the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 19 Oct 2026
Global Hourly Aerosol Optical Depth Retrieval from DSCOVR/EPIC Observation and Its Application in Extreme Wildfire Events
Abstract. Extreme wildfire events can substantially perturb the global radiative balance and cause severe transboundary air pollution. Their aerosol plumes evolve rapidly on hourly timescales, yet low Earth orbit satellites have limited revisit frequency, while geostationary satellites have restricted hemispheric coverage. The Earth Polychromatic Imaging Camera (EPIC) aboard NOAA’s Deep Space Climate Observatory (DSCOVR), located at the Sun–Earth L1 point, provides continuous observations of the sunlit Earth disk with high temporal and unique aerosol sensitivity from ultraviolet to oxygen absorption bands. Here, we use an end-to-end deep learning framework for hourly global AOD retrieval using a parameter-efficient BatchEnsemble-based architecture, enabling efficient large-scale inference and quantitative uncertainty estimation. The framework integrates multispectral across UV, visible, and O₂ A/B bands with meteorological and topographic information. Against AERONET observations, the retrieval achieves R = 0.909, InEE = 90.91%, MAE = 0.039, and RMSE = 0.066 in sample-based validation, and maintaining robust performance under spatial and interannual extrapolation, with InEE consistently exceeding 74%. SHAP analysis reveals physically meaningful feature contributions, with UV bands contributing 58.9%, followed by visible bands (18.7%), spatiotemporal features (10.0%), meteorological parameters (6.7%), and O₂ A/B bands (4.7%). The retrieval also outperforms operational EPIC products and provides substantially more valid retrievals. Case studies over Canada, Australia, and the western United States demonstrate the retrieved hourly AOD captures the intraday evolution and long-range transport of wildfire smoke. These results highlight the potential of DSCOVR/EPIC and deep learning for continuous global monitoring of rapidly evolving aerosol events.
- Preprint
(55007 KB) - Metadata XML
-
Supplement
(24677 KB) - BibTeX
- EndNote
Status: open (until 30 Nov 2026)
-
CC1: 'Comment on egusphere-2026-5788', Alexei Lyapustin, 25 Oct 2026
reply
The comment was uploaded in the form of a supplement: https://egusphere.copernicus.org/preprints/2026/egusphere-2026-5788/egusphere-2026-5788-CC1-supplement.pdfReplyCitation: https://doi.org/
10.5194/egusphere-2026-5788-CC1 -
AC1: 'Reply on CC1', Xiaohu Sun, 04 Nov 2026
reply
Thank you for your careful review of our manuscript and your valuable comments. Please find our point-by-point responses in the attached document.
-
AC1: 'Reply on CC1', Xiaohu Sun, 04 Nov 2026
reply
-
RC1: 'Comment on egusphere-2026-5788', Anonymous Referee #3, 03 Nov 2026
reply
The manuscript entitled “Global Hourly Aerosol Optical Depth Retrieval Based on DSCOVR/EPIC Observations and Its Application to Extreme Wildfire Events” presents a deep learning retrieval framework based on the TabM architecture to retrieve global aerosol optical depth (AOD) at 550 nm using multispectral observations from DSCOVR/EPIC. The retrieval results are validated against AERONET observations, and three representative wildfire events in Canada, Australia, and the western United States are further investigated. By exploiting the high-frequency full-disk observations from EPIC for global AOD retrieval, this study offers a certain degree of novelty and has potential value for investigating the diurnal variations in aerosols and extreme wildfire events. However, several issues remain regarding the reliability of comparisons with existing products, retrieval performance under high aerosol loading conditions, spatiotemporal generalization capability, the physical rationale for the input features, and the physical interpretation of the wildfire case studies. These issues require further analysis and clarification before the manuscript can be considered for publication.
Major Comments
- Lines 197–204: The authors report correlation coefficients of 0.458 and 0.387 between the EPIC UV and EPIC MAIAC products and AERONET observations, respectively. However, these results differ substantially from those reported in previous validation studies of the EPIC MAIAC product. For example, Lyapustin et al. (2021b) conducted a global validation and reported correlation coefficients of approximately 0.75–0.77 for the MAIAC EPIC product, which are considerably higher than those presented in this study. Although differences in product versions, validation wavelengths, and sample screening criteria may partly explain this discrepancy, such a substantial difference requires further clarification. In addition, Fig. 2 uses AOD at 500 nm for the EPIC UV product, whereas the proposed model and the AERONET reference data use AOD at 550 nm (Lines 90–95 and 209–213). The authors should clarify whether a wavelength conversion was applied to ensure consistency among the compared datasets. I recommend that the authors carefully re-examine the product versions, quality control procedures, and collocation methods, and reassess the performance of the different products under validation conditions that are as consistent as possible. This would help avoid overstating the advantages of the proposed method over existing retrieval algorithms.
- Lines 284–289 and Fig. 7a: The retrieval errors increase substantially with increasing AOD, with the error distribution becoming particularly broad when AOD ≥ 1. Meanwhile, the sample distribution shown in Fig. 7a is highly imbalanced: 141,281 samples fall within the AOD range of 0–0.1, whereas only 99 samples have AOD values greater than 2.0. Although the authors attribute the increased errors under high-AOD conditions to the imbalance in training samples and radiative transfer saturation, the specific effects of these factors on model performance have not been sufficiently explained. It is also unclear whether the sample distribution shown in Fig. 7a is consistent with the actual distribution of the training dataset. Furthermore, Lines 290–293 indicate that the model tends to underestimate AOD under coarse-mode-dominated conditions (AE < 0.5) and overestimate AOD under pure fine-mode aerosol conditions (AE > 2.0). Meanwhile, Lines 272–275 show that the RMSE values in regions such as India, North America/the Middle East, and Southeast Asia reach approximately 0.2 in the spatial extrapolation validation. These results suggest that the retrieval capability of the model under high-AOD conditions, extreme aerosol size regimes, and complex regional atmospheric conditions requires further examination. I recommend that the authors provide a clearer description of the actual training data distribution and conduct stratified validation or other targeted experiments to evaluate the retrieval performance and stability of the model under these specific conditions.
- Lines 165–169: The authors employ random sample splitting to conduct 10-fold cross-validation. However, given the pronounced spatiotemporal correlations in satellite–ground matched datasets, this splitting strategy may allow observations from the same AERONET site and temporally adjacent measurements to be simultaneously included in the training and validation sets, potentially leading to an overestimation of the model's predictive performance. Although the authors further conduct site-independent, temporally independent, and spatially independent validations, the independent-year validation results are considerably poorer than those obtained from random sample cross-validation. In particular, the correlation coefficient for 2023 is only 0.700, with an RMSE of 0.113 (Lines 276–278 and Supplementary Table S1). Although the sample compositions and evaluation objectives of these validation schemes are not entirely equivalent, this discrepancy still indicates a certain degree of performance degradation when the model is applied across different years. In addition, Lines 172–173 state that 2020 and 2023 were alternately used for training and validation, but the specific data partitions and model training procedures for these two experiments are not clearly described. It is worth noting that some regions experienced COVID-19-related lockdowns and associated changes in anthropogenic emissions in 2020, whereas large-scale wildfire events occurred in Canada and other regions in 2023. These differences may have resulted in substantially different aerosol backgrounds and loading distributions between the two years. Therefore, validation results based solely on these two relatively unusual years may be insufficient to demonstrate the model's general interannual generalization capability. I recommend that the authors further clarify the composition of the training and testing datasets for each validation scheme, analyze the differences in aerosol backgrounds among different years, and, if possible, include additional independent-year validations covering non-pandemic years and years without extreme wildfire events. This would provide a more objective assessment of the model's spatiotemporal generalization capability.
- Lines 45–51: The authors state that the EPIC ultraviolet bands are sensitive to absorbing aerosols, while the O₂ A/B bands exhibit sensitivity to the vertical distribution of aerosols. Lines 78–83 further indicate that all 10 EPIC spectral channels are used as model inputs, including the 317 and 325 nm channels, which contain absorption information related to ozone, SO₂, and other atmospheric constituents. However, Lines 90–95 indicate that the model is primarily trained using column-integrated AOD from AERONET as the supervised target, without incorporating corresponding constraints on aerosol absorption properties or vertical distribution. Therefore, it remains unclear whether these additional spectral features effectively improve AOD retrieval or potentially introduce interference associated with other atmospheric constituents and variations in aerosol vertical structure. Furthermore, Lines 97–103 indicate that the authors introduce several auxiliary variables, including DEM and ERA5-derived 10 m wind fields, surface pressure, planetary boundary layer height, total column ozone, and total column water vapor. However, the specific physical roles and individual contributions of these variables have not been sufficiently explained, nor have corresponding feature-selection or sensitivity analyses been conducted. For example, potential information redundancy between surface pressure and terrain elevation, as well as between total column ozone and the ultraviolet spectral channels, deserves further discussion. Meanwhile, in Lines 215–232, the authors mainly rely on SHAP feature contribution scores and their consistency with the expected physical mechanisms to argue that the model has good physical interpretability. However, feature contributions alone cannot directly demonstrate that the model has learned the underlying radiative transfer mechanisms.
- Lines 305–360: The authors select three representative wildfire events in Canada, Australia, and the western United States to demonstrate the capability of the model to monitor the spatiotemporal evolution of smoke plumes and their long-range transport. However, the current analysis focuses mainly on the spatial distribution of AOD, its diurnal variations, and comparisons with other products, while providing insufficient quantitative analysis of the meteorological processes and boundary layer effects underlying plume evolution. For example, Lines 335–338 attribute the diurnal variation in AOD during the Australian wildfire event to surface heating and boundary layer turbulence, but this interpretation is not supported by an analysis of meteorological variables such as boundary layer height. Lines 350–356 use wind fields and CAMS trajectories to investigate the transcontinental transport of the smoke plume from the western United States, but the vertical transport and evolution of the plume are discussed only to a limited extent. I recommend that the authors select at least one representative case and conduct further quantitative analysis of plume diurnal variations and transport processes by combining ERA5 wind fields, boundary layer height, and other relevant meteorological data. This would strengthen the physical interpretation and scientific significance of the case studies.
Citation: https://doi.org/10.5194/egusphere-2026-5788-RC1 -
RC2: 'Comment on egusphere-2026-5788', Anonymous Referee #2, 05 Nov 2026
reply
This manuscript presents a novel deep-learning framework for retrieving hourly aerosol optical depth (AOD) from DSCOVR/EPIC observations at the Sun–Earth L1 Lagrange point. The results demonstrate promising retrieval accuracy compared with existing operational EPIC AOD products, while also providing uncertainty estimates. These findings highlight the potential of the proposed approach for monitoring aerosol dynamics, particularly during extreme wildfire events. However, several methodological concerns remain, particularly regarding the robustness of the retrieval framework, the rigor of the experimental evaluation, and the transparency of the methodology. These issues should be adequately addressed before the manuscript can be considered suitable for publication. Specific comments are as follows:
-
Cross-validation and independent validation serve different purposes in assessing model performance, with the latter generally providing a more realistic assessment of generalization capability. Therefore, is it appropriate to compare the proposed method with existing operational AOD products primarily based on sample-based 10-fold cross-validation results? Such comparisons may lead to an overly optimistic assessment of retrieval performance. A more rigorous comparison using the same independent validation dataset and consistent evaluation criteria would be more convincing.
-
The authors state that the proposed model can provide prediction uncertainty estimates, which is a potentially valuable aspect of the framework. However, the manuscript does not sufficiently explain how these uncertainties are quantified or evaluate their reliability. Moreover, the estimates may be sensitive to the number of sub-models (k). The authors should provide a more comprehensive uncertainty evaluation, including the relationship between estimated uncertainties and actual retrieval errors, and clarify the rationale for selecting k and its influence on the uncertainty estimates.
-
Regarding the deep-learning framework, the concept of “independent weight matrices” in the linear ensemble layer at the output stage is not clearly explained. What is their specific role, and what advantages do they provide? In addition, the authors should clarify how the other model hyperparameters were determined and whether their selection was supported by systematic optimization or sensitivity analyses.
-
Although the manuscript highlights the unique observational geometry of DSCOVR/EPIC at the Sun–Earth L1 point, the role of its specific spectral bands, particularly the UV and O₂ A/B bands, in AOD retrieval is not sufficiently discussed. The authors should clarify the physical relevance of these bands and their contributions to the proposed retrieval framework.
-
The abstract should provide a balanced presentation of the strengths and limitations of the proposed method. I suggest reporting both random cross-validation and independent validation results to provide a more comprehensive assessment of retrieval performance and ensure a fair comparison with existing products.
-
Line 40: Please discuss the limitations of integrating observations from multiple geostationary satellites for global aerosol monitoring, including the GEO-Ring constellation, and clarify the specific advantages of DSCOVR/EPIC observations in this context.
-
Line 69: The term “continuous” should be used with caution, as this study focuses exclusively on clear-sky conditions. The authors should clarify the distinction between hourly temporal resolution and the actual spatial and temporal coverage of valid AOD retrievals.
-
Line 70: Briefly introduce the proposed deep-learning framework, including its key components and methodological innovations, to better highlight its advantages over existing approaches.
-
Materials: Provide a more detailed description of the datasets used in this study, including their spatial and temporal resolutions, spatiotemporal coverage, and the spectral bands selected for AOD retrieval, along with the rationale for their selection. In addition, were any quality control procedures applied based on observation geometry, such as solar and viewing zenith angles?
-
“3.2 Deep Learning Method” should be “3.2 Deep Learning method”.
-
Line 160: “10-4” should be formatted as “10⁻⁴”.
-
Figure 5: I suggest masking ocean areas in the MAIAC AOD maps to avoid potential misinterpretation.
-
Regarding the application to extreme wildfire events, the current evaluation relies primarily on visual comparisons. Are temporally matched AERONET observations available for these events? Including time-series comparisons with AERONET measurements would provide more quantitative evidence of the model’s ability to capture aerosol variations during extreme wildfire events.
-
Figure S3 in the Supplementary Material refers to “time-based 10-fold cross-validation”, whereas the Methods section describes a month-based 12-fold cross-validation strategy. Please clarify this inconsistency.
-
Line 212: “AEROENT” should be “AERONET”.
Citation: https://doi.org/10.5194/egusphere-2026-5788-RC2 -
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 371 | 323 | 33 | 727 | 81 | 35 | 35 |
- HTML: 371
- PDF: 323
- XML: 33
- Total: 727
- Supplement: 81
- BibTeX: 35
- EndNote: 35
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
