The operational glint detection algorithm is based on the approach first introduced in Marshak et al. (2017). The algorithm works by comparing two EPIC observations of the same scene taken at slightly different geometry: one image that captures any tightly focused specular reflection from horizontal objects in the scene, and another image that narrowly misses any such reflection, as the scene has a slightly different sun-view geometry in this second image. In glint studies, the sun-view geometry is often characterized through the glint angle (δ), which is defined as the angle between the actual satellite view direction and the direction that would look straight into the specular reflection from a perfectly horizontal surface. The algorithm reports a glint if the reflectance (R) is much higher for the observation that can capture the glint because of its smaller glint angle.

Earlier studies showed that the specular reflection from ice clouds extends to glint angles of about 2° (e.g., Várnai et al., 2020a; Kostinski et al., 2021) due factors such as: 1) the roughly 0.5° angular diameter of the solar disc, 2) wobbling that can tilt horizontally oriented crystals by up to 1° in ice clouds due to local turbulence effects (e.g., Katz, 1998; Bréon and Dubrulle, 2004), 3) diffraction that occurs when sunlight encounters very small ice crystals (e.g., Crawford, 1968), 4) small ripples or capillary waves in mostly smooth water surfaces. We note that if wind and currents make water rough, the wide range of wave slopes will spread specular reflection into a wide range of view directions, which makes glints appear larger but fainter in satellite images (Várnai et al., 2020a). The current operational product, however, aims at detecting glints from clouds and smooth water surfaces (for example from calm small lakes (a few km in diameter), as in Kostinski et al., 2021); extending the product to also detect diffuse glints from rough water surfaces (for example using the model developed in Cox and Munk (1954)) is left to the future.

As mentioned in the introduction, EPIC can provide observations of a scene at several glint angles because it takes multispectral observations using a filter wheel—which means that it captures images at different wavelengths at slightly different times. For example, the 388 nm ultraviolet (UV) images are taken about 5 min after blue images. During these 5 min, the Earth’s rotation changes the normal direction of the observed scenes by 1.25°. Considering the law of reflection, a 1.25° change in the orientation of the reflector will change the direction of specularly reflected light by 2.5°. Thus, if δ was 0° for a pixel in the blue image, δ will be 2.5° in the UV image. This means that if the pixel contains a large number of horizontally oriented ice crystals that cause a strong glint through specular reflection, EPIC data can reveal this by showing that R blue ≫ R UV (1)

We note, however, that EPIC can detect even faint glints for which Eq. 1 is not satisfied. Faint glints don’t satisfy Eq. 1 because, in the absence of glint, R _blue < R _UV due to the stronger Rayleigh scattering at shorter wavelengths—and so faint glints may lift R _blue only to be comparable or slightly larger than R _UV. As described below, the EPIC glint detection can identify such faint glints by using thresholds that specify what R _blue values imply glint for a given R _UV.

As a preparation for the in-depth description of the glint detection algorithm, let us mention three considerations:

• We will describe the detection process only for blue glints, as the operational processing uses the exact same method for green (551 nm) and red glints (except that we use in Eq. 1 R _green or R _red instead of R _blue as the “glint wavelength”, and we can use R _blue instead of R _UV as the “non-glint wavelength”).

Ultimately, the algorithm flags a pixel as blue glint if R _blue > T (R _UV), where T (R _UV) is a threshold function value determined by taking into account spectral variations in scene reflectivity. Since spectral variations can be markedly different over different surfaces, the operational algorithm uses separate T threshold functions for three basic surface types: water, desert, and non-desert land. The surface type at the location of each EPIC pixel is determined from the 0.05° latitude-longitude resolution global map of surface cover provided in the MCD12C1 product of the Moderate Resolution Imaging Spectroradiometer (MODIS) instrument (DOI: 10.5067/MODIS/MCD12C1.006). The exact method for determining the T (R _UV) threshold function values for the entire range of possible R _UV values is described in the next section.

Admittedly, the selection of glint detection thresholds is somewhat arbitrary and involves a trade-off between accuracy and sensitivity. As usual, if we prioritize accuracy and use very strict thresholds, we will miss detecting some not-so-clear-cut cases affected by relatively faint glints. On the other hand, if we prioritize sensitivity, we will capture not only actual glints but will likely have many false detections as well. The EPIC glint product uses T glint detection thresholds that result in a False Alarm Rate (FAR) of 10%. This means that in 10% of pixels flagged as glint, R _blue exceeds T (R _UV) not because of glint, but because of other factors such as strong spectral variations in surface albedo—for example if a red dust plume drifts over the blue ocean.

We determine the glint detection thresholds using a pragmatic statistical approach. This approach avoids the need for physical modeling, which would be greatly complicated by variations in surface spectral properties and in the areal extent, altitude, phase, and optical thickness of clouds (which may also vary within the roughly 8 by 8 km-size EPIC pixels). We note, however, that the absence of physical modeling does not mean the lack of physical basis or constraints: The physics of specular reflection is essential to the method as it is based on glints being limited to locations with suitable sun-view geometries and relies on the constraint that glint from clouds have a narrow angular spread (typically less than 2°).

In determining what T thresholds will result in a FAR of 10%, we will rely on the statistical probabilities (P) that, for a random pixel, R _blue exceeds T (R _UV) due to any (glint or non-glint) reason (P _any) or due to non-glint factors (P _ng), respectively. Specifically, we state that non-glint factors do not vary much at small glint angles (   P n g ( δ < 2 ° ) ≈ P n g ( 2 ° < δ < 5 ° ) ) , and that cloud glints rarely extend to glint angle δ > 2° ( P a n y ( 2 ° < δ < 5 ° ) ≈   P n g ( 2 ° < δ < 5 ° ) ) . Based on these considerations, we can express the total FAR of the glint detection algorithm (applied to pixels with δ < 2°) as FAR ( δ < 2 ° ) = P n g ( δ < 2 ° ) P a n y ( δ < 2 ° ) ≈ P n g ( 2 ° < δ < 5 ° ) P a n y ( δ < 2 ° ) ≈ P a n y ( 2 ° < δ < 5 ° ) P a n y ( δ < 2 ° ) (2)

We use Eq. 2 in determining the T (R _UV) thresholds for each possible R _UV value through a three-step statistical analysis of all EPIC images taken during 2018. First, for each possible R _UV value, we calculate the probability distribution functions PDF(R _blue | R_UV, δ < 2°) and PDF(R _blue | R_UV, 2° < δ < 5°)_. Second, for each possible R _UV, we calculate P _any for a wide range of candidate T values as P any ( T | R UV , δ < 2 ° ) = ∫ R ∗ ∞ PDF ( R blue |   R UV , δ < 2 ° ) d R blue (3a) and P any ( T | R UV , 2 ° < δ < 5 ° ) = ∫ R ∗ ∞ PDF ( R blue | R UV , 2 ° < δ < 5 ° ) d R blue (3b)

Third, for each R _UV, we select the T value that, in Eqs 3a, 3b, yields such P _any values that, in Eq. 2, will give a FAR close to 0.1 (Figure 1).

Finally, for robust glint detections, we need to consider that the T (R _UV) threshold functions determined through the procedure described above can feature random jumps or drops due to sampling noise caused by the finite size of our dataset (containing all EPIC images taken in 2018). To avoid such unphysical jumps and drops, the raw T values are adjusted through a series of three smoothing operations:

• Resetting to zero the PDF values for RUV bins containing three or fewer pixels,

The final T thresholds obtained at the end of this smoothing process are illustrated in Figure 2. The figure shows that the T threshold values are capped at the saturation level of the EPIC blue detector (R _blue ≈ 1.3), as it would be pointless to set a threshold higher than any reflectance reported by EPIC. At lower reflectances, T increases with R _UV; the discernible changes in the slope of this increase are likely caused by transitions from clear sky to partly and then fully cloudy situations as R _UV increases.

Using the basic approach described in Section Basic Approach and the thresholds discussed in Section Threshold Selection, the operational algorithm detects blue glints through the following process:

• Eliminate EPIC images if data for any of the bands used by the algorithm is missing.

Figure 3 shows glint detection results for four EPIC images, three containing glints from ice clouds and one from the smooth ocean surface. Additional examples can be found at the EPIC website at https://epic.gsfc.nasa.gov/science/products/glint.

Figure 3A shows the glint displayed in Figure 1 of Várnai et al. (2020b). The figure illustrates that in case of extensive ice clouds, EPIC images can feature a colorful pattern in which the eastern and western edges of the glint appear blue and red, respectively, while the middle appears green. Green and especially red glints appear to the west of blue glints because EPIC takes the red (green) images 4 (3) minutes after taking the blue image—and during this time, the rotation of Earth brings a more westward location into the position where EPIC can observe specular reflection from horizontal objects. The typical distance between blue and red glints is about 100 km.

Figure 3B illustrates that glints affecting different EPIC bands can overlap, for example the overlap of red and green glints creates a yellow glint.

Figure 3C shows that in rare occasions, cloud glints can extend all the way up to 2° glint angle. This may indicate intense turbulence that increases the wobbling of horizontally oriented ice crystals.

Figure 3D shows a bright glint from the smooth ocean surface. For two reasons, we believe that while clouds may contribute to the glint at a few pixels, the vast majority of this glint originates from the water surface and not from clouds. First, the glint extends to a much larger area (and larger glint angles) than any cloud glints EPIC observed over land. The presence of glint at even large glint angles indicates that while the reflecting object may be smooth for a water surface, it is not nearly as uniformly flat as the horizontal ice crystals are in clouds. Second, for almost all glint pixels, the oxygen A-band ratios (R _{764 nm}/R _{780 nm}) are below the 0.45 threshold that has been proposed to distinguish surface and cloud glints (Várnai et al., 2020b). We point out that the glint mask readily detects this bright glint, but it does not detect the relatively faint diffuse glints from rough ocean surfaces, which appear relatively dark grey in the image. We note that such diffuse glints from rough ocean surfaces have been detected in images of several satellite instruments such as MODIS or the Multi-angle Imaging SpectroRadiometer (MISR) (e.g., Feng et al., 2017; Knobelspiesse et al., 2021).

Finally, a comparison of glint latitudes in Figure 3 illustrates that, as discussed in Marshak et al. (2017), the latitude of glint observations shifts with the seasons, with glints being detected in the Northern Hemisphere in July (Panel A), in the Southern Hemisphere in January (Panel B), and around the equator in September and April (Panels C and D).

Following the individual examples discussed above, this section presents some basic statistics about the glints detected by the operational algorithm in all EPIC images taken during 2017. The goal of this initial statistical analysis lies not in providing physical insights on atmospheric or surface properties, but in getting an initial glimpse of glint product behaviors. (We note that statistics for 2018 were very similar to those for 2017, but the examination of interannual variability is left to a future study of the entire EPIC data series.)

Figure 4 shows what fraction of pixels were flagged as glint for various surface types and glint colors, as a function of glint angle. The figure reveals several notable features discussed below.

The dominant feature of all panels in Figure 4 is a sharp increase in the fraction of pixels with glint at small glint angles. This is consistent with glint reflectances being largest at small glint angles (e.g., Várnai et al., 2020a; Várnai et al., 2020b; Kostinski et al., 2021) and can be attributed to specular reflection from horizontally oriented ice crystals in clouds and from very smooth water surfaces (such as small, calm lakes, as in Kostinski et al., 2021).

Figure 4A reveals that glints are detected more frequently in the blue EPIC band than in the green and red bands. (The figure considers all glints regardless of whether they come from the surface or clouds.) Green and red glints may be detected less frequently because, similarly to green (or red) glints, glintless white clouds also increase the ratios of green to blue reflectances or red to blue reflectances—which makes it harder to identify glints with a high confidence. The panel also shows that glint detections do not drop to zero at larger glint angles, probably due to somewhat wavy water surfaces that spread specular reflection into a wider range of glint angles.

A comparison of the peak values in Panels a, b, and c reveals that, even for blue glints, glint detection is much less frequent over ocean than land. Considering that about a third of land pixels is classified desert in our dataset, the land-ocean difference is roughly 7-fold, which is roughly twice as large as the factor of 3.5 between glint-caused reflectance enhancements over land and ocean in Figure 2A of Várnai et al. (2020b). This may show that it is harder to identify individual glints over ocean than over land and calls for future improvements in the detection of ice cloud glints over ocean.

The combination of Panels b and c shows that for very small glint angles, glints are detected at about 22% of all observations taken over land. Considering that at the latitudes of EPIC glint observations, the average ice cloud cover above land varies between 20 and 50% during the course of a year (Várnai et al., 2020b, based on King et al., 2013), finding glints in 22% of pixels suggests that glint detection over land is highly effective and that horizontally oriented crystals are quite prevalent in ice clouds. It is important to keep in mind, though, that glint may be observed even if horizontal ice crystals occur only in a portion of a typically 8 by 8 km size EPIC pixel. This implies that horizontal ice crystals likely occur over less than 22% of the actual land area. Moreover, we note that while the sharp drop-off in glint probability with glint angle suggests a dominance of such almost perfectly horizontal surfaces that seem more likely in clouds than in lakes or rivers, future studies will be needed to estimate the relative frequencies of cloud and surface glints.

The comparison of Panels b and c reveals that glints are detected less frequently over desert than non-desert land, due to the lower frequency of ice clouds (and lakes). Finally, we point out that over desert, red glints are detected less frequently than green or blue glints. This seems to indicate that changes in mineral composition of the typically brownish desert can create red patches even without glint, which makes it more difficult to identify red glints with high confidence.

By displaying the number and typical size of glints, Table 1 shows that over all surface types, glints are more frequently detected for the EPIC blue band than for the green or red bands—but also that blue glints are systematically smaller than green or red ones. Overall, the total glint area is similar for all bands over non-desert land, whereas over water and desert, blue glints cover the largest areas despite their smaller sizes.

Over water, blue glints may cover the largest areas due to the higher effectiveness of blue glint detection mentioned above. Similarly, the higher detection effectiveness discussed at Figure 3C may explain why, over desert, blue glints cover larger areas than red glints. The difference between the total area of blue and green glints over desert comes from small blue glints detected at relatively large (>1°) glint angles (Figure 3C), but future studies will be needed to explain these glints.

Finally, we note that for all surface types and EPIC bands, the median of the glint size is much smaller than the mean. This implies that glint size distributions have long tails toward large sizes, indicating a significant occurrence of very large glints. Such large glints may occur in cases of extensive ice cloud cover or large calm water surfaces (as in Figures 3C,D) but, detector saturation artifacts may also play a role (as in the lower right panel of Figure 1 in Marshak et al., 2017). As discussed in the next section, the origin of large—and small—glints can be best explored using additional information not present in the current, initial version of the EPIC operational glint product.