This study designed an operational framework that uses the annual temperature cycle (ATC) and diurnal temperature cycle (DTC) models to generate global seamless daily mean land surface temperature (LST). The framework and generated product were validated with globally distributed in situ measurements. The validations show that the generated daily mean LST can correct the sampling bias caused by directly compositing the cloud-free MODIS LSTs. This is an interesting point for the thermal remote sensing community. Additionally, the authors discussed the uncertainties of the daily mean LST products, which are useful for further improvement. The authors clearly addressed the structure of the IADTC framework and comprehensively evaluated the generated daily mean LST product. This manuscript is generally well written and clearly organized. I recommend the paper for publication after the following issues are answered. 

Major comments

(1) The direct comparison results between the generated daily mean land surface temperature product and in situ measurements display systematically negative bias at most sites (Tables 1 and 2). The authors should provide more explanations about the negative bias. 

(2) The authors used the diurnal temperature range (DTR) to define different scenarios. In this paper, the calculated DTR can be affected by the accuracy of ATC model, then affecting the determination of which scenario is used to generate daily mean land surface temperature. I recommend the authors add more discussions about the uncertainties of ATC model to the daily mean LST estimation. 

Minor comments

(3) Line 138: I recommend the authors to add some descriptions about how they process the in situ measurement outliers.

(4) Line 176-178: Please add more examples or references about the LST change in low-latitude and high-latitude regions. 

(5) Line 218: Temporal normalization is a good way to handle the overpassing time fluctuations. Please provide more discussions about the role of temporal normalization in generating consistent LST products. 

(6) Line 242: Moving this sentence after the introduction of DTRfour would be better. 

(7) Fig. 4: I recommend the authors to add one subplot for the illustration of Scenario #1. 

(8) Line 317: “Lower accuracy” being compared to what needs to be clarified. 

(9) Line 394: Please provide more evidence about the link between ΔTsb and land cover type or DTR. 

(10) Line 414: Please clarify what’s the different information contained within the ΔTsb. 

(11) Fig. 11: I am wondering about the variation of error of Tdm_ATC_DTC versus DTRfour, which can provide more solid support for the necessity of defining Scenario #1. 

This paper describes an improved annual and diurnal temperature cycle-based framework method to generate global spatiotemporally seamless daily mean LST products from MODIS data with the support of reanalysis data. The developed dataset performs very well against global in-situ surface observations. Overall, this new method produces a 0.5 --degree daily product of daily mean LST over the globe. Given that this data has high spatial resolution at a daily time scale, it should be a useful tool for climate studies after its flaws are addressed.

Major comments:

1. The developed GADTC product has a spatial resolution of 0.5-degree, how to deal with the scale mismatch between the in-situ measurements and the product, the validation can be carried out at a higher spatial resolution, such as MODIS original resolution. Maybe, the authors can classify the in-situ sites to different levels according to the spatial heterogeneity of the site, to further analyze the errors at different sites.

2. The Surfrad site only has 7 sites, Why not merge the data from the Surfrad and Fluxnet networks when validating the Tdm product. Also, in section 5.1, the ΔDTR can be obtained using the Surfrad and Fluxnet data together.

3. The authors used MAE and bias, why not use the RMSE, which is typically used in the LST validation.

Minor comments:

Line 67, some latest papers about the C6 MODIS LST accuracy can be added, such as DOI: 10.1109/TGRS.2020.2998945, https://doi.org/10.1016/j.jag.2018.04.006

Line 104, the MxD11C1 was derived using the day/night algorithm and giving a reference

Line 139, how to get the hourly values?

Line 319, Scenarios #1 and #3, How many sites per scenario, the results can be analyzed by scenario, not by Surfrad and Fluxnet.

Line 360, Fig.8, combines data from the two networks.

Line 373, how to prove the large errors at these sites are related to the high spatial heterogeneity

spatiotemporally seamless land surface temperature at daily, monthly, and yearly scales are important for LST-related researches. This study presents a meaningful study with the use of MODIS LST product and reanalysis data to generate the mean LST value at different scales. It was well organized and the results were with good accuracy. Overall, the manuscript can be accepted with minor revision:

1. There are many other reanalysis data available and why you choose the MERRA2 dataset? What is advantage of this dataset?

2. The key steps are suggested to be clarified in in figure 2. The pre-processing is not included in this flowchart.

3. 175: A basic equation of the single-type and multi-type model is better to be provided here. 

4. Figure 3: multi-type ATC models are identical? Why there is no differences? It will be a little confused on the naming of the ATC models for single or multi-type model and single or double-sinusoidal ATC model?

5. Section 3.1.3: I think it should be the interpolation of the missing LSTs but not overpassing times.

6. Actually, the DTC model should be not applied to get the DTCdm when there are cloud-cover observations. 

7. Besides the direct validation of the estimated mean values at different temporal scales, there is a lack of the evaluation of the reliability of the trend detection based on the generated dataset. How about the performance of the dataset on identifying the area with significant trends. 

8. The threshold determination for the two criteria in Fig. 2 is a little objective. I think the determination can be automatically determined according to the differences between the average value from four observations and the fitted values. 

9. The LSTs of cloud cover pixels are generated with the reanalysis data at coarse-resolution. Currently, there are some other reconstruction methods without the use of the reanalysis data. How about the applicability of these methods in this study.

10. The dataset produced in this study has the resolution of 0.5 degree. However, to some extent, the LST product at 1-km and higher resolution will be useful. What is the key issue should be addressed at this high-resolution level.