Preprints
https://doi.org/10.5194/egusphere-2026-1987
© Author(s) 2026. This work is distributed under
the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
https://doi.org/10.5194/egusphere-2026-1987
© Author(s) 2026. This work is distributed under
the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
04 Jun 2026
| 04 Jun 2026
Twin Eyes in the Sky: Deep Learning-Based AOD Enhancement Using GOES-East and GOES-West
Abstract. High spatio-temporal resolution aerosol monitoring is critical to understand and mitigate air pollution and climate change. In this context, geostationary satellite instruments can be extremely beneficial, allowing fine-grained temporal characterization of aerosols over large regions. In this study, we combine data from the geostationary instruments Advanced Baseline Imager (ABI) on-board GOES-East and GOES-West, using Deep Learning methods to post-process NASA Dark Target ABI AOD and NOAA ABI AOD products and improve their accuracy and spatial resolution. We deploy a Transformer Encoder architecture, and compare it to a Multi Layer Perceptron (MLP) architecture predicting at single time step, showing how exploiting the temporal patterns in geostationary daily observations leads to improved accuracy and generalization in the post-process correction. Additionally, we show that further improvement can be obtained combining multi-view angles from different (though very similar) geostationary satellites. Our region of interest is the Contiguous United States (CONUS) in the years 2020–2022.
How to cite. Porcheddu, A., Kolehmainen, V., Lähivaara, T., Levy, R., Shi, Y. R., Zhang, H., and Lipponen, A.: Twin Eyes in the Sky: Deep Learning-Based AOD Enhancement Using GOES-East and GOES-West, EGUsphere [preprint], https://doi.org/10.5194/egusphere-2026-1987, 2026.
Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.
Download & links
Download & links
- Preprint
(4890 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
Comment types: AC – author | RC – referee | CC – community | EC – editor | CEC – chief editor
| : Report abuse
-
RC1: 'Comment on egusphere-2026-1987', Anonymous Referee #1, 21 Jul 2026
The comment was uploaded in the form of a supplement: https://egusphere.copernicus.org/preprints/2026/egusphere-2026-1987/egusphere-2026-1987-RC1-supplement.pdfCitation: https://doi.org/
10.5194/egusphere-2026-1987-RC1 -
RC2: 'Reply on RC1', Anonymous Referee #2, 25 Sep 2026
The manuscript presents a framework for improving geostationary satellite AOD products using deep learning, particularly by integrating GOES-16/17 observations and exploiting temporal information through a Transformer-based model. The study is potentially valuable for high-frequency aerosol monitoring and demonstrates the potential of deep learning approaches for satellite aerosol product improvement. However, several aspects require further clarification to better establish the scientific contribution and broader implications of the study.
- The contribution of the manuscript appears to be primarily an application of advanced deep learning architectures for AOD post-processing. The authors should better clarify the specific scientific limitations of existing GOES ABI AOD retrieval algorithms that the proposed framework aims to address. In particular, it remains unclear whether the model mainly corrects uncertainties associated with surface reflectance estimation, aerosol model assumptions, cloud contamination, viewing geometry effects, or temporal sampling limitations. Furthermore, the authors should provide additional analysis of what physical relationships or aerosol-related factors are learned by the network, in order to improve the interpretability of the deep learning approach and demonstrate its scientific relevance beyond improvements in statistical metrics.
- The manuscript claims improvements in both AOD accuracy and spatial resolution; however, the nature of the spatial enhancement requires further clarification. The NOAA ABI AOD product has a native spatial resolution of 2 km, while the NASA Dark Target ABI AOD product is provided at 10 km resolution with 10 km aggregation. In contrast, the proposed models operate on a 500 m grid. It is therefore unclear whether the output represents a genuine downscaled AOD product containing new sub-pixel spatial information, or whether it is primarily a bias-corrected post-processing product mapped onto a finer grid. The authors should clarify how the 500 m spatial variability is introduced, what additional information enables the model to provide finer-scale AOD estimates, and whether the resulting products should be interpreted as downscaled retrievals or resolution-enhanced corrections.
- More information is needed regarding the distribution of training and testing samples, as AOD observations are typically highly imbalanced, with a large number of low-AOD cases and relatively fewer high-AOD events. Such imbalance may affect model training and limit the ability of the model to generalize under extreme aerosol conditions. The authors should report the AOD distribution of the datasets, the number of samples across different aerosol loading ranges, and the model performance under different AOD conditions. This analysis is particularly important because the manuscript reports underestimation at high AOD levels, which may be related to insufficient representation of high-aerosol events in the training data.
- The manuscript concludes that exploiting temporal correlations through the Transformer architecture improves AOD retrieval accuracy; however, the benefits of temporal information may depend on aerosol variability. The authors should further evaluate the model performance during rapidly changing aerosol events, such as dust outbreaks, wildfire smoke episodes, and severe pollution events, where aerosol loading can change substantially over short time periods. Such analysis would help determine whether the temporal learning capability of the Transformer provides robust improvements under dynamic atmospheric conditions or whether it relies primarily on relatively stable aerosol temporal patterns.
Citation: https://doi.org/10.5194/egusphere-2026-1987-RC2
-
RC2: 'Reply on RC1', Anonymous Referee #2, 25 Sep 2026
Viewed
Total article views: 251 (including HTML, PDF, and XML)
Cumulative views and downloads
(calculated since 04 Jun 2026)
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 152 | 62 | 37 | 251 | 26 | 22 |
- HTML: 152
- PDF: 62
- XML: 37
- Total: 251
- BibTeX: 26
- EndNote: 22
Viewed (geographical distribution)
Total article views: 227 (including HTML, PDF, and XML)
Thereof 227 with geography defined
and 0 with unknown origin.
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
1
Latest update: 08 Oct 2026
Andrea Porcheddu
Department of Technical Physics, University of Eastern Finland, Kuopio, Finland
Ville Kolehmainen
Department of Technical Physics, University of Eastern Finland, Kuopio, Finland
Timo Lähivaara
Department of Technical Physics, University of Eastern Finland, Kuopio, Finland
Robert Levy
NASA Goddard Space Flight Center, Greenbelt, MD, USA
Yingxi Rona Shi
NASA Goddard Space Flight Center, Greenbelt, MD, USA
University of Maryland Baltimore County, Baltimore, MD, USA
Hai Zhang
Science and Technology Corporation at NOAA, College Park, MD, USA
Finnish Meteorological Institute, Atmospheric Research Centre of Eastern Finland, Kuopio, Finland
Short summary
We developed a new way to improve satellite measurements of atmospheric aerosols, which affect air quality and climate. By combining data from two geostationary satellites and using advanced machine learning methods that capture changes over time, we achieved more accurate and detailed results. This approach reduces errors in current satellite-based aerosol data and can better support monitoring and decision making related to air pollution and climate impacts.
We developed a new way to improve satellite measurements of atmospheric aerosols, which affect...