygon-based?

Technical corrections

. Figure 1: There are too many subplots which make it difficult to read, especially for the DoD (11.c)). I suggest splitting the figure in two: 1.a) and 1.d) together and 1.b) and 1.c) in another figure (maybe located in section 3.2 or 3.3). Also, because the DoD is kind of a result by itself. Also, not being familiar with the area, I think the figure 1.a) does not give enough information on the location, maybe add main cities would help to understand the location.

. L 84: “focusses” to “focuses”

. L 121: “… 10 landslides are located within the areas with high tree density.”. What’s high tree density? Please give some quantification.

. L 127-128: “… using a seeded region growing algorithm”. At this point the reader is wondering how this algorithm works. I suggest referring to section 3.2 here.

. L 151: “… by thresholding the canopy cover at 30%”. I did not understand 30% of what? Please can you clarify?

. L254: “TPR” and “FPR” are used without being previously defined. Please correct.