Review of “Full-coverage 250 m monthly aerosol optical depth dataset (2000-2019) emended with environmental covariates by the ensemble machine learning model over the arid and semi-arid areas, NW China” By Chen et al. This manuscript applies bagging trees ensemble methods to produce monthly full-coverage and high-resolution AOD product (FEC AOD). Compared with AERONET AOD, FEC AOD has good performance with an R2 of 0.79. A good analysis of spatio-temporal variability is then presented and the interpretation of environmental covariates on FEC AOD is explored using redundancy analysis. I would like to recommend minor revisions.

1.       Line 33, the expression is ambiguous since the bimodal pattern usually refers to the aerosol size distribution.

2.       The blank space before the reference is lacking.

3.       How did you get the FEC AOD at 250m resolution? Is it simply a matter of interpolating the original input data to a resolution of 250m and then inputting it into the model to get the FEC AOD?

4.       In Figure 3, please include a monthly comparison of MAIAC AOD with AERONET AOD for the same period.

5.       In Line 351-352, the author concludes that “FEC AOD products demonstrate a reliable accuracy and ability to capture local information, even superior to MAIAC and MxD08 AOD products”. However, the loess-based seasonal trend decomposition procedure (STL) in Figure 5 does not show the advantage of FEC AOD over MAIAC AOD. If the advantage is only the spatial resolution, as described in the third point, wouldn't we be able to get any resolution with interpolation?

6.       Please describe in detail the calculation of AOD uncertainty (lines 479-483).

High-resolution with full spatial coverage AOD dataset is important for air pollution-related studies, especially for bright surfaces. The authors have done a lot of work to generate a 250 m AOD dataset in arid and semi-arid areas in northwestern China. However, my biggest concern is the spatial resolution of current developed dataset since it is not clear how to generate the 250 m AOD dataset with two much coarse-resolution MAIAC AOD (1 km) and MxD08 (1 degrees) as main predictions although there are some high-resolution auxiliary data. It sounds incredible and I don't find any downscaling approach or descriptions in the paper.

Below are some other specific comments:

The authors are suggested to summarize previous published studies focusing on multi source AOD dataset fusion or AOD gap filling using different models to rich the Introduction since a lot of related work have been done.

Section 2.3: MxD08 AOD product is too coarse in the spatial resolution (1 degrees) to be used for comparison in such a small study region. I suggest using the MxD04 product with a high resolution 3 or 10 km.

Section 2.4: Please clarify the version and level of AERONET data, and the number of the stations (also suggest adding them in Figure 1) used in the study.

In addition, the author should highlight the novelty of their study and the differences compared to previous studies.

Lines 273-288: It is not clearly how to train and validate the model. Please clarify what are the inputs to the model, and what is the real/true vaule for target? What are the training samples and verification samples? Are they independent of each other?

Figure 4: I don't see much differences compared with 1km AOD, and I think the authors need show the advantages of 250m data set, e.g., may zoom in the image by looking at the AOD distributions at urban areas.

Lines 341-353: The results are pretty similar among different AOD products and difficult to distinguish the difference, and more quantitative comparison results are needed.

Sections 3.2 and 3.3: It is recommended to calculate the monthly and seasonal long-term trends and statistical significance by removing the seasonal cycles to have a look at how AOD changes throughout the study area since AOD data for nearly 20 years are available.