Data from sprites using a 16-channel, multi-anode photometer (MAP) have been recorded at 0.1 ms resolution. The majority of the sprites have onsets at an altitude of 70–75 km and subsequently propagate both upward and downward from this initial altitude. The statistical lifetime of the emissions measured by the blue-sensitive MAP is approximately 1.3 ms. The velocities of propagation are between 107 and 108m s−1, with larger velocities being measured at higher altitudes. The larger propagation velocities of the upward and downward sprite luminosity channels are consistent with the characteristics of highly over-voltage streamers in the mesosphere.
If you can't find a tool you're looking for, please click the link at the top of the page to "Go to old article view". Alternatively, view our Knowledge Base articles for additional help. Your feedback is important to us, so please let us know if you have comments or ideas for improvement.
 Sprites are luminous optical phenomena which occur predominantly above mesoscale thunderstorms, (see Rodger  for review). The first low-light-level TV observations of sprites from the ground were reported by Franz et al.  and were quickly followed with observations made from space [Vaughan et al., 1992] and aircraft, [Sentman and Wescott, 1993]. Lyons  showed that sprites have diameters of 20 km or more and are usually associated with positive cloud to ground lightning. Sentman et al.  used color TV cameras to demonstrate that the lower altitudes of sprites are dominated by blue emissions, while the emissions from the tops of sprites are red. Observations of the spectra of sprites were reported by Mende et al.  and Hampton et al. . Both of these papers reported that the spectra were dominated by emission from the first positive bands (1P) of N2, with no evidence of the N2+ Meinel bands, leading both to conclude that the production mechanism for sprites produced little or no ionization.
 All of these observations were made at TV frame rates, giving a temporal resolution of, at best, 15 ms. Winckler et al.  used a single-channel photometer with a time resolution of 10 μsec and noted that the differences in signal duration of the photometer (6–7 ms) and the TV images (≈30 ms) must be due to the differences in spectral response for the two different systems. This result was confirmed and expanded by Armstrong et al. [1998, 2000] and Suszcynsky et al. , who used a narrow-band blue-filtered photometer boresighted with a wide-band blue-filtered camera to look at the same optical events. Suszcynsky et al.  found that the duration of the blue-sensitive photometer was very short, with FWHM of 0.1–1 ms. Analysis of their data led both Armstrong et al. [1998, 2000] and Suszcynsky et al.  to conclude that contrary to the slower time resolution work that “…significant ionization occurs during sprite generation.”
 In addition to these single-channel photometers the first use of multi-anode photometers (MAPs) to obtain limited spatial and good temporal resolution of sprites, [Takahashi et al., 2000], and related optical events called elves [Fukunishi et al., 1996; Inan et al., 1997], were reported. Complementary to these observations was the first high speed imagery (1000 frames per second) of sprites reported by Stanley et al.  who found that sprites initiate at altitudes of approximately 75 km and propagate both down and up with velocities in excess of 107 m s−1. Stenbaek-Nielsen et al.  also used a 1000 fps camera to reveal that sprites occur in a highly structured mesosphere, and may modify the mesosphere. This report compares simultaneous high-speed imager (HSI) data with a MAP boresighted to look at the same events.
 We used a Hamamatsu 16-channel multi-anode photometer, model number R5900U-00-L16, as the basis for the photometer system developed for this project by the Physics Department at the Air Force Academy. A single reflex camera lens was used to give an effective field of view of 6.1 × 6.1 degrees. The 16 channels of the MAP are much wider than they are tall (20:1 aspect ratio), and were oriented in a horizontal fashion for this experiment. The output current of each anode was routed into one channel of a 16-channel trans-impedance logarithmic amplifier. A voltage proportional to the log of the intensity was subsequently read out as 12-bit digital data and recorded at 10 kHz per channel.
 We boresighted this system with the HSI described by Stenbaek-Nielsen et al. . This intensified CCD camera produces 256 × 256 8-bit images at 1000 fps with a field of view of 6.4 × 6.4 degrees, essentially the same field of view as the MAP. We find that the two systems are synergistic and together provide good temporal and spatial resolution for the observation of sprites. Wider angle (both 15 and 30 degree) low light level scene cameras at 30 fps with frame by frame GPS encoded time were also used to record the general morphology of the events. The wavelength responses of the HSI and MAP are quite different. Figure 1 shows the manufacturer quoted quantum efficiency for both instruments, with the blue sensitive MAP peak response at 350nm and the 0.1 response points at 300 and 600 nm respectively.
 All data presented in this report were obtained from the trailing edge of a mesoscale storm over Nebraska on August 18, 1999. Figure 2 shows the storm as recorded by the GOES satellite at 05:32 UT. The storm was moving slowly towards the northeast. Our observations were made from the Wyoming Infrared Observatory (WIRO) operated by the University Of Wyoming on Jelm Mountain located south of Laramie, Wyoming. This site, at 9650 feet, latitude 41.10, longitude −105.98, provides excellent views to the east and south overlooking the mesoscale storms seen frequently over Kansas and Nebraska during the summer months. Lightning activity was recorded on the National Lightning Detection Network (NLDN) [Cummins et al., 1998], and the location of 28 positive cloud to ground strikes associated with sprites reported on in this paper are shown in the figure. All lightning locations referred to in this paper are from the NLDN data set.
Figure 3 (top) shows a comparison of 6 frames from the HSI at 1 ms intervals and, (bottom) 6 ms of data from the MAP centered on the same sprite. The sprite followed a large (90 kA) positive cloud-to-ground lightning strike recorded by NLDN at 05:10:38.668 UT, and located at latitude 40.55N, longitude 99.37W for a range of 570 km from our site at WIRO. The azimuth to the strike is shown in the HSI images with a vertical white line. The strike is the second of three positive cloud-to-ground lightning strikes recorded by NLDN causing four sprite groups within 1 s. The event in Figure 3 starts with a halo, a flat, horizontally extended region of luminosity, [see Barrington-Leigh et al., 2001], followed by the development of a large sprite (labeled “b” in the second frame). The sprite reaches maximum luminosity in frame 3 and then fades rapidly. Additional sprite activity is seen in the center region of the HSI images starting with the relatively faint structures (labeled “a” in the first frame) below the halo. Tendrils develop in the direction of the main sprite and the NLDN strike location to form two additional sprites within this group (labelled “c” in frame 4). The initial luminous structures, primarily beads, were shown by Stenbaek-Nielsen et al.  (their figure 3) to be associated with a sprite that had occurred 72 ms earlier. This led them to suggest that sprites may have longer lasting local effects. A similar example of a sprite developing at the location of an earlier sprite has also been reported by Armstrong et al. .
 The bottom portion of Figure 3 shows the MAP data. The 16 channels are displayed with a grayscale based on the log of the intensity. Since the MAP channels are aligned horizontally this type of display clearly shows the altitude-time development of sprite luminosity.
 The altitude scale, shown on the left, has been derived using the location of the associated lightning strike recorded by NLDN and the elevation angle to the sprite. The elevation angle was determined by fitting a star field to the stars in the HSI images. The MAP was boresighted with the HSI in the field. Lyons  and Wescott et al.  found the locations of sprites to be within, on average, 42 km and 25 km, respectively, of the causative lightning. At the elevation angles of our observations the 25 km range uncertainty translates to an uncertainty in altitude of approximately 5 km.
 The MAP and HSI computers only recorded system time, and we cross referenced all sprite events with the TV rate scene camera which has GPS time encoded on each frame. Thus the absolute time uncertainty for each sprite event is 1/60th of a second. Due to this limitation the time axis on the MAP data have been centered on the pixel with maximum signal amplitude. We normalized all pixels in the MAP data set to the maximum signal pixel within the 6 ms window. The MAP data are recorded sequentially at 10 kHz causing pixels in each channel of the MAP image to be offset from each other. The offset between adjacent channels is thus δt (δt = 0.1 ms is the time between sequential samples). This offset in time between channels is considered when calculating sprite velocities in the next section; however unshifted data are shown in all figures since the change is too small to resolve in the figure.
 In Figure 3 the halo in the first frame of the HSI is clearly identifiable at approximately 80 km (channels 6–7) and 1 ms in the MAP. There is a definite decrease in intensity in channels 6–8 at 1.5 ms, prior to development of the sprite at 2.5 ms. This is in contrast to the HSI data which show the halo continuing on into the third frame when the sprite is fully developed. The sprite initiation altitude at 70–75 km (channels 8–10 of the MAP) is slightly below the halo. The propagation of luminosity is visible in the time delay of the channels in both the upward and downward directions from this initiation altitude. This example shows the tendrils starting before the branches. The luminosity in the branch (going towards higher altitude) and tendril (going towards lower altitude) MAP channels are shorter in time than the luminosity in the initiation channels (9–10) as is more clearly demonstrated in Figure 4.
Figure 4 shows the MAP data for the same event as a stacked densitometer plot versus time. Again the data have been scaled to the maximum signal value in the 6 ms window. The black asterisks on each channel show the location of the maximum signal intensity for that channel, while the black “x” show the location of the maximum rate of change. In Figure 4 the darker gray regions show the time periods when each channel is above an arbitrary threshold of 0.67 of the maximum signal (corresponding to the top 50% of the maximum scene intensity due to the logarithmic response of the amplifier). This threshold is also used in the following section to develop a statistical description of the distribution of sprite occurrence in altitude and time.
Figure 4 clearly shows the halo, (channels 6 and 7 near t = 1 ms). The onset of the sprite is at lower altitudes (channels 8 through 10). The maxima in these channels are at 3 ms, but the maxima of the downward propagating tendrils in channels 13–16 occur between 2.5 and 3 ms.
 The time duration of intensity in channels 8–11 above the 0.67 threshold is longer than that of the tendrils (channels 13–16), and the branches (channels 3–6). This demonstrates that the propagating branch and tendril emissions are shorter in duration than the initiation channels of the sprite. The longer duration at initiation altitudes is complicated in this example by the presence of several sprites. In addition, changing the threshold value of 0.67 in this analysis will change the resulting durations in the differing channels. If a very low value for the threshold is selected then the duration will be closer for all the channels.
 By identifying a common feature we are able to calculate the propagation velocity of luminosity between two channels. We identified channels associated with tendril or branch emission by looking for channels where the feature did not have any discontinuity in time from one channel to the next. In Figure 4 we identify channels 13 through 16 as tendrils, and channels 7 through 3 as branches. Examination of the maximum signal shows a discontinuity between channels 12 and 13, which is probably due to the multiple sprites in this event.
 Tracking features across channels allows us to identify uncertainties in altitude and time, thus giving us the uncertainty in propagation velocity. We find the uncertainty in time to be 0.1 ms, our sampling rate. We chose the uncertainty in altitude as the width of one channel. While the range to each sprite event varies, we found that an altitude uncertainty of 5 km well represents this one channel width.
Figure 5 shows a sprite at 05:10:39 UT with different characteristics. NLDN located the lightning at latitude 40.59N, longitude 99.73W. Again the NLDN azimuth is aligned with the sprite. This sprite is the last of the four groups of sprites occurring within 1 s. The second group was shown in Figure 3. The upper panel of Figure 5 shows the low-light-level scene camera image of this sprite (the HSI data terminated shortly before this sprite). There is no observable halo in the MAP for this particular sprite. The bottom panel of Figure 5 shows the photometer data in a format similar to that in Figure 3. Notice that the duration of signal intensity at the initiation altitude (channel 8) is considerably longer in duration than the upward propagating branches and downward propagating tendrils. In this example the branch propagation begins before, and propagates noticeably faster, than the tendrils. Figure 5 illustrates very well the typical features of sprites we see in the MAP data. The initiation altitude is between 70–75 km followed by both upward and downward propagation of the luminosity and the optical emission lasts longer at the initiation altitude.
Figure 6 is a group of three sprites with NLDN time and position: 04:36:09.230 UT, latitude 41.28N, longitude 98.83W. Looking at the upper panel of HSI frames, the three distinct sprites are clearly visible. The apparent propagation velocity seen in the MAP data are similar for all three sprites. The downward propagating tendrils and upward propagating branches, seen around 2 and 3 ms, are shorter in duration at a constant signal level than the luminosity at initiation altitudes. Cases such as Figure 6 show the complicated nature of multiple sprites occurring within a few milliseconds. None of the sprites were located directly over the NLDN azimuth, but the tendrils of the leftmost sprite do curve towards this azimuth. The curvature of the tendrils towards the NLDN azimuth in this sprite was first pointed out by Stenbaek-Nielsen et al. . This curvature of the tendrils is not seen in all cases, but may be an important physical phenomena which has not been addressed or explained in the literature.
 The rapid 10 kHz sampling rate of the MAP allows us to establish high fidelity measurements of the altitude and time histograms of the blue emissions from sprites. The reader should note that the statistics developed in this section were obtained from the blue sensitive MAP, and may be significantly different from statistics developed for the lower energy red emissions. We extracted 28 events from the 18 August 1999 data set for which we have NLDN lightning strike location and good signal-to-noise in the MAP. We constructed a 2D matrix for each sprite with 16 altitude and 61 time pixels, i.e., 16 altitude rows, by 61 time columns. We centered the 61 columns on the time of the single pixel with the largest signal amplitude to give a 6 ms wide temporal window.
 We normalized the signal intensity to the largest signal in the matrix. We then removed pixels below 0.67 (corresponding to 50% signal intensity in the scene) of this maximum, leaving only pixels above this threshold (e.g., the gray areas in Figure 4). This results in our counting only pixels which are due unambiguously to sprites; we eliminated pixels in which the light is likely from such spurious sources as light reflecting from clouds. We have confirmed that changing the threshold value (which will include or exclude certain pixels within the 2D matrix) does not substantially change the occurrence distributions.
 To determine the distribution of sprite occurrence in altitude we counted the number of pixels within an altitude bin (rows), irrespective of where they occurred in time (column) during the sprite. We then summed the number of sprite pixels in common altitude bins for all 28 sprites, and divided this number by the total number of pixels above the threshold.
 The altitude width of a channel depends on the range to the sprite, and thus it is different for all events. To obtain combined statistics for all 28 events we then apply a spline interpolation on the individual distributions to make a common set of altitude bands. The bin width of this common altitude band is 4 km, which is of same order as the uncertainty in altitude (5 km). We note that the main effect of the altitude uncertainty is to slightly increase the width of the altitude distributions derived from our data. The distribution is intentionally normalized with the sum over all the altitude bins equal to unity. The resulting distribution is shown in Figure 7.
 The reader should note since we have normalized each sprite matrix to the maximum signal for that sprite we are not measuring the average brightness of the ensemble. We chose this analysis method since the logarithmic amplifier changes slowly with temperature during the course of data acquisition. The analysis effectively gives equal weight to each sprite, regardless of its intrinsic brightness.
 This distribution effectively is a measure of where in altitude the sprites are brightest and most persistent. Examining Figure 4 we see that channel 8 has pixels above the threshold (colored gray) for a longer period of time than the lower altitude channels 13–16, which only exceed the threshold when the tendril propagates through their field of view.
 The maximum persistence is observed at 71 km. This result is similar to the findings of Stanley et al. , Figure 2b, who showed an individual example of a carrot sprite with the initiation altitude between 70 and 75 km. The altitude distribution has an asymmetric shape. Luminosity is seen down to 50 km, and up to 90 km at the 4% occurrence rate. We do see instances of sprite emission at altitudes lower than 50 km. All such cases correspond to downward propagating tendril emission. This supports Fernsler and Rowland , who calculate that quasi-static enhanced electric fields due to the long thin downward propagating tendrils can still exceed the breakdown voltage below 50 km.
 To better understand the temporal evolution of the 28 sprite events we can likewise determine the distribution of sprite occurrences in time. To do this we count the number of pixels within each time bin (column), irrespective of altitude (row), for each sprite event. We then sum the number of sprite pixels in common time bins for all 28 sprites, and divide this number by the total number of pixels above the threshold. This results in a normalized distribution shown in Figure 8.
Figure 8 shows the blue sensitive MAP distribution of sprite occurrences in time. The peak of the distribution is not at 0 ms, but somewhat later because of the increasing size of the sprites following the initial bright onset used for time alignment. The FWHM of this distribution is 1.3 ms.
 The MAP data can give propagation velocities of the branches and tendrils as previously discussed in relation to Figure 4. Due to the measurement method, the apparent propagation velocity is the vertical component only. Since most of the sprites observed have both upward propagating branches and downward propagating tendrils, and some events have more than one sprite (for example events in Figures 3 and 6), we often get multiple upward and downward velocity measurements from the same event.
Figure 9 shows all the sprite luminosity propagation velocities versus altitude. The downward propagating tendrils are shown with an asterisk, while data from the upward propagating branches are displayed with a period. We find typical values of 107–108m s−1. Similar velocities have been reported by Stanley et al. . The vertical error bars in velocity were derived by propagating the error with a time uncertainty of 0.1 ms, and an altitude uncertainty of 5 km. The horizontal bars, shown on five random data points, are representative of the altitudes over which the sprites were tracked when determining the velocity. The designation of upward propagation as a negative streamer and downward propagation as a positive streamer is explained in the discussion section to follow.
 Several features common to sprite propagation velocities are clearly shown in Figure 9. First, the tendency for downward and upward propagation away from a central initiation altitude is seen with 15 of 20 downward propagating tendrils having center altitudes below 75 km, and all but one of the upward propagating branches having center altitudes above 75 km. Second there is a general tendency for larger velocities at higher altitudes. Finally, note the larger scatter in upward propagation velocities compared to the downward propagation velocities. This may be due to the shorter distance the luminosity is tracked for upward compared to downward propagation.
 The upward propagating branches have mean and median velocities of 5.3 × 107 m s−1 and 4.6 × 107 m s−1, respectively. The downward propagating tendrils have slightly lower mean and median velocities of 3.3 × 107 m s−1 and 2.8 × 107 m s−1. Statistically, however the probability that the data sets are consistent with a single distribution is 99% using the Kolmogrov-Smirnov test [Press et al., 1989, p. 472].
 When looking at the high speed imagery the early time development of sprites is very reminiscent of streamer discharges in a low pressure gas. Indeed, recent observations and theoretical papers on sprites, [Barrington-Leigh et al., 2001; Pasko et al., 2000; Pasko et al., 1998; Raizer et al., 1998] all support in different ways the idea that sprites can be described by the streamer electrical breakdown process. Streamers are sometimes called ionization waves. The propagation velocity is that at which the potential of the streamer tip is pushed ahead of the streamer by the large space charge inside the streamer tip [Raizer et al., 1998; Raizer, 1997, p. 344; Lagarkov and Rutkevich, 1994, p. 1]. The initial ionization can be due to either previous ionization, or an electron avalanche transitioning into a streamer at the beginning of the ionization wave and then propagating in both directions, [Lagarkov and Rutkevich, 1994 page 5]. With the electric field pointed down from the ionosphere due to the positive cloud to ground lighting strike, the upward propagating branches are negative streamers, and the downward propagating tendrils are positive streamers [Raizer et al., 1998; Raizer, 1997, p. 334].
 The measurements shown in Figure 9 are consistent with laboratory measurements of ionization streamer discharges in low pressure gas [Suzuki, 1977]. To compare our velocity measurement with laboratory measurements we ran the standard Mass Spectrometer and Incoherent Scatter (MSIS) model, [Hedin, 1991], for atmospheric densities and temperatures at 49N, 99W and 6 UT on 18 August 1999. Treating the mesosphere as an ideal gas, this results in a total pressure at 60 km of 23 Pa (0.17 torr), and gives a breakdown voltage for dry air of approximately 80 V/m.
 Propagation velocities for ionizing waves measured in the laboratory for highly over-voltaged streamers at a few torr are a few times 107m s−1, (see Figure 9 of Suzuki ). Our measurements are in agreement and in addition, indicate that the velocity of propagation increases at higher altitudes (lower pressures). This inverse relationship is also in good agreement with Figure 9 of Suzuki .
Pasko et al.  show that the characteristic streamer radius at 75 km altitude would be on the order of a few meters, with smaller streamer radii at lower altitudes, and larger streamer radii at higher altitudes, and develops a scaling relationship showing that the streamer velocity vs varies as the square of the streamer radius, rs2. Using this relation, Pasko et al.  predicts streamer velocities of 105 < vs < 106 m s−1 which are between one to two orders of magnitude lower than our measurements. However Gerken et al.  observed streamer scale sizes of 20–50 m at altitudes between 76–80 km, but found larger streamer scale sizes of 50–150 m at altitudes both above and below the 76–80 km range. The scaling relations of Pasko et al.  indicate these larger streamer scale sizes at altitudes above and below the initiation altitude would bring the predicted streamer velocity estimates into the same range as our velocity measurements. This leads us to suggest that our upward and downward propagating sprite velocities are consistent with highly over-voltage streamers in the mesosphere.
 The propagation velocities measured by Suzuki  are dependent upon the initial ionization level, with increasing amounts of initial ionization leading to increased velocities. Such increased ionization in the mesosphere might be left over from previous sprites in the same general location, as suggested by Stenbaek-Nielsen et al. . Another suggestion for a source of initial ionization is from Wescott et al.  who speculate that a micrometeor with sufficient mass leaves an ionized trail which could be the source for the streamer formation. Wescott et al.  calculate that some 19 micrometeors per second within a 25 km radius with sufficient mass to cause ionization impinge on the upper atmosphere. Since Figure 5 shows three sprites occurring all within a few ms this would indicate the subsequent sprite ionization probably comes from a source other than micrometeors.
 Our velocity measurements are in good agreement with Raizer et al.  who predict the maximum velocity of a sprite streamer is ≈2 × 107m s−1. These high velocities imply a large potential difference between the streamer tip and the potential external to the streamer due to the large polarization space charge within the ionization wave. In contrast to Raizer et al. , Lagarkov and Rutkevich [1994, p. 1] note that the streamer formation in the laboratory often begins in the middle of the gap and propagates towards both the anode (the ionosphere), and the cathode (the cloud top). Vitello et al.  also indicate that double headed filaments with both positive and negative streamers form from small enhancements in the electron density if they are far enough from the anode. Generally we find sprite initiation forms around the maximum persistence altitude of 71 km (Figure 7) and propagates both upwards and downwards, in agreement with Stanley et al. .
Pasko et al.  points out that the critical field for positive streamer propagation is a factor of 2–3 less than that for negative streamers. This may play some role in our observations that downward propagation (positive streamer), in most cases, starts before upward propagation (negative streamer). We have only a few instances where we see only downward or upward propagation.
 Since our measurements were broad band there are significant contributions from both neutral N2 (2P) and ionized N2+ (1N) bands within the MAP spectral response. Armstrong et al.  used a set of three filtered photometers and determined the relative contributions of both the N2 (2P) and N2+ (1N) bands to the blue emissions from sprites. They found that not all sprites exhibited measurable emission from ionized species. Further, even when ionization was present, the blue emissions were dominated by the neutral N2 (2P) emissions. They found the duration of these neutral emissions was approximately 1 ms in one specific case (see Figure 1 of Armstrong et al. ). This is in good agreement with the FWHM of 0.1–1 ms by Suszcynsky et al.  for a photometer with a narrow band filter centered in the blue at 427.8 nm. Finally, Takahashi et al.  used a set of MAPs similar to ours and show an example where the broad band emissions are approximately 2 ms duration. All of these examples are in good agreement with our statistical data shown in Figure 8.
 Figure 17 of Pasko et al.  predicts that N2 (2P) dominates the ionized N2+ (1N) in instantaneous emissions from sprites. Further, in a study of the spatial structure of sprites, Pasko et al.  calculate the intensity of emissions from streamers. They find that the neutral N2 (2P) again dominates except in the very narrow region of the streamer tip. At the streamer tip the emissions from the ionized first negative dominates. The MAP signals are thus consistent with the measured emissions being primarily from neutral nitrogen.
 As demonstrated in Figure 4, the MAP has a decrease in signal intensity between the halo and the sprite. We can find 4 of 28 instances where the MAP shows this decrease in signal between the halo and sprite. In these cases the HSI measures the halo persisting through the first few ms of the sprite. This is presumably due to the lack of temporal resolution in the HSI combined with the longer persistence of the red N2 emissions measured by the HSI. This is consistent with Winckler et al. , who report that the red emissions from sprites last longer than the blue emissions, a result based on energetics. Barrington-Leigh et al.  also report on variable delay times between the sprite halo and streamer initiation. Modelled temporal and spatial development of sprite halos by Barrington-Leigh et al.  predict that halos will extend over one or two pixels of a MAP, and several tenths of ms; which is in good agreement with our measurements. Further observations, preferably from two separate locations, are needed to develop the statistics associated between sprite halos and streamer emission.
 Several important questions for future research remain concerning the implications of our measurements. If streamer propagation governs sprite propagation, as seems plausible from our velocity measurements, then sprites ionize and deposit charge and energy into the mesosphere. The critical question is how much ionization, and how much energy are deposited? Given the large volumes associated with the streamer regions of sprites, the charge and energy deposited above a mesoscale thunderstorm complex may be large enough to change the local mesospheric parameters.
 Our results show that sprite initiation occurs between 70 and 75 km, followed by both upward and downward streamer propagation. The downward propagation generally occurs slightly before the upward propagation. Propagation velocities are between 107 and 108m s−1, and increase with altitude. We do not measure any statistical difference between the negative (upward) and positive (downward) streamer propagation velocities.
 Sprite signal intensity above a threshold of 50% of maximum scene intensity persists longer at the initiation altitude compared to the propagating streamers. The measured propagation velocities of 107 and 108m s−1 are consistent with laboratory measurements and theoretical predictions of streamers at similar pressures. The large propagations velocities lead us to believe that both upward and downward propagating sprite luminosity is consistent with the characteristics of highly over-voltage streamers in the mesosphere.
 MGM would like to thank Martin Johnson and Duane Dunlap for helping develop the high speed photometer. This work was supported by the United States Air Force Office of Scientific Research and the Dean of the United States Air Force Academy, as well as NASA grant NAS5-5125 to the Geophysical Institute, University of Alaska Fairbanks. We are grateful to the University of Wisconsin-Madison Space Science and Engineering Center (SSEC) for the use of GOES imagery in Figure 2. The authors would like to thank both reviewers for useful comments and suggestions.
 Janet G. Luhmann thanks the referees for their assistance in evaluating this paper.