the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
AOD data fusion with Geostationary Korea Multi-Purpose Satellite (Geo-KOMPSAT) instruments GEMS, AMI, and GOCI-II: Statistical and deep neural network methods
Abstract. Aerosol optical depth (AOD) data fusion of aerosol datasets from the Geostationary Korea Multi-Purpose Satellite (GEO-KOMPSAT, GK) series was undertaken using both statistical and deep neural network (DNN)-based methods. The GK mission includes an Advanced Meteorological Imager (AMI) onboard GK-2A and a Geostationary Environment Monitoring Spectrometer (GEMS) and Geostationary Ocean Color Imager-II onboard GK-2B. The statistical fusion method corrected the bias of each aerosol product by assuming a Gaussian error distribution. The Maximum Likelihood Estimation (MLE) fusion technique accounted for pixel-level uncertainties by weighting the root-mean-square error of each AOD product for every pixel. A DNN-based fusion model was trained to target Aerosol Robotic Network AOD values using fully connected hidden layers. The statistical and DNN-based fusion results generally outperformed individual GEMS and AMI AOD datasets in East Asia (R = 0.888; RMSE = −0.188; MBE = −0.076; 60.6 % within EE for MLE AOD; R = 0.905; RMSE = 0.161; MBE = −0.060; 65.6 % within EE for DNN AOD). The selection of AOD around Korean peninsula, which is incorporating all aerosol products including GOCI-II resulted in much better results (R = 0.911; RMSE = 0.113; MBE = −0.047; 73.3 % within EE for MLE AOD; R = 0.912; RMSE = 0.102; MBE = −0.028; 78.2 % within EE for DNN AOD). The DNN AOD effectively addressed the rapid increase in uncertainty at higher aerosol loadings. Overall, fusion AOD (particularly DNN AOD) most closely matched the performance of the Moderate Resolution Imaging Spectroradiometer Dark Target algorithm, with slightly less variance and a negative bias. Both fusion algorithms stabilized diurnal error variations and provided additional insights into hourly aerosol evolution. The application of aerosol fusion techniques to future geostationary satellite projects such as TEMPO, ABI, and GeoXO may facilitate the production of high-quality global aerosol data.
- Preprint
(3996 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
-
CC1: 'Comment on amt-2023-255', Ding Li, 18 Dec 2023
This document employs two distinct methods, Maximum Likelihood Estimation (MLE) and Deep Neural Networks (DNN), to integrate pixel-level uncertainties in the fusion of three Aerosol Optical Depth (AOD) products. The result is an improved hourly AOD dataset compared to individual AOD products, evident in its superior validation against ground-based AERONET AOD over East Asia. The approach meticulously addresses potential sources of errors and various challenges, positioning the resulting dataset to provide high-precision hourly Aerosol Optical Depth (AOD) products. Moreover, the article maintains a coherent logical flow and is enriched by visually appealing charts, rendering it an outstanding paper.
Major comments:
- The amount of data is a barrier to machine learning models. Can we consider all types when there are not many AERONET sites in eastern Asia? It would be interesting to see a discussion on how the model performance varied with different volumes of data. Did the model performance improve with more data? If the data volume was a limitation in this study, it would be worth discussing how future work could overcome this. Are there plans to gather more data or use techniques like data augmentation?
- For MLE (235 line): Based on this analysis, the bias of each AOD product was subtracted according to the NDVI value, selected aerosol type, and observation time. For DNN (245 line): This involved standardization of the NDVI, hour, and aerosol type index. Both models have taken into account the parameters mentioned above. However, in the figures presented in this document, there is an absence of accuracy analysis for diverse parameter combinations, with the focus solely on overall data analysis. Enhancing the article's significance can be achieved by incorporating analysis results for various parameter combinations and providing explanations for the observed outcomes.
Citation: https://doi.org/10.5194/amt-2023-255-CC1 -
AC3: 'Reply on CC1', Jhoon Kim, 21 Feb 2024
The comment was uploaded in the form of a supplement: https://amt.copernicus.org/preprints/amt-2023-255/amt-2023-255-AC3-supplement.pdf
-
RC1: 'Comment on amt-2023-255', Anonymous Referee #1, 07 Jan 2024
The comment was uploaded in the form of a supplement: https://amt.copernicus.org/preprints/amt-2023-255/amt-2023-255-RC1-supplement.pdf
-
AC1: 'Reply on RC1', Jhoon Kim, 21 Feb 2024
The comment was uploaded in the form of a supplement: https://amt.copernicus.org/preprints/amt-2023-255/amt-2023-255-AC1-supplement.pdf
-
AC1: 'Reply on RC1', Jhoon Kim, 21 Feb 2024
-
RC2: 'Comment on amt-2023-255', Anonymous Referee #2, 19 Jan 2024
Overview
The paper compares AOD products from the GEMS, AMI, and GOCI-II instruments aboard the GEO-KOMPSAT and two fusion products of the single instrument retrievals with AOD AERONET observations and with MODIS DT data. The fusion products are optimised to match the AERONET observations using a deep neural network (DNN) or a maximum likelihood estimate (MLE). The error analysis is detailed and distinguishes between different AOD loads, NDVI, observation time and aerosol types. The authors find that the GOCI-II retrievals have the lowest error of the one-instrument retrievals and that the fused products using DNN has overall the smallest errors.
General remarks
The paper is very detailed and provides a lot of quantitative information about the error of the evaluated AOD products. But, the paper should provide more scientific information to help the reader to understand or interpret short-comings of the products or the choices for the fusion approach. For example, the choice of NDVI (and not other candidate parameters) as predictor of error needs to be discussed in more detail. Likewise, the determination of the aerosol type should be better explained.
The two fusion products have smaller error against AERONET observations than the single-instrument retrievals. This is perhaps not surprising because the fusion approaches were designed to match the AERONET observations and a prior bias correct of the single-instruments retrieval was performed. Such a correction procedure could also be applied to the individual satellite data sets. So, it remains unclear if the added benefit of the fusion approach is the AERONET-based error correction or the synergistic benefits of the MLE or DNN based methods to merge the products.
The paper uses a lot of acronyms for different versions of the retrievals and it is difficult for the reader to follow. For improved readability I suggest 1) to spell out more of the acronyms in the figure captions, 2) to add a table that that summarises the data sets and 3) to add a table that summarised the error measures (bias, RMSE etc) for all considered single or fused products to give a better overview of the accuracy.
Specific comments:
L 140 Please provide more detail on the aerosol type classification. Why is GEMS not affected by the misclassification of the type?
L 159 How is the aerosol type derived?
L 164 Please comment on the differences and biases between AERONET version 3 level 2 and level 1.5
L 178 Please motivate better the choice of NDVI. Fig 5, 6 and 7 (b) do not show a distinct relation between NDVI and error.
L 225 What is the procedure if an instrument product has no data?
L 226 All retrievals come from the same satellite. So, index i should represent the instrument or product.
L 235 From which instrument was the aerosol type obtained?
L 245 This type classification should be explained earlier.
L 265-280 It remains unclear what type of cloud masking was applied for the different products and if all problems related to cloud masking could be resolved. Please provide a summary in this section.
L 287 Please provide more detail on EE. What is its purpose? What is tau. Why does it make sense to use the MODIT DT approach here.
L 343 Why were only the fused data processed and evaluated for the two different domains EA and KO? Which was the domain for the single instrument data?
L354 “The statistical fusion approach thus effectively accommodated nonlinearity in retrieval uncertainty, despite possibly not capturing all complexity in the data.” Please explain better what you mean. The fused data have the advantage of being optimized to match the AERONET data.
L 361-415 This section is perhaps to detailed and complicated. It would be sufficient to simply compare the AOD products and MODIS DT against AERONET and compare the errors for the different situations or locations. It remains slightly unclear if the new fusion products have smaller errors than the MODIS DT retrievals over the study area.
L 363 please explain the retrieval error. Is that the “theoretical” retrieval error provided by the retrieval algorithm or the error of the product against AERONET. How is the theoretical retrieval error of the fused data set calculated.
Figures:
Please include the statistical error measure of Fig 4 and Fig 8 in a table.
It is not obvious what Fig 9 shows.
Tables:
See general comment
Citation: https://doi.org/10.5194/amt-2023-255-RC2 -
AC2: 'Reply on RC2', Jhoon Kim, 21 Feb 2024
The comment was uploaded in the form of a supplement: https://amt.copernicus.org/preprints/amt-2023-255/amt-2023-255-AC2-supplement.pdf
-
AC2: 'Reply on RC2', Jhoon Kim, 21 Feb 2024
Viewed
HTML | XML | Total | BibTeX | EndNote | |
---|---|---|---|---|---|
315 | 107 | 26 | 448 | 18 | 16 |
- HTML: 315
- PDF: 107
- XML: 26
- Total: 448
- BibTeX: 18
- EndNote: 16
Viewed (geographical distribution)
Country | # | Views | % |
---|
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1