2.1.

Surface Sky Observations

Distilling surface sky cover observations into a percentage based on coded Meteorological Terminal Aviation Routine Weather Reports (METARs) cannot be done in a precise manner for multiple reasons. First, while surface weather stations have ceilometers to help with the estimation of cloud base height, human observers may opt to augment these reports at some stations, but there is no report indicator or practical method to assess whether this is done regularly or if the skill of different human observers at augmenting the instrumentation is good. Though darkness can increase challenges for observers estimating cloudiness overnight compared with during the day, an observer’s skill during a partial eclipse would likely compare with other daytime observations. Given that totality lasted under 3 min, the careful observer could easily assess cloudiness before and after the total eclipse if the unique illumination of the horizon made it challenging to ascertain clouds. Second, automated ceilometer reports use a time average from a point, which is typically preferential to low clouds, or clouds directly over the ceilometer, instead of capturing the entire celestial dome. A stationary cloud over a surface station ceilometer may lead to an overrepresentation of sky cover in the respective observation. Third, while cloud cover is observed in oktas, or one-eighth increments of fractional coverage, coded METARs constrain the observation to one of the five groups, CLR (clear), FEW (few), SCT (scattered), BKN (broken), and OVC/VV (overcast, or vertical visibility obscured).¹

For this study, the routine and special METARs were decoded and binned to a 10-km Lambert Conformal grid and, by bin, averaged over a 1-h window bounded by the top of the hours adjacent, with sky conditions assigned 0%, 25%, 40%, 75%, and 100% for the five progressively greater METAR cloud coverage groups, following an approach from the General Meteorological PAcKage, or GEMPAK software.⁶ The binning occurred in both time and space; however, the selection of the 10-km grid generally precluded the inclusion of more than one observation site in a grid cell, with very limited exception. Both automated and manual reports were binned, but routine observations for this study do not include those at a frequency of 1 or 5 min. These binned observations were then compared with visible satellite imagery from the eastern United States Geostationary Operational Environmental Satellite (GOES-13), valid at or shortly after the top of the hour, when most of the observations were made, as shown in Fig. 1.

Fig. 1

Surface sky observations. This four-panel plot shows decoded, binned, and averaged surface sky observations, in units of percent (%) coverage, on a 10-km Lambert Conformal grid over a portion of the southeastern United States for the 1-h period following (a) 1700 UTC, (b) 1800 UTC, (c) 1900 UTC, and (d) 2000 UTC on August 21, 2017, in the respective panels. Similarly, the underlying visible satellite images from a Geostationary Operational Environmental Satellite, GOES-13, are valid at (a) 1700 UTC, (b) 1815 UTC, (c) 1900 UTC, and (d) 2000 UTC on August 21, 2017. The satellite images are darker in the top-right and bottom-left due to the proximity in time to totality from the solar eclipse, which decreased incoming solar radiation, though less cloudiness was also evident.

Visible images from the GOES-16 Advanced Baseline Imager (ABI), which was not yet operational at the time of the eclipse but has two visible wavelength bands for monitoring clouds during the day, are similar. The lunar umbra and penumbra from the eclipse were also clearly evident in the ABI near-infrared bands, particularly the $0.86\text{-}\mu \mathrm{m}$ vegetation band,⁷ though the relatively high reflectance of land surfaces at that wavelength makes cloud detection more challenging. The presence of cloud on satellite imagery adequately corresponds to the gridded surface sky observations, though the coverage accuracy is difficult to assess other than subjectively.

2.2.

Comparisons and Trends

An hourly comparison of the mean of the stations shown in Fig. 1 is shown in Table 1 for each 1-h window between 1700 and 2100 UTC, along with the satellite-surface RTMA mean and human NDFD 1-h forecast grid mean for the corresponding four times. Between 1700 and 1800 UTC, the mean sky cover reported in surface observations was consistent with the mean value from the NDFD 1-h forecast grid, though that correspondence diverged for the remainder of the period of interest. The time of totality in the vicinity of Nashville, Tennessee, was around 1830 UTC. There was a minimum in mean total cloud cover from the RTMA at 18 and 19 UTC that was not reflected in the NDFD 1-h forecast grid mean and that preceded the minimum in the surface sky observations mean. The latter is likely due to the collection of many observations from surface stations solely at the top of the hour.

Table 1

For 1-h increments between 1700 and 2100 UTC on August 21, 2017, this table summarizes the computed mean, in percent (%) coverage, for surface sky observations, the RTMA total cloud cover analysis, and the NDFD 1-h sky cover forecast grid. The geographic area of interest from which the mean was computed is shown in Fig. 1. The valid time range for each column is 1 h following the header time. Each NDFD forecast was issued 1 h prior to the header time; the mean was computed for the valid time.

A value-binned comparison of the surface sky observations and RTMA total cloud cover analysis mean with the NDFD 1-h sky cover forecast grid is shown in Figs. 2 and 3, respectively, for the 19 UTC time window, following the departure of the total eclipse for this area of interest. For this case, a high fraction of nonzero NDFD forecast sky cover values correspond to observed clear skies (i.e., a cloud cover of $<5\%$), though this is particularly evident with the comparison with the RTMA. Almost all of the NDFD forecast values in the area of interest are greater than the corresponding RTMA cloud cover values at this time period.

Fig. 2

Comparison of surface sky observations to the NDFD 1-h sky cover forecast. There were 110 comparisons of gridded automated and manual surface sky observations with the NDFD 1-h sky cover forecast in the domain of interest (as shown in Fig. 1) for the 1 h following 1900 UTC on August 21, 2017. Comparisons are plotted in bins of 5% that represent the percent of grid points in the corresponding range; each cell on the plot has edge dimensions of 5%. Cells to the left of the dashed line indicate where surface sky observation values are greater than NDFD sky cover forecast values; cells to the right of the dashed line are lesser than NDFD sky cover forecast values.

Fig. 3

Comparison of the RTMA cloud cover to the NDFD 1-h sky cover forecast. There were 6696 grid cell comparisons of RTMA total cloud cover field with the NDFD 1-h sky cover forecast in the domain of interest (as shown in Fig. 1) for the 1 h following 1900 UTC on August 21, 2017. Comparisons are plotted in bins of 5% that represent the percent of grid points in the corresponding range; each cell on the plot has edge dimensions of 5%. Cells to the right of the dashed line indicate those sky coverage categories where the NDFD sky cover forecast exhibits a high bias compared with the RTMA.

This mean decrease in the RTMA is likely a result of the dense satellite observations, particularly because the RTMA mean was less than the surface sky observations mean. The location-specific 1-h differences between time-adjacent RTMA total cloud cover analyses are shown in Fig. 4. There was a decrease in cloud cover over northern Mississippi per the RTMA total cloud cover analysis from 17 to 18 UTC, as indicated in blue colors in Fig. 4(a), as the eclipse approached, and an increase in cloud cover from 19 to 20 UTC, as indicated in red colors in Fig. 4(c), when the eclipse effects waned. Cloudiness decreased substantially, in excess of 45%, over portions of northern Mississippi during the passage of the eclipse, and then rebounded later in the afternoon. The trend over time contrasts with the NDFD 1-h forecast grid adjacent hour differences, shown in Fig. 5. There are slight differences between NWS forecast areas of responsibilities but no evidence of occurring or anticipated eclipse effects in the NDFD forecast grid for the area of interest. This suggests that the NWS meteorologists monitoring the NDFD on the eclipse day either relied strongly on the operational NWP guidance that did not account for decreased incoming solar radiation during the eclipse, or they were uncertain about the magnitude and timing of the decrease in sky cover that the eclipse would produce, even as it was imminent and observable in visible satellite images (Fig. 1). The NWS does not provide information about the composition of specific NDFD grids and meteorologist modifications; those likely vary from office to office.

Fig. 4

RTMA total cloud cover analysis temporal differences. This three-panel plot shows the RTMA total cloud cover analysis temporal differences, in units of percent (%) coverage, on a 10-km Lambert Conformal grid between (a) 17 and 18 UTC, (b) 18 and 19 UTC, and (c) 19 and 20 UTC on August 21, 2017, in the respective panels. In each of the three panels, the earlier analysis valid time is subtracted from the later analysis valid time. Yellow areas show slight increases to the cloudiness during the respective time windows, whereas blue areas show decreases. The collection of analyses depicts decreasing cloudiness ahead of totality and increasing cloudiness following totality.

Fig. 5

NDFD sky cover forecast temporal differences. This three-panel plot shows the NDFD 1-h sky cover forecast grid temporal differences, in units of percent (%) coverage, on a 10-km Lambert Conformal grid between (a) 17 and 18 UTC, (b) 18 and 19 UTC, and (c) 19 and 20 UTC on August 21, 2017, in the respective panels. In each of the three panels, the earlier forecast valid time is subtracted from the later forecast valid time. Yellow areas show slight increases to the sky cover forecast during the respective time windows, whereas blue areas show decreases. A trend pattern between panels is not apparent. Each NDFD forecast was issued 1 h prior to the valid time. Differences aligned to political boundaries likely reveal different forecasts along adjacent office areas of responsibility.

2.3.

Future of Satellite Sky Cover

NWS meteorologists have access to weather satellite imagery and satellite-derived products. An infrared window band provides an alternative to a visible band for monitoring a solar eclipse, despite the nominal 2-km spatial resolution, because the radiating temperature of land surfaces decreases without incoming solar radiation and clouds are still evident regardless of the amount of solar illumination. The additional infrared window spectral bands and finer spatial resolution of the ABI on GOES-16 will improve the detection and characterization of clouds from the geostationary orbit.⁸ To examine the quantitative value of increased-resolution infrared window radiances, the adjusted average cloud top emissivity (AACTE) fields from GOES-13, computed with data from the legacy imager, and GOES-16, with data from the ABI,⁹ are compared, as shown in Fig. 6.

Fig. 6

Adjusted average cloud top emissivity. This four-panel plot provides a comparison between the adjusted average cloud top emissivity (AACTE), in units of percent (%), from two Geostationary Operational Environmental Satellites, (a) GOES-16 and (c) GOES-13, at 2000 UTC on August 21, 2017. A (b) visible image from GOES-13 at the same time assists in characterizing the scene, and the (d) AACTE difference of GOES-13 from GOES-16 reveals local contrasts between (a) GOES-16 and (c) GOES-13. Yellow (positive) areas generally show mostly clear areas where GOES-16 imagery input increased the areal emissivity calculation, whereas blue (negative) areas show cloudier areas where GOES-16 imagery input decreased the areal emissivity calculation. Some semiopaque cirrus clouds also have a greater AACTE with GOES-13 than GOES-16.

Infrared window radiances are used to calculate cloud top emissivity. Cloud top emissivity is also referred to as effective cloud amount because a cloud top emissivity of less than unity indicates that a cloud occupies a fraction of the pixel, the cloud is semiopaque, or a semiopaque cloud occupies a fraction of the pixel. The AACTE quantifies an otherwise subjective assessment of sky cover with a single time spatial average containing corrections in cases of multilayered cloud decks and some near-overcast homogeneous cloud decks.¹⁰ The spatial average contains all pixels within a 25-km radius around a point. Near the satellite subpoint, there are more adjacent pixels incorporated into the AACTE for a given point than at latitudes and longitudes further from the subpoint. Because the cloud top emissivity is itself a spatial average, though becoming less so for higher resolution imagers such as the ABI, the additional smoothing effect of the AACTE does not lead to significantly degraded calculations at large viewing angles where pixels cover a broader spatial area than at the satellite subpoint. Although it is not possible to run the RTMA for this day with GOES-16 input, the AACTE would have been the principal input to the RTMA total cloud cover analysis had it not been preoperational.

The grid mean AACTE values over the interest area for 2000 UTC on the eclipse day, $\sim 90\min$ after totality, are similar; the GOES-13 imager grid mean AACTE value is 16%, 3% greater than the mean from the GOES-16 ABI. Local differences are evident, most commonly and likely resulting from the nominal 2-km spatial resolution of the ABI infrared bands.⁶ For mostly cloudy scenes, subpixel cloud breaks are less evident with the lower, 4-km resolution of the legacy imager. Conversely, the ABI can capture small clouds in otherwise mostly clear scenes. The result of this, as evident in Fig. 6(d), is that the mostly clear areas are cloudier in terms of AACTE when viewed with the ABI, such as over southwestern Kentucky, and partly/mostly cloudy areas are clearer, such as over northern Mississippi. There are also differences where semiopaque cirrus clouds are present, such as over southeastern Kentucky, based on improved cloud top height assignments with the additional ABI infrared window spectral bands.⁹