This manuscript first recalibrated the coefficients of generalized split window algorithm based on a comprehensive atmospheric profile database, then validated the accuracy of LST estimations based on in situ measurements, and finally conduct some analyses about LST trends over the Pan-Arctic region.

In my opinion, this manuscript is well organized and conducts some meaningful analyses over the Pan-Arctic region, demonstrating the potential for acceptance. However, I have some comments and problems, which can be found as follows:

1. In [Page 7, Table1], the unit of latitude, longitude, and elevation in the table head should be indicated.
2. The 2m-air temperature is derived from the ERA5-Land product, but skin temperature and total column water vapor are derived from MERRA-2. I would like to know why different reanalysis data are used, as ERA5 seems to provide these parameters as well, with a higher spatial resolution.
3. The accuracy validation of LST estimations is conducted over SURFRAD and KIT sites, while the application and analysis mainly focus on the Pan-Arctic region. If there are available sites in the Arctic region to provide a more intuitive validation?
4. In [Page 8], the authors mention that “The Copernicus digital elevation model (DEM) GLO-90 upscaled to 0.05° spatial resolution is used for the RT modelling”. The geopotential height has been included in the used atmospheric profile dataset which can be used to calculate elevation, why still using additional data sources?
5. In [Page 12], the authors mention that “Pixels that have a cloud fraction higher than 0.1 are removed, and the average of the remaining pixels is computed”. As far as I am concerned, in order to eliminate potential cloud contamination, LST averaging should be performed only when all pixels within the window are clear.
6. In [Page 13, line 265], the explanations of symbols in equation (4) to (6) are missed.
7. In [Page 14, line 295], only nighttime observations from EVO site are used to bypass the directional effects. While for several SURFRAD sites covered by vegetation, possible directional effects may also exist, but why both daytime and nighttime data are used?
8. It seems that only the overall performance is shown in Figure 4. It is recommended to add accuracy metrics for daytime and nighttime, respectively, to provide a better comparison.
9. The trend analysis is conducted based on monthly averaged LST of each pixel (There are fewer descriptions about this in the manuscript, thus I guess maybe all clear-sky daytime and nighttime observations of the pixel are used to calculate). However, the averaged LSTs may be seriously affected by the frequency of cloud cover. For example, for pixel A, daytime observations account for 50% proportion, whereas in pixel B, they constitute 70%. Since daytime LST tend to be larger than nighttime LST, the averaged LST of pixel B tend to be larger than A, but that may not be the true situation. Therefore, the cloud cover may lead to the incomparability between pixels. Even for the same pixel, changes in cloud cover frequency between different months may also result in temporal incomparability. Therefore, is it possible to reduce these incomparable effects, such as ensuring a balanced distribution between day and night LSTs to calculate averaged LST? Besides, this restriction should be briefly explained in the discussion section.

This paper describes a long dataset of LST covering the Artic based on the AVHRR series. I believe there are several aspects that need to be address before publication:

1. It is not completely clear what is the added value of having this specific dataset derived for the artic region if the authors are simply averaging all observations within a day. There are already datasets available based on AVHRR that provide daily composites (e.g. GLASS, LSA-SAF). In my opinion, it would have been more beneficial to explore the multiple passages of the different AVHRR to try to reconstruct the diurnal cycle. That would have made the dataset more unique and more useful. Having averages of whatever observations exist in a day can create high instabilities in day-to-day variability, depending on what sensors are available and cloud coverage.
2. For the same reason, I’m not convinced the dataset is appropriate for trend, and specially not for anomaly analysis. If the time of observation that goes into the average keeps changing, then there is just too much instability in the series.
3. Also, in terms of algorithm calibration, here there was a unique opportunity to explore an algorithm more suited for the specific conditions of the Artic. That maybe would allow using a higher range of view angles, resulting in an even larger sampling of observations through the day. The same in terms of the calibration database, why not tailor the database to the more specific conditions of the Artic? Using a generic algorithm and database that are valid over the whole globe is something that is already available in other products.
4. With respect to the LST validation, the authors only used KIT and SURFRAD stations. None of the stations is within the study area and therefore are not representative of the presented LST dataset. This is very clear when looking at figures 4 and 10. These stations’ LSTs lowest values are around 260K, while most of the Artic is well bellow this value. There is a very with range of surface flux stations within the considered area (AmeriFlux, Fluxnet, BSRN) or even in Antarctica, which has much more similar conditions. The authors should have tried to use more stations that encompass the specifics of the polar climate. It’s true that these stations tend to be more heterogeneous, but the SURFRAD stations are also very heterogeneous.
5. There is a long discussion on whether the problems in stability seen when comparing Tair with T2M and LST being related to day/night problems. It’s not clear to me why the authors did not separate daytime from nighttime observations. This would make comparing with Tair_max and Tair_min more easy to interpret. For T2M, it’s not clear from the text but it seems the authors are averaging all hours of the day? The ERA5-land provides a seamless diurnal cycle with hourly frequency, why not compute the daily max and min to obtain variables comparable to Tair?
6. Why do you use ERA5 in some cases and MERRA-2 in other? ERA5 has better spatial and temporal resolution.
7. Is/will this dataset be made available publicly? What is the format? What is the projection? More technical details about the dataset are needed.