This manuscript proposes a long-term (1981-2005) AVHRR land surface temperature (LST) dataset that includes outcomes at both daytime and nighttime. The algorithm is the generalized split-window (GSW) algorithm while in the production, this dataset also considered annual land cover change. Overall, the accuracy of the proposed dataset is promising, and it filled the gaps regarding long-term global LST datasets, especially at nighttime. Therefore, I would recommend it be published on ESSD after a major revision.

Major:

1. Positive bias issue. Based on site validation and inter-comparison with MYD11 and the other two AVHRR LST products, the proposed GT-LST shows a clear positive bias (>1 K) nearly in all results. The authors claim the bias is due to the emissivity difference (Line 370), however, the proposed GT-LST has a clear bias than the other three products, and it seems that the emissivity used by GT-LST is not accurate. The authors mention that the dataset will be calibrated to remove the bias in the future (Line 436). I am thinking if it would be better to solve this issue in this paper as it doesn’t need to be done in a separate paper.

2. Large RMSE (4.1 K) of the monthly mean LST result. The GT-LST is claimed to have the strength to generate gap-free monthly mean LST; however, the outcome has an RMSE of 4.1 K which is too large at a monthly scale compared to other studies (Line 395). This part weakened the statement of the advantage of GT-LST for temporal upscaling based on the logic chain. I would suggest either removing this part or quantifying the impact of orbit drift, in other words, comparing the accuracies of samples that have not and have suffered from orbit drift, and then claiming the potential of this data after orbit drift.

3. The impact of annual land cover change. This is an interesting part of the study, whereas the study didn’t pay attention to the performance of such change. Traditionally people mainly utilized a land cover climatology map rather than annual changes to retrieve global LST. I would suggest including additional analysis to find some examples and compare with LST from Ma et al. (2020) to demonstrate the progress using annual land cover maps.

4. Some processes were not introduced clearly.

1) why does not GT-LST cover 1981 to 2022? GAC raw data is still updating.

2) why did the authors only employ the site observations from 1995 to 2000? If you can extend it to 2005, you can include one more SURFRAD site.

3) Regarding the site validation, 6 sites seem not enough to represent the accuracy of the global product. I would recommend adding some BSRN sites that also have good data quality.

4) why Fig 9(b) has some considerable scattered samples? Those cases should be discussed in the context.

5) Line 350: as MODIS has been spatially aggregated to match with GT-LST, why spatial heterogeneity is still an issue here?

6) Fig10: I would suggest changing Fig10 to another format: consider RMSE and bias as the two dimensions of the plot, and mark each dot by their names as using color to show the bias is not easily quantified.

7) Line 357: why do savannas and cropland show considerable bias?

Minor:

Line 35: Some of them used surface air temperature rather than LST to detect climate change and it should be not mixed.

Line 71: remove ‘the’

Line 94: polar-orbiting

line 101: the first

Line 179: Especially

Line 298: identifier

Line 301: difference

Line 317: due to -> because

Line 327: RMSEs

Line 403: remove ‘in’

Line 404: ‘due to’ should be followed by a noun rather than a sentence, suggest revising the whole manuscript for this issue.

Line 411: considers

Line 446: open-source

Line 451: cloud mask

Summary: In the study titled “A global historical twice-daily (daytime and nighttime) land surface temperature dataset produced by AVHRR observations from 1981 to 2005”, the authors produce a global LST product from 1981 to 2005 at 0.05 degree using AVHRR observations. The study is potentially useful for understanding changes in surface climate over a longer time period than what we can currently examine using most existing LST products. However, I have several concerns that should be addressed before the paper is considered for publication

Comments:

1. The biggest issue I have is that the dataset is restricted to 2005. Given that AVHRR products have large biases compared to MODIS Aqua and use different inputs (such as the dynamic emissivity estimates used), one cannot combine MODIS and AVHRR to perform long-term analysis. Since the AVHRR is still operational, the dataset needs to be extended to more recent years.
2. As an addendum to the previous point, since one of the most important use cases of long-term datasets is time series analysis, the long-term changes in GT-LST should be compared against equivalent changes from MODIS products. If the orbital drift has a significant impact on long-term trends, we should be very cautious about the suitability of this data product for this use case. This issue needs to be quantified more clearly instead of just discussed in text in one section. This can potentially avoid misleading results from future uses of this dataset.
3. A second major source of concern is the dynamic emissivity method used. There are several vegetation-adjusted emissivity methods available, which can give different values, different enough to account for some of the biases seen. Of note, at 0.05 degree, you would start resolving larger urban areas, which is a major use case for satellite-derived LST (Voogt & Oke, 2003). Different emissivity methods perform differently over urban surfaces, which impacts this important use case (Chakraborty et al. 2021). Ideally, this issue needs to be tested further using different emissivity methods.
4. The comparison with MODIS MYD11 is somewhat difficult because of the different emissivity method used. The comparison should be done against MODIS MYD21, which uses the same observations, but a temperature-emissivity separation method instead of classification-based prescribed emissivity.
5. For comparison with SURFAD stations, did the authors check that the emissivity used to generate the LST in the ground observations is same as the LST in the GT-LST product? If they are different, would be good to adjust by the emissivity difference and check if that improves the accuracy.
6. Finally, given the view angle of AVHRR, a broader discussion needs to be added about thermal anisotropy (DUffour et al. 2015). Satellites only provide a 2d directional view of LST, and this is not directly comparable across satellites (Landsat vs MODIS) or against ground observations that have a downward pointing radiometer. This is of particular concern over heterogeneous terrain, such as mixed forests and over cities.

References:

1.    Voogt, J. A., & Oke, T. R. (2003). Thermal remote sensing of urban climates. Remote sensing of environment, 86(3), 370-384.

2.    Chakraborty, T. C., Lee, X., Ermida, S., & Zhan, W. (2021). On the land emissivity assumption and Landsat-derived surface urban heat islands: A global analysis. Remote Sensing of Environment, 265, 112682.

3.    Duffour, C., Olioso, A., Demarty, J., Van der Tol, C., & Lagouarde, J. P. (2015). An evaluation of SCOPE: A tool to simulate the directional anisotropy of satellite-measured surface temperatures. Remote sensing of environment, 158, 362-375.

Why was the MERRA-2 atmospheric dataset selected considering its coarse spatial resolution?

In my opinion, the SURFRAD measurements are not the best option for LST validation, especially in the case of evaluating medium/coarse spatial resolution LST, considering the substantial spatial heterogeneity of the sites. Moreover, the measured longwave radiations by pyrgeometers are different from the directional radiance collected by satellites, which has been reported in different studies.

Why was the MYD11 LST product selected instead of the MYD21 LST, or geostationary LST products that have closer spatial resolutions to the GT-LST product? In the inter-comparison, the MDY11A1 LST was spatially aggregated to the spatial resolution of the GT-LST product with a simpler arithmetic mean. I doubt the validity of the MYD11 LST after the simple aggregation.

Typo in Eq. 5. It should be ‘AVH’ rather than ‘AST’ on the left side of the equation.

The description of the emissivity retrieval process is unclear. How were the soil type data used?

Why were the RMSEs over savannas and croplands the largest amongst different land surface types?

Any explanations for the higher uncertainties in spring and summer?

Line 370, why is the ASTER GED-based emissivity retrieval used in the GT-LST product lower than the classification-based emissivity used in the MYD11 product?

Line 400, why is the emissivity of GT-LST lower than that of RT-LST? The analysis is too simple to understand the intercomparison between different LST products. More in-depth investigations are needed for the comparison with the existing AVHRR LST data.

There are quite some minor grammatical errors, e.g., Line 411, ‘an improved method that consider annual changes’. Please check them carefully.

The authors mentioned the increased uncertainties of AVHRR LST with time due to the orbital drift. It would be useful to add some analyses of the variation in the accuracy of the GT-LST product in the time series.

This paper developed a global historical twice-daily LST dataset (GT-LST) with a spatial resolution of 0.05° from 1981 to 2005. I believe this is an important study and it does make sense for earth science communities. The data and methods are clearly described, and the main results are well presented. However, there are some issues that need to be addressed or clarified before the paper can be published. Therefore, I recommend a major revision.

Some major comments:

1. This paper has inter-compared the GT-LST and MODIS LST over a variety of land cover types such as savannas and cropland/natural vegetation, permanent snow and ice, water bodies and etc., yet I wonder how much are the accuracies (such as RMSD and bias) over urban surfaces? 

2. Why did you choose January 15 and July 15, 1997 for the GT-LST and RT-LST comparison over continental Africa? Please clarify the selection criteria.

3. I just suggest combining Figs. 3 to 7 into a single figure for clarity. 

4. To what extent the differences in the emissivity between MODIS LSTs and GT-LST will influence their inter-comparison results, can you provide some quantitative results?

5. As you stated, the LSTs for a long period such as > 40 years are important for monitoring and evaluating global long-term climate change. Thus, the validation of tendency consistency for the generated GT-LST products is also of vital importance in addition to its spatiotemporal pattern. Could you test the accuracy of time series GT-LSTs over several typical regions, as I guess the orbit drift of AVHRR could also introduce uncertainty for the tendency estimation. 

Some minor Comments:

1. Line 125: Is there a writing mistake on this sentence? “we used 54 land surface emissivity spectra to represent different land surface types, including 41 soil types, four vegetation types, four water body 125 types and five ice/snow types were selected.” 

2. Line 165: “The instrumental error of the SURFRAD station give rise to uncertainty in the retrieved LST value of less than 1 K”. Should be “gives rise to”.

3. Line 266 to 268: “Therefore, to obtain relatively accurate emissivity values, we developed an improved method that consider annual changes in land cover from the GLASS-GLC dataset and combines ASTER GED data with the NDVI threshold method to estimate the emissivity” The verb forms need to be unified.

4. Line 313 to 315: This sentence seems redundant, please write it in a more explicit way. 

5. You can use either RMSE or RMSD, but keep consistency throughout the paper and all figures. 

6. Please unify the format of all references.