Intercomparison of nighttime aerosol optical depth retrievals from both reflectance-based and city light-based methods using VIIRS DNB data

Schlief, Aanan H.; Marquis, Jared W.; Zhang, Jianglong; Reid, Jeffrey S.; Miller, Steve

doi:10.5194/egusphere-2026-2225

Preprints

https://doi.org/10.5194/egusphere-2026-2225

Preprints

04 May 2026

| 04 May 2026

Intercomparison of nighttime aerosol optical depth retrievals from both reflectance-based and city light-based methods using VIIRS DNB data

Aanan H. Schlief, Jared W. Marquis, Jianglong Zhang, Jeffrey S. Reid, and Steve Miller

Abstract. Using observations from the Visible Infrared Imaging Radiometer Suite (VIIRS) Day/Night Band (DNB), two different nighttime aerosol optical depth (AOD) retrieval methods were evaluated and inter-compared. The first approach is a lunar reflectance-based retrieval method, using reflected moonlight in a manner similar to daytime retrievals. The second approach utilizes changes in light patterns over regions with artificial light sources due to the upward diffusion of light by aerosol particles. Both retrieval methods were implemented over Dakar, Senegal for 2017 and 2018. Retrievals from both approaches were evaluated against ground-based solar and lunar AErosol RObotic NETwork (AERONET) data, as well as daytime AOD retrievals from the Moderate Resolution Imaging Spectroradiometer (MODIS). Additionally, impacts of using the Miller and Turner lunar model for estimating Top-Of-Atmosphere (TOA) lunar spectral flux for AOD retrievals were also studied. Findings suggest that while both retrieval methods show skill in retrieving nighttime AOD by qualitatively identifying over-ocean aerosol plume locations and quantitatively comparing with solar and lunar retrieved AERONET data, cloud contamination and variations in lunar properties are factors that need to be carefully quantified in future studies for accurate nighttime aerosol retrievals using VIIRS DNB data. This study suggests that there are sampling issues from both approaches, but the combined use of both retrieval methods can increase the sampling rate for nighttime aerosol retrievals by more than 50 %.

Received: 17 Apr 2026 – Discussion started: 04 May 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

Aanan H. Schlief, Jared W. Marquis, Jianglong Zhang, Jeffrey S. Reid, and Steve Miller

Status: final response (author comments only)

RC1:
'Comment on egusphere-2026-2225', Yingxi Shi, 01 Jun 2026
Paper Review: Intercomparison of Nighttime AOD Retrievals from VIIRS DNB Data
This paper evaluates and intercompares two nighttime AOD retrieval methods, a lunar reflectance-based approach and a city light-based approach, using VIIRS DNB data over Dakar, Senegal for 2017–2018. Both methods are validated against AERONET and compared with MODIS, demonstrating qualitative skill in identifying aerosol plumes. The work is timely and significant, as nighttime aerosol retrieval remains an underserved area, and the finding that combining both methods increases nighttime sampling by over 50% is a practically valuable contribution to the remote sensing community. The paper is well-written, well-organized, and the multi-source validation framework is solid.

Major Issues
The most critical weakness is the extremely small coincident sample size, only 8 valid retrieval pairs across two years, which fundamentally limits the statistical credibility of the intercomparison. The authors should consider extending the time period or expanding to additional sites to build a more meaningful dataset. Additionally, the relationship between the lunar reflectance method presented here and the previously published algorithm from Wang et al. (2020) and Zhou et al. (2021, 2024) is never clearly established, raising concerns that whether what is presented here shows the similar quality as what is published. Finally, while moon phase is acknowledged as a major uncertainty source, its systematic impact on retrieval count and bias for each method is never quantified, a dedicated analysis of retrieval performance as a function of lunar illumination fraction is strongly recommended.

Minor Issues
It is unclear whether the DT aerosol model parameters are spectrally averaged across the DNB wavelength range or selected at a single central wavelength, this should be explicitly stated.

Figure 2: panels (d) and (e) appear to include a glint mask while panel (f) does not; if intentional, this should be explained.

Figure 3a is too cluttered to convey a clear message, if the intent is to show sampling differences, consider a bar chart or availability matrix; if showing time-series agreement, merge daytime and lunar AERONET into a single reference series.

The rationale for separating 2017 and 2018 data in Figure 3 is unexplained; the terms *waxing* and *waning* should be briefly defined for general audiences.

Line 337 contains what appears to be an incomplete placeholder sentence regarding the standard deviation threshold for cloud masking.

It is counterintuitive that the city light method yields fewer AERONET-paired data points than the lunar reflectance method when against lunar AERONET comparisons. This discrepancy should be explained.

Table 1 is overly dense with redundant statistics; a simplified chart showing data count and one representative error metric would be more effective.

Figure 4 (June 1, 2018): the flagged feature near 17–20°N, 17.5°W visually resembles a dust plume rather than cloud, additional supporting evidence (e.g., IR imagery or lidar) is needed to justify the cloud screening decision.

Figure 5 would be more informative if MODIS AOD were compared against lunar AERONET rather than solar AERONET, to better contextualize diurnal AOD differences.

Wang, J., Zhou, M., Xu, X., Roudini, S., Sander, S.P., Pongetti, T.J., Miller, S.D., Reid, J.S., Hyer, E. and Spurr, R., 2020. Development of a nighttime shortwave radiative transfer model for remote sensing of nocturnal aerosols and fires from VIIRS. Remote sensing of environment, 241, p.111727.; Zhou, M., Wang, J., Chen, X., Gomes, J., Levy, R.C. and Miller, S.D., 2024, January. Link Day and Night: A Deep Learning Framework to Retrieve Global Nighttime Aerosol Optical Depth from VIIRS DNB. In 104th Annual AMS Meeting 2024 (Vol. 104, p. 435849
Citation: https://doi.org/10.5194/egusphere-2026-2225-RC1
CC1: 'Comment on egusphere-2026-2225', Thomas Eck, 02 Jun 2026

One minor comment:
I suggest that the authors read and consider the AMT paper (currently in discussion) on the AERONET lunar AOD database:
Schafer, J., Eck, T. F., Slutsker, I., Gupta, P., Sorokin, M., Sinyuk, A., and Smirnov, A.: An Assessment of Lunar Photometry in AERONET, EGUsphere [preprint], https://doi.org/10.5194/egusphere-2026-1506, 2026.
https://egusphere.copernicus.org/preprints/2026/egusphere-2026-1506/
This paper provides deeper insight into characteristics of the AERONET lunar AOD database thereby enabling more knowledgeable use for validation or other purposes.

Citation: https://doi.org/10.5194/egusphere-2026-2225-CC1
RC2: 'Comment on egusphere-2026-2225', Anonymous Referee #2, 07 Jul 2026

This study evaluates two methods to quantitatively retrieve mid-visible aerosol optical depth (AOD) at night from the VIIRS Day-Night Band (DNB), a highly sensitive pan-visible band this sensor carries which is mostly used for nighttime imaging (e.g. city lights). The study area is off the west coast of northern Africa during 2017 and 2018, which is motivated by the availability of both solar and lunar photometer AOD data from AERONET at a site near Dakar, Senegal. The nighttime retrievals are also compared with daytime retrievals from MODIS. The first retrieval method is an analogue of the daytime reflectance-based over-ocean technique often applied to satellites (albeit restricted to the one DNB and using lunar rather than solar irradiande to calculate reflectance). This provides fairly broad spatial coverage. The second is based on darkening of artificial light sources (from cities) through aerosol scattering/absorption, and appears more of a point estimate (maps based on this are not presented).

The article is in scope for AMT and has sufficient novelty (it’s an application and evaluation of two techniques which the authors have previously worked on to a new area, and nighttime AOD remains a data gap so there is a scientific need for this type of retrieval). It is fairly well written (with a few exceptions I try to cover below) and referenced, and the quality of language and figures is also quite good (a standard journal copy-editing would suffice for language). There are some points where I think clarification would be beneficial and I do think the authors could have pushed the analysis a bit further. I recommend minor revisions and would be happy to review the next version. My comments supporting this recommendation are as follows:

General comment: The analysis gets a bit limited by the small volume of matchups with AERONET. I understand that this paper is led by a student (and it is good work) but scientifically it would be stronger if the data volume could be expanded. AERONET’s lunar photometer data records are fairly extensive (https://aeronet.gsfc.nasa.gov/new_web/webtool_aod_v3_lunar.html ) so to focus on one site for two years is quite limiting. There is a valid argument to be made about different optical models and cloud screening thresholds being needed for different parts of the world, but even within the region mapped within the article (for which this would likely not be a concern) there are multiple unused AERONET sites. The authors use the site Dakar from 2017-2018 but I believe it has data through 2020. There’s also Dakar_Belair (about 60 km away) from 2019 onwards, and a little further out Capo Verde (2020-2025) and Mindelo OCSM (2022 onwards). And several more in the Canary Islands further North. These are all coastal sites so the LUT-based method could be applied, even if the city light one can’t (and I’m not sure on that second front). I encourage the authors to consider extending the space and/or time period (a little – I acknowledge adding all of the above would be too much) when revising the manuscript, to see if the data volume can be increased, and make some of the quantitative conclusions more robust (retrievals don’t necessarily need to be presented site-by-site in the paper so this should not cause much bloat).

A related point is that we now have multiple VIIRS sensors in orbit (SNPP returning data since 2012, NOAA20 since 2017, NOAA21 since 2022). I’m not sure if the DNB data from them are radiometrically equivalent, but adding a second VIIRS (NOAA20 seems the simpler choice given the time period) would also be an option. Given the platforms are not in identical orbits (same nominal time but staggered), observing on the same nights from different sensors would help give a look at how the difference in viewing geometry (the platforms will see the same areas at different view zenith/azimuth angles) affects the retrieval for the same scenes.

General comment: the paper jumps back and forth a lot between describing the radiative transfer, the AERONET data, the ocean approach, and the city lights approach. Some of the information is repeated between parts, and there was some stuff I found myself asking when reading a section about one of the data sets/approaches which is only answered pages later when the manuscript jumped back to the topic in question. This made it difficult to read in places. I wonder if the authors can find another way to organize the paper. Maybe pull all the algorithm stuff up front and put the AERONET stuff in a section just before the analysis of AERONET comparisons.

General comment: I would appreciate a bit more information about the Miller/Turner lunar model since it is key to the approach. A paragraph or two about: where is it from (I’m assuming semi-empirical), what is the valid range of wavelengths/geometries, what is its uncertainty (understanding that may be a function of the previous), is this something which we expect can be improved, what is the relation between this and ROLO and what are their differences?

Line 33: Somewhere in the manuscript I would write about why you chose to only investigate two of these three methods.

Line 73: At several places (e.g. also line 121) the text says something like “Miller and Turner model (Miller and Turner, 2009)” where it gives author names and a paper citation close by. This feels redundant, just cite the paper, or define the acronym “MT” for the model after first citation and call it by that. This will make the text more readable.

Line 74-76: Unless I am misunderstanding, this sentence seems out of place or not needed. Lines 68-70 already say that the two methods are applied so why the “Finally, we explored” here – this implies some third thing besides the other two?

Lines 93-95: More information is necessary explaining the different spatial resolutions of these data products.

Line 110: the start of this line says “daytime” but I think this is an error, as later text and the analysis describe use of both daytime and nighttime AERONET data.

Line 112: Somewhere around here I would add a citation to the AERONET paper on lunar photometry (which is currently a preprint): https://egusphere.copernicus.org/preprints/2026/egusphere-2026-1506/

Line 126: Somewhere around here, more information is needed about the radiative transfer simulations. The DNB is a very wide band. Were simulations run at some fine spectral resolution (if so, what) and then aggregated according to the spectral response? Or was representative central wavelength calculated (and if so, for which quantity) and effective monochromatic calculations done for that?

Line 134: Note that the MODIS algorithm this model is taken from assumes this coarse mode is spherical. Spheroids are a more realistic assumption in terms of phase function (although the aspect ratio distribution assumed should be mentioned in the text) but the paper should state that this shape assumption is different from what is done by the MODIS team. It should also be noted in the paper that this therefore means the retrieval is assuming purely coarse mode dust-like aerosols with these properties, e.g. no fine mode aerosols, no sea spray aerosols. The fidelity of this assumption is probably worst when AOD is low (minimal dust) and best when AOD is high (minimal relative contribution from non-dust). I suggest that the text is expanded to go into detail about these assumptions a bit more.

Line 158: Somewhere here I would add a sentence about why the ocean case is easier than the land case. My assumption is that this is because ocean is comparatively dark and spectrally and spatially uniform, while land has more spectral and spatial variation, so becomes a larger potential error source from mischaracterizing it.

Line 163: I think the comment and citation about water vapor being a marginal absorber needs a clarification here that this is only true for a broad band such as the DNB (and the single visible band of the earlier GOES sensors). For narrower bands used in multispectral retrievals (e.g. VIIRS/MODIS), we absolutely need to account for variations in water vapor absorption otherwise it induces a spectral bias.

Line 187: It’s good to remind the reader that lunar phase is the same everywhere for a given night, but this also means that the bit about selecting the prime meridian is not relevant and can be deleted (I am guessing this means you used a calculator and plugged in those coordinates because it required you plug in some coordinates?).

Line 198: I am assuming that lunar glint is the direct analogue to Sun glint, but I would add a note to state this and say exactly how this is calculated in the present study (e.g. do you use a Cox-Munk model to estimate strength or is it purely geometric)? How is the threshold for exclusion defined in the approach?

Line 200: I would say high “errors” and not high “biases”. Presumably the SBDART model is able to model the glint, so it is there albeit modeled with large uncertainty (leading to large errors) and it is not that SBDART has no glint (I would say “bias” only if it’s a physical feature which is missing from the system)?

Figure 2: I think panel (a) should be reflectance not reflectivity (it’s what is observed not what is inherent). I suggest panel (b) is retitled “Cloud Fraction”. Panel (c) also needs a better title and an indication of units. Is this glint angle?

Section 3.2: I think the text should be clearer on whether this is the “direct estimation” or “variance” method mentioned in the introduction. Papers cited here are associated with both methods in the Introduction, and both rely on artificial (e.g. city) lights.

Line 299: Why 1 degree (about 100 km)? This is much coarser than the 25 km distance often used for daytime data validation, and so will introduce additional uncertainty related to spatial heterogeneity. Does it really increase data volume that much (and if it does, doesn’t the fact that you don’t have retrievals close by but do have them far away inherently mean there is heterogeneity in the atmosphere)? Smaller areas could potentially also lead to less cloud contamination as there is a smaller area for a missed cloud to creep in. I would also write how the space/time aggregation is actually done, I am guessing simple averaging?

Lines 311, 314: I understand aggregating the MODIS data set and your VIIRS retrievals from level 2 to 0.5 degrees for a common comparison. But why then compare MODIS with AERONET at 0.5 degrees instead of either 1 degree (as was done for VIIRS) or 25 km (as is normal, as stated above)? I suggest 25 km for both level 2 comparisons. Also, why choose 0.5 degrees for the MODIS/VIIRS comparison over 1 degree (I think the choice is fine just wondering why a third spatial scale is introduced for this part of the analysis)?

Line 320: It would be interesting to present the standard deviation of these five domains. This would have contributions from aerosol heterogeneity and from retrieval uncertainty. One could examine whether high variability between domains is associated with worse agreement with AERONET or not. I would also like some justification for the choice of 25 km x 25 km domains. Since the comparison is already restricted to times close to a full moon, I would not have thought that too much averaging for noise reduction is needed? Ok, I see this topic comes up again lines 420-425 (one of my points earlier was about information jumps around the paper) but this still doesn’t say why this decision was made, discuss variability between regions, and introduces a new threshold (40 light sources) without justification. These later lines also switch from calling them “domains” to calling them “grids” (please be self-consistent with terminology).

Line 326: Do you mean average difference or average ratio here?

Lines 402-403: I don’t think that the data volume is high enough to make statements about what optimal thresholds are. Sure, report on what was seen for the sample of data you have, but I don’t think this can be generalized.

Lines 410-414: I am not sure I understand what is being said here. Is this saying that even though the threshold for minimum lunar illumination was 0.5 (line 191), in practice the minimum value found for the AERONET matchups was 0.73 and all except for that one had 0.85 or above? I would rephrase this text.

Table 1: I am not sure “error” is the right term here. Maybe for the lunar AERONET data if we can treat this as a reference. But the comparison to daytime AERONET and to MODIS is really a “difference” not an “error” since the daytime data are quite different in time, and the MODIS retrieval has non-negligible uncertainties itself and is also from daytime not nighttime.

Section 4.4: I found this section quite weak. From my understanding the authors have the two years of data available, why is the comparison restricted to two scenes? Were they cherry picked because they look good? How about making e.g. seasonal composites from the two sensors across the study region and showing that as well? (These could also be restricted to common grid cells from the same day/night to reduce some of the sampling related artefacts.) How about a time series analysis over the domain or something? That would also help look at differences in less vs. more dusty periods.

Lines 459-461: this is another example of information that was already given earlier being repeated (too much jumping around – I suggest confining the info about MODIS/VIIRS aggregation for comparison to one another being kept all together, maybe here).

Figure 4: See earlier comment that “Cloud Screening” as a plot caption could be “Cloud Fraction”. I don’t know that these panels are needed for this figure anyway?

Line 500: The word “agreement” here is not very meaningful. Correlation is a measure of linear association, so this means that the two move in the same direction with some commonality. It doesn’t say anything about how similar the magnitudes are. Is the timing of changes more important than their magnitudes (which is what using correlation for agreement implies), and if so, for what application? Are these correlations strong enough for some practical purpose or is it just nice that they are non-negligible? I suggest using some more precise wording.

Section 5: This feels misplaced, and there’s only one subsection so it doesn’t need a number. I feel it would be better earlier on (when talking about lunar illumination fraction thresholds), or in an appendix, or else merged into the conclusions (also when talking about coverage relationship to lunar illumination).

Line 541: how much of this is really “random noise” vs. physical signals which are hard to model? Just saying “random noise” feels like hand waving away complexity. Sure, Figure 7 shows noise increasing when illumination drops but also a change in the baseline value (also seen in Figure 6) which implies some real signal. If it were really noise then this could be overcome by spatially aggregating the VIIRS data. I suggest expanding the discussion here and being more careful with word choice.

Figure 7: why is the land white in all these panels? Is it that reflectance was just not computed for land (in which case “value of 1” and “missing data” should be distinguished somehow by the color bar), or that it is always above 1?

Lines 568-570: I suggest rewording this, it feels a bit redundant (starts and ends talking about reflectance definition) unless I am misunderstanding something (in which case it should be reworded because it leads to reader misunderstanding).

Line 595: A link is missing for the T-Matrix software.

Citation: https://doi.org/10.5194/egusphere-2026-2225-RC2

Aanan H. Schlief, Jared W. Marquis, Jianglong Zhang, Jeffrey S. Reid, and Steve Miller

Viewed

Total article views: 390 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
292	70	28	390	21	21

HTML: 292
PDF: 70
XML: 28
Total: 390
BibTeX: 21
EndNote: 21

Views and downloads (calculated since 04 May 2026)

Month	HTML	PDF	XML	Total
May 2026	216	41	13	270
Jun 2026	44	8	9	61
Jul 2026	28	18	2	48
Aug 2026	4	3	4	11

Cumulative views and downloads (calculated since 04 May 2026)

Month	HTML	PDF	XML	Total
May 2026	216	41	13	270
Jun 2026	44	8	9	61
Jul 2026	28	18	2	48
Aug 2026	4	3	4	11

Viewed (geographical distribution)

Total article views: 373 (including HTML, PDF, and XML) Thereof 373 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 16 Aug 2026

Short summary

Two methods for retrieving nighttime aerosol optical depth, with the first method using reflected moonlight and the second method using emissions from artificial lights were intercompared using the Visible Infrared Imaging Radiometer Suite (VIIRS) Day-Night-Band data. This study indicates that both methods have skills in retrieving nighttime aerosol optical depth, and the combined use of both methods can improve data sampling rates.


Total:	0
HTML:	0
PDF:	0
XML:	0