Local PM_2.5 Hotspot Detector at 300 m Resolution: A Random Forest–Convolutional Neural Network Joint Model Jointly Trained on Satellite Images and Meteorology

Round 1

Reviewer 1 Report

The manuscript describes a new random forest – CNN model for estimating PM2.5 in urban areas, building on a recent paper by the authors. Tests are centered on air quality monitors in two cities: Delhi and Beijing. The use of machine learning modeling is topical and inclusion of 3 m satellite imagery novel. Three meteorological variables are included as inputs. I found the writing a little confusing, and after reviewing am not as convinced that this is the best approach for advancing prediction capabilities. Please address the following:

1. The cited references [8-15] (which are referred to in the abstract as “very limited works”) are important to help describe the prior literature on high-resolution PM2.5 from satellites. However, it needs to be clear the current work is fundamentally different from those papers (other than the authors previous work, Zheng et al. 2020), in that they all used new satellite AOD retrievals. Currently, this is only mentioned with regard to one of them (Zhang et al. [14]). That is an important methodological distinction compared to using satellite imagery.

2. One thing I am wondering is what the imagery looks like (and how much temporal evolution there is) specifically over the AQM stations used for training and testing the model, to get a better idea of what the range of land cover inputs the model is evaluated on.

3. It needs to be clear earlier on (L84) why the 300 m (100x100 pixel) resolution was chosen. Saying we [the authors] previously did 200 m and now present 300 m resolution seems somewhat random. It is not until L255 that this question is addressed.

4. Prediction of temporal variability is not much discussed - something I wonder is the approach of withholding AQM stations instead of time periods. Is the main goal to use CNN to spatially fill in where there no stations? Is temporal variability independently predicted from the meteorological averages and daily satellite images, or the test stations benefits from there being measurements from other training stations at the same time period?

5. Prediction of spatial variability - since much of the variability in air quality data is temporal rather than spatial, does this explain a lot of the performance shown in the regressions? It is easy to get bogged into details, therefore it could be helpful to show performance relative to one-day persistence (if temporal variability is of interest) and performance relative to daily averages of the AQM stations (for spatial variability) as a simple metric

Specific

L17 “in these very limited works” – I don’t think this is appropriate. Neither the model formulation here which excludes AOD or the LCN hotspot detection seem particularly more compelling

L43 diameter => aerodynamic diameter

L50 “Despite being a major threat to human health, the intra-urban spatial gradients of PM_2.5 are still not well understood due to the sparseness of rather costly regulatory air quality monitoring (AQM) stations, even in megacities” -  there are other alternatives besides AQM stations and satellite monitoring. Mobile monitoring has produced a lot of information, as well as mentioned at the end, there are a lot of cities that have rather dense Purple Air networks

L55 “resorted to” – suggested “used’. Resorted sounds like an unnecessarily negative connotation

L72 downweighed is not a word

L196 “completeness of 100%” for meteorological data. This is not necessarily bad to mention, but it would be very odd if there were not coverage from at least 1 of the AQM stations at a given time. More interesting since these are used individually (rather than as an average) to train the model with regard to PM2.5, it may be worth adding in Table 1 the % uptime for each station

L222 How can there be no significant differences in meteorology across a city? What is the criteria for ‘significant’?

Figure 3 – why does the final output in the bottom left appear to have a different shape than the other PM2.5 maps in the figure?

L538 suggest deleting ‘excellent’ – PM2.5 is not driven by meteorology alone

L541 “using meteorological conditions, are not restricted to the most polluted Delhi” – check grammar

Figure 6 and 7 meaning of ‘Trim line’ in the plot?

Figure 7 – the satellite images are a little too small to really see/interpret

L819 this is not a PM2.5 satellite retrieval

L838 ‘causally somewhat strapped’ - is somewhat informal

Figures in Appendix C, D, and E – the axes and labels load but the data did not load in multiple PDF readers for some reason

Figure F – so much space is used for the colorbars I can’t discern anything from this figure

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

The authors do a thorough job of walking through the limitations and expectations of their modeling approach. The veins of future research topics that derive from this research are scientifically sound and noteworthy. The references were coherent, recent and applicable.

This reviewer ponders how this research approach could scale in different air quality regimes that are dominated by larger particulates - this could serve as potential test case that would interrogate a different limitation of the modeling methodology.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

This paper developed a random forest–convolutional neural network–local contrast normalization (RF-CNN-LCN) pipeline that detects local PM2.5 hotspots at a 300 m resolution. The RF-CNN joint model in the pipeline uses three meteorological variables and daily 3 m/pixel resolution PlanetScope satellite imagery to generate daily 300 m ground-level PM2.5 estimates. Overall, this research is meaningful and interesting, but it is a little lengthy and dense. I recommend be accepted after revision. Please make the below revisions:

1. The Abstract Section is so long. Please remove some details, mainly keep the methods and main results you got.
2. Please use the International Units for an international journal such as Miles (please use km or others).
3. For the Webpage access, please add the access date in the Reference section not the main text such as Line 187. Please also check other places.
4. For the maps, please add essential map elements such as north arrow, scale bar... in Figures 6 & 7. Please also check the whole paper.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Thanks to the authors for a thorough response to my comments, including adding additional supplementary information. I believe the work in now ready for publication.