Reply on RC2

The paper of Finch et al. entitled ‘Automated detection of atmospheric NO2 plumes from satellite data: a tool to help infer anthropogenic combustion emissions’ examines the potential of using a deep learning method to detect plumes in satellite NO2 retrievals. This paper is a nice piece of work with a novel approach. Although their work on the plume detection is very solid I do have some critical remarks on the relationship with CO2 emissions. Nevertheless, the developed method seems promising and with more satellite instrument coming into place this manuscript is very relevant for the scientific community.


General comments
The paper of Finch et al. entitled 'Automated detection of atmospheric NO2 plumes from satellite data: a tool to help infer anthropogenic combustion emissions' examines the potential of using a deep learning method to detect plumes in satellite NO2 retrievals. This paper is a nice piece of work with a novel approach. Although their work on the plume detection is very solid I do have some critical remarks on the relationship with CO2 emissions. Nevertheless, the developed method seems promising and with more satellite instrument coming into place this manuscript is very relevant for the scientific community.

Specific comments
In the introduction the authors describe the importance of establishing a national CO2 emission baseline as starting point for climate mitigation efforts. Although I agree this baseline is very important, I would like to point out that the reported annual country-level emissions of fossil fuel CO2 are very accurate. However, when looking at specific (urban) areas or facilities and/or at shorter time scales the uncertainties increase. As in the remainder of the manuscript the focus is on plumes from urban centres and industrial facilities I would stress this difference to clearly describe the importance of this work. Also on pg. 2, lines 32-33 'Compiled inventories, which rely on self-reporting, provide estimates on these emissions but rely on assumptions that can sometimes lead to inaccurate values.' We have added the following to the introduction: "The importance of accurate emission estimates becomes even more prevalent at smaller geographical and temporal scales. Reported annual country level emissions of CO$_2$ tend to be reasonably accurate, but are typically not sufficiently detailed to support targeted policy development. Given the importance of establishing accurate national and sub-national emission baselines from which to reduce emissions as part of the PA, it is essential we have a robust measurement-based approach to estimate emissions of CO$_2$ and methane to complement inventory estimates." At several places in the manuscript the authors say that NO2 is a tracer of incomplete combustion, but strictly speaking this is not true and I would rather say a tracer of fossil fuel combustion. The authors explain this well on pg. 2, lines 51-53. Also in the discussion on why plumes from natural gas flaring are lacking this is mentioned and I doubt whether this conclusion is valid. This is a good point. We have changed the text to now read "a tracer of combustion" rather than "a tracer for incomplete combustion" as NO2 is also emitted during complete combustion. We have also amended the discussion so it does not say that the reason flaring produces large amounts of NO2 is due to it being an inefficient form of combustion. It now reads: "We find no correlation between NO$_2$ plumes and the location of natural gas flaring, which is unexpected since this will be an major form of combustion and therefore should result in a significant source of NO$_2$" Pg. 5, lines 153-154: The authors describe that the plume coordinate is determined by looking for the maximum value. What does this mean for images which contain multiple plumes? Also, I'm wondering whether a difference in performance exist for images with one vs. multiple plumes. Could the authors also indicate how many of these images contain multiple plumes?
Our method cannot currently differentiate between images that contain one or more plumes. However, our method of defining the plume centre that uses the maximum value would lead to plumes not being identified in images where there are multiple plumes. This is something we will address in further work and requires substantial model development which is outside the scope of this paper. We have amended the following sentence: "We acknowledge this method could lead to inaccuracies as the maximum pixel value in the image will not necessarily correspond to the origin of the plume and may not identify all plumes where images contain multiple plumes, but we consider this as a minor source of error." Pg. 7, lines 180-181: I think it's also likely that the reverse is true, namely that anthropogenic emissions are incorrectly discarded as biomass burning emissions, while they can easily be co-located. This leads me to the question what the goal of this exercise exactly is. Do the authors aim to detect plumes that are almost certainly anthropogenic and use that for verification of those specific locations? Or is the goal to detect as many anthropogenic plumes as possible for a full verification of global or national emissions? This is also related to their decision to remove images with a <75% confidence that a plume is present in that image. Could the authors reflect on this?
Agreed. We already acknowledge that anthropogenic plumes may have been incorrectly labelled as biomass burning emissions. As stated in the text, our method of differentiating between biomass burning and anthropogenic plumes is imperfect. This problem will especially be a problem where both emission types are co-located. We have adjusted the text to make this clearer.
The goal of this study is to find as many plumes as possible, as accurately as possible, which represents a trade-off. Once this has been achieved (or close to it), we show a method that can help distinguish anthropogenic emissions which can be used to locate, and potentially verify sources across the globe. We envisage that most steps in our methodology will be refined over time and subsequent studies. There are numerous options of how to use these data. There is potential to collate the information on a national and global scale to verify emissions but also to find and verify local emissions at any location (satellite permitting).
The <75% confidence limit is also something that can be refined at a later date or adjusted depending on the way the data is being used for a particular study. We made the decision to use 75% as it generated a large enough dataset that we could demonstrate the usefulness of the tool but not introduce too many errors. We have added the following text to explain this: "This confidence limit can be adjusted to change the ratio of number of plumes spotted to the confidence in the results."

Pg. 9, lines 200-201: 'Discrepancies between known sources and the NO2 plumes, especially over China and India suggest that inventories being used to identify power plants are out of date.' This could be one explanation, but given the authors' conclusion
that 92% of the CO2 emissions are covered with their methodology it also seems likely that the missing sources are rather small and therefore more difficult to detect.
We have amended the text to make it clearer that we refer to areas where there is a large number of plumes -reducing the chance of it being due to a small source. We have also amended the text to say that power plants, identified by the inventory, that do not have any associated plumes could also suggest the inventory is out of date and that further discrepancies maybe due to sources outwith the inventories used in this analysis (e.g. small settlements with large industrial emissions).
Pg. 9, lines 219-220: 'Persistence of plume detection locations (Figure 4) provide confidence that we observing point sources.' Could the authors indicate how often the same location is sampled on average and is there a seasonality in the detection of certain sources? Later in the manuscript the authors compare the detected sources and total emissions with monthly CO2 emissions and therefore the timing may play a role. I also wonder whether the 92% of CO2 emissions covered by the plumes are based on annual emissions?
Locations will be sampled daily as the satellite passes overhead (occasionally more often when swaths overlap). As mentioned earlier in the paragraph, these observations will be subject to quality control and cloud cover so cloudier locations will have an accurate observation less frequently. There is also a seasonality at high latitudes where sunlight becomes an issue during winter months.
It is unlikely that a single source will have a plume detected every day as there might not be an adequate observation or the shape of the plume may not be obvious and not captured by the model. An exploration into what can be considered the same plume occurring repeatedly and what can be considered as "persistent" would be of interest in future work for determining what sources are producing the plumes.
The only marked seasonality observed in our detection of plumes was due to biomass burning. This is to be expected and for our study was driven by the seasonality in the VIIRS dataset that was used to determine which plumes are biomass burning. We found no seasonality in anthropogenic plumes.
We report 92% of the ODIAC emissions covered by the plumes on a monthly basis, i.e. when a plume is detected, are there ODIAC emissions in that location that month? We believe this will account for any seasonality on the ODIAC dataset.
Pg. 14, line 276: 'The impetus for our study is using NO2 as a tracer for anthropogenic emissions of CO2 and methane.' Methane has not been mentioned before in the manuscript (except for the introduction) and I would like to point out that the conclusions drawn here for CO2 may not apply to methane. The emission sources of methane are very different and therefore also the relationship between NO2 plumes and CH4 emissions. More nuance is needed in this statement.
Agreed. Methane emissions would have a very different relationship with NO2. We have amended the text to emphasise that this work would provide a basis to explore further pathways in developing emission estimates of CO2 and methane from combustion.
Pg. 16, lines 299-300: I agree with this statement, but I would rather move this to the introduction. Now the introduction seems to suggest that the authors want to establish a CO2 baseline emission, which is in fact not true.
We have now included a version of that sentence to the introduction: " Although the NO2 plume detection algorithm does not quantify anthropogenic emissions of CO2 or methane, it provides a method to refine the development of future MVR systems which can directly feed into policy decisions."

Technical corrections
Pg. 2, line 39: Please update the reference in this line (International Marine Organization).
This has been corrected.
Pg. 5, lines 124-125: '… were individually normalised to remove the influence the magnitude of image NO2 features, …' Please correct this sentence.