Overall Comment:

The paper presents an improved automated avalanche terrain classification model (AutoATES v2.0) which is a valuable contribution in the specialized field of describing and mapping the severity of avalanche terrain. Many novel and appropriate scientific techniques have been applied in the development and assessment of the model which are worthy of publication. However, I recommend revisions be completed before publication. I hope that this review helps to improve the overall quality of this excellent manuscript.

Specific Comments:

The writing in the paper suffers greatly from what can be described as expert familiarity. That is, the authors are obviously experts with and extremely familiar with the subject matter, which has led to a natural assumption that the readers have similar understanding. I recommend the authors have the manuscript reviewed by a non-expert to identify areas where the manuscript can be clarified.

The manuscript would benefit greatly from fixing the many comma splices found throughout. I have identified many in the technical corrections section, however I recommend a thorough review.

Lines 27-28: Confirm that the two significant digit accuracy is valid and useful, i.e. does the level of accuracy increase the reader's understanding? Further, the authors use both one and two significant digits in similar results. Please explain why varying significance is valid or be consistent across the manuscript.

Abstract: The abstract requires significant improvement. Overall, it assumes the reader is familiar with the Avalanche Terrain Exposure Scale (ATES) and AutoATES v1.0, and thus does not adequately introduce the overall topic of rating the severity of avalanche terrain. Suggest that the distinction between linear and spatial ATES ratings be introduced. Suggest introducing AutoATES v1.0. I'd also recommend expanding on the improvements from AutoATES v1.0 by explaining the forest data improvements to the PRA algorithm, the use of Flow-Py, and the post forest classification.

Lines 41 and 42: Assessing avalanche risk indeed is a complicated task for backcountry recreationalists. Lines 41 and 42 imply that spatial and temporal variability are the reasons for the complexity, however these factors certainly contribute to the complexity, but are not the sole reasons. Suggest changing the wording of this sentence.

Lines 43-44: The sentence when interpreted with the preceding two implies that backcountry recreationalists make detailed assessments of instabilities at a regional scale (i.e. > 10⁴ km²). Backcountry travellers typically travel on the mountain or drainage scale (i.e. > 10² km²), therefore they are unlikely to make detailed assessments at the larger regional scale. Regional scale assessments are often done by public forecasters. Further, avalanche risk when travelling in avalanche terrain is managed by more than just assessments of weather, snowpack, and signs of instability. Other risk management techniques include terrain selection, travel techniques to reduce exposure to hazard (e.g. not stopping on an avalanche slope), the use of airbags, and the use of rescue equipment (e.g. transceivers, probes, shovels). Suggest revising the sentence.

Lines 37 and 38: Please explain how the 90% triggering statistic is relevant to the topic.

The introduction needs to describe the distinction between the original use of ATES to rate linear routes (i.e. backcountry trips) and the development of spatial ATES ratings. It is critical for readers to understand that ATES was developed initially to describe the severity of avalanche terrain one would be exposed to when travelling on linear routes through mountain terrain. The discussion and/or the introduction sections would benefit from a discussion of the implications of using ATES spatially which is a deviation from its original intention.

Lines 212 and 213: Should this sentence refer to "stems per hectare" rather than "stem density per hectare"? Stems per hectare is commonly used to describe forests. If stem density per hectare is correct, please define the unit. Also, it is not possible to have infinite stems / hectare as this implies infinitely small stem diameter.

Suggest referring to the new 5-class ATES rating system recently published by Statham and Campbell (2023) as "ATES v2.0". Then refer to the AutoATES algorithms as AutoATES v1.0 and AutoATES v2.0 throughout. For example, the clarity of the sentence on Line 401 could be improved if v2.0 was replaced with AutoATES v2.0.

The study evaluates the output from the AutoATES algorithms against two benchmark maps: 1) Bow Summit, and 2) Connaught Creek. An improvement to this study and future studies would be to expand the amount of terrain in the benchmarks to allow a more robust evaluation of the output. For example, avalanche terrain in the coastal mountain range differs in character to the interior and Rocky mountain ranges in Canada, and it would be useful to see how the model performs in the coast. Also, terrain in different countries would be valuable to evaluate (e.g. the Alps, Japan, Sierras). Please speculate on how the model would perform in different snow and avalanche climates.

It does not appear as though the authors include Class 0 - Non Avalanche terrain as an output of the AutoATES v2.0 algorithm. Please provide a robust explanation of why this was excluded and the Class 4 - Extreme terrain was included. Class 0 - Non Avalanche terrain is a critically important classification. I strongly suggest the authors revise the algorithm and manuscript to include this classification level which would increase the value of the scientific and practical contribution.

Lines 89 to 90: The term zoning is often used for avalanche hazard zoning where the zones specify requirements for land use (e.g. building not permitted in red zone). Suggest using “avalanche terrain classification” here. Also, I’d suggest making very clear distinctions between hazard zoning or mapping and terrain classification / mapping. Avalanche terrain is classified mostly independent of hazard and is much different than hazard zoning. Consider removing any reference to hazard zoning from the manuscript for clarity.

Line 156: Recommend ending the introduction with the proposed solution that solves the identified knowledge gaps and improves on the identified shortcomings of AutoATES v1.0.

Line 158: Section 2. Model Motivation appears to include descriptions of the model methodology and development. Consider a more appropriate title for this section - perhaps “2. Model Methodology” or “2. Model Development”.

Lines 195 to 204: Forest canopy also impedes wind transport of snow reducing the formation of wind slabs. Suggest adding this to the description of the effect of forest canopy on avalanche formation.

Line 307: Please clarify if the PRA is calculated from the DEM only or is forest density considered as well.

Technical Corrections:

Many technical corrections were found. I recommend the authors thoroughly review subsequent versions of the manuscript prior to submission.

Lines 22-24 - clarify the sentence. It is not clear what is meant by "open-source software" in relation to addressing the significant drawbacks of AutoATES v1.0. A comma in Line 24 after "overhead exposure" would help.

Line 30: Perhaps a missing word at the end of the sentence? Should this read "decision-making".

Lines 24 and 31: Suggest defining ATES before using the acronym or not using the acronym in the abstract at all and writing out the full “Avalanche Terrain Exposure Scale”.

Line 42: Suggest replacing "snow" with "avalanche hazard".

Line 52: ATES is an ordinal scale. Ordinal scales are not typically described as quantitative measurements. Replace "quantifies" with something like "classifies".

Line 44: Delete ",i.e.,"

Line 48: replace the "," before "however" with a semicolon, OR delete the "," after "however".

Line 50: ATES communicates the severity of avalanche terrain and not the risk. Avalanche risk is often described as a function of expected avalanche size, likelihood, exposure, and vulnerability; therefore I'd suggest deleting "risks" from this sentence. Parks Canada originally used the system to describe the exposure to avalanche terrain one would experience on a particular backcountry trip.

Line 51 and 52: Appears to be a missing word between scheme and worldwide. Consider: “ATES is a terrain classification system commonly used worldwide to describe avalanche terrain using easy-to-understand ratings: Simple (1), Challenging (2), and Complex (3). Also, suggest capitalizing Simple (1), Challenging (2), Complex (3), and Extreme (4) here and elsewhere for clarity that these are output variables from the model.

Lines 55 to 57: Rewrite sentence to clarify that the new ATES system is called ATES v2.0. Consider: “Recently, the ATES classification system has been updated to include two new additional ratings: 1) Non-avalanche terrain (0), and 2) Extreme (4). The updated system complements the existing system and is now referred to as ATES v2.0 (Statham and Campbell, 2023).” Suggest using consistent ATES terms with capital first letter and paratheses containing the level.

Lines 59 to 70: The connection between avalanche hazard mapping (i.e. magnitude x return period) and terrain classification is vague. Consider removing the hazard mapping description unless the authors feel it adds value to the topic of avalanche terrain classification.

Line 61: Suggest replacing "i.e.," with "e.g.,". The authors are providing an example and not further describing the definition of specific return periods. That is, hazard mapping is often done for other return periods depending on the context.

Line 92: Consider replacing “, i.e., terrain traps and forest density,” with (e.g. terrain traps, forest density). These are examples of other factors but not a comprehensive list.

Line 95: Comma splice between width and accounting.

Line 112: Add comma after “et al.”. Here and throughout.

Line 113: Missing parentheses around “2016”.

Line 123: Replace “is” with “are”. Common error throughout the manuscript.

Line 133: Remove the “s” in the word “defines”.

Line 162: Comma splice after the Sykes reference. Add a conjunction such as "which limits".

Line 163: Comma splices. Consider: “For AutoATES v2.0 to be a viable option for largescale ATES classification, the model performance should be at least as accurate as manual mapping.”

Line 169: Suggest re-writing this sentence. It is not clear what is meant by "integrating the parameter into track and deposition area".

Lines 166 to 175: Regarding the model description. Doesn’t the model also include post-forest-classification? Consider clearly describing the main model additions with a description of each. This will help add clarity to the model assessment and discussion sections.

Line 181: Delete "Sykes" from the reference.

Lines 191 and 192: Confirm that the algorithm uses all of these forest data types at once OR if the algorithm uses only one of those forest data types. If the algorithm uses one of the forest types, then replace the word "and" in line 192 with "or". Same comment for Line 194 in the parameter list.

Line 225: replace "is" with "are".

Line 237: add comma after "et al.". Here and elsewhere in the manuscript.

Line 239: delete the "s" on the word "values".

Line 241: Suggest using semicolons between the main list items: "where u(x) is the Cauchy membership value; x is an input variable; and a, b, and c are parameters ...".

Line 255: Comma splice. Re-word the final sentence in the caption.

Line 256: Replace "is" with "are".

Lines 363: Suggest referring to the new ATES system as “ATES v2.0 (Statham and Campbell, 2023)”. Here and throughout for clarity.

Line 370: Figure 3 – the inset labels are ineligible which makes placing the study areas difficult for readers who are unfamiliar with the locations.

Line 381: The reference should be (Statham and Campbell, 2023) because there are only two authors. Check all references and citations.

Lines 388 to 392: Use a consistent level of significant digits for the model accuracy values OR explain why different levels of accuracy are relevant.

Table 4: The headings for SIMPLE (1) and CHALLENGING (2) look to be misaligned.

Line 400: Replace "is" with "are". Please check the entire manuscript as this error repeats.

Line 408: Should the reference be “(Statham and Campbell, 2023)”.

Line 413: Avoid words like "dramatic" when describing scientific findings.

Line 445: Define AAT3.

Table 5: Suggest defining SAT34 and AAT3 in the table caption so that the table can stand alone from the text.

Lines 456 and 457: Comma splice. Suggest re-writing this sentence with connectors. For example: Initial attempts by Larsen et al., (2020) compared AutoATES v1.0 to available linear and spatial ATES ratings in Norway, however the validity of these layers was uncertain because they were developed over multiple years by numerous experts with limited review.

Line 459: delete the commas after (2023) and deficiencies.

Lines 460 - 463: Re-write the sentence please because it runs on. Suggest: "Their approach - which used three experts to map each study area and then create benchmark maps based on their individual output - is a more comprehensive methodology to address this issue."

Line 466: Does the term consensus matrices refer to the confusion matrices? If so, please adjust.

Line 471: Suggest replacing "human mappers" with something more suitable like "terrain rating experts" or "avalanche experts". If agree, do so here and elsewhere. The change will improve the sentence by reducing the repetitive use of "mappers" and "maps".

Line 475 and 476: suggest replacing the ", like Western Canada" with "(e.g. Western Canada)".

Line 484: Change "investigate" to "investigated". Also, fix the comma splice after "model".

Line 490: Change "measure" to "measured".

Line 510: Does the model identify other common terrain traps such as cliffs, crevasses, forest? If not, please provide rationale and speculate on the effect on model performance if all common terrain traps were included.

Line 532: Missing parentheses.

Line 565: VRI is not defined.

Line 575: Re-write sentence. Consider replacing "reevaluate" with "re-evaluation".

Lines 577 to 579: Re-write sentence to reduce the comma splices and increase clarity.

Lines 635 to 640: Write out the abbreviated names for clarity.

This paper details the latest advances to the spatial ATES mapping approach, called AutoATES v2.0, to improve an automated terrain classification system for snow avalanche hazards at a regional scale. The authors use two study sites to demonstrate the improvements to the earlier version of AutoATES as well as a reference classification based on manual mapping by experts.

The paper topic is of interest to the natural hazards community and specifically the snow avalanche community. It is generally well-written, however the structure of the paper needs to be improved to aid in interpretation, and there is important context missing from sections. The paper is written as a companion article, but this is not made clear enough and probably explains some lack of broader context/background on ATES and spatial ATES. A review from a non-expert would help ensure broader interpretation of the research.

Some significant revision is required before the paper will be ready for publication. Please find general comments below followed by specific comments/technical points.

## General comments

The paper is generally well-written, but the structure of the paper needs significant re-organisation and improvement. Sections in the discussion repeat what is written earlier and, in some cases, provides more detailed information that would aid the reader had it been introduced earlier. For example, the benchmark sites are referenced early but not properly introduced until the Discussion.

The abstract needs to be reformulated to bring better context to the paper. It currently reads more like a description of the algorithm, but as this is a general hazards journal, it should be written so someone can get all the context needed to understand the contents of the paper before diving in. In the first instance, please provide a (brief) introduction of ATES and AutoATES. I believe you intend ‘large-scale mapping’ to refer to regional or landscape (10 or 100s km²) mapping products, but the term actually conveys a map covering a small geographic extent. ‘Regional-scale mapping’ may be more appropriate.

The writing feels like the reader should already be an expert on all aspects of ATES and AutoATES development. You do not need to repeat everything covered in the companion article, but context is missing throughout. For example, the forest data layers are a big part of the focus of this paper. There is no detail on what dataset was used, how any uncertainty on the rasterization of those data was accounted for in their use in the PRA tool or the classifier, etc, how users without your forest data product should approach the use of other forest data.

Overall the paper feels like it was written for an audience of expert ATES users who might be trying to run AutoATES v2.0. This should be broadened for a general audience of the journal NHESS. Some sections (like Application) are important for broadening the reach of the paper but currently feel like they were dropped in without connecting to the rest of the paper. For example the statement in Lines 588-589 is a good idea but more specificity is needed on how a user would fine-tuning. What new forest density data are you referring to?

The goal of this work is to create a product that is more standardized and globally applicable than what can be created ad-hoc by experts, but the product is tested against a reference from experts interpreting largely the same data (but with field observations and oblique images etc as well). This approach is justified but take care in how this is expressed. Should the AutoATES replace the expert ATES mapper or is it meant to be a tool for the expert to use when doing ATES mapping in a new location? Or something else? This should be made clear.

One main audience for AutoATES is those wishing to classify terrain in data sparse regions. The paper should make clearer how to use AutoATES v2.0 in these data sparse regions. What global forest product is recommended? How do we handle the uncertainty that is created when resampling these data to match the resolution of the DEM used?

There is a general tension in the paper between fine-scale terrain features and broad-scale terrain classification. Recommendations are made for fine-tuning to local conditions (adjusting DEM resolution, forest data, the use of PRA, etc) but this suggest the AutoATES v2 is sensitive to scale and terrain shape. A section that details (or speculates) on how the performance changes in different kinds of terrain and forested landscapes would aid the application to other regions.

The forest density approach is an intuitive improvement over binary forest presence/absence, but there is no demonstration of how these density products (and the thresholds used) transfer to the avalanche hazard. Please refer to other research/literature that helped inform the values used in thresholds. There are several density measures used here. Help the reader to understand your preferred option (pcc, stems etc). Why include all three? These answers may be straight-forward, but they should be made explicit to the reader.

Relatedly, please clarify why a binary forest product is not an option since this will be a more widely accessible global product over forest density and the density values are collapsed into three encoding classes.

Please specify what forest product you used (only an acronym is provided). From the companion paper I can see it was created from a vector polygon dataset. What was spatial scale used to create the vector dataset and is rasterizing to 10m appropriate based on the scale or is there additional uncertainty generated? The best global forest density product I am aware of is 30m. This uncertainty may be incidental but the process should be documented.

Scale matters for the DEM as well. While a DEM sensitivity test showed little change to the final ATES mapping (Sykes et al., 2023), this seems to be mostly related to increasing the slope window size to only look at broad landscape-scale features rather than fine-scale features. Flow-py modelling will be sensitive to DEM resolution and whether a DTM or DSM is used. If one advantage of Flow-py is the ability to capture terrain traps and confined terrain then a higher-resolution DEM is needed. Can you clarify if these features are captured in the final ATES class, i.e. do you see terrain with fine-scale features like terrain traps reclassed as a result of using Flow-py? Since ATES mapping is more about broad-scale terrain I suggest reviewing the discussion on fine-scale terrain.

The distinction made between various simulation approaches (Lines 120-140) should also clarify Flow-py is modelling the dense core of a mass movement not a powder cloud, as some (e.g. RAMMS-Extended) do, which would obviously affect the total runout delineation.

The justification for removing roughness from the PRA calculation is not compelling enough (4.1.3). Including parameters that asses terrain shape over several spatial scales (3x3 window vs 10x10 etc) can help distinguish between steep features surrounded by other steep terrain and steep terrain that is surrounded by flat terrain (e.g. in a runout zone).

The Dev6 Post-forest-classification is not clear enough. Please explain Lines 550-551 where sections had to be reclassified. How should other users approach this manual intervention?

With regards to ATES comparison (Section 3), could you provide an uncertainty estimate from the area mis-classified as percent of total area mapped?

“Avalanche zoning” (used several times) is usually applied to engineering situations. “Avalanche terrain classification” is clearer.

## Specific comments

Lines 17-18: This note seems out of place here. Check with style guide, but I would suggest introducing the interchangeable use of ‘model’ and ‘algorithm’ in the introduction. Ideally you would only use one term in the paper if they are supposed to mean the same thing.

Line 25: These sites will be unknown to many – generalize e.g., ‘…two study sites in W Canada’

Line 27: These values are very precise for an abstract.

Line 28: “Large-scale mapping” refers to a small geographic area. I think you mean a regional-scale mapping approach.

Introduction

Lines 42-43: Sentence “This results in…” needs to be re-phrased/placed in more context.

Lines 110-111: Explain why Veitinger PRA output is 0-1 (based on fuzzy membership approach). And explain the terrain parameters used in these two approaches.

Lines 120-121: Cite some examples.

Line 152: briefly explain what “intensity of avalanches” means here, as this is not self-evident nor the same as impact pressure, for example, that comes from the depth-averaged modelling approach.

Section 2

Lines 198-200: Provide citation for this statement.

Line 209-210: Provide citation for this statement.

Line 213: Stem density is stems per ha?

Lines 245-246: More justification, clarification is needed on scale issues with roughness.

Lines 247-248: Please explain these new membership functions.

Lines 248-29: Please explain what you mean by the values could be fine-tuned. Did you choose not to fine-tune them, or do you mean they could be fine-tuned based on local conditions in other regions? This is explained better in the caption to Figure 2 than in the text.

Figure 2: Please explain the a,b,c legend in the caption.

Line 292: Previous model here referring to AutoATES v1?

Table 1: are the Range values for Overhead exposure in number of cells or area? Please specify and use area rather than cells on account of varying spatial resolution of DEMs.

Table 2: Provide units for basal area.

Table 2 Caption: Please clarify what you mean by encoding is identical for all three forest density types when we the encoding values are different as presented in the table under Encoding column.

Table 3: I don’t see the red coloured text is needed here. If you really like it, make note of it in the caption.

Lines 337-338: This sentence needs a citation at minimum. How is accuracy determined here—based on the reference expert map?

Results and validation

Lines 363-364: More detail is needed on the study sites used as refence/benchmark. If this is covered in another paper, please refer reader for more detail.

Figure 3: state this figure was adapted from Sykes et al., 2023. The inset maps are too small to be useful. Ensure updated manuscript has higher-resolution image. What are data sources (e.g. forest, DEM) and image credits?  

Figure 4: Ensure high-res image is included in the revised manuscript.

Lines 384-385: Please specify how you prepared the ATES benchmark. Was this already in a 10m raster format? Was your test conducted with a 10m DEM? Was is the same DEM used / available for the ATES benchmark? This may be better placed in Section 3.

Table 4: Adjust column headers to fit with accuracy scores.

Figure 5 and Lines 398-404: I am not sure how useful this section is given that AutoATES v1 does not have the Extreme category. Unless it was compared to a previous version of ATES with only three categories?  Please clarify.

Discussion

Lines 466-469. Consensus matrices have not been introduced yet and this sentence needs re-phrasing.

Section 4.1.3 heading does not match discussion of roughness and should be separated from forest density

Line 565: VRI introduced for the first time here. It should be introduced earlier.

Line 575: rephrase ‘reevaluate’