General comments:

This study presents an annual glacial lake dataset spanning 2017 to 2024, mapping lake extent and quantifying lake area change across an 8-year study period using imagery from both Sentinel-1 and Sentinel-2. Image processing was conducted in Google Earth Engine, with Sentinel-1 annual median mosaics derived from July and August acquisitions and Sentinel-2 composites covering May to November with a cloud cover threshold of 60% at 10 m resolution. Model training employed the DeepLabV3 architecture within ArcGIS, with a training dataset constructed from manual digitizations for the year 2020 from previously published datasets. Post-processing is applied to remove unnatural water bodies based on a digital elevation model. Validation was performed against both manually digitized lakes and published datasets. Sentinel-2 derived results showed higher lake counts compared to Sentinel-1 derived lakes. Hence, the lake area detected based on optical imagery was higher than from SAR imagery. Both sensors performed better for larger lake sizes with smaller errors. The authors' planned glacial lake observatory, which aims to monitor lakes on an annual basis, is a commendable and timely initiative that has the potential to make a valuable long-term contribution to the glaciological community.

The manuscript is well written, clear, and methodologically detailed, and provides useful analyses of lake number, lake area time series, and elevation distribution. Nonetheless, several aspects of the manuscript would benefit from further clarification and discussion, particularly a more thorough examination of dataset limitations and the underlying reasons for the observed differences between lakes mapped from optical and SAR data.

Specific comments:

Figure 1: Consider putting the coordinate system only around the map and the plots outside the map frame.

Table 1: Consider to add your own dataset into this table and highlight what makes your dataset novel/different from existing ones.

L126ff: Outline more clearly (or create a map) how the training, evaluation and manually digitized lake data is spatially distributed. Not being familiar with the GTN-G regions it is hard to understand where the data is located and if the training and evaluation data is spatially overlapping in any kind.

L133 & L140: Annual mosaics for Sentinel-1 (Jul-Aug) and Sentinel-2 (May-Nov) are created over different time periods. Are there any intra-annual/seasonal variations in the lake area that could be mapped differently due to this difference in time period? Maybe something difficult to quantify but worth discussing in a limitations section.

L153ff: Justify in more depth why you chose DeepLabV3 as deep learning model. E.g. did other studies come to the conclusion that this model works best? Maybe give the reader a bit more in-depth information about DL methods that have been used for lake mapping and how well they performed (e.g. by extending from line 57 onwards on the DL methods).

L165: How did you decide on training for 50 epochs? Did you use the model weights after 50 epochs or after early stopping based on validation loss? It would be very helpful to have accuracy and loss curves of the training in the appendix to see how the model converged and whether it over/underfitted. Usually, it is recommended to use the model with the lowest validation loss instead of a model trained after a fixed number of epochs.

L170: At this point it is difficult for the reader to understand how the power-law model was derived without having seen the numbers of the manual validation dataset in the later section. Consider re-structuring sections or give a bit more in-depth information already from L170 onwards to clarify.

L185: What is the acquisition date of the DEM you used? Is it possible that anomalously elevation deviations you are masking out could have occurred due to glacier retreat that occurred after the DEM was acquired? Maybe also something for a discussion in the limitations section.

L182: You use the RGI (which version?) dataset to distinguish between glacier fed and non-glacier fed lakes. The RGI provides glacier outlines for the year 2000. Discuss how this influences your categorization and which uncertainties arise from this.

L187: It would be helpful to either provide a visual example about the errors in overlapping regions or describe in more detail what kind of errors there are and how you remove them.

L215: Intersection over Union (IoU) could be an interesting additional evaluation metric to see how well the different lake datasets geometrically overlap.

L237ff; L339ff & L499ff: Discuss in more detail why lakes have been mapped differently by Sentinel-1 and Sentinel-2. Why are accuracies lower for Sentinel-1? Why remained lake numbers more stable based on Sentinel-1 compared to Sentinel-2? A discussion including sensor specifics would probably be very helpful so the user of the dataset knows better what kind of lakes are mapped best by which sensor. E.g. do you have visual examples of lakes that were mapped in Sentinel-2 imagery but not Sentinel-1 imagery? How do these lakes look like in both images and why wasn’t it possible to map them from Sentinel-1 imagery? What is the recommendation for the user regarding the S1/S2 datasets? Use both in combination or prefer one over the other in specific cases?

L440: A limitations section in the discussion would be very helpful so the user knows what the dataset can/cannot provide.

This study presents a new annual glacial lake dataset for Nepal and adjacent transboundary catchments covering the period 2017–2024. The authors use multi-sensor satellite imagery from Sentinel-1 and Sentinel-2 to map lake extent and quantify changes in lake area over time. Image processing was conducted in Google Earth Engine, and lake detection was performed using a DeepLabV3 deep learning model trained on previously published manual inventories. The dataset is validated against manually digitized lakes and existing inventories, showing generally strong performance, particularly for larger lakes. 

I really enjoyed reading this paper. I think it is a strong contribution and will be very useful for the community. The manuscript is well written, clearly structured, and the methods are explained clearly.

That said, I believe there are a few aspects that should be clarified so that users of the dataset can better understand the methodological limits and avoid potential misinterpretations.

1 – it would be helpful to be more explicit about detection limits and methodological biases. While the nominal threshold is 0.001 km², the validation shows that small lakes have much higher relative uncertainty. It would strengthen the paper to clearly state what lake sizes can be reliably monitored for change analysis, and where uncertainty becomes too large for robust interpretation. Similarly, lake surface ice and partial freezing are mentioned as sources of error, but it would help to specify when and where this is most problematic (e.g., elevation bands, specific basins, certain months). Without this, users might overinterpret trends in small or high-elevation lakes. A short paragraph summarizing recommended use and limitations would be very useful.

2- the different compositing windows for Sentinel-1 (July–August) and Sentinel-2 (May–November) introduce a potential sampling bias. Even if justified by cloud cover and image availability, the two sensors are not sampling the same seasonal window, which could influence median lake extent and comparability between sensors. A short discussion of how much this might affect inter-sensor comparisons would increase confidence in the results.

3- Sentinel-1 in mountainous terrain is affected by layover, foreshortening, and radar shadow, which can strongly influence backscatter and geometric representation in steep valleys. It would be helpful to clarify whether SAR visibility constraints were explicitly accounted for, for example by masking radar shadow/layover areas or using terrain-flattened products. In landslide studies, we have seen that training SAR-based models using inventories digitised from optical imagery can degrade performance if SAR visibility is not considered. There are approaches to compute SAR visibility masks in GEE (e.g., https://doi.org/10.3390/rs12111867). This might not be essential here, but acknowledging this limitation and potentially considering it in future refinements could strengthen transparency.

Related to this, since the training data are derived from optical inventories, it would be useful to briefly comment on potential cross-sensor discrepancies in geometry or boundary definition and whether these could influence Sentinel-1 training quality.

4- The reliance on ArcGIS Pro for model training and deployment may pose practical limitations for some users, particularly given the strong open-data framing of the study. Since a Python file and trained model weights are provided, it would be helpful to clarify how portable these resources are outside the ArcGIS ecosystem. For example, can the trained DeepLabV3 model be readily deployed in standard open-source environments (e.g., PyTorch/TensorFlow workflows), or does it require ArcGIS-specific dependencies?

5 – personal curiosity. Can you comment if you had any false positives and in which occasions? Deploying over such large areas is quite challenging.

Please find very few more comments in the attached PDF.