1. Please add a reference for Table 2, and justify why the thresholds are reasonable and reliable in the study area. 

2. Figure 3. It is true that the NDSI and NDVI of vegetation, soil, and snow endmembers show much difference. However, I am curious about whether the approach works well for some vertically mixed pixels. For example, forest or shrub-covered snow represent a special mixture of spectral information. Please discuss the performance of the algorithm and products in these regions. Some studies have suggested that NDSI does not show many changes with or without snow under the canopy of boreal forest.

3. L212-214, L230-232. The authors indicated that the snow cover in the area change rapidly, and thus selected a small time window for the temporal interpolation (3 day). However, in the following step, they applied a 9-day window to remove the rest cloud pixels (PCHIP method). It does not make sense to me. Did this PCHIP approach brings too much errors?

4. L363-365. The authors argued that increase in the amount of station data and observations lead to better accuracy assessment. It does not make sense. Were the station data only used for verification of the FSC products? Were they used for the training of the algorithm? If not, I think more station data for verification would not increase the accuracy of the developed products. Declined cloud cover should have contributed to the better verification as MODIS has higher accuracy in cloud-free days. I am not sure whether changes of snow cover days also affected the accuracy indices. It seems the high OA and CK values after 2015 were mainly driving by higher PA index. It likely means the omission errors decreased. Was there less snow cover in the AWT area in these years?

5. L492-494. It is true that there are many cloud/snow confusion errors in MODIS data. Some researchers (e.g. Dong and Menzel, 2016, Journal of Hydrology; Remote Sensing of Enviroment) have conducted some research on this topic and developed some algorithm to remove overestimated (misclassified) snow pixels on MODIS snow maps using station data. It seems the proposed algorithm here did not consider this problem. It would be helpful to add some discussions about this.

Pan’s paper produced a daily cloud-free daily Cloud-gap-filled fractional snow cover products with 0.005°spatial resolution based on MODIS surface reflectance data in the Asia Water Tower (AWT) region. The core of the algorithm contains two parts, one part is the automatic endmember extraction (MESMA-AGE) technique for producing FSC, and the other part is the multistep spatiotemporal interpolation algorithm (MSTI) for filling gaps. However, before this article is accepted by ESSD, the following questions are still required.

General comments:

1.    The author conducted a comprehensive evaluation of the product using Landsat8 and sites, but comparison with existing products was lacking. It is recommended to increase comparison with other 500-meter fractional snow cover products. The author also mentioned that SNOW CCI data can be compared before going to the cloud to clarify the accuracy of the MESMA-AGE algorithm. In particular, the MOD10A1F product is global daily Cloud-gap-filled fractional snow cover dataset. This study does not mention or compare with it. 

2.    The selection of pure end-menmber is the key to the MESMA-AGE algorithm. What is the basis for selecting pure end-menmber in TABLE2? Qinghai-Tibet soil and rocks are quite different. Is the current selection very representative? Further explanation is needed.

3.    In 3.2 "Multistep spatiotemporal interpolation algorithm", the author's last two steps are "piecewise cubic hermite interpolating polynomial (PCHIP) for the 19-day period, and further spatial interpolation using a 10×10 window." However, the snow cover on the Tibetan Plateau changes rapidly and has strong spatial heterogeneity. The author also mentioned in "4.2.3 Performance of spatiotemporal reconstruction algorithm" that "The results show that the accuracy of the fractional snow cover based on clear sky observations is significantly better than that of the spatiotemporal reconstruction. "The 19-day time series and the 10 ×10 window” interpolation may not be suitable for the Tibetan Plateau, and the author needs to further consider the rationality of the Multistep spatiotemporal interpolation algorithm.

4.    In previous studies, FSC set thresholds greater than 10, 15... to consider snow, and site verification set thresholds greater than 0cm, 1cm, 2cm, 3, cm, 4cm, 5cm... both, this study Please explain further the reasons for using FSC>15 and sd>3cm

5.    It is recommended that the author add a schematic diagram of Asian Water Tower Region's products so that readers can more intuitively understand the algorithm effect of each step.

Minor comments:

1.    There are two RMSE indicators in Figure 2. It is recommended to modify one of them to MAE.

2.    Line 207 mentioned “further spatial interpolation using a 10×10 window”, and Line 241 mentioned “In this study, the observation information from the 11*11 interpolation window centered on the cloud pixel was used based on the inverse distance weight ", did the author use "10×10" or "11×11" window in the fourth step of the multistep spatiotemporal interpolation algorithm?