the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 21 Apr 2026
An AI Based Algorithm for Retrieving Aerosol Optical Depth and Single Scattering Albedo Using All-Sky Imager Observations
Abstract. Accurate measurement of aerosol optical properties is critical for understanding their radiative and environmental impacts. Currently, the most accurate retrieval of aerosol properties comes from the multi-channel surface sun photometer, but with relatively high cost and deployment/maintenance requirements. Here we develop a novel AI based method for retrieving daytime aerosol optical parameters, namely aerosol optical depth (AOD) and single scattering albedo (SSA) using images acquired by All-Sky Imagers (ASI). Surface based AOD and SSA retrievals from surface sun photometers are used as the training targets. Algorithm training and retrievals were performed for two sites in East China and Central US respectively. Independent validation against ground-based measurements demonstrated high consistency between the ASI-retrieved and sun photometer measured AOD and SSA, with Pearson correlation coefficients (r) exceeding 0.86 for AOD across all wavelengths at both sites and Root Mean Square Errors (RMSE) below 0.25. For SSA, r values reached 0.67 at the Beijing_PKU site and 0.84 at the SGP site, with RMSE remaining below 0.09 across all spectral channels, demonstrating the feasibility of simultaneous AOD and SSA retrieval from low-cost all-sky imagers. This method not only overcomes the high computational cost associated with traditional radiative transfer iterative algorithms, but also provides great potential for denser surface aerosol measurements by leveraging the low-cost and easy-maintenance advantages of the all-sky imager.
Competing interests: The authors declare that a patent application related to the method described in this manuscript has been filed by Peking University (Application No. CN202511448418.3). The authors have no other competing interests.
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.- Preprint
(1920 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2026-2019', Anonymous Referee #1, 09 Jun 2026
-
AC1: 'Reply on RC1', Jing Li, 05 Aug 2026
Major comments:
1. While the method has been applied to two independent sites, Beijing and SGP, I noticed different versions of AERONET data were used. Specifically, the authors state that Level 2.0 SSA retrievals at Beijing are too sparse due to strict quality control, so they use Level 1.5 instead. But Level 1.5 SSA certainly has larger uncertainty, the authors should explain how this uncertainty impact the retrieved SSA results, and if possible, provide some of the retrieved SSA uncertainty.
Response: Thank you for pointing this out. We have clarified that AERONET Version 3 Level 2.0 SSA inversion products are not available for the period overlapping with the Beijing_PKU ASI observations, rather than being merely sparse or unavailable for the site in general. This is mainly because the post-deployment calibration required for fully quality-assured Level 2.0 inversion products could not be completed in time, due to pandemic-related restrictions on international shipment and calibration of the instrument. Therefore, Beijing SSA retrievals were trained and validated against Version 3 Level 1.5 cloud-screened inversion products. We agree that Level 1.5 SSA has larger uncertainty than Level 2.0.
Furthermore, we performed a Monte Carlo sensitivity experiment by adding Gaussian random perturbations to the AOD and SSA training labels. The results show that SSA retrievals are more sensitive to the assumed label uncertainty than AOD retrievals. For SSA, the propagated prediction uncertainty increased from 0 to 0.0130 and 0.0214 under the low- and high-uncertainty scenarios, respectively. Correspondingly, SSA RMSE increased from 0.0318 to 0.0331 and 0.0384, while the correlation decreased from 0.608 to 0.580 and 0.489. In contrast, AOD RMSE and correlation remained nearly unchanged, indicating greater robustness of AOD retrievals to the assumed label uncertainty.
Figure S2. Sensitivity of Beijing_PKU AOD and SSA retrievals to assumed AERONET label uncertainty at 675 nm. Gaussian perturbations were added to the AOD and SSA training labels under three scenarios: Baseline (σSSA = 0, σAOD = 0), Low uncertainty (σSSA = 0.03, σAOD = 0.02), and High uncertainty (σSSA = 0.06, σAOD = 0.04). The model was retrained 100 times for each scenario. The first column shows the mean standard deviation of the predicted values, representing the propagated retrieval uncertainty. The remaining columns show the mean RMSE, MAE, and Pearson correlation coefficient across the Monte Carlo runs, with error bars indicating ±1 standard deviation. The perturbation amplitudes are used as assumed uncertainty scenarios rather than official AERONET Level 1.5 or Level 2.0 uncertainty values.
To reduce the potential uncertainty of the Level 1.5 SSA records, this revised version has further applied additional quality-control criteria to the Beijing_PKU inversion data. Specifically, only SSA retrievals with solar zenith angle greater than 50° and sky residual less than 5% were retained. These criteria remove less stable sky-radiance inversion cases associated with unfavorable observation geometry or relatively large fitting residuals, for which SSA retrievals are expected to be more uncertain. We have revised Section 2.2 to make this quality-control procedure explicit. We also clarify that the Beijing SSA validation should be interpreted as agreement with additionally quality-screened Level 1.5 AERONET inversion products, rather than with fully quality-assured Level 2.0 products.
2. The practice to simultaneous retrieve AOD and SSA from ASI is indeed novel. Butthe SSA performance—especially at Beijing (r as low as 0.33 at 1020 nm, Fig. 4 and Table 1)—is relatively weak. I wonder how this performance compares with existing SSA products, perhaps from satellites or other remote sensing techniques? Such comparison is important in understanding the capability and limitions of ASI in retrieving aerosol properties.
Response: We agree that comparison with existing SSA products is important for interpreting the capability and limitations of the proposed ASI approach. So we have added a comparison with representative ground-based and satellite-based SSA retrieval products in Section 3.1. The revised manuscript compares the reported correlation coefficients and RMSE values with AERONET, PARASOL/POLDER-GRASP, GEMS, and Gaofen-5 DPC products. We also clarify that these comparisons should not be interpreted as evidence that the ASI retrieval outperforms established products, because the products differ in wavelength, viewing geometry, aerosol loading, spatial representativeness, and reference-data quality. Instead, the results suggest that ASI observations may provide complementary, low-cost, high-temporal-resolution local SSA information.
Minor Comments
1.Inconsistent figure labeling in caption of Figure 1
The caption for Figure 1 lists (a)-(d) but the text refers to (a-c) and (d-f) in the following paragraph. Please check and harmonize the labels. Also, the source attribution for (c,d) is given, but not for (a,b) – add "source: authors" for clarity.
Response: We have corrected the panel labels in the text and caption of Figure 1. We have also added source attribution for panels (a) and (b), which are images from our team’s own all-sky imager observations at the Beijing_PKU site. The source attribution for panels (c) and (d), which are from the ARM user facility, has been retained.
2.Lack of detail on the cloud and lens flare masking algorithm
Section 2.3 mentions "iterative threshold segmentation" (citing Anon, 1978; Gonzalez and Faisal, 2019) but does not specify the actual threshold criterion (e.g., Otsu? fixed percentile?). Given that cloud masking quality directly affects sky radiance extraction, please provide the specific algorithm or pseudo-code, or at least cite a more recent and relevant ASI cloud masking method.
Response: We have added a detailed description of the cloud and lens-flare masking procedure in Section 2.3. specifies the solar-region detection procedure, mask radius, initial threshold, iterative threshold-update rule, and convergence criterion. We have also replaced the former “Anon, 1978” citation with the complete reference to Ridler and Calvard (1978).
The revised text in section 2.3: “Clean-Sky Image Processing: Upon acquiring time-stamped raw sky images, we first mask areas that interfere with the extraction of cloud-free sky radiance—namely buildings, clouds, and glare spots caused by the sun and lens artifacts. The specific steps are:
(a) Masking buildings at the image periphery.
(b) Identifying and masking the brightest circular region (sun and its immediate surroundings): the solar center was identified from the grayscale image using an integral-image-based search. A square window with a radius of 45 pixels and a step size of 5 pixels was used to locate the brightest region, and a circular mask with a radius of 220 pixels was applied around the detected solar center. For the TSI images, an additional fixed circular/elliptical mask was applied to remove the instrument support structure.
(c) Cloud and lens flare masking: all-sky images were first converted to grayscale for masking. Cloud-contaminated and other bright regions were removed using an iterative thresholding scheme (Ridler and Calvard, 1978)(Anon, 1978)(Anon, 1978)(Anon, 1978). The initial threshold was set to 150. At each iteration, pixels with non-zero intensities were separated into two groups according to the current threshold: pixels with intensity values greater than the threshold were assigned to the background group, whereas pixels less than or equal to the threshold were assigned to the foreground group. The mean intensities of the two groups were then calculated, and the threshold was updated as the average of the two means. The iteration was stopped when the threshold change was less than or equal to 1. Pixels classified as background were set to zero in the RGB image, producing a cloud-masked image. Iterative threshold segmentation is applied until the threshold stabilizes, after which a black mask is used to occlude high-brightness regions associated with clouds and flares (Anon, 1978; Gonzalez and Faisal, 2019; Zhang and Xiao, 2014).”
3. The diurnal cycle analysis (Section 3.3) would benefit from a quantification of the 5-minute smoothing effect.
The authors apply a 5-minute smoothing to the ASI retrievals in Figure 6 but then report statistics on unsmoothed data. This is acceptable, but the magnitude of high-frequency noise should be characterized (e.g., standard deviation of the difference between smoothed and unsmoothed). Without this, readers cannot assess whether the fine-scale variability in Figure 6a is signal or noise.
Response: We have quantified the effect of the 5 min smoothing by calculating the difference between the raw ASI retrievals and the 5 min moving average. For the case shown in Fig. 9, the mean raw-minus-smoothed difference is 0.0008, with a standard deviation of 0.0, mean absolute difference of 0.0076, and 95th percentile absolute difference of 0.027. We have added these values to Section 3.5 and included a supplementary F showing the raw series, smoothed series, residual time series, and residual distribution. We also clarified that all validation statistics were calculated using unsmoothed retrievals.
Figure S1. Quantification of the 5 min smoothing effect for ASI-retrieved AOD at 440 nm on 15 June 2017. Panel (a) compares the raw ASI retrievals with the 5 min moving average, the middle panel shows the difference between the raw and smoothed retrievals, and the lower panel shows the distribution of the difference. The standard deviation of the raw-minus-smoothed difference is 0.0126, with a mean absolute difference of 0.0076 and a 95th percentile absolute difference of 0.0265.
4.Typographical and formatting issues
- Line 30: "Almuantar" should be "almucantar"
- Line 135: "XGBoost" appears twice in close succession – rephrase for readability.
- Table 1: The column headers "440nm, 675nm..." lack spaces (use "440 nm" for consistency with text).
- Line 195: "alucantar" → "almucantar"
- Reference "Anon, 1978" – please replace with the actual author names (it is a well-known paper by Ridler and Calvard or similar).
Response: Thank you for these careful checks. We carefully checked all occurrences of “almucantar” in the revised manuscript and did not find the spellings “Almuantar” or “alucantar” in the current text. To ensure consistency, we standardized the term as “almucantar” throughout the manuscript. We carefully checked the relevant sentence and rephrased the nearby model-description text to avoid unnecessary repetition of “XGBoost” and improve readability. We also revised the wavelength labels in Table 1 from “440nm”, “675nm”, etc. to “440 nm”, “675 nm”, etc., and replaced the incomplete “Anon, 1978” citation with the full reference to Ridler and Calvard (1978).
-
AC1: 'Reply on RC1', Jing Li, 05 Aug 2026
-
RC2: 'Comment on egusphere-2026-2019', Anonymous Referee #2, 25 Jun 2026
This study developed a novel method to retrieval aerosol optical depth and single scattering properties from low cost sky cameras. It serves as a potentially useful supplement for the more sophisticated sun photometer approach. The manuscript is also in general clearly presented and well constructed, which is a great contribution to the community by providing AOD and SSA simultaneously. I suggest several points need further clarification before it can be accepted by AMT.
Specific comments:
- The manuscript states that previous studies mainly focused on AOD, while this study additionally retrieves SSA. This is an important point and should be emphasized more clearly in the Introduction and Summary. It would be helpful to briefly compare the present method with previous ASI-based AOD studies in terms of input features, retrieval targets, validation strategy, and computational efficiency.
- The feature set includes RGB means, RBR, SZA, and RGB values sampled along an annulus. These choices are reasonable, but the manuscript could better explain how each feature relates to aerosol extinction, scattering, or absorption. This would make the AI model less like a black box and help readers understand why the method can retrieve both AOD and SSA.
- The SSA retrieval is an important contribution, but its performance varies strongly by site and wavelength. For example, the Beijing SSA correlation decreases at longer wavelengths. The manuscript notes the narrow dynamic range of SSA values, which is reasonable, but a more detailed discussion would be useful. The authors could consider showing the SSA value distributions, sample sizes, or uncertainty ranges by wavelength to support this interpretation.
- The chronological split within each month is a good attempt to reduce temporal leakage. However, because aerosol conditions can be strongly autocorrelated, adjacent observations may still be similar. The authors may consider adding a more stringent sensitivity test, such as day-based, week-based, or event-based splitting, to demonstrate that the model is not mainly learning local short-term persistence.
- The manuscript suggests that the method can help establish dense aerosol observation networks and bridge gaps between satellites and ground stations. This is a promising direction, but the current validation is limited to two sites and selected time periods. I suggest slightly moderating these claims and presenting them as future potential, pending further validation across more environments, seasons, instruments, and aerosol types.
- The reported correlation, RMSE, and MAE provide useful validation metrics, but additional uncertainty analysis would strengthen the manuscript. For example, the authors could discuss how retrieval errors depend on solar zenith angle, aerosol loading, season, image brightness, or instrument type. This would be particularly helpful if the method is intended for future low-cost aerosol monitoring networks across diverse environments.
Citation: https://doi.org/10.5194/egusphere-2026-2019-RC2 -
AC2: 'Reply on RC2', Jing Li, 05 Aug 2026
Specific comments:
1. The manuscript states that previous studies mainly focused on AOD, while this study additionally retrieves SSA. This is an important point and should be emphasized more clearly in the Introduction and Summary. It would be helpful to briefly compare the present method with previous ASI-based AOD studies in terms of input features, retrieval targets, validation strategy, and computational efficiency.
Response: Thank you for pointing this out. We agree that the simultaneous retrieval of AOD and SSA should be emphasized more clearly. We have revised the Introduction and Summary sections to distinguish the present method from previous ASI-based studies, which mainly focused on AOD or size-related aerosol parameters. We have also added Table S1, comparing representative studies in terms of input features, retrieval targets, validation/reference data, and retrieval methodology.
2. The feature set includes RGB means, RBR, SZA, and RGB values sampled along an annulus. These choices are reasonable, but the manuscript could better explain how each feature relates to aerosol extinction, scattering, or absorption. This would make the AI model less like a black box and help readers understand why the method can retrieve both AOD and SSA.
Response: We have expanded Section 2.3 to explain the physical motivation for each feature group. The revised manuscript discusses how the global RGB means relate to aerosol extinction, how RBR reflects the spectral dependence of aerosol radiance, how annulus RGB values provide directional scattering information, and how SZA accounts for solar-illumination geometry and optical path length.
3. The SSA retrieval is an important contribution, but its performance varies strongly by site and wavelength. For example, the Beijing SSA correlation decreases at longer wavelengths. The manuscript notes the narrow dynamic range of SSA values, which is reasonable, but a more detailed discussion would be useful. The authors could consider showing the SSA value distributions, sample sizes, or uncertainty ranges by wavelength to support this interpretation.
Response: In the revised manuscript, we have provided additional support for this interpretation in Figure 5 and the accompanying discussion. Because you specifically highlighted the decreasing long-wavelength SSA correlation at Beijing_PKU, Figure 5 focuses on the Beijing_PKU clean-regime samples (AOD440 < 0.2). It presents the wavelength-dependent distribution of AERONET reference SSA values, together with the sample size, mean value, 95% bootstrap confidence interval, and interquartile range. The sample size is identical at all four wavelengths (N = 618), indicating that the decrease in correlation toward longer wavelengths is not caused by unequal sampling. The figure also shows that the reference SSA values are concentrated within a relatively narrow range. Under such restricted variability, even small absolute retrieval differences can substantially reduce the Pearson correlation coefficient. This helps explain why the Beijing_PKU SSA correlation decreases toward 1020 nm while the RMSE remains low.
4. The chronological split within each month is a good attempt to reduce temporal leakage. However, because aerosol conditions can be strongly autocorrelated, adjacent observations may still be similar. The authors may consider adding a more stringent sensitivity test, such as day-based, week-based, or event-based splitting, to demonstrate that the model is not mainly learning local short-term persistence.
Response: We agree that adjacent observations may remain temporally autocorrelated under a within-month chronological split. We therefore added a day-block sensitivity experiment in Section 3.4. All observations from a given calendar day were assigned exclusively to either the training or test set, so no day contributed data to both subsets. The model architecture, features, and hyperparameters were kept unchanged. The results show broadly consistent performance with the baseline split, indicating that the retrieval skill is not primarily caused by short-term temporal persistence.
5. The manuscript suggests that the method can help establish dense aerosol observation networks and bridge gaps between satellites and ground stations. This is a promising direction, but the current validation is limited to two sites and selected time periods. I suggest slightly moderating these claims and presenting them as future potential, pending further validation across more environments, seasons, instruments, and aerosol types.
Response: We agree that the claims regarding network-scale application should be moderated because the current validation is limited to two sites and selected observation periods. We have revised the Abstract and Summary sections to describe the study as a proof of concept and to state that broader validation across instruments, seasons, surface types, aerosol mixtures, and cloud-screening conditions is required before operational deployment.
6. The reported correlation, RMSE, and MAE provide useful validation metrics, but additional uncertainty analysis would strengthen the manuscript. For example, the authors could discuss how retrieval errors depend on solar zenith angle, aerosol loading, season, image brightness, or instrument type. This would be particularly helpful if the method is intended for future low-cost aerosol monitoring networks across diverse environments.
Response: We have added a stratified uncertainty analysis in Section 3.3, examining the dependence of retrieval errors on solar zenith angle, aerosol loading, season, and image brightness. The results show that the retrieval uncertainty is conditionally dependent, with larger errors under lower aerosol loading and lower image brightness. Because the present dataset contains only two sites and two instrument configurations, the independent effect of instrument type cannot be robustly isolated.
-
EC1: 'Comment on egusphere-2026-2019', Omar Torres, 05 Aug 2026
Authors have satisfactorily addressed all issues raised in the review phase. They can now proceed with the submission of the revised article.
Citation: https://doi.org/10.5194/egusphere-2026-2019-EC1
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 297 | 139 | 24 | 460 | 31 | 31 |
- HTML: 297
- PDF: 139
- XML: 24
- Total: 460
- BibTeX: 31
- EndNote: 31
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Review for “An AI Based Algorithm for Retrieving Aerosol Optical Depth and Single Scattering Albedo Using All-Sky Imager Observations" by Ni et al.
This study presents a machine learning (XGBoost) approach to retrieve Aerosol Optical Depth (AOD) and Single Scattering Albedo (SSA) from all-sky imager (ASI) data, using collocated AERONET measurements as training targets. The method is tested at two contrasting sites (Beijing, China and Southern Great Plains, USA) and demonstrates promising skill, particularly for AOD. The work addresses an important gap—low-cost, high-temporal-resolution aerosol monitoring—and offers a clear advancement over traditional radiative transfer-based methods by avoiding iterative inversion.
Overall, the manuscript is well-structured, the methodology is sound, and the results are presented clearly. However, several issues—ranging from methodological transparency to the interpretation of SSA performance—should be addressed before publication.
Major comments:
Minor Comments
The caption for Figure 1 lists (a)-(d) but the text refers to (a-c) and (d-f) in the following paragraph. Please check and harmonize the labels. Also, the source attribution for (c,d) is given, but not for (a,b) – add "source: authors" for clarity.
Section 2.3 mentions "iterative threshold segmentation" (citing Anon, 1978; Gonzalez and Faisal, 2019) but does not specify the actual threshold criterion (e.g., Otsu? fixed percentile?). Given that cloud masking quality directly affects sky radiance extraction, please provide the specific algorithm or pseudo-code, or at least cite a more recent and relevant ASI cloud masking method.
The authors apply a 5-minute smoothing to the ASI retrievals in Figure 6 but then report statistics on unsmoothed data. This is acceptable, but the magnitude of high-frequency noise should be characterized (e.g., standard deviation of the difference between smoothed and unsmoothed). Without this, readers cannot assess whether the fine-scale variability in Figure 6a is signal or noise.