This study aims to integrate remote sensing data with ground-based tempreature measurement to preict canopy heat island at a large spatial extent. Overall the idea is interesting, but the data size is very limited and the method seems to be a typical approach. Below are my major comments:

1) Line 55, "meteorological measurements and high-density observations", it is not clear what the authors are referring to here. Does it mean rural station against microclimate observations? Please rephrase here.

2) Line 104, ""Nanjing's UHI was observed to be 0.5 in 2005". It is important to clarify how the 0.5 is calculated, as it does not contain any spatial variation, and differs from the results in this study.

3) Line 110, my major concern about this study is the usage of only 3 snapshots of the satelliate measurement. Only 3 days are selected over a 5-yr period from 2013 to 2017. Such data availability is suprising low. On top of this, this performance of RF model is only marginal, with R2 about 0.5 in the cross-valiadtion. How would the authors justify the potential or accuracy of the model in predicting the urban heat island for pratical usage, if the model is to be extended to more days under complex weather conditions?

4) Line 113, what does 0.5 intervals mean? 

5) Line 124, Does the AHF data vary diurnally and seasonally? If it does not, then the AHF data may be less meaningful to be incorporated into the RF model. Instead of the LULC map, I will suggest authors to add the AHF map here.

6) Line 150, a table summarizing the predictors with their sources and resolution used in the RF model will be very useful. 

7) Line 180, Different variables are used on different days, and the optimal Mtry and Ntree also change substantially. This essentially means the built model can only apply for a specfic day. Given that the study only focus on Nanjing under clear sky conditions, I am concerned about the appcaliability of the model. If we utilize this method to study CUHI in the future, that means one will need to run it for every hour, and the accuracy is not guranteed even under the clear sky condition. What I suggest may be a lot of work, but I think using more models and comparing their performance, and come up with a consistently well-perform model is really needed to enhance this paper to a higher quality/level.

8) Line 197, from my perspective, using a buffer size of 2-km to predict temperature at 30-m resolution is not scientifcally sound. As expected, the estimated spatial variablity of air tempearture/CUHI is small. This contradicts the local climate zone framework that local urban landscape may dominate the air tempreature under a similar weather condition. Did the authors check the spatial map of SUHI? I believe the spatial variablity of LST will be much larger. I suggest authors add a map of LST and compare it with the estimated CUHI. That can help facilitate the discussions on your model.

9) Figure 6 and Figure S2, why is the number of stations different in three subplots?

10) Line 406. It is good that authors summarize the existing approach of predicting AT in the liteature. Given that RF model is a very typical approach adopted by many studies, what is the nolvety of this study? Maybe the authors can elaborate more here.

In the manuscript the authors proposed a framework to train the random forest (RF) model so as to output high resolution air temperature map. The RF model takes various environment parameters (mostly derived from remote sensing data) around a dense network of meteorology stations as independent variables and it is trained using the air temperature data observed by those stations. The work starts with a good motivation to provide high resolution air temperature map, which is important for a better understanding of thermal environment of the city and helps for heat exposure assessment.

The method propose by the authors shows the capability of generating high resolution air temperature map. With this temperature map the authors further studied the fine scale canopy layer urban heat island as well as its relationship with driving factors. This, from my point of view, matches well the original motivation of this work.

The workload of this study involves quite a bit effort, especially regarding the preparation of the input variables that are fed to the RF model, and the identification of the proper buffer size, as well as the comparison of the model performances between the proposed RF model and other regression models.

As I did not really give a fast review before the discussion started, in the ‘fast review’ I raised several questions, such as increasing the presentation quality, better organizing the technical stuff, giving more space to technique details, I appreciate the authors’ great effort in address my concerns, I found they have been addressed mostly. So I personally do not see fundamental issues in the current version of the manuscript.

There are a few minor issues remaining:

L34: “relatively” is unnecessary.

L83: I suggest replacing ‘Meanwhile’ with ‘However’, and moving this passage to the following paragraph.

L104-105: I find the last sentence is confusing and is not necessary to put it here.

L107: a link directing to ‘geospatial data cloud’ would be fairer.

L269-277: Please be aware that due to different weather and climate conditions, it is very risky to directly compare the UHI intensities, even though you seem to have carefully selected the same barely impacted rural reference area. The rural area and urban area respond to different parameters very differently. If you really want to compare UHI intensities from different, I would suggest classifying the UHI intensities into several levels using the quantile instead of the uniform scalar values. If not so, please at least avoid direct comparison, while just plainly describing the table content is fine.

L406: I would suggest putting it this way: “the RF prediction framework proposed in this work not only can dynamically predict CUHII in detail within highly heterogeneous cities, but also can be built…..”. Besides, I see the potential that the model can be used cross a short period as most of the environmental parameters fed to the model probably can remain stable for some time, but I would say that it is rather bold to claim a period like one month or even longer.

L733: in the caption of figure3, it should be 'canopy layer urban heat island' instead. Please note that in the figure you no longer have Abbreviations like DEM, AHF, LULC, etc.

L764. In the caption of figure10, I suppose by ‘standard value’, you mean standard deviation.