The paper is quite interesting, but I still have some comments on it

The major comments: 1. how could you identify the improvement comes from your new methods or just spliting data to 6 groups? spliting data into spatial groups also can help model more easier to capture the variation. 

    2.why you split the sites into 6 categories? and when you apllied your final models, how could you define the predicted location/grid belong to which categories?

    3.please add the spatial cross-validation results to check your model spatail predict ability

    4.for the temporal validation, .the test data are only in winter? have you tried to use rolling temopral validation? use previous 4 seasons as training, 1 season as validation.

    5.The study only use 2 year data, I wonder whether model can be predicted in the following year.

    6.the study period did not include 2020, so will the model forecast be affected by covid-19 when we applied this model for the real early warning system? Have you test the models with 2020 or 2021? this will be the important issue for early warning system.

    7.the ground-level pollutant is also included as input? how could you predict the value for thoes location outside these sites?

Minior comments

1.why the max depth of RF is Unknown?

2.why each region choose 15 sites for features selection? have you tried 10,20, or some other numbers?

3.why do you pick only top 3 features? not 5 or more? will they affect the results?

Technique points:

1."Cubic imputation  was applied to fill in the missing data because the 3 h resolution of the CAMS reanalysis data is too coarse for interpolation. After interpolation, 0.75o x 0.75o grid data were imputed to each monitoring station" here you mean impute the missing or downscale? if downscale, from 0.75 to which resolution?  

2."The data were interpolated to the monitoring  station locations for use in machine learning." It is still confusing.

General Comments:

This paper presents the development of an improved machine learning based air quality nowcasting system. Instead of using all possible related features in the model training and predicting, they selected those general important and effective features. The feature selection is done using a computationally efficient ensemble method. Their nowcasting system is tested on the PM2.5 forecast on a national scale and validated to be superior than a CTM model and conventional MLs. Generally speaking, the paper is clearly written and well structure, their results are scientifically solid. I recommend accepting it after a minor revision. I also have questions and comments for the author that could help to improve their manuscript.

Major comments:

They have tested their regional feature selection-based ML nowcast system at a national scale and using several common ML models (RF, GB and MLP), which makes their results very sound. However, it is only tested at a 2019 winter season. I understand winter is the most severe polluted season there. Their system should be able to operate in a rolling forecast way. If extra training and testing are conducted at a less-polluted period/season, this study would be an excellent paper.

They should also explain the current machine learning model cannot fully replace model-based air quality forecasting systems, as ML models could not be trained and operated without inputs from the historical measurements. While for many rural regions, they are unavailable. The authors should explain this point clear. 

Specific comments:

Page 6, Table 1: esolution to resolution

Page 12, line 5: computational complexity?

Page 17, line 1: and at a predicting horizon?