This paper compiled a comprehensive inventory of rock glaciers across the entire Tibetan plateau except Himalaya and Hindu Kush. The inventory of rock glaciers is of significant importance for the study of periglacial landforms. However, the claim of the article is to compile an inventory of rock glaciers across the entire Qinghai-Tibet Plateau, yet it lacks a systematic inventory of the Himalayas and the Hindu Kush, where there are numerous rock glaciers and the most extensive development of such landforms. The article mentions data limitations as the reason for not completing inventories in these two regions. However, it is feasible to achieve inventory in these regions through visual interpretation using integrated data from ESRI, Bing, Google Earth, etc. Moreover, some researchers have already achieved inventory in the Himalayas (Jones). Therefore, the main issue with this inventory is its lack of completeness. If the article is to be published, the first step should be to complete the inventories in these two regions. Other technical issues are as follows:

1. To include contextual information around rock glaciers, a buffer zone of 1500 meters was set. How the size of this buffer zone is chosen and its impact on model performance?
2. Rock glaciers are minimally represented in imagery, which theoretically poses a severe issue of data imbalance and could result in numerous false positives. This concern was also indicated by validation results. However, upon examining the prediction outcomes, there weren't too many false positives among the 48,767 candidate polygons and 44,273 rock glaciers. How was this issue addressed?
3. There are significant differences in F1 scores among different sub-regions, and although some regions with fewer rock glaciers also have large areas, how does this result support the earlier conclusion about the model having good generalization ability?
4. In lines 75-77, the authors summarized the inventory of rock glaciers in the Qinghai-Tibet Plateau region, which contained a large amount of available inventory data (>20,000 records). However, the authors ultimately selected only a small portion of these as training data (<2,000 records). Why was this the case? Additionally, does the training data (4,085 records) mostly consist of rock glaciers from the Alps region? Is there evidence to suggest that rock glaciers in the Alps region share similarities with those in the Qinghai-Tibet Plateau?
5. In lines 107-109, please explain the unique characteristics of the Hindu Kush-Himalayas compared to other neighboring mountain ranges, as well as the differences between rock glaciers in this region and those in other areas.
6. In line 370, please verify the data on rock glaciers within the Qinghai-Tibet Plateau. As far as I know, multiple field expeditions have failed to find rock glaciers near the source of the Yangtze River.
7. In line 80, the authors emphasized the inconvenience of compiling a rock glacier inventory through visual interpretation, as it requires strong geomorphological expertise and is labor-intensive and time-consuming. However, in the end, all data were still inspected and modified through visual interpretation, making it inevitable.
8. Typically, we consider active rock glaciers as evidence of permafrost existence. However, in the southeastern part of the Qinghai-Tibet Plateau, a large number of rock glaciers seem to be outside the permafrost zone. Please explain the reliability of identifying rock glaciers in these areas.
9. In Figure S4, why did this study not separate multiple rock glaciers based on other datasets?
10. The latest research progress in the Qilian Mountains region is missing. Please cite: Hu, Z., Yan, D., Feng, M., Xu, J., Liang, S., & Sheng, Y. (2024). Enhancing mountainous permafrost mapping by leveraging a rock glacier inventory in northeastern Tibetan Plateau. International Journal of Digital Earth, 17(1), 2304077.

Review of “TPRoGI: a comprehensive rock glacier inventory for the Tibetan Plateau using deep learning” by Sun et al., (2024)

General comments

This article describes a comprehensive inventory of rock glaciers on the Tibetan Plateau using deep learning techniques. The authors also used the recent baseline and guidelines developed by the IPA Action Group Rock Glacier Inventories and Kinematics (RGIK). However, they primarily employed the geomorphological approach outlined in these guidelines, which requires strong expertise from the mappers.

The authors utilized a large volume of optical images, mainly from Planet images with a resolution of  4.7 m. The deep learning algorithm processed three bands of these Planet images, which surprisingly produced acceptable rock glacier outlines without considering other components (e.g., slope, aspect, solar radiation, surface roughness) in the model or the movement of areas (displacement). The kinematic approach, also proposed by RGIK, was not considered in delineating the rock glacier areas.

To validate the deep learning outputs, existing rock glacier inventories were used, which are assumed to be of high quality. The manuscript presents very interesting findings; however, this study needs some clarification. Therefore, I recommend that this paper undergo major revisions.

Specific comments

1. It seems that the deep learning algorithm is not the most efficient since it cannot support multiple datasets. What is the reason for choosing this method? What is the advantage of using it?
2. The authors claim to have used the RGIK baselines. They mainly utilized the geomorphological approach, but it is unclear how this approach was applied to the entire inventory of 44,273 rock glaciers. Please provide more information in the text. It would be beneficial to incorporate the kinematic approach (e.g., InSAR) in the near future, as both approaches are complementary. Reinosch et al. (2021) utilized InSAR time series to generate a rock glacier inventory for the western Nyainqêntanglha Range. It would be great if you could conduct a thorough comparison with this study and others that used InSAR data (kinematic approach). This could help to understand if your results are comparable with those using other data and techniques.
3. The authors used data from only one year (2021). To me, one year is not enough to characterize such a large region. Moreover, some mass movements can be erroneously mapped, especially in areas with poor previous rock glacier inventories, making comparisons difficult. Perhaps a comparison with RGI V6 (despite its limitations) could shed some light. While including another year of data may require substantial effort, it could help remove potential discrepancies
4. Overall, no details are mentioned about the uncertainty analysis, what is the uncertainty or error estimation of the rock glacier inventory? Is it +/- 10% or >20% of the total area? Is it possible to quantify this using your methodology?
5. It would be beneficial to also include other areas (Himalaya and Hindu Kush) in their study.
6. How did you manage the cloud cover? You mentioned that in some regions, such as Nyainqêntanglha, you found that problem

185-> For the model training 70% and 30%, any specific reason for these values?

186-> You chose a 1,500 m buffer. Is there any technical reason for this choice?

405-> “However, rock glaciers are frequently found in regions characterized by poor image quality due to factors associated with cloud cover, shadows, and distortions, which are common in mountainous areas”. This is why it would be useful to incorporate more than one year of data into your dataset