the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 07 Dec 2023
Detecting ship-produced NO2 plumes and shipping routes in TROPOMI data with a deep learning model
Abstract. Ship emissions are important contributor to air pollution and the climate through interacting with clouds. They are the dominant anthropogenic source over the oceans. However, their magnitudes still have large uncertainty. Here we develop a deep learning model to detect ship-emitted NO2 plumes at the pixel level in TROPOMI tropospheric NO2 data. The trained model performs well and, when applied to a year of data, it finds major shipping routes, but misses several other routes. We show that high cloudiness in these shipping routes is the culprit because clouds block signals from reach the sensor. Indeed, detected shipping routes in this study complements shipping routes detected using ship-tracks that is only available in cloudy regions. Our method can find application in several areas such as improving ship emission estimates and compliance verifications. Our method will benefit from improved tropospheric NO2 retrievals since their quality is critical for plume detection.
Status: final response (author comments only)
-
RC1: 'Comment on Egusphere-2023-2469', Anonymous Referee #1, 16 Dec 2023
- AC1: 'Reply on RC1', Tianle Yuan, 30 Apr 2024
-
RC2: 'Comment on egusphere-2023-2469', Anonymous Referee #2, 02 Jan 2024
Review of ‘Detecting ship-produced NO2 plumes and shipping routes in TROPOMI data with a deep learning model’ by Tianle Yuan et al.
The manuscript by Yuan et al. presents a method that uses machine learning to identify NO2 plumes from ships in individual TROPOMI orbits. The main result of the application of the machine learning method is a climatological map of ‘NO2 pixel frequencies’ obtained from one year of TROPOMI NO2 data. The map itself is convincing, and in line with what has previously been observed in satellite NO2 data: the main shipping routes between Europe and Asia can be well-identified, but ship plumes between Asia and North America, and between Europe and North America are much less detectable. The authors claim that this is mostly due to persistent cloud cover over the Atlantic and Pacific Ocean. In MODIS cloud data they have previously detected ship tracks (as bright clouds) right over these oceans. In my opinion, this complementarity of MODIS-derived ship tracks and TROPOMI-derived ship NO2 plumes is a nice idea that has the potential to push ship emissions research from space further in the long run. This manuscript does not do that, but is mostly an AI-effort to find NO2 plumes from ships, and derives a climatology from the result.
For a non-expert like me, I found section 3.1 where the human labelled training of NO2 ship plume detection is explained with useful imagery (Figure 3) quite useful. However, a bit more details about the capabilities and limitations of the method are needed and strengthen the paper, as detailed below. Should this manuscript be published I encourage the authors to make available their results of positive plume identifications in open access, digital format, as electronic asset to this paper.
Major comments
1. The authors show some convincing examples of massive 50-100 km long NO2 plumes. These are easily identified by eye. But the authors should also give some examples of where their method is just at the level of positive detection. How elongated are such plumes in terms of pixels – a one pixel enhancement should never qualify as a plume. What are the NO2 column values, and to what extent are such ‘just positive’ plumes above the background NO2 values?
2. The main metric shown is ‘NO2 pixel frequency’, but it is not entirely clear what this means. Is it the accumulated number of (gridded?) pixels identified by the machine learning from all qa_value-filtered over one year? Would it not be more appropriate to express this frequency as a normalized frequency, to account for differences in the available cloud-free scenes?
3. It does not become clear how the machine learning approach handles plumes originating from land that flow out over sea. If unaccounted for, this leads to frequent misclassifications of plumes over coastal seas.
4. The claim that the methodology of the authors could be used for ‘verification of compliance and effect of emission control policies’ is not substantiated in the manuscript. The study deals with recognizing NO2 plumes in noisy data, which could be a first step in a system that estimates ship NOx emissions. Beyond attributing NO2 plumes to individual or multiple nearby ships, much more is needed in terms of inverse modelling, considering atmospheric chemistry and transport which influence the relationship between NOx emissions and NO2 columns.
5. That tropospheric NO2 is not well detectable from space for pixels (partially) covered by clouds is obviously not a new finding and has been well-documented since the early 2000s (Martin et al. (2002); Palmer et al. (2001); Boersma et al. (2004)-papers). In lines 222-223 the impression should be avoided that this is a new result from this manuscript. Furthermore, there are new insights that in situations of sun-glint, screened out by the authors now given their qa_value threshold, the NO2 retrieval is performing better (Riess et al., 2022). Including these scenes rather than excluding them could improve the performance of the Machine Learning model.
6. The authors have chosen to use version 1.2.2 of the official TROPOMI NO2 data product. This is a pity since there have been a few important retrieval improvements and reprocessing efforts since then. These have been discussed in van Geffen et al. (2022) and Riess et al. (2022), and especially the more accurate FRESCO+ cloud heights from version 1.3 onwards are important for the author’s purpose. In version 1.2 the too low cloud heights led to a substantial low bias in the TROPOMI NO2 columns. This low bias has been partly resolved in the later versions, and results in better detectability of ship NO2 plumes. I think it would be a sorely missed opportunity to not apply the method on version 1.4 or later.
7. The most pressing concern I have with the study is the lack of uncertainty analysis and the lack of validation. Although the (climatological) results presented appear plausible, a success metric is missing – what fraction of the ‘NO2 pixel frequency’ could have been misclassified? Besides retrieval uncertainties, also one/two-pixel size plumes, mis-training or mislabeling by the human, and the influence of plumes from land likely play a role. The authors should do more to address this issue, for example by comparing their (relative) frequencies to the locations of actual ships as stored in AIS-data or inventories on ship emissions,
Some relevant literature is not cited or discussed:
Kurchaba, S., van Vliet, J., Verbeek, F. J., Meulman, J. J., & Veenman, C. J. (2022). Supervised segmentation of NO2 plumes from individual ships using TROPOMI satellite data. Remote Sensing, 14(22), 5809.
Kurchaba, S., van Vliet, J., Verbeek, F. J., & Veenman, C. J. (2023). Anomalous NO2 emitting ship detection with TROPOMI satellite data and machine learning. Remote Sensing of Environment, 297, 113761.
Riess, T. C. V. W., Boersma, K. F., Van Vliet, J., Peters, W., Sneep, M., Eskes, H., & Van Geffen, J. (2022). Improved monitoring of shipping NO< sub> 2</sub> with TROPOMI: decreasing NO< sub>< i> x</i></sub> emissions in European seas during the COVID-19 pandemic. Atmospheric Measurement Techniques, 15(5), 1415-1438.
Riess, T. C. V. W., Boersma, K. F., Van Roy, W., de Laat, J., Dammers, E., and van Vliet, J.: To new heights by flying low: comparison of aircraft vertical NO2 profiles to model simulations and implications for TROPOMI NO2 retrievals, Atmos. Meas. Tech., 16, 5287–5304, https://doi.org/10.5194/amt-16-5287-2023, 2023.
van Geffen, J., Eskes, H., Compernolle, S., Pinardi, G., Verhoelst, T., Lambert, J.-C., Sneep, M., ter Linden, M., Ludewig, A., Boersma, K. F., and Veefkind, J. P.: Sentinel-5P TROPOMI NO2 retrieval: impact of version v2.2 improvements and comparisons with OMI and ground-based data, Atmos. Meas. Tech., 15, 2037–2060, https://doi.org/10.5194/amt-15-2037-2022, 2022.
Technical corrections
L17: “Straight” - Strait
L35: “over the global” – throughout the world?
L36: “the aggregated plumes … miss other routes” – plumes cannot ‘miss’ routes. The absence of plumes could be an indication that the method is incapable of detecting plumes in some routes,.
L36: “block the signals” – make clear what signals are meant here
L45: “show up as a long linear and bright” … as a long linear and bright cloud?
L63: here a discussion of NO2 retrieval characteristics for ship plumes that were the in-depth topic of investigation in the papers by Riess et al. (2022; 2023) should be discussed.
Figure 4: top middle panel states ‘NO2 emission’ which should be (annual mean) NO2 column.
L280-285: a discussion of the findings in Riess et al. (2022; 2023) is missing here.
L302-305 and L324-326: first names of authors should be replaced by their last names for these references.
L392-395: this paper has already been published in AMT and should be cited accordingly.
Citation: https://doi.org/10.5194/egusphere-2023-2469-RC2 - AC3: 'Reply on RC2', Tianle Yuan, 30 Apr 2024
-
RC3: 'Comment on egusphere-2023-2469', Anonymous Referee #3, 10 Jan 2024
In this manuscript, Yuan and co-authors describe a machine learning approach to identifying NO2 from ship emissions in daily images of TROPOMI tropospheric NO2 data. The model is trained by human classification of test data over known shipping lanes and then applied to a full year of TROPOMI data. The resulting map of frequencies of pixels identified as being affected by ship emissions is consistent with TROPOMI NO2 observations and earlier work on NO2 from ships.
The topic of the manuscript (a machine learning tool for the identification of NO2 from ships) is relevant for atmospheric applications and fits into the scope of AMT. The manuscript is clearly written but would benefit from another round of English proof reading. The level of details is low, and I recommend to provide more details on the method, the training and the performance of the model to make the manuscript more useful for the readers. It would also be good if the authors could give an example of a possible application of their method, as this is not clear to me.
Major comments:
- In the title and throughout the manuscript, the authors write about NO2 plumes. At first, I understood this to be plumes from individual ships. However, after reading the full manuscript, I think the algorithm looks for large elongated regions of enhanced NO2 as they are present over certain shipping routes. These NO2 enhancements are the accumulated result of many individual plumes, and in my opinion, they should not be called "plumes".
- As in all satellite NO2 maps, only some of the shipping lanes can be identified in this study. The authors explain this by clouds, which certainly are a factor, but probably not the main one. In my opinion, NO2 signals from ships can only be detected if a large enough number of ships operates on a narrow track at not too high wind speeds, as dilution can reduce the signal below the detection limit of the satellite measurements. These aspects need to be discussed, and the statement “we show that cloudiness in these shipping routes is the culprit …” needs to be changed.
- The authors discuss a shipping route from the coast of Africa to Madagascar. I’m sceptical about this finding as this is also a well known area of pollution export from South African power stations. My guess is, that these have elongated shapes and are therefore misclassified as ship emissions. Please discuss.
- The authors claim, that their method is useful for emission inventory and emission compliance studies. These claims need to be substantiated or removed. As it is presented in the manuscript, the counting of pixels classified as containing ship emissions is not quantitative relative to the NOx emissions of ships in general and even less for individual ships.
- Very little information is given on the model and the training. For example, from which regions are the training data sets? From line 99/100 it appears, as if only data over the Indian Ocean is used? What is the proportion of “empty” training data? What is the performance of the classification?
Minor comments:
Line 32: Satellites do not retrieve NO2 concentrations but NO2 columns
Line 34/35: Language
Line 47/48: These references are more than 20 years old and no longer representative
Line 76/77: See major point about “plumes”. I don’t think that your algorithm is actually identifying the plumes from individual ships.
Line 99: Which shipping route map?
Figure 1: Was this scene part of the training data set? If so, then please replace by a scene from the test data set.
Line 126: Please provide more details about the input matrix. Why did you choose 400 x 400? Does this mean, that you are skipping the outermost 25 pixels on each side of an orbit?
Line 128: lever => level
Figure 2: It would be nice to update this figure to show NO2, not clouds
Line 145: Does that imply two flips and 3 rotations? Are these transformations already included in your counting of training data?
Figure 3: Were these scenes part of the training data set? If so, then please replace by scenes from the test data set.
Figure 4: TROPOMI does not retrieve NO2 emissions as stated in the figure and in the text in line 218. It also does not retrieve NO2 concentrations as stated in the caption. It’s NO2 columns.
Figure 4: Some aggregation has been applied in these figures which appear to have low spatial resolution (otherwise, no more than 365 counts per year are possible in a given location, while frequencies here are up to 10000. Please explain what is shown here.
Line 222: Why low clouds? Aren’t high clouds even more of a problem? See also my general comments on clouds.
Line 282: I think this sentence is a bit mixed up. It’s not the uncertainties which add subpixel variabilities, but sub-pixel variability which leads to uncertainties.
Line 293: See general comment on clouds
Data availability: It would be good to make the model and the derived masks available on a repository
Citation: https://doi.org/10.5194/egusphere-2023-2469-RC3 - AC2: 'Reply on RC3', Tianle Yuan, 30 Apr 2024
Viewed
Since the preprint corresponding to this journal article was posted outside of Copernicus Publications, the preprint-related metrics are limited to HTML views.
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 307 | 0 | 1 | 308 | 0 | 0 |
- HTML: 307
- PDF: 0
- XML: 1
- Total: 308
- BibTeX: 0
- EndNote: 0
Viewed (geographical distribution)
Since the preprint corresponding to this journal article was posted outside of Copernicus Publications, the preprint-related metrics are limited to HTML views.
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1