the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 04 Apr 2024
A bias-corrected GEMS geostationary satellite product for nitrogen dioxide using machine learning to enforce consistency with the TROPOMI satellite instrument
Abstract. The Geostationary Environment Monitoring Spectrometer (GEMS) launched in February 2020 is now providing continuous daytime hourly observations of nitrogen dioxide (NO2) columns over East Asia (5° S–45° N, 75° E–145° E) with 3.5 × 7.7 km2 pixel resolution. These data provide unique information to improve understanding of the sources, chemistry, and transport of nitrogen oxides (NOx) with implications for atmospheric chemistry and air quality, but opportunities for direct validation are very limited. Here we correct the operational level-2 (L2) NO2 vertical column densities (VCDs) from GEMS with a machine learning (ML) model to match the much sparser but more mature observations from the low Earth orbit TROPOspheric Monitoring Instrument (TROPOMI), preserving the data density of GEMS but making them consistent with TROPOMI. We first reprocess the GEMS and TROPOMI operational L2 products to use common prior vertical NO2 profiles (shape factors) from the GEOS-Chem chemical transport model. This removes a major inconsistency between the two satellite products and greatly improves their agreement with ground-based Pandora NO2 VCD data in source regions. We then apply the ML model to correct the remaining differences, Δ(GEMS-TROPOMI), using as predictor variables the GEMS NO2 VCDs and retrieval parameters. We train the ML model with collocated GEMS and TROPOMI NO2 VCDs, taking advantage of TROPOMI off-track viewing to cover a wide range of effective zenith angles (EZAs) for the GEMS diurnal profiles. The two most important predictor variables for Δ(GEMS-TROPOMI) are GEMS NO2 VCD and EZA. The corrected GEMS product is unbiased relative to TROPOMI and shows a diurnal variation over source regions more consistent with Pandora than the operational product.
-
Notice on discussion status
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
-
Preprint
(4045 KB)
-
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
- Preprint
(4045 KB) - Metadata XML
- BibTeX
- EndNote
- Final revised paper
Journal article(s) based on this preprint
Interactive discussion
Status: closed
-
RC1: 'Comment on egusphere-2024-393', Anonymous Referee #1, 10 Apr 2024
Review of “A bias-corrected GEMS geostationary satellite product for nitrogen dioxide using machine learning to enforce consistency with the TROPOMI satellite instrument” (EGUsphere-2024-393) by Oak et al.
Recommendation: Minor Revision
Summary Statement: This article demonstrates a machine learning model can help to reduce the bias of GEMS geostationary satellite product of nitrogen dioxide compared to the TROPOMI product. This manuscript is well-written and presents a clear and concise approach to obtain bias-corrected GEMS product.
One concern is about the paragraph discussing the SHAP results (lines 195-200). While the contribution of different input variables to the model's performance is an important aspect, I would recommend delaying this discussion until after the performance of the ML model itself has been addressed. This would allow the reader to understand the overall effectiveness of the model before delving into the details of model and specific input variables.
Minor comment:
Line 248: It’s hard to understand the meaning of “ML correction increases the ocean background” Do you mean that ML correction increases product population over the ocean?
Citation: https://doi.org/10.5194/egusphere-2024-393-RC1 -
AC1: 'Reply on RC1', Yujin J. Oak, 16 May 2024
1. One concern is about the paragraph discussing the SHAP results (lines 195-200). While the contribution of different input variables to the model's performance is an important aspect, I would recommend delaying this discussion until after the performance of the ML model itself has been addressed. This would allow the reader to understand the overall effectiveness of the model before delving into the details of model and specific input variables.
We moved the paragraph to lines 220-229, after evaluating the model’s performance, and switched the order of figures accordingly.
2. Line 248: It’s hard to understand the meaning of “ML correction increases the ocean background” Do you mean that ML correction increases product population over the ocean?
We clarified the sentence to (lines 207-209):
“…GEMS product increases VCDs in the remote ocean background in the southeastern part of the GEMS scan domain by up to 200% and decreases VCDs in Central Asia by up to 40%, regardless of season.”
And also in lines 255-256:
“ML correction increases VCDs in the remote ocean regions by up to 200% and decreases VCDs in Central Asia by up to 40%.”
Citation: https://doi.org/10.5194/egusphere-2024-393-AC1
-
AC1: 'Reply on RC1', Yujin J. Oak, 16 May 2024
-
RC2: 'Comment on egusphere-2024-393', Anonymous Referee #2, 13 May 2024
Oak et al. present an interesting study in which they compare GEMS and TROPOMI NO2 total vertical columns. They find that by recalculating the AMF using consistent GEOS-CHEM profiles for GEMS and TROPOMI, the differences between GEMS and TROPOMI columns are greatly reduced. Furthermore, the comparison with PANDORA data is also improved by this step, both for TROPOMI and GEMS. Finally, they use a ML model to further improve the agreement between GEMS and TROPOMI columns. Their work shows how TROPOMI data can be correctly use as a transfer between the different geostationary instruments.
The results are clearly and honestly presented. It is appreciated that the comments addressed during the quick report have been included. I recommend publication after minor revisions. I would like to read more details/discussion about the points below.
1/ It is an interesting result to show that the main reason for the differences between GEMS and TROPOMI NO2 VCD lies in the AMF calculation. The relatively good agreement between the reprocessed columns shows that the NO2 SCD retrieval are consistent. Concerning the GEMS NO2 AMF, since the AK are taken from the GEMS CHOCHO L2 product, we cannot exclude another issue than a wrong use of the GEOS-Chem vertical coordinates.
P6, line 166: “much of the discrepancy in the L2 products stem from different vertical shape factors”. Please remind the reader that a large part could also come from an incorrect use of the vertical coordinates in the GEMS NO2 operational product.
2/ It is not shown that the ML model improves the diurnal variation comparison with the PANDORA (mainly from Figure 3). There is no evidence that including the TROPOMI VZA up to 50° actually helps to “build an ML model relevant to GEMS observations at different times of day”, as stated p6, line 185, in the abstract and in the conclusions. Please comment on the possibility to further improve the GEMS diurnal variation using ML technique.
Related to this point, it is not clear why the diurnal variation is more affected by the ML model during warm months than during cold months. I expect larger angles during cold months, and therefore a larger correction. Maybe it is because the days are longer during warm days?
3/ Figure 2: The GEMS NO2 columns seem to be cut for negative values. Is it an effect of the GEMS quality flags? This cutting effect seems to be amplified by the reprocessing and ML correction steps. The correlation is degraded from step 1 to 2. Have you tried to apply an improved quality filtering for GEMS? Or would it make sense to filter negative TROPOMI columns as well?
4/ Figure 3a: The corrected GEMS columns do agree better with PANDORA than the reprocessed GEMS columns. However, it is not obvious that they agree better with the reprocessed TROPOMI columns (it is the case for Jan/Feb, but not in May or June). It looks like the ML model tend to decrease the GEMS columns but has difficulties to increase them even when there is a negative difference with TROPOMI. Can you comment on this?
5/ Legend of figure 3: It should be mentioned explicitly that all the NO2 columns are total VCD, including the PANDORA columns.
Citation: https://doi.org/10.5194/egusphere-2024-393-RC2 - AC2: 'Reply on RC2', Yujin J. Oak, 16 May 2024
Interactive discussion
Status: closed
-
RC1: 'Comment on egusphere-2024-393', Anonymous Referee #1, 10 Apr 2024
Review of “A bias-corrected GEMS geostationary satellite product for nitrogen dioxide using machine learning to enforce consistency with the TROPOMI satellite instrument” (EGUsphere-2024-393) by Oak et al.
Recommendation: Minor Revision
Summary Statement: This article demonstrates a machine learning model can help to reduce the bias of GEMS geostationary satellite product of nitrogen dioxide compared to the TROPOMI product. This manuscript is well-written and presents a clear and concise approach to obtain bias-corrected GEMS product.
One concern is about the paragraph discussing the SHAP results (lines 195-200). While the contribution of different input variables to the model's performance is an important aspect, I would recommend delaying this discussion until after the performance of the ML model itself has been addressed. This would allow the reader to understand the overall effectiveness of the model before delving into the details of model and specific input variables.
Minor comment:
Line 248: It’s hard to understand the meaning of “ML correction increases the ocean background” Do you mean that ML correction increases product population over the ocean?
Citation: https://doi.org/10.5194/egusphere-2024-393-RC1 -
AC1: 'Reply on RC1', Yujin J. Oak, 16 May 2024
1. One concern is about the paragraph discussing the SHAP results (lines 195-200). While the contribution of different input variables to the model's performance is an important aspect, I would recommend delaying this discussion until after the performance of the ML model itself has been addressed. This would allow the reader to understand the overall effectiveness of the model before delving into the details of model and specific input variables.
We moved the paragraph to lines 220-229, after evaluating the model’s performance, and switched the order of figures accordingly.
2. Line 248: It’s hard to understand the meaning of “ML correction increases the ocean background” Do you mean that ML correction increases product population over the ocean?
We clarified the sentence to (lines 207-209):
“…GEMS product increases VCDs in the remote ocean background in the southeastern part of the GEMS scan domain by up to 200% and decreases VCDs in Central Asia by up to 40%, regardless of season.”
And also in lines 255-256:
“ML correction increases VCDs in the remote ocean regions by up to 200% and decreases VCDs in Central Asia by up to 40%.”
Citation: https://doi.org/10.5194/egusphere-2024-393-AC1
-
AC1: 'Reply on RC1', Yujin J. Oak, 16 May 2024
-
RC2: 'Comment on egusphere-2024-393', Anonymous Referee #2, 13 May 2024
Oak et al. present an interesting study in which they compare GEMS and TROPOMI NO2 total vertical columns. They find that by recalculating the AMF using consistent GEOS-CHEM profiles for GEMS and TROPOMI, the differences between GEMS and TROPOMI columns are greatly reduced. Furthermore, the comparison with PANDORA data is also improved by this step, both for TROPOMI and GEMS. Finally, they use a ML model to further improve the agreement between GEMS and TROPOMI columns. Their work shows how TROPOMI data can be correctly use as a transfer between the different geostationary instruments.
The results are clearly and honestly presented. It is appreciated that the comments addressed during the quick report have been included. I recommend publication after minor revisions. I would like to read more details/discussion about the points below.
1/ It is an interesting result to show that the main reason for the differences between GEMS and TROPOMI NO2 VCD lies in the AMF calculation. The relatively good agreement between the reprocessed columns shows that the NO2 SCD retrieval are consistent. Concerning the GEMS NO2 AMF, since the AK are taken from the GEMS CHOCHO L2 product, we cannot exclude another issue than a wrong use of the GEOS-Chem vertical coordinates.
P6, line 166: “much of the discrepancy in the L2 products stem from different vertical shape factors”. Please remind the reader that a large part could also come from an incorrect use of the vertical coordinates in the GEMS NO2 operational product.
2/ It is not shown that the ML model improves the diurnal variation comparison with the PANDORA (mainly from Figure 3). There is no evidence that including the TROPOMI VZA up to 50° actually helps to “build an ML model relevant to GEMS observations at different times of day”, as stated p6, line 185, in the abstract and in the conclusions. Please comment on the possibility to further improve the GEMS diurnal variation using ML technique.
Related to this point, it is not clear why the diurnal variation is more affected by the ML model during warm months than during cold months. I expect larger angles during cold months, and therefore a larger correction. Maybe it is because the days are longer during warm days?
3/ Figure 2: The GEMS NO2 columns seem to be cut for negative values. Is it an effect of the GEMS quality flags? This cutting effect seems to be amplified by the reprocessing and ML correction steps. The correlation is degraded from step 1 to 2. Have you tried to apply an improved quality filtering for GEMS? Or would it make sense to filter negative TROPOMI columns as well?
4/ Figure 3a: The corrected GEMS columns do agree better with PANDORA than the reprocessed GEMS columns. However, it is not obvious that they agree better with the reprocessed TROPOMI columns (it is the case for Jan/Feb, but not in May or June). It looks like the ML model tend to decrease the GEMS columns but has difficulties to increase them even when there is a negative difference with TROPOMI. Can you comment on this?
5/ Legend of figure 3: It should be mentioned explicitly that all the NO2 columns are total VCD, including the PANDORA columns.
Citation: https://doi.org/10.5194/egusphere-2024-393-RC2 - AC2: 'Reply on RC2', Yujin J. Oak, 16 May 2024
Peer review completion
Journal article(s) based on this preprint
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 436 | 144 | 32 | 612 | 19 | 29 |
- HTML: 436
- PDF: 144
- XML: 32
- Total: 612
- BibTeX: 19
- EndNote: 29
Viewed (geographical distribution)
| Country | # | Views | % |
|---|---|---|---|
| United States of America | 1 | 249 | 40 |
| South Korea | 2 | 85 | 13 |
| China | 3 | 57 | 9 |
| Germany | 4 | 39 | 6 |
| Japan | 5 | 31 | 5 |
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
- 249
Cited
Daniel J. Jacob
Nicholas Balasus
Laura H. Yang
Heesung Chong
Junsung Park
Hanlim Lee
Gitaek T. Lee
Eunjo S. Ha
Rokjin J. Park
Hyeong-Ahn Kwon
Jhoon Kim
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
- Preprint
(4045 KB) - Metadata XML