We report on upward lightning observations from ten tall towers (91–191 m) in Rapid City, South Dakota, USA and compare with National Lightning Detection Network (NLDN) data. A total of 81 upward flashes were observed from 2004–2010 using GPS time-stamped optical sensors, and in all but one case, visible flash activity preceded the development of the upward leaders. Time-correlated analysis showed that the NLDN recorded an event within 50 km of towers and within 500 ms prior to upward leader development from the tower(s) for 83% (67/81) of the upward flashes. A preceding positive cloud-to-ground stroke (+CG) was detected in 57% (46/81) of the cases, and a preceding positive intracloud flash (+IC) in 23% (19/81) of the cases. However, 8 of the 19 NLDN-indicated +IC events were actually +CG strokes based on optical observations. Preceding negative intracloud flashes (−IC) were recorded for 2% (2/81) of the cases. Analysis also showed that for 44% (36/81) of the upward flashes, the NLDN reported subsequent negative cloud-to-ground (−CG) strokes and/or −IC events at one or more tower locations. Of the 151 subsequent events, 70% (105/151) were −CG reports and 30% (46/151) were listed as −IC events. The geometric mean/median location accuracy and peak current for subsequent events were 194 m/206 m and −12.9 kA/−12.4 kA respectively. These correlated observations suggest that a majority of the upward lightning flashes were triggered by a preceding flash with the dominant triggering type being the +CG flash.
Berger was the first to raise the question of whether initiation of all upward lightning is caused by preceding flashes. He stated that, “Very often a sharp noise is audible at the tower tops at the instant of a distant flash, as with the well-known impulse corona. Field impulses may then cause the development of upward leaders.”Berger and Vogelsanger also felt that the high electric field needed for the initiation of upward leader is rapidly created by an in-cloud discharge, rather than from the slower charge buildup produced by cloud electrification. Even though this question was raised four decades ago, the initiation of upward lightning is an issue that has yet to be fully investigated and explained.
 Reports from recent field observations are beginning to address the nature of upward lightning initiation and suggest various initiation mechanisms. As part of a sprite research effort, Stanley and Heavner reported on two negative cloud-to-ground (−CG) flashes recorded by the National Lightning Detection Network (NLDN) at the location of a 457 m tall tower, and in each case, these flashes occurred within one second after a preceding positive cloud-to-ground (+CG) flash that was detected within 20 km of the tower. They suspected that the −CG strokes were dart leader/return stroke sequences that followed upward positive leaders (UPLs) initiated at the tower by the preceding +CG flashes. They then analyzed 405 −CG flashes that followed within one second and within 40 km of preceding +CG flashes. They found that −CGs which follow +CGs do not show a statistically enhanced probability of striking structures less than 400 m tall, but do show a more than threefold enhanced probability of striking tall structures greater than 400 m. They went on to speculate that +CG flashes with large charge moments (i.e., those suspected of triggering sprites) were also spawning upward lightning from the tallest towers.
 Specifically addressing the nature of upward lightning initiation, Wang et al. described two types of the electric field changes associated with upward lightning from a windmill and its lightning protection tower. The first type exhibited a rapid negative change in electric field (physics sign convention) following initiation of the upward leader. This type represented a self-initiated UPL forming from the tall object without preceding nearby flash activity. The second type of electric field change had first a positive change in electric field that lasted tens of milliseconds followed by a negative change similar to that observed for the first type.Wang et al.  felt that the initial rise in the electric field was associated with a lightning discharge occurring in clouds. This supported the idea of Berger that a preceding discharge rapidly changes the ambient electric field over a tall object and triggers an upward leader from that object. Lu et al. observed two sequentially occurring and opposite-polarity upward leaders from tall objects that were separated by 375 m and argued that this supported the hypothesis that a preceding upward leader can trigger a second, given a favorable thunderstorm charge structure.
 More recently, Wang and Takagi reported on 53 upward flashes from a windmill and its protection tower. 47% of the upward flashes were self-initiated and 53% were other-triggered (i.e., “triggered by a discharge that occurred in other places which could be either a cloud discharge or a cloud-to-ground discharge” [Wang et al., 2010]). Although not quantified, they noted that some storms produced exclusively self-initiated upward lightning while others produced only other-triggered upward flashes. A combination of self-initiated and other-triggered upward flashes was also observed during some storms. Their analysis suggested that taller, more electrically active storms favored other-triggering primarily from the taller protection tower (i.e., 75% of the other-triggered initiated from the protection tower). Less active storms favored self-initiation with only a slight preference for the upward leaders to initiate from the protection tower (56%) rather than the rotating windmill (44%). They also noted that self-initiation occurred with higher observed wind speeds (or a rotating windmill) compared with other-triggered upward flashes. Interestingly, the only observed upward flash with leaders from both the protection tower and windmill was other-triggered. In this paper, we will refer to “other-triggered” as defined byWang et al. as “lightning-triggered” since it is preceding lightning activity that triggers the upward lightning.
Zhou et al. reported on 205 upward flashes observed from 2005 to 2009 at the Gaisberg Tower near Salzburg, Austria. Their analysis of coordinated electric field and current measurements, along with lightning location data, showed that 179/205 or 87% of the upward flashes were self-initiated whereas only 13% (26/205) were initiated by nearby triggering lightning discharges (i.e., lightning-triggered). Ten positive cloud-to-ground (+CG) flashes, one negative cloud-to-ground flash (−CG) and 15 cloud flashes were identified as the triggering flashes. A majority of the self-initiated upward flashes (79%) occurred during the non-convective season (September–March), whereas a majority of the lightning-triggered upward flashes (85%) occurred during the convective season (April–August). They note, however, that the 4 lightning-triggered flashes occurring outside of the defined convective season took place in early September implying that they might also have been initiated during a more convective regime. They also suggest that the lower cloud base and freezing level experienced during the winter months may be a factor in the high percentage of self-initiated upward flashes experienced during this period.
 In this paper, we present observations and analysis of upward lightning from multiple towers in the northern High Plains of the United States and relate these observations to NLDN reports. These observations, made during the summer convective season, provide for a comparison with observations made in other regions of the world with varying weather patterns and storm modes. The comparison of NLDN data with high-speed optical recordings provides a unique opportunity to evaluate a lightning location system's performance during upward flashes and to understand how NLDN reported events relate to the initiation and development of upward leaders. Furthermore, the varied spacing and heights of the multiple towers in this study domain provides a unique database of observations where tower interaction and preferential development can be investigated.
 The focus of this research are 10 communication towers located on a north-south ridge that runs through Rapid City, South Dakota, USA. Rapid City is located in the western part of South Dakota and on the eastern edge of the Black Hills as shown inFigure 1. The Black Hills are an area of forested uplift roughly forming a north-south oriented oval with the north-south length approximately 150 km and the east-west length approximately 80 km. This elevated forest is surrounded by rolling prairie grassland. The terrain in the Black Hills rises to its highest point of just over 2,200 m above mean sea level (MSL), which is about 1,200 m above the surrounding plains on the east side. The elevated terrain in the Black Hills and resulting orographic lifting produces frequent afternoon thunderstorms during the summer months, which lasts from April through September. Thunderstorms typically dissipate once they move off of the Black Hills unless aided by synoptic or mesoscale support. When this support is present, strong multicellular clusters and supercells can occur along with occasional mesoscale convective systems (MCSs) that propagate through the region. The number of thunderstorm days for Rapid City is between 45 and 50 per year [Changnon, 2001, Figure 5], and the detection-efficiency-corrected flash density for this region is 2–3 flashes/km2/yr (K. Cummins, Vaisala, personal communication, 2009).
 The 10 communication towers range in height from 91 m to 191 m, and all but two are guyed using multiple grounded guy wires to surrounding attachment points. The remaining two towers are free standing and grounded. Table 1 summarizes the tower parameters including their heights above ground level (AGL) and base and top elevations relative to mean sea level (MSL).
Table 1. Tower Information
Tower ID Number
Base Elevation MSL (m)
Height AGL (m)
Top Elevation MSL (m)
Tower height lowered to 54.9 m (1200 m MSL) on 4/22/2005.
 The tower locations are shown relative to surrounding MSL elevation in Figure 2a. The north-south oriented ridge upon which the towers sit rises approximately 180 m above the elevation of the creek that runs east-west through the ridge between Towers 1 and 2. The distance from the northern-most tower to the southern-most tower is 7.9 km. Tower 1 sits on a separate ridge segment and has the steepest elevation gradient surrounding the tower.Figure 3 shows Towers 1–6 as viewed from one of the primary observing locations looking northeast. The gap in the ridge between Towers 1 and 2 created by the creek is visible in the middle of the image.
 Towers 2–6 are clustered together spanning 553 m and their relative spacing is shown in Figure 2b. Towers 4 and 5 are separated by only 56 m with the tip of Tower 5 10.6 m lower in elevation than Tower 4. The tallest tower in height above ground is Tower 6 at 190.8 m AGL, however, the tip of this tower sits 18.0 m lower in elevation than Tower 4 which is the highest tower in elevation at 1341.4 m MSL.
 Towers 7–10 sit on a wider and less steeply varying portion of the ridge and their spacing is less clustered. Tower 9's height was lowered from 115.8 m to 54.9 m AGL in early 2005, and no upward flashes have been observed from this tower since.
3.1. Optical Instruments
 GPS time-stamped standard-definition (SD) video camera systems were used to observe upward lightning flashes during each of the seven years of observations. These camera systems provided video fields with 16.7 ms exposure times (interlaced frame recording) and a time-stamp accuracy to 1 ms. In 2005, the addition of a SD camera with a field-of-view greater than 90° allowed for continuous monitoring of all 10 towers and the overlaying clouds out to the eastern horizon. Three other SD cameras employing narrower fields-of-view focused on individual clusters of towers to provide higher resolution details associated with upward leader development. In order to avoid confusion between camera recording systems utilizing interlaced frame recording which produce two images (i.e., two fields) per recorded frame and progressive frame recording which produces one image per frame, we will express camera recording speeds in images per second (ips). Therefore, a standard-definition interlaced frame recording video camera records at 60 ips. Standard-speed cameras are defined as those that record at 60 ips or less and high-speed cameras as those that record at greater than 60 ips.
 In 2008, a Vision Research Phantom Miro 4 and v7.1 high-speed camera were added and observations were obtained between 1,000 and 7,200 ips. A Vision Research Phantom v12.1 high-speed camera on loan from Texas A&M University (Richard E. Orville, facility manager) resulted in observations at recording rates up to 67,000 ips starting in 2009. In addition, a progressive frame scanning high-definition video camera was added that allowed for higher resolution (1280 × 720 pixels) at the 60 ips rate. For this last year of reported observations, 2010, a Vision Research Phantom v310 replaced the v7.1, and a second HD video camera along with two additional standard-definition video cameras providing “all-sky” coverage were added.
 The high-speed cameras all have monochrome sensors with 12-bit dynamic range. The ISO ratings are 4,800 for the Miro 4 and v7.1, 6,400 for the v12.1 and 7,000 for the v310. Sensor pixel size for the Miro 4 and v7.1 are 22μm and 20 μm for the v310 and v12.1. Nikon F-mount and surveillance C-mount fixed focal length, manual focus lenses were used based on the desired field-of-view. Camera speeds were selected between 1,000–67,000 ips and corresponding sensor resolutions varied from 1280 × 720 to 256 × 256 pixels. Slide film and digital still cameras were also utilized during the seven years of observations. These cameras were infrared-triggered during the day and took sequential 20 s exposures at night so that entire flashes could be recorded onto single images.
 The continuously monitoring, autonomously operated standard-definition cameras were located at fixed sites whereas the high-speed cameras were operated manually from a vehicle that could be positioned at various observing locations. This vehicle also housed standard- and high-definition video cameras and a digital still camera. A second filming vehicle was added in 2010 in an effort to get correlated high-speed observations from multiple locations.
 The National Lightning Detection Network (NLDN) cloud and cloud-to-ground (CG) stroke data for a 200 km radius from Rapid City were provided by Vaisala, Inc. for this analysis. The system's description and performance are summarized inCummins and Murphy . These data include the estimated stroke time with an accuracy of ∼1.0 μs, latitude and longitude of the strike location with an expected median error of less than 500 m, peak current estimate with an RMS uncertainty of ∼15%, and a number of quality parameters [Cummins et al., 1998]. Accuracy of these measurements for cloud pulses has not been verified, and the peak current estimates for cloud pulses should be viewed as an “equivalent peak current” had the discharge been a return stroke.
 In regards to classification by the NLDN, positive events are those in which positive charge is “effectively” lowered toward ground. For ground strokes, a +CG indicates an impulsive return stroke resulting from a positive leader connecting with ground. For an intracloud flash, a +IC indicates an impulsive event following the bipolar development of downward propagating positive leaders and upward propagating negative leaders [e.g., Weidman and Krider, 1979], and reflects the polarity of the initial deflection of the pulse-like waveforms that are detected by the NLDN sensors. This orientation would be associated with a normal polarity intracloud flash. Specifically, a normal polarity intracloud flash is one that initiates between an overlying positive charge area and underlying negative charge area with positive (negative) leaders initially moving downward (upward) from the initiation point [Krehbiel et al., 2008]. An inverted intracloud flash would have oppositely oriented charge areas (i.e., negative over positive) with positive leaders moving upward and negative leaders moving downward. An impulsive event recorded during an inverted intracloud flash would, therefore, be classified as an −IC. An NLDN-indicated −CG results from the return stroke formed by the connection of a downward moving negative leader with the ground. Based on our analysis of high-speed camera observations of cloud flashes with leader segments visible outside of the cloud, NLDN-indicated IC events correlate with rapid/impulsive leader growth or redevelopment, recoil leader connections with luminous channel segments or the connection of two opposite polarity leaders. The NLDN locations of these events represent the 2-dimensional positions on the ground over which these events took place.
 For further clarification, the NLDN would currently classify an impulsive event during the upward development of a positive leader from a tall object as negative (i.e., −CG or −IC). Similarly, an upward flash with upward propagating positive leaders is commonly defined in the literature as upward negative lightning [e.g., Berger, 1967; Rakov and Uman, 2003, p. 5; Diendorfer et al., 2009].
4. Analysis and Results
 In this section, we first present the optical observations, followed by a comparison with NLDN reports. Analysis of the correlated optical observations and NLDN data was carried out in two timeframes: (1) prior to upward leader initiation and (2) after upward leader initiation. NLDN events recorded before upward leader initiation are classified as “preceding events” and those associated with an upward leader after its initiation are considered “subsequent events.”
4.1. Optical Observations
 Analysis of recorded video included determination of upward leader initiation and end times as well as quantifying the beginning time and type (i.e., cloud or ground) of any nearby flash activity that preceded upward leader development. There were a total of 81 upward flashes observed over the 7 years of observations, and in all but one of the cases, there was visible brightness associated with nearby flash activity prior to initiation of the upward leader(s). Frequently, video showed a bright flash (cloud or ground) away from the towers followed by in-cloud brightening that propagated from the apparent flash origin toward the towers. As this in-cloud brightening passed near or over the towers, upward leaders would then initiate. This recurring observation suggests that leader development associated with preceding flashes (seen as horizontally propagating in-cloud brightening) created conditions favorable for upward leader initiation and development.
 The time from the first indication of in-cloud visual brightness increase to upward leader initiation was determined for those 80 flashes displaying this preceding activity. The resulting mean, GM, median, maximum and minimum time between the first visual sign of flash activity and upward leader initiation was 181 ms, 112 ms, 109 ms, 918 ms and less than 1 ms respectively. It is interesting to note that the maximum delay was almost an entire second. For the 37 cases (46%) in which there was a preceding visible ground stroke, the mean, GM, median, maximum and minimum time between the visible ground stroke and upward leader initiation was 50 ms, 32 ms, 36 ms, 267 ms and less than 1 ms respectively. Upward flash duration was determined from the first visible evidence of an upward leader until the last visible luminosity in the leader channel following either the initial upward leader development or subsequent return strokes if they occurred after initial upward leader decay. The mean, GM, median, maximum and minimum upward flash duration was 456 ms, 381 ms, 408 ms, 1934 ms and 21 ms respectively.Table 2 summarizes these measurements.
Table 2. Time Summary
Time from incloud visual brightness to upward leader (uncertainty) (ms)
Time from visual CG to upward leader (uncertainty) (ms)
Upward flash duration (uncertainty) (ms)
Time from preceding NLDN event to upward leader (uncertainty) (ms)
Time from upward leader to first subsequent NLDN event (uncertainty) (ms)
Time from upward leader to last subsequent NLDN event (uncertainty) (ms)
 Uncertainty in the time in which an event started or ended depended on the length of the exposure used in the camera that recorded the event. Since cameras operating at different speeds were used, the exposure times (and therefore the uncertainty) ranged from 34 ms to 0.015 ms. The mean and GM uncertainty values were as follows: time from visual brightness to upward leader (25 and 28 ms), time from visual CG to upward leader (23 and 11 ms), and upward leader duration (22 and 28 ms).
Figure 4 shows the number of flashes each tower experienced along with a) the tower heights above ground level (AGL) and b) tower tip elevations above mean sea level (MSL). Tower 4 experienced the most upward flashes at 40 (49%). This tower extends to the highest elevation at 1341.4 m MSL as shown in Table 1, but is 6.1 m shorter in height than the tallest tower (Tower 6). The tip of Tower 6 is 18.0 m lower in elevation than Tower 4 and has experienced 26 upward flashes (32%). Interestingly, Tower 1's tip, which is lower in elevation than Tower 6 by 14.3 m and is lower in height by 27.7 m, has experienced 34 upward flashes (42%), which is 8 more than Tower 6. The ridge topography near Tower 1 has much steeper elevation change near the tower when compared to the other towers (see Figure 2). It is also isolated from the location grouping of Towers 2–6. Tower 1 may have an effective height [e.g., Eriksson, 1978; Pierce, 1971] that is greater than Tower 6 and there may be some shielding interaction among Towers 2–6 due to their close relative locations. Shielding from Tower 4 seems the likely cause for the low number of upward flashes for Tower 5. Even though Tower 5 is the third tallest in height and second tallest in elevation, it has only experienced one upward flash. Tower 4, which is 56 m away and 10.6 m taller in elevation than Tower 5, is likely acting as a protection tower for Tower 5. Furthermore, Tower 4 sits closer to the steeper west side of the ridge, which might result in greater field enhancement for Tower 4 relative to Tower 5 (see Figure 2b). A scatterplot of tower height and corresponding number of upward flashes is shown in Figure 5. This plot does not include Tower 5 due to the apparent shielding provided by Tower 4. A best fit trend line is a second order polynomial with an R2 value of 0.7371. Interestingly, the trend line goes to zero near a height of 100 m, which is consistent with previous studies that provide an expected minimum height for upward flashes [Eriksson, 1978; Rakov and Uman, 2003]. When Tower 5 is included in the data, the R2 value drops to 0.4728.
 More than one tower initiated an upward leader during a flash in 37 cases (46%). Figure 6 shows the number of flashes involving more than one tower by the number of towers involved. The maximum number of towers that developed upward leaders during a single flash was seven. Figure 7shows how many times each tower participated in a multiple upward leader flash MULF, (i.e., a flash with upward leaders that develop from at least two towers). Analysis of high-speed camera observations from nine MULFs between 2008–2010 indicate that the multiple upward leaders initiated simultaneously or nearly simultaneously and were of the same polarity (i.e., positive). A detailed analysis of one MULF is given inWarner .
 Recoil leaders, which have been shown to be associated with positive leader development [e.g., Mazur, 2002; Saba et al., 2008], were identified optically in 73 (90%) of the upward flashes suggesting that UPLs developed in those cases. This high percentage of upward leader positive polarity is consistent with previous observations [e.g., Berger, 1967; Diendorfer et al., 2009]. Furthermore, upward leaders in the remaining 10% of the flashes did not exhibit optical characteristics of negative leaders (e.g., erratic leader tip directional change, localized branching and sustained stepping) suggesting they too were positive polarity.
4.2. NLDN Comparison
 Once the optical characteristics were quantified, a comparison with NLDN data was made. NLDN data included both cloud events and cloud-to-ground (CG) strokes within a 200 km radius of the towers. For each upward flash, NLDN-indicated events from 2 s prior to 3 s after the upward leader initiation time were identified and plotted relative to the tower locations. The resulting NLDN events were then correlated with standard- and high-speed video to specifically identify those that appeared to be associated with the upward flashes. This process resulted in the identification of 67 of the 81 upward flash cases (83%) in which a flash with a correlated NLDN event preceded upward leader development from the towers. A preceding positive cloud-to-ground stroke (+CG) was detected in 57% (46/81) of the cases, and a preceding positive intracloud event (+IC) in 23% (19/81) of the cases. However, 8 of the 19 NLDN-indicated +IC events were actually +CG strokes based on optical observations. A preceding negative intracloud event (−IC) was recorded for 2% (2/81) of the cases. All but one video-validated preceding CG stroke had a corresponding NLDN report. For the 14 cases (17%) in which there was no preceding NLDN-indicated event, optical evidence showed there was preceding flash activity in all but one case. In one of these 14 cases, there was a non-detected CG stroke observed by a standard-speed camera, however the polarity could not be determined from the optical assets.
Figure 8shows a plot of the preceding NLDN-indicated events relative to the towers. +CGs are indicated as a red circle, +ICs as a light red circle and −ICs as a light blue triangle. The 8 misclassified +ICs, which were +CG strokes based on optical evidence, are shown as +CGs in the plot. The symbol size is proportional to the estimated peak current as shown in the legend. The mean distance of these NLDN-indicated events from the closest participating tower was 17.5 km and the GM, median, maximum and minimum distance was 14.7 km, 14.8 km, 50.5 km and 3.5 km respectively. The mean time between the preceding NLDN-indicated event and upward leader initiation was 72 ms with the GM, median, maximum and minimum being 39 ms, 46 ms, 404 ms and less than 1 ms respectively (seeTable 2). Table 3summarizes the estimated peak current and distance parameters for each type of NLDN-indicated preceding event. Distances are from the nearest tower that developed an upward leader.
 A subjective look at the distribution of preceding events does suggest events are grouped to the east and southeast of the towers with a more widely spaced distribution to the north and west. However, an objective evaluation of the distribution pattern would require a correlated analysis with weather data to determine the storm characteristics (e.g., type, coverage and movement) present during the upward flashes. This analysis is beyond the scope of this paper but is ongoing and will be reported in the future.
 There were no NLDN-indicated events that coincided with initial upward leader development. This is not surprising since the current associated with initial upward leader development rises slowly compared to that of the impulsive dart leader/return stroke processes that the NLDN is optimized to detect [Cummins and Murphy, 2009; Cummins et al., 1998]. However, there were a total of 151 negative cloud-to-ground strokes (−CGs) and negative cloud events (−ICs) recorded at the tower locations after initial upward leader formation. These subsequent events, as they will be referred to here, were present in 36 (44%) of the upward flashes.
 The process of determining subsequent NLDN events first involved identifying those detections that occurred within the duration of the optical observations, and second, those that were also located within 1 km of a tower that developed an upward leader. For those flashes observed with high-speed cameras, the subsequent events were further correlated with the higher time-resolved optical observations. The high-speed video analysis showed that these subsequent events were detections resulting from either 1) the connection of a recoil leader, which formed on a cutoff UPL branch, with a main luminous upward channel at the branch point or the tower tip when the main upward channel was weakly luminous or 2) a dart leader connection with the tip of a tower after complete current cutoff.
Figure 8shows a plot of the NLDN-indicated subsequent events relative to the towers. All the subsequent events are clustered around the 10 towers as identified by the dark blue triangles (−CG) or light blue triangles (−IC). There were 105 −CG and 46 −IC events recorded representing 70% and 30% of the 151 respectively. Since the subsequent events involve known locations (i.e., the towers), an evaluation of the NLDN location accuracy was possible. The mean, GM, median, maximum and minimum distances for all the subsequent events were found to be 246 m, 194 m, 206 m, 2280 m, and 7 m respectively. Notice that the maximum distance (2.28 km) was greater than the 1 km used in selecting subsequent events. There was a single subsequent event that was clearly correlated with high-speed video and so this was included in the location accuracy calculations. All other subsequent events fell within 1 km of the towers.Table 3summarizes the estimated peak current and distance parameters for each type of NLDN-indicated subsequent event. Distances are from the tower that developed the upward leader.
 The results from statistical analysis in the NLDN-indicated subsequent event occurrence times relative to upward leader initiation is given inTable 2. The mean and GM time delay between upward leader initiation and the first NLDN-indicated subsequent event was 324 ms and 271 ms respectively with the shortest differential being 66 ms. The mean and GM delay between upward leader initiation and the last NDLN-indicated subsequent event was 459 ms and 389 ms respectively. The longest time differential was 1.741 s.
 An NLDN-indicated event density analysis was made to determine if there was enhanced event density near the tower locations due to the upward flashes. Given that the mean distance of the subsequent events was 246 m, the NLDN-indicated events that fell within a 500 m radius of each of the towers during the seven summer storm seasons (Apr–Sep, 2004–2010) were identified and used to determine an annualized number of events/km2/summer season. The density was calculated for all of the 500 m radii cumulatively and for Tower 1's radius individually as this radius did not intersect with any of the other towers' radii. Figure 9shows the tower relative locations and 500 m radius circles. For comparison, NLDN-indicated events that fell within a 10 km radius of Tower 6 but outside of the cumulative 500 m radii around the towers were also identified. The number of events/km2/summer season was determined for this area as well and is considered representative of the event density not influenced by upward lightning from the towers [e.g., Diendorfer and Schulz, 1998].
Figure 10 shows the annualized density of all events as well as individual event categories for a 500 m radius around Tower 1, a cumulative 500 m radii around all the towers and a 10 km radius around Tower 6 (not including the cumulative 500 m radii around all the towers). The positive polarity event density did not show a statistically significant difference between the three selected areas. However, the negative polarity event density was significantly higher in the Tower 1 500 m radius area as well as the cumulative 500 m radii area around all the towers. Specifically, the −CG density in the Tower 1 500 m radius area was 4.9 times greater than the exclusive 10 km radius area around Tower 6, and the −IC density was 14.4 times greater. The −CG density in the cumulative 500 m radii around all the towers was 2.8 times greater than the exclusive 10 km radius area around Tower 6, and the −IC density was 6.5 times greater. Based on this analysis, the subsequent events resulting from the upward flashes clearly enhanced the event density (specifically the negative event density) near the towers.
 A total number of +CG and −CG NLDN-indicated strokes for a 50 km radius of Tower 6 for the entire observation period was determined for comparison. There were 20500 +CG strokes and 280895 −CG strokes. The −CG to +CG stroke ratio was 13.7 to 1 and the percentage of +CG strokes to the total number of CG strokes was 7%. With 54 of the 81 upward flashes preceded by +CG strokes, the percentage of triggering +CGs within 50 km for the entire observation period is only 0.29% for all +CGs and 0.0008% for all CGs.
 Interestingly, there were no upward flashes associated with preceding −CG return strokes. Zhou et al. reported that one of 26 observed lightning-triggered upward flashes was preceded by a negative ground flash. Although the development of upward positive leaders following positive triggering flashes appears significantly more common than upward negative leaders following negative triggering flashes (e.g., this study andZhou et al. ), it is unclear why there have been no observed triggered upward negative leader cases in Rapid City. One possibility, based on visual observation, is that negative ground flashes tend to lack the horizontally extensive leader development frequently seen with positive flashes. Negative flashes also tend to occur in the convective core of mesoscale convection systems whereas positive flashes are frequently seen in the trailing stratiform precipitation area, which has been shown to contain horizontally stratified charge layers favorable for horizontally extensive flashes [e.g., Lang et al., 2010].
5.1. Preceding Events
 The fact that visible flash activity preceded upward leader initiation in all but one of the 81 upward flashes suggests that a component of a nearby flash triggered the upward leader(s) in most cases. Analysis of correlated optical and NLDN data indicated that a +CG return stroke occurred during this preceding flash activity in at least 67% of the cases. Therefore, a majority of the upward leaders appeared to initiate during a lightning process that followed a +CG return stroke. Given the variability in time delay between the return stroke and upward leader initiation (see Table 2), the triggering processes likely involved are the return stroke itself as it travels up and through the bipolar leader network that formed prior to connection with the ground or negative leader development that may follow a +CG return stroke once it reaches the upper and outer extent of initial leader development prior to the return stroke. This subsequent leader development (negative in the case of a +CG) is time-coincident with the continuing current seen in the return stroke channel.
 Since 8 of the 19 +IC events were actually +CG strokes based on optical evidence, it is possible that some of the remaining 11 +IC events were also +CG strokes that were outside the fields of view of the cameras. There were correlated in-cloud brightness pulses with each of the 11 events, and 5 had exceptionally intense brightness pulses similar to those cases with confirmed CG strokes. Regarding the two preceding NLDN-indicated −IC events, one correlated to an intense brightness pulse just northwest of the towers and upward leaders were visible in the standard-speed image (17 ms exposure) that followed. This event was not captured with high-speed cameras. In the second case, an in-cloud brightness pulse was again visible in standard-speed video, but there was a delay of 404 ms before an upward leader developed. During this delay, negative leaders were visible propagating near the towers as seen by high-speed cameras (18,604 ips, 53μs exposure).
 For those IC events that may have been correctly identified by the NLDN as part of a cloud flash, the significance of their location relative to the towers (see Figure 8) is unclear as the impulsive event may not be associated with the triggering flash component that caused the upward leader initiation. However, visual observations showed that horizontally propagating in-cloud leader development followed the bright NLDN-indicated IC events. Therefore, the IC events recorded by the NLDN seemed to correlate with a cloud flash component that caused horizontal leader development to pass near the towers, which was followed by upward leader initiation.
 For the 13 cases where there was preceding visible flash activity but no NLDN-indicated events recorded, a CG stroke was visible in one case and a rapid in-cloud brightness increase observed in nine cases. In-cloud brightness of fairly steady intensity propagated toward the towers in the remaining three cases. Interestingly, these latter three events occurred within an 11 min period during the same storm. The fact that the NLDN did not record any events during these three flashes and that they did not exhibit any intense brightness increases suggests that they may have been intracloud flashes.
 Fourteen of the upward flashes reported in this paper had electric field change instrumentation near the towers. Although analysis is ongoing, initial results indicate that UPLs were triggered by either 1) the approach of horizontally propagating negative stepped leaders associated with either intracloud development or following a +CG return stroke or 2) a +CG return stroke as it propagates through a previously formed leader network that is near the towers [Warner, 2012; Warner et al., 2012a].
5.2. Subsequent Events
 The fact that all of the NLDN-indicated subsequent events were negative polarity suggests UPLs developed in those cases (i.e., 44% of the upward flashes). The optically observed recoil leader activity in 90% (73/81) of the flashes further suggests UPL development in a large majority of cases. Recoil leaders were seen to developed on weakly luminous positive leader branches that became cutoff from the main upward channel from which they branched [Mazur and Ruhnke, 2011; Mazur et al., 2011]. A bright bidirectional leader (recoil leader) would initiate at a point that lagged behind the tip of the advancing positive leader branch. The bipolar/bidirectional leader would expand with the negative end propagating toward the branch point along the main channel and the positive end propagating toward the positive leader tip. When the recoil leader negative end (RLNE) connected with the main luminous channel at the branch point, a bright luminosity pulse would result that traveled from the branch point outward toward the branch tip. Correspondingly, the upward leader segment from the branch point downward to the tower tip would also brighten. The segment from the branch point upward (above) the branch point would not brighten indicating it did not participate in the connection and resulting pulse. During the later portions of upward flashes in which the main channel luminosity had weakened significantly, the luminosity front associated with the RLNE would reach the branch point and continue down the weakly luminous main channel without causing a luminosity pulse at the branch point. The RLNE would maintain a clearly defined luminosity front while propagating on the weakly luminous main channel, although the front tended to lengthen and increase in speed [Warner et al., 2012b]. A luminosity pulse would occur when the front reached the tower tip (similar to a dart leader/return stroke sequence following complete current cutoff in the main channel). Both of these connection types would result in a current pulse exceeding the slowly varying current resulting from the propagating upward leader, i.e., initial stage (IS) pulses [e.g., Flache et al., 2008; Diendorfer et al., 2009]. If the impulsive electromagnetic field change resulting from these connections where to exceed the NLDN's sensing threshold, then a negative polarity event (−CG or −IC) would be recorded. This was in fact the case for those connections exhibiting only the brightest pulse luminosity as seen by the high-speed cameras.
Figure 11shows the channel base brightness measured at the tip of the tower for an upward flash recorded by a high-speed camera at 9,000 ips. The trace is representative of the current flow to the top of the tower [e.g.,Diendorfer et al., 2003] and shows that the flash exhibited a slowly varying current with numerous IS pulses. The moderate brightness variation at point a lasting 40 ms correlated to the rapid and intense growth of one of the upward leader branches during the initial upward leader development. There were 22 pulses resulting from recoil leader connections, three of which were reported as −CG strokes by the NLDN. The NLDN recorded a −9.4 kA estimated peak current stroke at point b and two closely spaced strokes at point c with estimated peak currents of −8.8 kA and −13.9 kA respectively. Camera sensor pixel saturation resulted in peak brightness values being limited to 255.
 There were a total of 54 NLDN-indicated subsequent events recorded by at least one high-speed camera operating between 7,200 and 67,000 ips (135μs and 14 μs exposure). The initiation and development of a bright bidirectional recoil leader that would connect with either a luminous main channel or tower tip was visible in nine of these recordings. For the remaining cases, the bright leader propagated into the high-speed camera's field of view prior to connection. Six of the 54 recordings showed the impulsive connection occurred when the RLNE reached the main luminous channel at the branch point and the remaining 48 connections resulted when the RLNE luminosity front reached the tower tip. However, there was still luminosity visible (i.e., current) in the main upward leader channel as the recoil leader traveled downward toward the tower tip in 40 of the 48 latter cases.
Figure 12 shows two connections where the recoil leader initiation point was visible (7,207 ips, 135 μs exposure). The NLDN-indicated an −8.1 kA estimated peak current −IC event for the first connection shown inFigure 12a and a −15.1 kA estimated peak current −CG for the second shown in Figure 12b, which occurred 13 ms later along the same branch. The −IC event, and most other subsequent events classified as a −IC, exhibited NLDN waveform parameters much like subsequent return strokes, but with a slightly narrower pulse width.
Figure 12a Image 1 shows the initiation of the bidirectional recoil leader. The negative end propagates to the left toward the tower whereas the positive end only proceeds a short distance to the right. A connection is made with the tower as shown in Image 9, and a luminosity increase travels back up the path traversed by the recoil leader and illuminates a lower branch segment near the end of the positive leader as indicated by the arrow in Images 11 and 12. In Figure 12b the recoil leader again initiates bidirectionally in Image 1. As the left (negative) end travels to the left toward the tower, the right (positive) end illuminates a short forked segment at the apparent end of the positive leader branch (Image 2). The forked branches fade in the next image with only a single branch visible until the recoil leader connection with the tower. A connection with the tower is made in Image 4. The brightness pulse that traverses the recoil leader's path back is much brighter than in the previous case shown in a) as seen by the saturating luminosity in Image 5. Two arrows mark the tips of two branches illuminated by the arrival of the luminosity front in Image 5. Their brightness fades in the subsequent images, however, another branch illuminates as seen in Image 7. As this branch fades, a final fourth branch illuminates in Images 9 and 10. Due to the long exposure time of 135 μs, it is unclear if the last two branch illuminations are the result of new recoil leaders that formed on the decayed branches due to the potential increase experienced from the newly reconnected branch. The first author, using higher recording speeds, has observed additional recoil leaders form on secondary decayed branches of newly reconnected positive leader branches. When these connect with the newly reionized and still luminous branch, additional brightening along the branch and main channel results [Warner et al., 2012b]. For the flash highlighted in Figure 12, the 2-dimensional speeds of the RLNEs ranged between 4.5 × 106 m/s and 1.4 × 107 m/s with an arithmetic mean of 1.4 × 107 m/s and geometric mean of 9.1 × 106 m/s.
 The subsequent events that occurred later in the upward flashes' total duration tended to initiate outside the high-speed cameras' field of view or within the cloud. Frequently, in-cloud brightening could be seen by the wider field of view cameras and this brightening traveled toward the tower along a recurring path taken by previous subsequent events with a luminosity front emerging into the high-speed camera field of view prior to connecting with a main luminous channel or tower tip. These events exhibited the same characteristics and behavior as earlier events in which the initiation point was visible.
 Only 8 of the 54 high-speed recordings showed no luminosity in the upward leader channel prior to the recoil leader's connection with the tower tip. These 8, therefore, represent a dart leader/return stroke sequence occurring after complete current cutoff in the upward leader channel. The observed initiation and development of the connecting leader was the same whether the leader developed during the initial stage or after complete current cutoff at the tower tip. Specifically, a bidirectional/bipolar leader developed on a cutoff positive leader branch and the negative end of the leader connected with either a luminous channel or tower tip. This adds evidence that recoil leaders and dart leaders are in fact the same physical process serving to reionize a decayed positive leader channel/branch.
 Since the NLDN classified 30% of the impulsive connections as intracloud flashes, we investigated why the misclassification occurred. Figure 13 shows the distribution of IC and CG subsequent events based on their risetime (Figure 13a), peak-to-zero time (Figure 13b) and pulse duration (Figure 13c) as determined by the NLDN computational algorithm. The algorithm considers the recorded parameters from each station that sensed the event and classifies the event based on these values. One risetime and peak-to-zero time is then calculated for the event, using the closest reporting sensor. The risetime values (Figure 13a) did not show a statistically significant difference between IC and CG events. As seen in the peak-to-zero time (Figure 13b) and pulse duration distributions (Figure 13c), IC events formed a lower tail on a unimodal distribution of widths for the events classified as −CG events.
 The NLDN detection efficiency (DE) for IC pulses was increased in early 2006 [Cummins and Murphy, 2009]. Prior to that date, there were 6 out of 14 observed upward flashes (43%) in which the NLDN reported a total of 30 NLDN-indicated subsequent events, 7 of which were reported as IC events (23%). After that date, there were 30 out of 67 upward flashes (45%) in which the NLDN reported a total of 121 subsequent events, 40 of which were classified as IC events (33%). Therefore, after the DE change, the percentage of subsequent events classified as IC events increased 10% and the percentage of upward flashes with reported subsequent events increased 2%.
 Channel geometry may be a possible cause for misclassification as some of these channels had portions that were significantly angled from vertical. This could result in a short peak-to-zero crossing time characteristic of cloud flashes [e.g.,Fleenor et al., 2009]. Alternatively, short peak-to-zero crossing times for the subsequent events may be due to their connection with an elevated object [Diendorfer et al., 2010]. A quantitative evaluation of possible misclassifications causes will be conducted in the future.
 The comparative analysis between optical observations of upward flashes and NLDN data provides an opportunity to increase understanding of the triggering flash activity that is frequently responsible for upward leader initiation from tall objects. This comparison suggests that nearly all of the upward flashes from towers in Rapid City, South Dakota were triggered by nearby flashes (i.e., lightning-triggered upward lightning). The most common triggering flash type was the +CG flash, which preceded the initiation of upward leaders in 67% of the cases. For many cases, in-cloud visual brightness associated with leader development propagated from the triggering flash toward the towers prior to upward leader initiation. In the case of +CGs, this in-cloud leader development would be negative polarity and would cause a negative field change (atmospheric electricity sign convention) at the tower locations resulting in the initiation of upward propagating positive leaders. The presence of recoil leaders in 90% of the upward flashes and the fact that all NLDN-indicated subsequent events were negative polarity suggests that the upward leaders were positive. All upward leaders observed with high-speed cameras exhibited characteristics of positive leaders [Warner et al., 2012b], and for those cases analyzed with correlated electric field data obtained near the tower, a positive field change resulted during the development of the upward leader(s) [Warner, 2012; Warner et al., 2012a].
 The NLDN recorded subsequent −CG and/or −IC events at the tower locations for 44% of the upward flashes. When correlated with high-speed camera observations, these subsequent events were the result of connections made by bipolar/bidirectional recoil leaders forming on weakly luminous branches that became cutoff from the main upward channel. A connection made by a RLNE at the branch point of a brightly luminous main upward positive leader channel or at the tower tip in the case of weakly luminous main channel resulted in a luminosity pulse (i.e., IS pulse) and a corresponding electric field change that was detected by the NLDN. In only 8 out of 54 (15%) cases observed with high-speed cameras was there no apparent luminosity in the main channel when the inbound leader connected at the tower tip (dart leader/return stroke). Therefore, a majority of the NLDN-indicated “strokes” (−CGs and −ICs) occurred during the initial stage as IS pulses before complete current cutoff of the upward positive leader channel.
 Location accuracy for all the subsequent events was within expected values for the current NLDN configuration [Cummins and Murphy, 2009] (i.e., mean and GM, 246 m and 194 m respectively). The NLDN classified 46 (30%) of the subsequent events as intracloud flashes even though a connection was made with either a main luminous channel or the tower tip. A majority of the misclassified events had peak-to-zero times between 9 and 15μs suggesting further evaluation and adjustment of the cloud/ground classification algorithm for strokes associated with upward flashes.
 In contrast to observations in Rapid City, reports from other locations suggest that a significant fraction of observed upward lightning results from self-initiation [e.g.,Wang and Takagi, 2012; Zhou et al., 2012]. Wang and Takagi reported that 47% of the 53 observed upward flashes from a windmill and its protection tower between 2005 and 2010 where self-initiated and 53% (28) were other-triggered (i.e., lightning-triggered). They noted that the lightning-triggered flashes occurred during taller and more active storms, whereas the majority of self-initiated flashes took place when there was not significant lightning activity. They also noted that higher wind speeds or an operating windmill was more favorable for self-initiated flashes. This seems reasonable since a corona-produced screening layer present near the tip of the tower will act to inhibit upward leader initiation unless wind can effectively remove this layer. In the lightning-triggered flashes, wind should be less of a factor due to the high rate of change of the electric field compared to the production of space charge [e.g.,Chauzy and Rennela, 1985; Becerra et al., 2007], and Wang and Takagi's results showed that the lightning-triggered flashes did occur during lower wind speeds. Between 2004 and 2010, there have been no direct observations of upward lightning from the towers in Rapid City during the winter months.
 The storms in Rapid City tend to have cloud bases between 2–4 km, which may be an additional reason why self-initiation does not seem to occur with the towers. Cloud bases in winter storms in Japan and in Europe are reported to be much lower than those observed during summer, and frequently the clouds envelope the tower tips [Diendorfer et al., 2009; Wang and Takagi, 2012]. Furthermore, the towers in Rapid City have heights less than 200 m AGL. There is some apparent enhancement in the effective height due to the ridge upon which the towers sit, but the enhancement may not be significant enough to promote self-initiation as suggested byZhou et al. . Specifically the 100 m Gaisberg tower, which mostly experiences self-initiated flashes, sits on an isolated mountain that is 800 m taller than the surrounding terrain. However, a much taller tower height along with a high cloud base and an active summer storm may still favor lightning-triggered upward flashes. This appeared to be the case for the 553 m tall CN Tower in Toronto, Canada during an intense storm on 25 August 2011 UT. A leading-line/trailing stratiform region MCS passed over the tower and produced at least 34 upward flashes in under an hour and a half. Standard-speed video (30 ips) recorded approximate 3 km from the CN Tower and analyzed by the first author showed in-cloud brightening associated with preceding flash activity prior to upward leader initiation for all 34 upward flashes.
 We would like to thank Vaisala, Inc. for their contribution of NLDN data for this analysis. This study is part of a lightning program funded by the National Science Foundation (ATM - 0813672), and we thank Bradley F. Smull for his interest and support.