Deep Space Observations of Terrestrial Glitter

Deep space climate observatory (DSCOVR) spacecraft drifts about the Lagrangian point ≈1.4–1.6 × 106 km from Earth, where its Earth polychromatic imaging camera (EPIC) observes the sun‐lit face of the Earth every 1 to 2 hours. At any instance, there is a preferred (specular) spot on the globe, where a glint may be observed by EPIC. While monitoring reflectance at these spots (terrestrial glitter), we observe occasional intense glints originating from neither ocean surface nor cloud ice and we argue that mountain lakes high in the Andes are among the causes. We also examine time‐averaged reflectance at the spots and find it exceeding that of neighbors, with the excess monotonically increasing with separation distance. This specular excess is found in all channels and is more pronounced in the latest and best‐calibrated version of EPIC data, thus opening the possibility of testing geometric calibration by monitoring distant glitter.

tained by comparing EPIC observations with measurements taken by low Earth orbit satellite instruments such as moderate resolution imaging spectroradiometer (MODIS) and (OMI) ozone monitoring instrument (Geogdzhayev & Marshak, 2018;Herman et al., 2018). Several "superglints", with ρ ≥ 1.3, occurred over land in Peru and Bolivia, high in the Andes (the only high mountains in tropics, where EPIC viewing geometry permits glints) and caught our attention.
As a case study, consider a single color (blue) example of an intense detector-saturating glint (superglint) over cloud-free land shown in Figure 1. We chose blue glint because of the highest spatial resolution, ≈8 km . To demonstrate cloud-free conditions, in Figure 1 we present "before" and "after" images of the region by Terra (10:50 a.m.) and Aqua (1:40 p.m.) MODIS for the same day. These natural color composite RGB images (MODIS bands one (red, 0.65 μm); four (green, 0.55 μm), and three (blue, 0.47 μm) show nearly cloud-free conditions over the Andes, except for marine stratocumuli observed by Terra (see panel 1b), far away from the specular spot (red pixel in Figure 1) and a few low-level liquid water small cumulus clouds, observed by Aqua (see panel 1c). It is quite unlikely that ice clouds were absent at 10:50 a.m., appeared at noon and disappeared at 1:40 p.m. See additional evidence in panels (d) Bréon & Dubrulle, 2004;F. Bréon, 1993;F. Bréon & Deschamps, 1993) but no data of sufficient quality was obtained to document it (Bréon, private communication.) In contrast, large lakes such as Titicaca in Peru have been used as calibration targets by the Hyperangular Rainbow Polarimeter cubesat (Vanderlei Martins, private communication) and airborne observations of lake glints have been well documented (Gatebe & King, 2016).
To examine the physics of the superglint more closely, we recall that multicolor glint-causing conditions (composite image) must persist for at least 4 min (the lag between blue and red channels), implying considerable spatial extent as the Earth spins ∼100 km during such time. Conversely, single color glints such as the one in Figure 1 had been caused by briefer and more fleeting conditions. We therefore draw the reader's attention to the snapshot, panel (d) of Figure 1, depicting the raw, Level 1A, 443 nm data alone. Locations of particular interest are: the specular position (marked by yellow star), a stand-alone pixel of maximal reflectance, marked by the pink star, (ρ ≈1.31), and another superglint where saturation appears to have occurred as indicated by the thick straight red line, due presumably to the CCD spillage along the detector columns, for example, (Gorkavyi et al., 2020).
Let us examine the physics of specular (singly scattered) reflection in the context of panel (d) of Figure 1. For a single color image, most of the pixel-to-pixel reflectance (ρ) variation can be attributed to: (i) illumination conditions; (ii) differences in the effective index of refraction; and (iii) angular concentration of reflected radiation. While the first two vary by an order of magnitude or so (Rees, 2013), the last can vary by three orders of magnitude (hence, frequent saturation from glints). Indeed, Lambertian surface scatters perfectly diffuse radiation into an entire hemisphere, i.e., a solid angle of dΩ = 2π, whereas a smooth glint-producing facet can concentrate that singly scattered radiation into a solid angle dΩ ≈1 × 1° or ∼10 −4 fraction of 2π. We use the 1° estimate because the angular size of the sun is ≈0.5° and so is the spread of incident sunlight. In addition, the surface normal change across a single pixel is ≈0.1°, affecting both incidence and reflection angles (pixel extent of ≈(3 × 10 −4 )° is negligible).
KOSTINSKI ET AL.   , during which the Earth spun ≈13 m. Pink square marks second reflectance (ρ) peak, likely not caused by CCD saturation spillover. The red squares in panels (b) and (c) center on the specular spot (ρ = 1.29), demarcating EPIC pixels of ≈8 km size that day. The extended glint is, likely, due to the mountain lake, but the other glint, pink star in panel (d), with the maximum reflectance ρ max = 1.31, is perplexing. (See also Figure 2 images occasionally attain dynamic range values of ∼10 3 , the upper bound being set by the 12 bit (=2048) detection system. It is the minimal ρ that varies most from image to image and in the raw (level 1A) images of the blue channel it is set by the atmospheric Rayleigh scattering (optical depth of 0.24 at 443 nm) for pixels near the image edge, by the dark side of the Earth. In contrast, the upper bound on ρ is set by detector saturation, the latter depending on the exposure time. The exposure time is set so that bright cloud pixels (ρ ≈0.9) fill the potential well of the EPIC CCD detector to about 80% of its 95,000 electrons saturation limit (Jay Hermann, private communication). The EPIC maximal signal-to-noise ratio (SNR) = (0.8 × 95,000) 1/2 = 275 (at near saturated pixels), with the exposure time adjusted accordingly. Using this SNR ∼10 2 for a ballpark estimate, one then expects (conservatively) the glints to saturate the detector at ∼10 2 of KOSTINSKI ET AL.
10.1029/2020EA001521 4 of 8 Pixels are geographically positioned, with their areas demarcated by the colored squares, (7.45 km on a side, the EPIC resolution for the day of observation). The red pixel covers the specular spot and is centered at the yellow star, whereas the pink star flags the puzzling spot of maximal reflectance where a rocky river bed appears, possibly with streams at the time of EPIC image. The areal extent of the mountain lake is comparable to the pixel area (scale bar in the lower right corner). Note more than a pixel separation between the red line (aligned with CCD columns) and the second superglint, suggesting that perhaps the pink star reflectance is not merely an artifact of CCD spillage. Panels (b) and (c): Aqua and Terra MODIS cloud optical thickness (τ) maps of the scene, superimposed on the natural color panels 1b and 1c. Infrared and visible channels contribute to τ-maps and show small subpixel clouds (yellow and red) detected there by Aqua 1:40 p.m. local time but absent at 10:50 a.m. local time in Terra images. These clouds were classified as liquid phase by Aqua despite the relatively high altitude. No ice clouds are found (no purple, blue, or green). Panels (d) and (e) the perfectly diffuse Lambertian pixel value near the image edge. Therefore, an important conclusion in the present context is that a mere one percent fraction of the 100 km 2 pixel area, occupied by a perfect specular reflector with the right geometry, saturates the detector. Thus even a small calm mountain lake of ∼1 km 2 area can cause a superglint. Figures 1 and 2 confirm that we can add small lakes to cloud ice and ocean surface as potential causes of EPIC superglints.
Discovering that small lakes can beam focused light all the way into deep space is rendered plausible by the above back-of-the-envelope estimate. However, the second peak, flagged by the pink square in panels 1b-d and 2a-c, (ρ = 1.31) is perplexing. Could it possibly be the tiny subpixel warm cloud in the Aqua images of panels 1c and 2c? Liquid water drops, although conceivably retro-reflectors, have never been implicated in specular reflections to space, to the best of our knowledge. Nor is it likely that the tiny cloud, if misclassified by Aqua and mixed phase in reality, would have contained enough ice crystals to yield a superglint. Indeed, it is tempting to dismiss this superglint as an instrumental artifact such as a spillage of the incoming radiance along the columns of the CCD detector, for example, (Gorkavyi et al., 2020). Closer inspection of panels 1 e, f reveals unrealistically steady ρ values in columns on either side of the column, containing the main glint, with low values of ρ in the next-neighbor columns and slightly higher but still unsaturated ρ in the second column on either side. However, panel 1 e also shows that the CCD spillage in the second column to the left caused two neighboring pixels to have identical mid-range ρ values whereas the second peak pixel has a much higher, saturated ρ. This permits the scenario that some neither lake nor ice cloud feature in the area of the second peak pixel increased ρ well over the value implied by the CCD spillage.
Therefore, we raise the admittedly far-fetched possibility that superglints could also be caused by small features, capable of constructive interference. For example, salty deposits in Bolivia and snowy mountains in Ecuador are within the DSCOVR observational band of latitudes. Could a superglint be a reflection from stationary elements such as snow or salt deposits or even fleeting ones such as facets of capillary waves on ponds (Lynch et al., 2011)? In contrast to incoherent glints, caused by angular concentration of scattered radiation, the coherent ones would depend greatly on the exposure time, i.e., constructive interference relies on maintaining zero relative phase between pairs of scattering elements (e.g., facets of the lake surface, or ice crystals in clouds), throughout the time of observation.
To that end, we ask: how far should the detector be from the terrestrial image pixel, for the latter to be considered a point source? This spatial coherence (all parts radiating in phase) condition is given by, for example, (Crawford, 1968 where the two transverse lengths d and D are the detector and source (pixel) sizes, respectively and the two longitudinal length, λ and L are wavelength and Earth-satellite distance, respectively. From this perspective, EPIC/DSCOVR deep space observations are particularly interesting because of the sheer magnitude of the observation distance. Inserting numbers in meters gives 3 × 10 −1 × 10 4 ∼ 0.5 × 10 −6 × 1.5 × 10 9 , yielding ∼10 3 for both sides. Thus, partial coherence can occur and specular areas must be smaller than the mountain lake to satisfy (1), e.g., 1% of the pixel area. Next, we examine the temporal coherence of scattered light. One condition is the constancy of relative positions of scattering facets during the exposure time of ∼10 ms. These can be readily satisfied by stationary elements on the ground but not necessarily by moving ice crystals in clouds or capillary waves on the lake surface. The hallmark of coherence is that the scattering intensity is quadratic rather than linear in the number of scatterers n, e.g., number of cloud ice crystals. Indeed, the total intensity I, averaged over the exposure time (averaging over the observation time is denoted by angular brackets) is given by where a i are the scattering amplitudes of n elements such as facets of a capillary wave. These amplitudes may be considered comparable for the purpose of this argument. In the incoherent case (e.g., radiative transfer), n 2 − n off-diagonal element time-averages vanish and only n diagonal terms remain so that < I > is indeed additive. Not so when the exposure time is comparable to the coherence time T. Even if a typical relative cross-term magnitude ϵ is relatively small, ϵ × n 2 ∼ n may still result. Although the natural coherence time of the filtered (width: 10 nm) incident blue light is on the order of microseconds, this discussion pertains to the coherence acquired via scattering.
For instance, the number of ice crystals in a cloud volume of 100 m (one mean free path of a tenuous ice cloud) × 1 km × 1 km ∼10 8 cubic meters, with number density of 10 3 per cubic meter, yields n ∼10 11 while the decorrelation time, estimated by setting the wavelength λ to the product of relative speed δu ∼1 mm/sec and the sought coherence time T, yields T ∼10 −3 sec. This msec is still short compared with the 46 msec exposure time (Geogdzhayev & Marshak, 2018) for the EPIC blue channel but even in this case, considering huge n 2 , partial coherence effects may be discernible if ϵn ∼1. The case is much stronger for stationary elements such as salt grains on the ground. EPIC/DSCOVR exposure time can vary via the adjustable slit widths and spinning rates. This allows for testing the coherent glint conjecture, should opportunity arise, because, in contrast to incoherent glints, even relative coherent glint KOSTINSKI ET AL.
10.1029/2020EA001521 6 of 8 intensity within the scene depends nonlinearly on the exposure time so that changes in the saturation statistics would be a revealing symptom. Such coherence-induced many-fold increase in glint intensity could lead to detection of exosolar glitter.
We now return to the time series of glints (terrestrial glitter) but, in contrast to earlier work (Marshak et al., 2017), here we use neither detection algorithms nor any subjective criteria. Instead, relying on the fundamental physics that glints can occur only where an observation of specular reflection off properly oriented facets is possible, we monitor the reflectance versus time at these specular spots (one per each image). Specular reflector at a given instant of observation may be absent (e.g., the specular pixel is covered by forested land with no lakes nor cloud cover) and then the usual diffuse component is responsible for the observed pixel brightness. However, insofar as specular reflections (glints) are always brighter than the diffuse ones, one expects time-averaged brightness at these specular spots to exceed the rest of the pixels. The key ingredients here are statistics and geometry and we propose the notion of specular excess, denoted δρ as the quantity of interest: one expects δρ maximum at the specular position and decrease toward the typical diffuse reflectance as the distance from the specular spot increases. Hence, all pixels located, say, five pixels away from the global specular spot are averaged over with the associated < ρ > below the peak value. This specular excess hypothesis is confirmed by Figure 3, panel (a) where the red and black cross-hairs are nearly coincident. Panels (b) and (c) of Figure 3 further show that the positioning of the maximum is within a fraction of one pixel from the specular position.
The argument for geo-calibration by specular reflection excess statistics is further buttressed by the comparison of EPIC data versions. For the latest (third) version, coastlines were cross-checked with MODIS and GOES-17 to produce superior geolocation (Karin Blank, personal communication) as illustrated in Figure 4. Again, both images are centered to within a pixel of the reflectance maximum. Note that while the mean location of the intensity peak does not change much from second to third version, the variance about the mean location decreases substantially in the third version. This is a welcome finding as the smaller variance implies that it is now less likely for random uncertainties in geolocation to cause discrepancies with the geometric localization of specular points.
Lastly, Figure 5 demonstrates that the notion of specular excess holds for all 10 of the EPIC channels, rendering the finding statistically robust. What determines the order of glint specular excess curves? Presumably, the primary cause of the (land) excess is the reflection by ice clouds. Rayleigh scattering above clouds as well as absorption by oxygen and ozone diminish δρ as the increased absorption weakens the contrast. In other words, δρ caused by the glint contribution is least obscured by the weakest atmospheric absorption or scattering above the glint. For example, B-band oxygen absorption (688 nm) is much weaker than A-band absorption (764 nm) and indeed, the 688 nm curve is above the 764 nm one. Interestingly, the 688 nm curve is also close to 443 nm curve: one due to Rayleigh scattering (443 nm) and one due to oxygen absorption (688 nm). Aside from oxygen absorption, the rest of the curves are ordered nearly by the wavelength, in agreement with the Rayleigh background interpretation. Our most recent findings on seasonal and spectral aspects of glints can be found in (Várnai et al., 2020).

Summary
In summary, we found that even small lakes can project superglints into deep space and that such features, in addition to being calibration targets, cause statistically steady terrestrial glitter enhancement δρ and, thereby, can serve as beacons for monitoring geo-location. Should the coherent glint mechanism prove viable, it has the potential for exosolar glitter, linking geoscience with astronomy.
10.1029/2020EA001521 7 of 8 Figure 5. Specular Excess is spectrally robust δρ > 0 is observed in the vicinity (14 pixels) of the specular spot for all 10 EPIC channels. Hence, specular excess is a statistically robust finding. Data for the entire 2017: land surface and cloud glints. Rayleigh scattering above clouds reduces the excess more at the shorter wavelengths, and so does absorption by oxygen (at 688 nm and 764 nm in particular) and by ozone (at 317 nm and 325 nm). EPIC, earth polychromatic imaging camera.