Satellite remote sensing and ground-based stations play their unique advantages in global meteorological parameter retrieval, and near-surface temperature is of great significance to global and local climate. In this study, satellite and ground-based data are used to retrieve near-surface maximum and minimum temperatures in almost global land regions with 1-km resolution. Overall, this manuscript is clear and well written and presents interesting research, however I found there was a lack of details necessary to fully understand the methods. Some concerns are needed to address, below are my major comments. (Major revision)

1. Line 16: I think “ground station-based Ta” is better than “station-based ground Ta”.
2. The descriptionof the “global dataset” is not rigorous, as the authors have only achieved retrieval of near-surface temperatures for most of the global land region.
3. Lines 82-83: Different regional retrieval models have certain differences. Two different Tmaxs and Tmins will be obtained in the overlapping region. How does the author calculate the final Tmax and Tmin results?
4. The latitude and longitude grid should be added in Figure 1.
5. Please explain in detail the similarities and differences with the study by Zhang et al. (2022b), and whether there are other differences besides the study area.
6. Lines 107-114: The ground-based stations usually have a high temporal resolution, and the seamless global surface temperature data used by the author comes from the MODIS sensor, which has a limited number of transits. When the training sample library is constructed, the author only describes the spatial matching process and ignores the temporal matching, please give a detailed description.
7. Lines 124-125: Does the authors mean that the ground-based measurement result corresponding to mid-daytime is the Tmax of the site, and the ground-based measurement result corresponding to mid-nighttime is the Tmin of the site? If yes, I don't think it's reasonable, especially for the Tmin.
8. Due to the spatial correlation of the near-surface air temperature at the different stations and the temporal correlation between the training data at same station, the 10-fold cross-validation verification cannot truly reflect the accuracy of the model. The author needs to give independent verification results, such as the first 15 days of each month as training Set, the data of the last 5 days is used as the validation set or test set, or the data from 2003-2018 is used as the training set, and the data in 2020 is used as the validation or test set, or 80% of the site data is used as the training set, and the data from the remaining 20% sites is used as the validation or test set.
9. What does the Y-axis of Figure S1 represent? Is it the number of valid observation samples or the number of stations that only include 1 valid observation sample? Also, I did not find a related description of Figure S1 in the manuscript.
10. Comparing Figure 1 and Figure 3 confuses me. In my view, the author has constructed Ta estimation models in 5 study regions, but does not include Greenland (Figure 1). I have three questions. First of all, why can the Ta retrieval in the Greenland region be achieved? The parameters of the retrieval model are the same as in which regions Ta retrieval model? In addition, the standardized regression coefficient for the Greenland region in Figure S7 is also absent. Second, can the constructed retrieval model be used for the retrieval of Ta in the oceanic region? Third, what do the non-color-filled regions on land in Figure 3 (such as the Amazon region and south-central Africa) mean?
11. It is suggested to modify the color bar range of Tmax and Tmin in Figure 5 to be the same
12. I think that the validation of the results in this manuscript needs to be expanded further and needs to add scatter plots for the validation of the retrieval results, such as density scatter plots instead of just calculating RMSE and MAE.

Tmax and Tmin are both very important of ecosystem.  High quality of long period and high resolution of air temperature products estimated from satellites are still lacking. This paper provided a set of such grided data with elaborated accuracy evaluations. However, there are some key questions needed to answer. 

1. I suggest you change the title because you olny produced Tmax and Tmin. 'Daily' means too much. Besides,'global' is more than what you did.  

2. Line 77-78: I don't think you improved accuracies of spatial resolutions and temporal coverge. Many current studies have achieved 1km resolution products and the scale of daily even hourly. 

3. line 103-104: As you trained models for each day of the period 2003-2020 as well as for each of the five regions, you would got the same number of results from 10-fold cross validation with trained models. Because different models will produce different validation results, how did you got only one result for each region, each landtype, each climate type, each year....  Possibly, you validate the products using all the records, thus, the validation is not independent. Or you caculated the average of RMSE of all models' 10-fold cross validation, such as in one region in one day, thus the RMSE through the paper is debatable. 

4. Please describle how many records you filled because of missing values of weather station observations, and what percent of filled records to all records. Please note it's 'records', not number of sations with valid observations. Figure S1 and S3 would confuse readers because Y axis and title are inconsistent. Besides, how do you define valid station number? For example, Latin America in 2020 has no more than 150 stations before filling missing data. After filling, it is over 150. If one station had only one valid record in 2020, it didn't make sense. 

5. In equation (1), there is only one LST, however, as both Terra and Aqua observes in one daytime and nightime, there are two LST values. Did you use average value? or lower value? or higher value? 

A global dataset of daily near-surface air temperature at 1-km resolution (2003-2020)

Review Comments

General comments:

Near-surface air temperature (Ta) has extensive applications in climate and environment studies. This study, based on a newly developed Spatially Varying Coefficient Models with Sign Preservation (SVCM-SP) algorithm, generated a global dataset of daily maximum and minimum Ta (Tmax and Tmin) at 1 km from 2003 to 2020 by integrating ground Ta observations from weather stations and gridded LST and DEM data. The assessment shows that the employed algorithm can effectively capture the negative relationships between Ta and elevation and the positive relationships between Ta and LST. The cross-validation indicates the estimated Ta show satisfactory accuracies, and the RMSEs of Ta estimates range from 1.20 to 2.44 ºC for Tmax and 1.69 to 2.39 ºC for Tmin.

The study designed a global maximum and minimum Ta estimation scheme and developed an applicable time-series (2003-2020) daily Ta dataset. I think this work is important because the generated datasets are of great demand and value in practical applications (e.g., urban climate research). However, some issues in the manuscript still need to be addressed before being ready for publication. The specific comments are given as follows.

Specific comments:

1. Line 18, Please add the unit for ‘2.44’.

2. Line 20, There is ambiguity in the expression. The positive and negative relationship is suggested to be expressed separately.

3. Line 30, Why is the LST mentioned here?

4. Line 70, What does ‘these’ refer to? 

5. Line 88, It is recommended to add the region name represented by each color boundary, or to label the region name directly in Figure 1.

6. Line 90, Do these ground measurements provide hourly Ta observations?

7. Line 96, What are the DEM data years used?

8. Line 136-137, What does this sentence mean?

9. Line 216-227, There are too many introductions about previous studies, which are already discussed in the Introduction. It is suggested to simplify these contents.

10. Line 227-239, Some contents (e.g., Line 228-232) that have been mentioned similarly in the Introduction are also suggested to be simplified.

11. Line 228, What does the “which is not always true in other gridded Ta datasets” mean?

12. Line 244, Why is the UHI effect mentioned here? The previous description of the Ta dataset does not seem to refer to UHI.

13. Line 278-281, This is more appropriate for the discussion section than for the conclusion.