the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 12 Jun 2026
Method for Consistency Analysis of a set of space-borne Climate Data Records: Application to Aerosol Optical Depth
Abstract. The consistency of a set of satellite observations obtained with different algorithms from the same satellite instruments can be used as an indicator for their robustness or reliability. For this four metrics to evaluate those characteristics are defined and integrated into one overall consistency score per grid cell. These four metrics are: their median values, annual cycle, median values, decadal trends and correlation. This paper discusses how a combination of these metrics can be used to evaluate the overall consistency of different datasets for the same essential climate data variable.
This study performs an examplary application of this approach on Aerosol Optical Depth (AOD) Climate Data Records (CDRs) obtained from a series of 3 similar instruments from the Copernicus Climate Change Service (C3S). Assessing the consistency of those CDRs is highly relevant because space-based aerosol information and their climate impact have substantial uncertainties. The regional dependency of this consistency score shows the influence of surface properties and geographic regions. With the consistency score, regions with robust information can be separated from more challenging regions and the advantages and disadvantages of different instruments can be identified.
- Preprint
(3004 KB) - Metadata XML
- BibTeX
- EndNote
Status: open (until 02 Aug 2026)
- RC1: 'Comment on egusphere-2026-2885', Anonymous Referee #1, 05 Jul 2026 reply
-
RC2: 'Comment on egusphere-2026-2885', Anonymous Referee #2, 14 Jul 2026
reply
This manuscript provides a concrete example of a consistency method, based on a previous paper discussing consistency on a higher level (Popp et al 2020). The authors propose a simple method, then apply it to 3 data sets to study their consistency. I think that the idea to derive a concrete method is useful for the community, however I think that the proposed method is too simple for a dedicated paper. I would see it as support for an analysis in a paper focusing on the use of different data sets. For a paper focusing on the method itself (with an illustration of its use), it requires to be deeper than saying one should look at medians, cycle amplitude, trends and correlation which are all obvious to me. Such a methodology paper should go deeper in how to define and use those metrics for a correct and in-depth analysis. Here are some more precise comments about what I think is missing:
1. There is no use of the data sets uncertainties, although their importance is highlighted in the paper cited as foundation of the current manuscript (Popp et al, 2020).
2. There is no discussion of the difference between a constant bias between algorithms and truly inconsistent data sets, which would both reflect in the median value but have very different meaning in terms of data use.
3. Linear trends are used, with a robust estimator, but their uncertainties and significance are not considered. I am not sure to understand how to justify the comparisons of trends without their uncertainty range, and without ensuring their significance. I do understand the limitation of having less data if considering uncertainty and significance, but I think it is scientifically more sound to have less meaningful data than more data with uncertain meaning.
4. The seasonal amplitude is calculated using means and not medians, leaving a possible strong impact of outliers. In addition the mean seasonal cycle is obtained using different years for the different data sets (due to their different time coverage), but there is no discussion about the expectation of these seasonal cycles to remain constant over time. I also find the cycle amplitude definition to be very basic and dependent on outliers. Why not fitting a curve (such as a sinus, if this is the expected behavior) on the data and take the amplitude of that fitted curve?
5. The correlation should not be used bluntly. The possibility to have a good correlation depends not only on the quality of the data, but also on the variability / cycles with respect to the uncertainty / noise in the data. A bad correlation between 2 data sets (in a grid cell) could simply mean that the variability (in that grid cell) is below or close to the precision of the retrievals.
Other precise questions / comments:
line 66: why average on a 5° grid?
lines 69-70: what is this std dev of all three algorithms?
eq 1: although they may seem obvious, n and m should be defined
lines 99-100: does that mean that each observation has the same weight, but each day does not have the same weight? Then the mean depends on whether there were more observations when there was significant aerosol loading, for example. Please justify this approach for the determination of a mean annual cycle.
lines 101-105 are unclear
line 126: how does this impact the aerosol retrieval and why is it important to discuss here?
Minor comments / suggestions / corrections:
In general, "examplary" is used incorrectly; first I do not think it exists with an “a”, and “exemplary” does not mean “for example” at all.
line 3: median is listed twice
line 15: is -> are
lines 32, 38, 211: wrong citation command
line 68: consistency “is” fulfilled ? (instead of “if”)
lines 96-97: phrasing is a bit weird
Citation: https://doi.org/10.5194/egusphere-2026-2885-RC2
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 43 | 7 | 2 | 52 | 0 | 1 |
- HTML: 43
- PDF: 7
- XML: 2
- Total: 52
- BibTeX: 0
- EndNote: 1
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Overall Notes
This paper aims to assess dataset consistency by comparing four metrics applied to the data records from three instruments in the Along-Track Scanning Radiometer family, using three different algorithms. It is important to note that the results do not necessarily represent uncertainty; validation against a “ground truth” dataset, to the degree possible with known, quantified uncertainty, is required to assess uncertainty against standard values. This could be made clearer in the introduction and conclusions – it does not reduce the value of the work, but I think it places the results in proper context.
Another aspect that might be worth mentioning is the degree to which different instruments or algorithms might be sensitive to different aspects of the environment. In such cases, “consistency” might not be as meaningful as it might seem. Specifically, even if different products claim to be reporting the same quantity (e.g., AOD at a particular wavelength), some methods might be more sensitive to larger or more light-absorbing particles than others, e.g., depending on the actual wavelengths measured. Also, sensitivity might vary with observing geometry or surface properties differently for different instruments or algorithms. These differences might also be convolved with differences in pixel resolution, such that the “right value” for the variable in question depends to some degree on the measurement or retrieval technique. This type of ambiguity is an issue, for example, with the variable reported as “aerosol layer height,” where widely used products give different answers, all of which might be correct (within uncertainly limits), but they sample different aspects of the aerosol vertical distribution. This issue also applies, though less prominently, to AOD.
That said, this paper presents useful results; it provides important insight into the strengths and limitations of the AOD products studied, as well as providing an illustration of a “consistency” approach for data-record evaluation. And the conditions under which the observed AOD consistency is greater are likely to represent places where uncertainties at least in the interpretation of the observations are diminished. In places where consistency is low, the approaches might be sensitive to different things, or there might simply be less information in the measurements about the variable of interest.
More Specific Notes