About 100 years ago, the radiation of the hydroxyl (OH*) airglow layer was observed from the ground for the first time without knowing which molecule caused this radiation until Meinel (1950) identified OH* as the source. While he exposed photo plates for approx. 4 h, today two-dimensional InGaAs (Indium-Gallium-Arsenid) sensors manage with exposure time of the order of a second to get a sufficient OH* signal. Due to technical progress, OH* airglow observations, which refer to ca. 85–87 km height with some seasonal and geographical variation (Wüst et al., 2017, 2020; von Savigny, 2015; Baker and Stair, 1988), have now been used for several decades to analyse gravity waves (GWs, see, e.g., López-González et al., 2020; Wüst et al., 2018, 2016; Offermann et al., 2009; Taylor, 1997; Takahashi et al., 1985) whose intrinsic frequencies are lower than the (angular) Brunt-Väisälä frequency, which is ca. 2.0–2.3×10-2 s-1 (=4.6–5.2 min) for the OH* airglow layer with some seasonal and geographical variation (Wüst et al., 2020).

GWs are of great interest because they contribute significantly to the vertical and horizontal redistribution of energy and momentum in the atmosphere. However, it is still difficult to measure where and how much energy they deposit: the spatial resolution of today's instruments for observing turbulence is often insufficient. Breaking GWs and the transition to turbulence was already observed around the 2000s: Yamada et al. (2001), for example, sensed the OH Meinel bands between 680 and 1000 nm over an area of appr. 180km×180km with a 1024px×1024px array and detected wave fronts as well as turbulence-like features. Fritts et al. (2002) derived amongst others the momentum flux for this event based on additional meteor radar wind measurements. In 2021, Sedlak et al. (2021) and Hecht et al. (2021) were able to observe the breaking of GWs using two-dimensional OH* airglow measurements with a much higher spatial resolution and calculated the energy dissipation rates without the use of further measurement techniques. The spatial resolution of their observations is in the range of ca. 24–25 m px⁻¹, the derived energy dissipation rates are 0.08 and 9.03 Wkg-1. In order to achieve such high spatial resolutions, it is necessary to focus on a relatively small area of the sky (13.1–14.1km×13.4km in Sedlak et al., 2021, 47.4km×47.4km in Hecht et al., 2021) near the zenith. At the same time, the observation of an extended area (preferably all-sky) would be desirable to derive as much information as possible.

One, albeit not perfect, way to achieve this is to operate a scanning camera that measures all-sky but at the same time has the temporal and spatial resolution to at least recognise small-scale wave-like signatures, which are referred to as ripples in literature. Ripples are characterized by horizontal wavelengths of 5–15 km (Li et al., 2005; Taylor et al., 1995) or of 20 km at most (Takahashi et al., 1985). Their lifetime is in the range of 45 min or less (Hecht, 2004). In the past, such small-scale structures were often attributed to instability signatures, which may be related to or also be part of a breaking process of a GW (e.g. Li et al., 2005; Hecht, 2004; Fritts et al., 1997). If this is the case, they are supposed to move with the background wind (Hecht, 2004; Fritts et al., 1997). Their orientation with respect to the generating GW is sometimes interpreted with respect to the generating process (convective or dynamical instability). Today, it is known that such small-scale structures, especially if not moving with the prevailing wind, can also be secondary waves (Li et al., 2017). Unfortunately, wind information is often not available at OH* airglow measurement sites. This makes it difficult to differentiate between secondary waves and instabilities. However, in order to distinguish between both features, we propose to align a second camera with a higher spatial resolution to the ripple. If turbulence is observed, this is not only an indication of a breaking GW, the size and speed of the vortices relative to the surroundings can also be used to provide information about the dissipated energy (Sedlak et al., 2021). The first step in implementing this concept is the operation of such a scanning camera and the rapid analysis of its data with regard to small-scale wave-like structures in the spatial range of ripples. This part is the focus of this manuscript. Since the term ripples is associated with instability structures in literature, we call the analysed signals small-scale wave-like structures in the following.

At DLR Oberpfaffenhofen (48.09° N, 11.28° E), Germany, a scanning camera has been operated since 2019. It is based on the FAIM (Fast Airglow IMager) instrument already presented in Hannawald et al. (2016). Operated in the scanning mode, as presented here for the first time, it takes 35 pictures in consecutive order of different azimuth and zenith positions, which together cover an area of ca. 500 km diameter over Oberpfaffenhofen with an average spatial resolution of approx. 150 mpx-1. One complete scan takes 2:03 min. Every night, it delivers approx. 1.5–2.0 GB of image data. Due to this large amount of data, machine learning algorithms have already been tested or used to distinguish cloudless from cloudy or dynamically quiet from dynamically active episodes in camera images (Sedlak et al., 2023). However, in all these cases the cameras were static.

This publication now presents a machine learning method to identify spatially confined wave-like signatures in the scan images and to derive their horizontal wavelengths. Before machine learning became popular, such analysis was mostly done by eye, i.e. difference images were calculated to remove stationary or nearly-stationary structures or to reduce their signal. The wave signatures were then identified and characterized by eye (Li et al., 2017; Suzuki et al., 2011; Yue et al., 2010). We rely here on a modified YOLOv7 (You Only Look Once, version 7, Wang et al., 2023) algorithm to identify such signatures and characterize them with respect to horizontal wavelength and their direction of advection. Already Lai et al. (2019) presented a machine learning method to detect spatially confined wave signatures in OH* airglow images derived by a static camera. Those authors use a two-stage method, i.e. a classification model based on convolutional neural network (CNN) for the selection of clear-sky images and an object detection model based on faster region-based convolutional neural network (Faster R-CNN, Ren et al., 2017) for the localization of wave patterns. The horizontal wavelengths were then derived using a two-dimensional (2D) FFT. In contrast to their work, YOLO is a one-stage detector, i.e. the objects are identified and classified in one step, which makes it faster than two-stage detectors; at least that used to be the case. The development in this thematical area is very fast and today's versions of the algorithms have improved to such an extent that there is less difference in time. We compare the accuracy of our modified YOLO algorithm with the accuracy of a 2D FFT. Furthermore, we derive the direction of advection.

This paper is organised as follows. Section 2 introduces the scanning FAIM instrument and the pre-processing of the raw data. Section 3 contains a detailed description of the YOLOv7 algorithm and its modifications. Section 4 presents the results and Sect. 5 their discussion. The paper ends with a summary and outlook in Sect. 6.

In order to achieve the results shown in Sect. 4, a neural network is trained. Due to its speed and accuracy (Meng et al., 2023; Zhao et al., 2023), YOLOv7 is chosen for this purpose. To prevent overfitting and increase inference, the YOLOv7-tiny variant is chosen.

A neural network in general consists of different nodes arranged in layers. Each node receives different inputs, e.g. from other nodes in upstream layers. These inputs are weighted and added. A predefined non-linear activation function is applied to the result. The output of this activation function is forwarded to other nodes in subsequent layers, where the procedure is the same as just described. As part of the training of a neural network, the weights of the various nodes are optimised against a loss function until the output of the neural network is sufficiently good. This is therefore a non-linear optimisation problem.

The weights can be chosen arbitrarily at the start of training or a pre-trained network can be used. The advantage expected from a pre-trained network is that the optimisation problem can be solved more quickly. In addition, better results have also been observed in some cases (Hendrycks et al., 2019).

BYOL (Bootstrap Your Own Latent), an unsupervised learning approach, is used here for pre-training, i.e. to determine an initial set of weights for the YOLOv7 backbone, which consists of all but the last layers of the YOLO network. BYOL training is performed on a pretrained COCO backbone using colour jitter, random flip, random resize crop, normalization and gaussian blur augmentations. This BYOL pretraining increased convergence speed of YOLOv7. In these backbone layers, specific features of pictures are identified. In the last layers of such a network, in the so-called head, such features are used to identify the object, which is present in the picture. In the work by Sedlak et al. (2023), a neural network was trained for FAIM 3 (static camera in Oberpfaffenhofen) that can distinguish between dynamically quiet, dynamically active and cloudy episodes. The field of view of FAIM 3 lies within the field of view of FAIM 4. The field of view of FAIM 3 is only 5.5°×6.9° using a 100 mm lens; that means that it covers an area of approx. 10.5km×8.35km (in zenith direction) with an average resolution of 16 mpx-1 at the height of the OH airglow layer using a 640px×512px focal plane array (see Sedlak et al., 2016, for more details and especially their Fig. 7 for a comparison of FAIM 3 with a typical all-sky imager). For this study, BYOL is trained with 10 000 FAIM 4 images that were dynamically active according to the FAIM 3 analyses of Sedlak et al. (2023).

In contrast to “classical” object detection applications (e.g., looking for a cat in a picture), the small-scale wave-like structures we focus on here are less clearly separated from the background (Fig. 4). Therefore, it is often difficult to precisely define their boundaries. This makes it equally difficult to use conventional algorithms to detect the direction to which the structures move.

In the following subsections, the basic idea of YOLOv7 is described as well as its modifications for the identification of horizontal wavelengths and orientation. Different quality measures for YOLOv7 are presented and the derived wavelengths and the orientation of the wave fronts are compared to results based on a 2D-FFT. Finally, the algorithm applied for the derivation of the direction of advection of the identified small-scale structures is introduced.

As mentioned above, the focus of this manuscript is on small-scale wave-like structures with horizontal wavelengths in the range of ripples. This section is divided into two parts: the first one deals with the frequency of occurrence of such structures, the second one with the direction perpendicular to the orientation of the phase fronts (intrinsic direction of wave propagation) and the direction of advection. Some results are separated according to the different seasons. If a distinction is only made between summer and winter, then the term summer refers to the months of April to September. October to March are assigned to winter. If necessary, an additional differentiation is made between spring and autumn. In this case, spring includes March until May, summer June until August, autumn September until November, and winter December until February.

A total of 264 824 FAIM 4 images taken in 941 nights between October 2020 and September 2023 are investigated, of which 49.2 % are from the summer and 50.8 % from the winter season. During 133 nights in this time period, the measurements were not possible due to technical problems. So, the instrument worked properly in ca. 88 % of all cases. For all analyses, we use only images, whose cloud coverage probability, as it was described in Sect. 2, is less than 10 %. The number of detections of wave-like structures varies with the chosen confidence value. Starting with ca. 70 000 detections for a confidence value of 0.1, they drop sharply to less than 50 % for higher confidence values (Fig. 8). We choose a confidence value of 0.5. That means ca. 16 000 small-scale structures with horizontal wavelengths between 3 and 19 km are identified with a reasonable amount of confidence for our investigations. The distribution of these horizontal wavelengths is right skewed for both seasons and shows a maximum at 7 km for summer and 8 km for winter (Fig. 9). The number of detections is higher in summer than in winter.

If we no longer differentiate by horizontal wavelength, but calculate a kind of probability of occurrence, so add up the images with at least one small-scale wave-like structure and divide by the total number of images per season, seasonal variations become apparent (Fig. 10). The occurrence probability ranges between ca. 13 % and 24 % over all seasons with a clear maximum in summer, while the values for the other seasons do not differ all too much, taking their standard deviation into account. Seasons of one year with less than 900 cloudless images are excluded. This holds for winter 2020, summer 2021, spring 2023 and autumn 2023. The lower values during these seasons are a result of technical problems of the instrument and bad weather.

The analysis of the direction of advection and the intrinsic propagation direction is based on two different approaches: firstly, the approach described in Sect. 3.3 is used, i.e. in this case the direction of movement of the rectangle covering the wave structure (bounding box), the advection of the air parcel, is tracked. This approach is based on the analysis of FAIM videos. For the second approach, the orientation of the wave fronts is determined based on the horizontal wave parameters from individual images. GWs propagate in the direction of the wave vector or perpendicular to the orientation of their wave fronts, so 90° needs to be added to the wave orientation to give the propagation direction with a 180° ambiguity. The ambiguity is due to the use of individual images instead of a sequence of images: it is not possible to distinguish whether a north-south oriented wave is propagating eastwards or westwards, for example. For the intrinsic propagation direction, so based on the individual images, there is a clear north-easterly (±180°) component at all times of the year (Fig. 11). This is most evident in summer. During the other seasons there are additional weaker maxima with a southerly (±180°) component. The direction of advection derived from the videos shows more variation (Fig. 12). While a west-southwest component dominates in spring, the direction of advection seems to shift to the south in summer. In autumn and winter a clear east component can be observed.

In the following, we compare our results to other already published ones referring to ripples. As the extraction of ripples is time-consuming, there do not exist many publications that cover data of some years. We focus here on two publications referring to the Northern hemisphere (Yue et al., 2010; Li et al., 2017) and two referring to the Southern hemisphere (Suzuki et al., 2011; Rourke et al., 2017).

Yue et al. (2010) analyse data from two imagers, one at the Yucca Ridge Field Station, YRFS (40.7° N, 105° W), Fort Collins, Colorado, USA and another at two different locations in Japan (first Shigaraki, MU Observatory (34.9° N, 136.1° E) and then Misato Observatory (34.1° N, 135.4° E), Wakayama, Japan). The data from the US device covers five years (2003–2008), that from the Japanese four years (1999–2003). Both systems measure OH*, but unlike ours, not above 1 µm. This means that the signal comes from a slightly different altitude range (order of magnitude: some hundred metres) than that of FAIM 4; however, this should be irrelevant for further discussion. Both instruments are all-sky imagers with a 512px×512px sensor, the spatial resolution is not explicitly stated, but is estimated to be less than ours (see Fig. 2). The temporal resolution is 2 min for the US imager which is comparable to ours and 5.5 min for the Japanese imager. The authors calculate difference images in order to emphasize small-scale travelling structures and get rid of stationary ones. This means that periods of 4 min are amplified in the data from the US imager, while periods of 11 min are amplified in the Japanese data. Periods longer than 4 or 11 min are attenuated. Some attenuation of ripples cannot be excluded, as they move with the background wind. The calculation of difference images leads to a varying filter function depending on the background wind, the horizontal wavelength of the ripples and their orientation to the wind

Destructive interference appears for example, if the wind component which is perpendicular to the orientation of the wave is in the range of half a horizontal wavelength divided by 2 or 5.5 min. For a wave with 10 km horizontal wavelength measured by the imager in Japan this is the case when wind perpendicular to the wave orientation accounts for ca. 15 ms-1. The smaller the horizontal wavelength, the weaker the wind needs to be for the filtering to occur.

It appears that Li et al. (2017) partly use the same observations at the YRFS as Yue et al. (2010). However, they cover the shorter period of 2003 until 2005 and use additional measurements from a sodium lidar and an MF radar. They investigate an area of 200km×200km and compute difference images. In contrast to Yue et al. (2010), they make a selection of ripples depending on the availability of lidar and radar measurements. Thus, they analyse about 320 ripples. They define the probability of occurrence in another way than Yue et al. (2010): Li et al. (2017) divide the total ripples' lifetime to the total observed time under clear sky and no-moon conditions. They find a maximum of ripples in autumn (September–November) and a minimum in summer (June–August). This roughly compares with Yue et al. (2010). Higher values in autumn and winter and lower values in spring and summer can be read from their Fig. 2b, which shows the frequency of occurrence on a monthly basis. However, the maximum is probably in winter and the minimum in summer. This comparison may suffer from the slightly different definition of the parameter probability of occurrence and also from the calculation of seasonal values from monthly ones. Li et al. (2017) summarize the seasonal variation of the advection direction of the ripples with northward in spring and summer and southward in winter.

The analysis of Suzuki et al. (2011) and Rourke et al. (2017) both refer to measurements at or near the South Pole. The measurement setup and the data analysis of Suzuki et al. (2011) is quite similar to that of Yue et al. (2010). The authors use an all-sky camera (512px×512px) and calculate time difference images. The temporal resolution is slightly better than the one of our scan images with about 100 s and they observe a slightly higher airglow layer (sodium, ca. 90 km height). Due to polar night conditions, their data cover the months April–August of the years 2003–2005 and they observe 213 cases of ripples. For them, ripples have horizontal wavelengths shorter than 17 km. These structures move mainly towards 90–120° E and 300–330° E.

Rourke et al. (2017) refer to wavelengths shorter than 15 km as ripples. In contrast to the other studies mentioned above, they do not use an all-sky imager but a scanning radiometer (UWOSCR, University of Western Ontario SCanning Radiometer), which is based on an InGaAs photodiode. The UWOSCR is sensitive to the spectral range between 1100 and 1650 nm and provides information about an area of the night sky of 24km×24km divided into 16px×16px. The authors use a Fast Fourier transform (FFT) to determine the wave period between 2 and 16 min and correlation analyses in orthogonal directions of the image to derive the zonal and meridional velocity components. Horizontal wavelengths are calculated from the period and the horizontal phase velocity. Ripples are mainly observed from March to June and very few ones from July to October. They have a horizontal wavelength of 8–10 km; their preferred observed propagation direction is westward and poleward, for the intrinsic propagation direction the westward component dominates. This analysis covers the years 1999 to 2013 and includes about 400 cases.

Compared to the results of the authors above, ours are based on about two orders of magnitude more small-scale wave-like structures, although the number of observation years is of the same order of magnitude. Since Oberpfaffenhofen, which is North of the Alps and in the densely populated area of Munich, is not a station with outstanding good observation conditions in terms of cloud coverage and artificial illumination, this effect can be attributed primarily to the use of the developed machine learning algorithm. Other reasons are the better spatial resolution compared with all-sky imagers and possibly also the use of original images instead of difference images. The latter can lead to a reduction of the data basis.

It is not possible to compare the absolute probability or frequency of occurrence, as the various authors define these slightly differently. However, one can compare the relative course. Figure 13 shows the values read from Yue et al. (2010) and our own. The values for the southern hemisphere are not included here, as these authors either do not provide them or it is not possible to read them from their figures. Even if there are differences between the individual courses, the data sets show increased values from May or June to July or August. Additionally, the US observations are characterized by significantly higher values from October to February. This does not hold for the Japanese data and our own. When mentioning the period May to July or August, one should note that this is the time when the OH* measurements actually address the mesopause. The rest of the year, the mesopause is higher (at approx. 100 km, She et al., 2000, Berger and von Zahn, 1999). The temperature gradient in the mesosphere is thus steeper from May to August, which leads to an increased probability of convective instability. As far as dynamic stability is concerned, one can see in Fig. 1 of Jaen et al. (2023) that the vertical shear of the horizontal wind measured at Collm, Germany, approximately 380 km northeast of Oberpfaffenhofen at the height of the OH* airglow layer is most pronounced from May to September with a maximum around July. Therefore, we conclude that the probability for dynamic instability is enhanced for the OH*-airglow layer height in the summer season. This increased tendency to instability, convective or dynamic, could therefore contribute to or explain the higher probability of instability features (showing up as small-scale wave-like signatures) at this time of year. Even if the instability tendency remains constant, a higher probability of such instability features may also occur during May to July/August, if more gravity waves, which might break, are observed at this time. This is indeed the case as Fig. 14 shows. The number of wave events in this figure with horizontal wavelengths longer than 15 km are derived according to Hannawald et al. (2016) from non-scanning FAIM data sets recorded in Oberpfaffenhofen and some neighbouring Alpine regions (SBO: Sonnblick Observatory (47.05° N, 12.95° E), Austria and OTL: Otlica Observatory (45.93° N, 13.91° E), Slovenia). They are characterized by a maximum in spring, especially in May. The smaller-scale instability structures with wavelengths shorter than 15 km have a maximum in the summer months May to July. Similar results were obtained by Wüst et al. (2016) and Sedlak et al. (2021), however, these authors do not focus on the number of waves, this parameter contributes indirectly to their results. Based on OH*-airglow spectrometer data of different European stations (in the vicinity of the Alps), Wüst et al. (2016) show that gravity waves with periods greater than 60 min transport more potential energy not only in winter, but possibly also in May and June. Sedlak et al. (2020) who evaluated FAIM data in the Alpine region with regard to their gravity wave activity and depending on their period, see a similar behaviour for medium-frequency waves. Low-frequency waves have a maximum in winter, high-frequency waves are characterized by a relatively constant activity. This could be interpreted as an indication that the observed small-scale wave-like structures are due to medium-frequency waves.

Due to the work of Li et al. (2017) at the latest, however, it is clear that such small-scale structures could equally be secondary gravity waves. These waves are generated through breaking primary GWs (see e.g., Becker and Vadas, 2018, and citations therein). In our measurements, as in many others, it is not possible to distinguish between secondary and primary waves. Further discussion at this point is therefore superfluous.

Instability features are supposed to move with the background wind (see e.g. the argumentation in Li et al., 2017, in order to discriminate instability features from secondary GWs), GWs do not need to do this. In the intrinsic frame, GW move perpendicular to their wave fronts. This does not necessarily hold when they are observed while not moving with the background wind; if the instrument is static, the observed wave propagation direction is the superposition of the intrinsic wave propagation direction and the background wind direction. In the case of an intrinsic phase velocity that is low compared to the background wind speed, the wave is primarily moving with the background wind, it is mainly advected. This is more likely the case for waves with small horizontal wave lengths as they are often associated with low horizontal phase velocities (see e.g. Wachter et al., 2015). Therefore, the term “direction of the advection”, which is used in Fig. 12, is not entirely correct but refers to the main reason for this movement. In order to compare the intrinsic wave propagation direction and the direction of advection, the latter are artificially degraded, i.e. 180° are added to all directions with a westward component so that they also lie in the range between north and south-southeast (Fig. 15). Since a wave-like structure must be present in several images in succession in order to capture its direction of movement in the video, the number of observed directions of advection is lower than the number of the intrinsic propagation direction retrieved from individual images. The results are therefore normalized for better comparability. Due to the different amount of data, the results are not compared quantitatively. Qualitatively, however, the difference between the intrinsic propagation directions and the directions of advection is most pronounced in summer.

For discriminating between secondary waves and instability features, the direction of advection and the direction of the background wind are compared. According to Jacobi et al. (2015), who measured at Collm, approximately 380 km northeast of Oberpfaffenhofen, the averaged wind is for SW (NE-ward) for spring, NW (SE-ward) for summer, and N/NE (S/SW-ward) for autumn and winter. The dominating directions of advection in our measurements (see Fig. 12) is SW in spring, S in summer, NE in autumn and E in winter. So, only in summer a significant amount of the observed small-scale wave-like features seems to be advected with the wind. Therefore, the probability that an observed small-scale wave-like structure is an instability feature (secondary GW) is highest (lowest) during summer.

Concerning the direction of advection, we observe in our data changes from south-west in spring to south in summer, east and/or south-east in autumn and east in winter (see Fig. 12). In general, all directions occur at all times of the year. Li et al. (2017) also observe all directions of advection except in summer. In spring, their analyses show a broad southerly component as well as a clear peak to the North. This peak persists in summer. It can no longer be observed in autumn and winter. In autumn, it is in general difficult to determine a dominant direction. A slight preference for a meridional direction may be observed. In winter, a minimum towards the west and a southern rather than a northern component can be recognised. In contrast to Li et al. (2017), we do not observe this strong northern component at any time of the year. However, we agree in the broad maximum to the south in spring. In winter, we observe a preferred eastward component, while Li et al. (2017) derive a minimum in westward direction. This is at least not contradictory. However, in summer, our results and those of Li et al. (2017) are contradictory and while the results of Li et al. (2017) are less clear in autumn, we observe a clear eastward and poleward component. In the southern hemispheric autumn (March to June), a strong westerly and at least to some extent poleward intrinsic propagation direction is observed (Rourke et al., 2017). A comparison with Suzuki et al. (2011) is not made here, as the station is located directly at the South Pole and is therefore subject to specific meteorological conditions.

So, there appears to be less agreement with regard to the direction of propagation than with the seasonal frequency of occurrence. One reason for this could be that local effects dominate. If the wave-like structures are interpreted in terms of instability features, one could argue that they move with the local wind and the wind is variable in space and in time, even during one night since it shows strong variations due to tidal activity. As already mentioned above, the interpretation as instability features is most likely correct in summer and indeed in this season, our advection directions and those of Li et al. (2017) are contradictory.

In this study, we analyse ca. three years of OH* airglow all-sky images for spatially confined wave-like structures with horizontal wavelengths of ca. 20 km and smaller. The data were recorded by a scanning OH* airglow camera, which is operated operationally every night at DLR Oberpfaffenhofen (48.09° N, 11.28° E), Germany since 2019. It provides nearly all-sky images (diameter ca. 500 km) with a spatial resolution of ca. 150 mpx-1 (at 30° zenith angle) and a temporal resolution of ca. 2 min. In order to identify those small-scale and spatially confined structures, we adapt and train YOLOv7, a deep convolutional neural network, to derive the position and the horizontal wavelength. Our results can be summarized as follows:

We identify ca. 16 000 small-scale wave-like structures. Studies mentioned in literature cover two orders of magnitude fewer events. The main reason is that former studies are based on the visual inspection instead of machine learning. Further reasons might be that we use a scanning imager that provides a better spatial resolution compared to non-scanning all-sky imagers. Scanning photodiodes cover a smaller part of the sky or work with lower resolutions than we do. Our mode of operation therefore increases the probability of observing small-scale wave-like structures.

In order to obtain more precise information about the reason for the occurrence of these small-scale structures in future, the scanning FAIM system is to be combined with a camera with better spatial resolution. This will make it possible to visualise turbulent structures and thus not only distinguish between instabilities and secondary waves but also derive information about the amount of energy transferred (e.g. following the approach of Sedlak et al. (2021) that needs to be adapted to large data sets). This means that the possible position of the wave-like structure is to be recorded using FAIM 4, the position is forwarded to the higher-resolution camera, which then aligns itself with it (operating-on-demand). The work presented in this manuscript is considered a first step in this direction.

Finally, here are some ideas on how to improve the modified YOLO algorithm: The most obvious would be the addition of more labels. This generally leads to more robust and universal detections. A major issue in creating the labels and in subsequent training is the subjective boundary between an event worthy of annotation and one that is not. Here, it would be advisable to also create confidence values for each box while labelling, and to modify the YOLOv7 algorithm so that it predicts not just zero or one for the object value, but the labelled confidence value. The same data set could also be labelled by multiple scientists to calculate confidence values based on the intersection of labels. Another possibility would be the labelling of “ignored areas”, where the scientist is not sure about the existence of wave structures. These areas could then be ignored by the loss function. However, each of these approaches is pretty time-consuming and may be done in the future.