Automated Analysis of Ripple-Scale Gravity Wave Structures in the Mesosphere Using Convolutional Neural Networks

Hu, Jiahui; Liu, Alan; Feener, Adriana; Dong, Wenjun; Li, Jing; Li, Tao

doi:10.48550/arXiv.2603.03669

Preprints

https://doi.org/10.48550/arXiv.2603.03669

© Author(s) 2026. This work is licensed under
the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

https://doi.org/10.48550/arXiv.2603.03669

© Author(s) 2026. This work is licensed under
the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Preprints

29 May 2026

| 29 May 2026

Automated Analysis of Ripple-Scale Gravity Wave Structures in the Mesosphere Using Convolutional Neural Networks

Jiahui Hu, Alan Liu, Adriana Feener, Wenjun Dong, Jing Li, and Tao Li

Abstract. All-sky OH airglow imaging provides two-dimensional observations of mesospheric gravity wave structure near ~87 km altitude. Ripple-scale instability signatures, characterized by 5–15 km horizontal wavelengths and short lifetimes, are particularly difficult to identify consistently using manual inspection. In this study, we develop a reproducible, automated detection framework based on a squeeze-and-excitation convolutional neural network (SE-CNN) trained on 41 x 41 pixel image patches, to identify ripple-scale structures in 512 × 512 pixel all-sky airglow images acquired at Yucca Ridge Field Station (40.7^o N, 104.9^o W). The time-differenced images are normalized using a robust median-absolute-deviation (MAD) scaling procedure to mitigate star contamination and background variability. The model is trained and validated on manually annotated ripple and non-ripple patches, then evaluated using independent test subsets. The automated detection is performed using a sliding-window approach with spatial and temporal clustering criteria for event definition. At the patch level, the classifier achieves 92% F1-score with high precision and recall. At the event level, automated detections recover approximately 90% of manually identified ripple events while identifying additional low-amplitude occurrences. Validated against previous manual identification study, the automated detection catalog enables objective quantification of ripple occurrence frequency, seasonal modulation, and lifetime distributions. By emphasizing methodological transparency, calibration considerations, and validation metrics, this framework establishes a scalable measurement technique for systematic detection of mesospheric instability signatures in long-term airglow image archives.

Received: 25 Mar 2026 – Discussion started: 29 May 2026

Jiahui Hu, Alan Liu, Adriana Feener, Wenjun Dong, Jing Li, and Tao Li

Status: final response (author comments only)

RC1:
'Comment on egusphere-2026-1700', Anonymous Referee #1, 09 Jul 2026
This manuscript addresses the application of a convolutional neural network for the identification of ripples in OH airglow all-sky images. This is an up-to-date subject that is relevant to AMT. The introduced method is very interesting and the manuscript is clearly structured. I recommend publication once some revisions have been made.

Major points:
This manuscript is intended for publication in AMT. The focus of AMT is on new measurements, methods/algorithms, etc. The manuscript's innovative aspect is its use of a machine learning approach to derive small-scale wave structures, also known as ripples, from OH airglow measurements. This means that the focus is on the algorithm. It would be helpful for readers and other airglow scientists operating similar instruments who might want to apply this or a similar approach, if the algorithm was described in more detail and in an easier-to-follow way.
Even though the focus is on the algorithm and the results do not provide any more information than has already been extracted manually, the discussion could be improved. There is no comparison with other publications, and statements are not supported by citations. Consequently, readers may get the impression that the discussion was put together rather hastily. A similar impression is given by the results section. Figure 4 and 5 are described but not referenced in the text. Figure 6 is not described at all (or I was not able to attribute the description to the figure), and in the case of figure 7, it seems that a different figure is described or that information are lacking in the figure.
Finally, for a long time, ripple structures have been interpreted as being part of the gravity wave breaking process. Li et al. (2017) demonstrated that these small-scale wave structures can also be secondary waves. These authors used not only airglow images, but also additional measurements. As these are not available at most airglow measurement sites, it is not possible, as far as I am aware, to discriminate between instability features and secondary gravity waves based on the horizontal wavelength alone. The manuscript neglects the possibility of small-scale waves being secondary waves. It would be worth mentioning this.

Details:
General: A line numbering would have facilitated the review.
Introduction:
Formal:
Parentheses around citations are lacking from time to time (e.g. page 2, end of first paragraph Fritts and Alexander (2003))

Abbreviations are not explained in all cases (e.g. YOLO, ASI, CCD)

In terms of content:
As already mentioned above, ripples are not necessarily part of the breaking process of gravity waves, they can also be secondary gravity waves as nicely shown by Li et al. (2017). It would be great, if the authors could clarify that.

There are not many studies that use machine learning to extract small-scale waves from airglow images. As far as I am aware, there exist three such studies (Lai et al., 2019; Prasad et al., 2025; Wüst et al., 2026). It would be worth mentioning all three studies and comparing your approach (maybe in the intro section) and results (maybe in the discussion section) with theirs.

Methodology:
Formal:
Figure 3 and in parts 4: I printed the manuscript and had severe difficulties to identify the symbols and lines properly. If you zoom in, it is much better in most cases, only the white line on a black-white background is not a good choice. Can the authors please improve these figures? In most cases, making the lines thicker and the symbols larger will help.

In terms of content:
Section 2.1: Given that Taylor and Hill (1991) did not investigate the height of the OH airglow layer, are there not more suitable citations for its height? The most well-known citation in this context is probably that of Baker and Stair (1988), who investigated the height of the OH airglow layer using 34 rocket flights at mid- and high latitudes. Shepherd et al. (2006), for example, used UARS/WINDII data for high latitudes during winter 1972 (DOI: 10.1016/j.jastp.2006.06.006). Information about the latitudinal and seasonal dependence can be found in publications using satellite-based and multi-year datasets (e.g. von Savigny (2015), mainly low latitudes using Envisat/SCIAMACHY (doi: 10.1016/j.jastp.2015.02.001), Wüst et al. (2020), low and mid latitudes using TIMED/SABER (doi:10.5194/amt-13-6067-2020)).

Section 2.1: In the abstract, the authors mention that they use time-difference images, however, in section 2.1 this is not noted. Could the authors please clarify that point? If the authors are using time differencing, could they also provide information on which waves are enhanced and which are damped?

Formula (1): The median is calculated over space (512 px 512 px image), over time or over both?

Section 2.1: the authors mention that they use an all-sky camera, however, it would also be helpful to specify the diameter of the field of view (as this is normally not identical to 180°) to give a rough idea of the conversion of the 41 px × 41 px field into kilometres. Additionally, the maximal zenith angle would be good to know as the horizontal resolution becomes very coarse for very large angles.

Section 2.2: As mention above, the explanation of the neural network needs to be improved. In the following, I give some examples. At the beginning of section 2.2, the authors mention channel-wise features, however, I cannot find an explanation or further information on both terms, “channel” and “feature”. Can the authors provide it please? Furthermore, different filters are mentioned. How do these filters look like (2D sinusoidal oscillations of different wavelength and orientation, for example)? The authors mention ReLU without explaining what it is. Could they please do so? The authors mention the sets C, H, and W without explaining what it is. Could they please do so? Etc. So, could the authors please try to formulate section 2.2 in a way that also a scientist who is not an expert in machine learning can follow it better? Could the authors please also provide references where these readers can find more information on the topic (not a single one is given) or why a specific choice is a good one or where it was already used?

Figure 2: What is a/the vanilla CNN? The term appears first in figure 2 (there with an orographic mistake) on page 4 and is mentioned in the text on page 5, however, it is not explained. Out of curiosity: The F1 score is very low for some epochs. However, this score does not indicate whether recall or precision is low, or if both are low. What causes these low F1 values?

Results and discussion:
Formal:
Figures 5 – 8 are plotted in red and green: that might not be the best choice for people with a colour deficiency and it is easy to correct as only two colours are used.

Figure 4 – 6 are not addressed in the text, i.e. they are described but I cannot find a hint such as “see figure 4” or similar in the text.

In terms of content:
The authors state that their machine learning approach provides probabilities that a detected anomaly is a ripple structure. They consider ripples with a probability of more the 70%. Are the training data also providing probabilities? If this is not the case, how reliable are the given probabilities or should they just be considered as threshold that needs to be passed to enter statistics?

The results are not compared to other publications. There are not many as identifying ripples is time-consuming and machine learning approaches have been barely used, but there exist publications such as Rourke at al. (2017) and Suzuki et al. (2011), both referring to the South Pole, and Yue et al. (2010), who address data recorded at Yucca Ridge Field Station but also in Japan, and Wüst et al. (2026), who use European data [Yue et al. (2010): 10.5194/angeo-28-1401-2010, Suzuki et al. (2011): 10.1029/2011JD015882, Rourke et al. (2017): 10.1002/2017JD026998, Wüst et al. (2026): 10.5194/amt-19-3539-2026].

Figure 5:
The detection probabilities of ripples published by Yue et al. (2010) do not show a distinctive maximum in autumn and winter. The same holds for Wüst et al. (2026) (Rourke at al. (2017) and Suzuki et al. (2011) do not have data all over the year). How does this compare to figure 5 of this manuscript? It is clear to me that figure 5 shows ripple counts and not ripple probabilities. But is that the reason for the differing results?

Only during autumn, the neural network identifies more ripples than humans do. This is a very pronounced effect. Do the authors have any idea why this is the case? In the text, the overall detection bias of +32 % is explained by the higher sensitivity of the CNN to faint ripple structures. If this holds, what could be the reason for the over-representation of these faint structures during the autumn months? By the way: which months belong to which seasons?

Figure 6: This figure is nowhere described or did I overlook it?

Figure 7: The lifetime distribution does not show any manually detected events lasting more than 5 minutes. Therefore, saying that the distribution of manually detected events and events detected by the model is similar is slightly misleading. Could the authors please rephrase it? Furthermore, ripples are supposed to live some 10 min. Why are they identified over longer periods? Is this a misdetection by the CNN, or does it indicate secondary waves or something else? Later in the manuscript, it is written that a marginally higher fraction of long-duration events is detected in winter in the manual catalogue. I do not see this in figure 7.

This is related to the point mentioned before: the authors state that the seasonal occurrence of ripples is linked to the atmospheric background, to gravity wave and tidal activity. I have two points here:
As mentioned before, ripples can be attributed to breaking gravity waves but they do not have to. If the authors argue in a relative general way with the atmospheric background, it seems that they neglect the possibility of ripples being secondary gravity waves.

Why do other authors observe other results? The relative generic argumentation that the authors present reads as if this must be the same case everywhere.

Can the authors please provide evidence for their arguments, e.g. tidal and gravity wave perturbations are relatively strong during autumn. This might, for example, depend on the period of the waves and it is not a worldwide feature.

What is the event recovery rate? It is given by 90% at the end of the result section but it is defined nowhere or did I miss it?

Concerning the discussion section, I would change the title to 'Discussion and Summary'.
Citation: https://doi.org/10.5194/egusphere-2026-1700-RC1
RC2: 'Comment on egusphere-2026-1700', Anonymous Referee #2, 16 Jul 2026

Review of "Automated Analysis of Ripple-Scale Gravity Wave

Structures in the Mesosphere Using Convolutional Neural Net-

works" by Jiahui Hu et al.

1. Summary

The authors present a convolutional neural network to detect ripple-

scale structures in airglow images taken at Yucca Field Ridge Sta-

tion. The network is trained using manually selected scenes of

ripple patches and non-ripple patches. The general performance

of the network is quite good and it achieves values of about 90% in

all relevant quantities such as precision and accuracy. Finally, the

results of the network are compared to manually derived results

for the period 2003 and 2004.

2. Overall assessment

The general topic of the paper is very interesting and well suited

for a publication in AMT as it describes a new method/algorithm

to detect ripple-scale structures in airglow images an important

part of the gravity wave spectrum. I recommend the publication

of the manuscript after revision of several points.

3. Major comments

1. Please insert line numbers the next time.

2. The overall impression is that the manuscript was put together

rather quickly as it shows several deficiencies as missing ref-

erences, not discussed figures, misleading statements etc. (see

details below).

3. There is a large difference between the number of selected

ripple events from the network (952) compared to the manual

catalog (720). This difference is explained by a better selection

capability of the network. I have several questions regarding

this:

(a) The authors write that a visual inspection afterwards show

that many additional model detections correspond to weak

but coherent banded structures. Can you show examples

such that readers can judge on there own.

(b) Is the change of selection capability caused by the network

only or is it mainly changed by the normalization using

MAD? Are the ripples also better visible after applying

the normalization for a human eye?

(c) Why is the difference present in one season (autumn) only?

Do you have an explanation for that?

(d) Why does this overestimation not occur in the test data

set? Are there any comparable patches included in the

test data set? If not, what changes would you expect if

you include such patches, which are manually marked as

non-ripple patches, in the training set and the test set? In

other words: How robust in the outcome of the network

against changes in the preselected ripple and non-ripple

patches?

4. Minor comments

Introduction

1. At some points the authors claim that the network is objective

or reduces a subjective bias. I find these statements inappro-

priate as the network is trained with a manual and therefore

subjective selection. It should also recall a former subjective

bias. Please rephrase these statements.

Methodology

1. There are better references for the height than Hill and Taylor

(1991) such as Baker and Stair (1988) and others (von Savigny

et al. (2015), Garcia-Comas et al. (2017), Wüst et al. (2020)).

2. Can you provide more information regarding the selection of

the ripple and non-ripple patches. Are only perfect ripple

patches taken or are they randomly chosen from the whole

catalog?

3. The vanillaCNN is first mentioned below Fig.2. What is the

difference between vanillaCNN and SE-CNN? Is the vanil-

laCNN used in the following? Otherwise you might delete

these parts from the manuscript.

4. Can you explain the large drops in the F1 score at some par-

ticular epochs.

Results

1. Can you give the criteria for spatial coincidence that have to

be fulfilled for the two selections (manual and automated) to

be classified as overlapping.

2. Fig. 4 and Fig. 5 are not referenced in text.

3. In the Figs. 5-8 the only two colors used are green and red.

This may be problematic for people with color blindness.

4. I find the Figure 4 is hard to read. Maybe two panels above

each other instead of one are better.

5. Figure 6 is not discussed in text. Please include the corre-

sponding text or delete the figure.

6. The text describing Fig. 7 is not precise enough as at some

points the statements are only true for one season or specific

situations and not in general. For example, only in autumn the

automated selection shows a higher fraction of very short-lived

ripples and the manual selection does not show more long-lived

ripples in winter. Please add more text here and improve your

statements.

7. Why are there no manual selections of ripple with lifetimes

longer than 5 minutes in Fig.7?

8. At the end of the section a statement of the more objective

network is given again. I find this inappropriate as the network

is trained with subjective manual selections.

Discussion

1. There are again wrong or at least misleading statements about

the observation frequencies as a function of lifetime. In total

the amount of detections of very short-lived ripples should be

very similar as only in autumn the automated selection shows

a higher fraction (compare Fig. 7).

2. The complete discussion sections lacks on comparisons to other

studies. Is it possible to add some comparisons to other studies

here?

Citation: https://doi.org/10.5194/egusphere-2026-1700-RC2

Jiahui Hu, Alan Liu, Adriana Feener, Wenjun Dong, Jing Li, and Tao Li

Viewed

Since the preprint corresponding to this journal article was posted outside of Copernicus Publications, the preprint-related metrics are limited to HTML views.

Total article views: 76 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
75	0	1	76	0	0

HTML: 75
PDF: 0
XML: 1
Total: 76
BibTeX: 0
EndNote: 0

Views and downloads (calculated since 29 May 2026)

Month	HTML	PDF	XML	Total
Jun 2026	38	0	38
Jul 2026	37	1	38

Cumulative views and downloads (calculated since 29 May 2026)

Month	HTML	PDF	XML	Total
Jun 2026	38	0	38
Jul 2026	37	1	38

Viewed (geographical distribution)

Since the preprint corresponding to this journal article was posted outside of Copernicus Publications, the preprint-related metrics are limited to HTML views.

Total article views: 65 (including HTML, PDF, and XML) Thereof 65 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 01 Aug 2026

Short summary

All-sky OH airglow imaging captures mesospheric gravity waves near ~87 km, but ripple-scale instabilities (5–15 km) are difficult to identify manually. We develop an automated SE-CNN framework trained on image patches to detect ripples in 512×512 images. Using MAD normalization and sliding-window detection with clustering, the model achieves 92% F1-score and recovers ~90% of manual events, enabling objective, scalable analysis of ripple occurrence and variability.


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