3.1 VHF lightning interferometer vs. the National Lightning Detection Network

3.1.1 Introductory material

Before commencing with the intercomparison of the ITF and the NLDN data, it is important to note that these platforms measure different aspects of lightning. To summarize, the ITF maps VHF emissions from lightning, both intra-cloud (IC) and cloud-to-ground (CG). However, as was noted in the last chapter, only certain parts of a lightning discharge are expected to radiate strongly in the VHF band. For example, the negative stepped leader process of a CG discharge is associated with VHF emissions, but return strokes and positive discharges emit only weakly at VHF frequencies. Thus, the ITF is expected to map the leader stages of a CG well, but will have a difficult time mapping return strokes and positive CGs. The ITF is additionally hampered by its reduced vertical resolution at low elevation angles, as discussed in Chapter 2. Furthermore, the added vertical resolution degradation that occurred during STERAO-A (also discussed in Chapter 2) compounds the task of properly detecting, classifying, and mapping CG lightning with the ITF.

In contrast, the NLDN detects only the return stroke stages of a CG. It is not intended to detect or locate IC lightning, though it is possible that some ICs are falsely identified as CGs by the NLDN, as discussed in Chapter 2. In contrast to the ITF, which locates the positions of the negative leaders (i.e., the flash channels), the NLDN only locates the ground strike positions of CG lightning, not the positions of the flash channels themselves.

For the purposes of this intercomparison then, CG data from the two sensors will be compared. Note that flashes are not classified as CGs by the ITF in real time. Instead, they are classified during post-processing with the use of an altitude threshold on negative leader processes, as mentioned in Chapter 2. This classification algorithm is still in testing at ONERA. Later versions of ITF data from 10 and 12 July 1996 (received from ONERA) need not have flashes classified the same. The NLDN data are not expected to change in the future, however.

3.1.2 Intercomparison of ITF and NLDN data from 10 July 1996

The storm of 10 July 1996 began as a multicellular line, but late in its lifetime one cell dominated and became supercellular in character, with mid-level rotation and a Weak Echo Region (WER). For much of its lifetime it stayed within the highest resolution eastern lobe of the ITF, making it an ideal case for comparing CG data from the ITF and NLDN. The storm was very isolated, with few unrelated cells nearby, so it was easy to remove CG data from these unrelated cells for the intercomparison. Five-minute CG rates were calculated for the remaining data from both sensors and the results are plotted in Figure 3.1. CG rates for each 5-minute period are plotted at the start of their respective time periods. Both sensors appear to capture the CG burst between 1600 and 1645 MDT, and the subsequent dearth of CGs in the approximately 1-hour period following this burst. Then after 1745, CG rates from both sensors begin to diverge, with significant disagreement after 1830 through the end of the storm. After 1830, ITF CG rates became quite large, but NLDN CGs were almost nonexistent. Also, the NLDN CG burst around 1810 MDT was not captured by the ITF. Overall, the ITF detects and classifies significantly more CGs than the NLDN.

This disagreement between the CG data from the two sensors becomes even greater when an attempt is made to match individual CGs from the two sensors, in order to gain a sense of which “CGs” are being detected by only one sensor, and which are being detected by both. When matching individual CGs between the two sensors, two problems arose. One was that, at the time the intercomparison was done, only 1-second resolution data were available for the two sensors, so it was difficult to estimate the time difference between the two sensors’ clocks. The second problem was that the ITF data contained the mean horizontal position of the radiating flash channel, whereas the NLDN data of course gave the ground strike location of the flash. CG lightning is known to travel horizontally, sometimes on the order of several km, before striking the ground. Thus the mean horizontal flash channel position and the ground strike location need not be identical, especially when the inherent location errors of the two sensors are considered. Thus, the criteria used to judge whether a given NLDN CG and a given ITF CG were the same were that they occurred at the same time, to the nearest second (due to the time resolution of the data, and keeping in mind the fact that the two clocks were not synchronized during STERAO-A), and that their horizontal positions be within 10 km of one another. In practice it was found that the time criterion was the strongest; that is, when an NLDN CG and an ITF CG were coincident, they were almost always very close to one another, from a distance perspective. Thus, the distance criterion for flash matching was not very important; increasing the distance to greater than 10 km would not change the results significantly.

Using this flash matching algorithm, the following were identified: CG flashes which were detected by both sensors, CG flashes which were detected and classified by the NLDN only, and CG flashes which were detected and classified by the ITF only. The results are plotted in the form of 5-minute flash rates in Figure 3.2. As can be seen, during the early period of the storm's lifetime, when Figure 3.1 appeared to show rough agreement between the two sensors, Figure 3.2 demonstrates this agreement in bulk flash rates to be misleading. During this time there is a significant number of CGs that are being detected and classified by only one sensor; the other sensor is either not detecting anything, or it is not classifying any detected flash as a CG. This problem only becomes worse after 1745 MDT. For the entire storm, the number of CG flashes detected and classified by both sensors is quite low, especially compared to the number of CGs detected and classified by only one sensor. After 1830, the number of CGs detected and classified by the ITF only is especially high, as might be expected based on Figure 3.1.

There is a small population of flashes detected and classified as CGs by the NLDN, but not by the ITF. An attempt was made to match these NLDN CGs to ITF ICs using the same algorithm for matching NLDN and ITF CGs. Interestingly, all of these NLDN CGs, save one, corresponded to a detected and classified ITF IC. If the CG intercomparison is specialized to only NLDN positive CGs, then 1 of 13 NLDN positive CGs corresponded to an ITF CG. Most of the remainder corresponded to ITF ICs.

3.1.3 Intercomparison of ITF and NLDN data from 12 July 1996

The same CG data intercomparison performed on 10 July data also was performed on the 12 July data. The storm of 12 July 1996 began as a cluster of several cells over the Cheyenne Ridge in southeastern Wyoming. As the storm moved to the southeast and entered Colorado, the storm remained multicellular. Substantial reflectivity (> 40 dBZ) continued to cover a large area throughout the remainder of the time of radar sector coverage. With such large areas of contiguous echo, the storm was not very isolated, so the ITF data were not edited to exclude unrelated cells. The NLDN data were edited to exclude flashes outside the ITF lobes, however. This intercomparison was done using older ITF data, like the 10 July intercomparison. These data covered the time period from 1730 to 1910 MDT. After this time span, the storm moved out of the lobes, so later data were impacted by these lobe effects. Figure 3.3 is a plot of 5-minute CG flash rates for both sensors throughout the intercomparison time period. Much like 10 July, overall the ITF detects and classifies many more CGs than the NLDN, but unlike 10 July this occurs throughout the analysis period, not just during a portion of it.

Individual CGs were matched using the method employed with the 10 July data. The result is Figure 3.4, which is the same as Figure 3.2 except now for 12 July. In general, very few CGs are detected and classified by both sensors. CGs detected and classified by only the ITF dominate the plot.

Of the 21 CGs that were detected and classified by only the NLDN for this time period, 12 matched up with detected and classified ITF ICs. Sixteen of these 21 NLDN-only CGs were positive CGs.

3.2 VHF lightning interferometer vs. flat plate antennas

For 12 July 1996, concurrent data from the flat plate fixed at the CHILL radar and from the ITF exist from 1747 to 1910 MDT. One-minute total flash rates were calculated during post-processing of the FCM data for this time period. By combining the ITF CG and IC data, 1-minute ITF total flash rates can be computed. However, in order to properly compare these two estimates of storm total flash rate, the issue of the flat plate’s limited range must be addressed. As stated in Chapter 2, the CHILL flat plate had a range of 35-40 km at the most sensitive gain setting, which is what it was set at during this time period. However, recall that the flat plate’s detection efficiency is a function of both range to flash and flash strength. Thus, the ITF data were filtered to exclude, from the flash rate calculations, flashes beyond a certain distance from CHILL. Because of the uncertainties about flat plate range and detection efficiency, this filtering distance was varied and sensitivity tests were performed.

Figure 3.5 shows 1-minute total flash rates from both the ITF and the CHILL FCM. The filtering distance in this case was 35 km, the lower limit of the flat plate maximum range estimate. Note that any trends seen in these data do not necessarily reflect true trends in storm flash rate. The large increase in flash rates from both sensors toward the end of the period is mostly due to the storm moving into range for the FCM, and moving into the 35 km filtering circle for the ITF. In general, the ITF detects many more flashes than the CHILL FCM, though the two flash rates seem to trend one another fairly well, especially toward the end of the intercomparison period.

Table 3.1 lists the square of the correlation coefficient, RHO², for the two flash rates throughout this intercomparison period, for different filtering distances ranging from 20 km to 40 km. The correlation between the two flash rates increases as the filtering distance becomes more restrictive, with RHO² reaching a maximum of 0.935 when ITF flashes with mean positions more than 20 km away from CHILL are excluded. A minimum of 0.774 occurs when ITF flashes with mean positions more than 40 km away from CHILL are excluded. This is strong evidence that the detection efficiency of the FCM is range dependent, if we assume that the ITF has a detection efficiency near 100% (or at least is mostly independent of range from CHILL). Correlations improve as the ITF range filter becomes more restrictive because distant flashes are detected by the FCM less often than closer flashes.

There is the possibility of some minor precipitation contamination of the CHILL FCM during the time period 1845-1900 MDT as the main storm grazed the radar site. Thus, FCM flash rates during this time may be artificially elevated, with an unknown but probably minor effect on the correlation analyses.

Despite its limitations, and assuming that the ITF has a very high detection efficiency, it appears that the FCM is a good tool for estimating total flash rate trends, based on this intercomparison. However, it is preferable that the storms sampled by the flat plate move slowly or be nearly stationary, in order to control false trends created by storm movement into and out of range.

3.3 Short-duration flashes in the VHF lightning interferometer data

Figure 3.6 shows normalized frequency distributions of ITF-derived flash durations for the storms of 10 and 12 July. These are the durations of the VHF emissions from each flash, and thus may not necessarily reflect the true flash durations (i.e., the time for which the lightning channel is ionized and current is propagating). Flash data have not been edited in any way. The plot limits extend from the lowest possible time resolution of the ITF, 23 microseconds, to the maximum allowed duration of a flash, 1 second (set by the flash-classification criteria). The distributions for both days are very similar. Why this is so is not certain. Both days’ data contain several thousand flashes each, so whatever is forcing the similarity of the two distributions is likely common to both days. The distribution shapes may be caused by the ITF or its classification algorithm, making the distributions entirely or partly instrument-related artifacts. Or, the spectra could be due to the physics of the lightning, which also would be common to both days, as long as the processes governing the electrification and electrical discharging of these two storms were the same. It is not known which of these two possibilities, if any, are true. However, ONERA researchers have no idea how the ITF or its flash classification algorithm could force the distributions presented here (P. Laroche, private communication, 1997).

The distributions themselves appear to be bimodal. The first mode which will be considered is centered around 10^-0.6 s, or 250 ms. This mode is distributed roughly lognormally, and appears consistent with other observed lightning duration spectra (Uman, 1987).

The second mode is a bit more complex, and is less easily understood. The maximum occurs at 10^-4.6 s, or 23 microseconds, the temporal resolution limit of the ITF. In fact, over 20% of flashes on either day last 23 microseconds or less, more than any other bin in the entire distribution. This corresponds to a single localization of VHF radiation as seen by the ITF. There are also a few gaps in the distributions for this particular mode, where no flashes with the given durations occur on either day. This is a curious phenomenon as the surrounding bins contain relatively large numbers of flashes.

The tails of the two modes appear to meet somewhere in the vicinity of 10^-3.0 s, or 1 ms. This duration, 1 ms, was chosen as the boundary between the two modes. While this decision is somewhat arbitrary, it is apparent from the spectra that, because so few flashes occupy the region from 10^-3.0 s to 10^-2.0 s, or 1 ms to 10 ms, choosing a different boundary between these two durations will not change the results significantly. Thus, the ITF flash duration frequency spectra were considered to be comprised of two different populations of flashes: a more standard flash population with approximately lognormally distributed durations, centered near 250 ms in this case; and a population of “sub-millisecond” flashes with durations on the order of individual strokes in a flash (Uman, 1987). The former population is probably what most researchers consider lightning flashes (Uman, 1987); the nature of the latter population is in question, but it is important to understand this nature as these sub-millisecond flashes make up close to one-third of the flash populations for either day. Maier et al. (1996) have reported observing similar short-duration “flashes” with the Lightning Detection and Ranging (LDAR) system, a time-of-arrival (TOA) VHF mapping system at the Kennedy Space Center in Florida.

There are a few possible explanations for these sub-millisecond (hereafter referred to as sub-ms) flashes. One is that they actually are components (i.e., individual strokes) of longer flashes – perhaps flashes that also are detected and classified by the ITF, and are part of the more standard flash population – but are being falsely classified as separate flashes by the ITF classification algorithm. The flash classification algorithm is very likely not perfect, and so some errors in flash classification are expected. Some of these sub-ms flashes do appear to be quite close, in space and time, to other, longer flashes. However, many sub-ms flashes are isolated in both time and space.

Another explanation is that, while the VHF emissions from these sub-ms flashes may last less than 1 ms, the actual flashes themselves last much longer. That is, perhaps many of these sub-ms flashes actually are longer flashes which do not radiate strongly in the VHF, like positive flashes and return strokes, so only a few VHF radiation localizations can be attributed to them, giving the false appearance of a short duration for the entire flash.

A final possibility is that these sub-ms flashes may be in fact true short-duration flashes lasting on the order of tens of microseconds, perhaps a subset of IC discharges. This possibility has been examined before, by Taylor et al. (1984), who used a VHF mapping system in a single storm to distinguish between long-duration (i.e., standard) IC flashes, which occurred near the middle of the thunderstorm, and short-duration IC flashes which were centered near the top of the storm, in the vicinity of the positive charge region (Williams, 1989). Unlike the sporadic long-duration IC events, the short-duration ICs occurred on a near-continuous basis. In a private communication, P. Krehbiel (1997) mentioned that these sub-ms flashes instead may be what he called “positive bi-polar discharges”. These discharges last on the order of 10-20 ms, and can be considered a kind of failed upward-propagating leader from the region of negative charge near cloud mid- levels (Williams, 1989). The “failed leader” does not reach the upper positive charge region, and thus no return strokes or recoil streamers occur.

In order to understand how to best interpret these sub-ms ITF flashes, a more detailed intercomparison of the ITF and FCM data was performed. Figure 3.7 is the same as Figure 3.5, except that ITF flashes with durations less than 1 ms also have been excluded from the analysis. Not surprisingly, ITF flash rates have been reduced, but the ITF still counts many more flashes than the CHILL FCM. Table 3.2 lists the results of correlation analyses of flash rates from the two sensors. The format is the same as Table 3.1, except now ITF flashes lasting less than 1 ms have been excluded. In all cases, correlations have improved over those presented in Table 3.1, which includes the sub-ms flashes.

As mentioned in Chapter 2, the CHILL FCM data are sampled at 1 kHz. Thus, the FCM cannot possibly resolve events associated with electrostatic field changes lasting less than 1 ms. Indeed, undersampling of events lasting longer than 1 ms, say up to 10 ms, will also be a problem with this sensor. Because the correlations improve when the sub-ms ITF flashes are excluded from the analyses, the sub-ms flashes likely are not being detected with good efficiency by the FCM. The FCM cannot detect anything shorter than 1 ms, so this implies that the sub-ms flashes also tend to be associated with electrostatic field changes lasting less than 1 ms. Another possibility is that these events may last lionger than 1 ms but are not associated with substantial electrostatic field changes, so that they do not register well on the flat plate.

If the sub-ms ITF flashes were actually parts of longer flashes that do not radiate strongly in the VHF, then these flashes probably should register electrostatic field changes lasting significantly longer than 1 ms, on average. Recall from previous discussions that such flashes could consist of positive discharges and return strokes. These events commonly are registered on this type of FCM (Carey and Rutledge, 1997), so a significant percentage of them should be detectable in the FCM data. A fellow member of the radar meteorology group at Colorado State University, J. Ryan, looked for evidence of sub-ms ITF flashes that were isolated in time and space in the CHILL FCM data from 12 July 1996. He considered 80 such ITF flashes, lasting between 23 and 500 microseconds, and having mean horizontal positions within 35 km of the CHILL FCM. He found no signals corresponding to these flashes, which is what would be expected if such flashes were associated with electrostatic field changes lasting less than 1 ms (J. Ryan, private communication, 1997). This, along with the observed trend of improved correlations when the sub-ms ITF flashes are excluded, implies that these sub-ms ITF flashes are not parts of longer, non-VHF emitting flashes, so this possibility is ruled out on a tentative basis.

If the sub-ms ITF flashes are actually parts of longer, detected discharges, and thus are being falsely classified as separate discharges by the ITF, then removing them from the correlation analyses should improve the correlations, as the ITF would be counting individual flashes twice or more, whereas the FCM, if it detects the flashes, would count them only once. Thus, this possibility cannot be ruled out because it is consistent with the observed improvement in flash rate correlations when the sub-ms ITF flashes are excluded.

If the sub-ms ITF flashes are in fact true short-duration discharges with electrostatic field changes lasting less than 1 ms, then the FCM could not detect them. Thus, when they are removed from the ITF/FCM intercomparison, correlations should improve, as is observed. Thus, this hypothesis is consistent with the data analysis results as well.

So there are two possible hypotheses consistent with the observed increase in correlations when the sub-ms ITF flashes are excluded from analysis. Note, however, that the hypothesis that these sub-ms flashes are parts of longer, detected and classified ITF flashes does not explain the large number of observed sub-ms flashes which are isolated in both time and space. These isolated flashes probably can be interpreted only as true individual short-duration IC flashes. Non-isolated flashes could be explained by either hypothesis, or perhaps a combination of both.

3.4 How the ITF data will be interpreted in this study

There are two issues to consider when interpreting the ITF data in terms of flash rates. The first is the IC/CG classification problem discussed in Section 3.1. The second is the sub-ms ITF flashes which were examined in Section 3.3.

The NLDN is reported to have close to a 90% detection efficiency in northeastern Colorado (Cummins et al., 1996). Additionally, no positive CGs with peak currents less than 7 kA were observed, so all NLDN CGs from both days are likely true CGs and not mis-classified ICs. In addition, the NLDN is generally accepted by most researchers to be the current “industry standard”, having been used in a large number of published research studies which are too numerous to mention here.

In contrast, the ITF to a certain extent is still under development, particularly its flash classification algorithm. The algorithm used to classify a given flash as a CG is essentially an altitude threshold on a leader process. Recall that the CG leader stroke resembles a spider lightning discharge to a VHF interferometer like the ITF. Spider discharges, being IC flashes, typically have higher mean altitudes than CG leaders. Thus, an altitude threshold would work to distinguish the two provided the ITF does a good job in resolving altitude differences. However, as discussed in Chapter 2, the ITF had unusually poor elevational resolution during STERAO-A. Compounding this problem is the generally poor low-elevation angle resolution of the ITF, regardless of any additional problems. Thus, it is best to suspect the ITF of error in classifying CGs rather than the NLDN.

The sub-ms ITF flashes, as stated before, are probably either components of longer, detected and classified ITF flashes, or they are individual short-duration flashes in their own right, or perhaps some combination of the two. In any case, eliminating them from consideration reduces the possibility of counting the same flash twice (or more), and reduces the possibility of counting short-duration flashes as part of the standard (i.e., long-duration) IC flash rate. Given the large uncertainty regarding the nature of these short-duration discharges, it is probably best to disregard them for the purposes of this study since it is geared toward using the ITF data to compute IC flash rates, not looking at the physics of the lightning discharge itself, or exploring the natures of the different types of lightning discharges. Also, removing these discharges from consideration leaves a flash duration spectrum more in accord with past studies involving IC lightning (e.g., Uman, 1987), so flash rates are more comparable to previous studies.

These short-duration flashes, if they exist in reality, are certainly worth studying in their own right, though they will not be the focus of this thesis. However, in order to better examine the implications of removing them from consideration, Figure 3.8 was developed using the latest version of the ITF data, the version used in the radar/lightning case studies (though not the same version used in the lightning sensor intercomparisons in this chapter). It shows IC and sub-ms 5-minute flash rates for 10 July 1996. No editing of the flashes has occurred (except for removing from consideration those flashes classified by the ITF as CGs); they have merely been categorized by flash type. The sub-ms category contains all flashes lasting less than 1 ms, and the IC flash category contains all IC flashes lasting 1 ms or more. Note that all sub-ms flashes were classified as ICs by the ITF. In general, the sub-ms flash rate trends the IC flash rate fairly well. The square of the correlation coefficient for these flash rates is 0.610. This is consistent with the sub-ms flashes being, at least in part, components of longer detected flashes, as they would be very related then. However, if the sub-ms flashes are, at least in part, individual short-duration IC discharges, this result is still consistent provided the short- and long-duration flashes are related somehow. Thus, the truth of either possible explanation for the sub-ms ITF flashes cannot be discerned completely. However, it is apparent that the relatively high correlation between the sub-ms and regular IC flash rates implies that removing the sub-ms flashes from consideration will not change IC flash rate trends significantly, though absolute flash rates will change. As this study is most interested in flash rate trends, removing the short-duration discharges from consideration will not make a major impact on results.

Thus, based on this chapter’s analyses and discussion, the following method will be used in this study to determine IC and CG flash rates. Flashes lasting less than 1 ms will be removed from the processed ITF flash data. Then, NLDN CGs will be individually matched to remaining ITF flashes, regardless of whether the ITF classified them as CGs or ICs, following the method outlined in Section 3.1. Those ITF flashes matching NLDN CGs in time will be removed from the ITF data as well. The remaining ITF flashes will be considered ICs, and the NLDN CGs will form the CG data set.

This algorithm thus removes the sub-ms flashes from any further consideration. Though this eliminates the ability to consider any possible individual short-duration ICs, it also avoids the problem of possibly overcounting ICs because some of the sub-ms ITF flashes may be components of longer, detected flashes. This algorithm also avoids relying on the suspect CG flash classification method of the ITF. Another problem that the algorithm avoids is cases where CGs are detected and classified by the NLDN, but are not detected at all by the ITF. The algorithm ensures that these NLDN-only CGs will remain in the CG data set.

One possible source of error in this algorithm is the fact that the NLDN’s detection efficiency is not 100% in northeastern Colorado. Thus, the NLDN may fail to detect a CG, but that CG may be detected by the ITF. Such a flash, by this algorithm, would be falsely considered an IC. However, the NLDN’s detection efficiency in northeastern Colorado is around 90% or more (Cummins et al., 1996), so it is quite high. However, this estimate has never been independently verified in this specific region, so any possible error in this estimate, or whether such an estimate is storm-dependent, is unknown. However, there is no other estimate available, and there is no reason – beyond healthy skepticism – to suspect this estimate to be in error. Hence, it must be used.

Based on this detection efficiency estimate, the error associated with this algorithm is not likely to be much more than 10% of the CG population, which should not seriously impact CG flash rate trends. Because both storms in question were very poor producers of CGs in the first place, this should not impact the (often high) IC flash rates significantly. Any error associated with this algorithm is probably minuscule compared to the error associated with applying the ITF CG classification method used by ONERA. Figures 3.1 through 3.4 show that fact very clearly.

This approach is different than that of Lang et al. (1997), who computed total flash rate for the 10 July 1996 storm by editing out of the ITF data only those flashes that were not associated with the storm of interest. Sub-ms flashes and ITF “CGs” were included in the total flash rate. However, as will be seen in Chapter 4, the flash rates computed by Lang et al. trend the results of this study quite well, mainly because the sub-ms flashes trend the long-duration ICs so well, and because NLDN CG rates were not high enough to impact total flash rates significantly.