Quantification and identification of lightning damage in tropical forests

Abstract Accurate estimates of tree mortality are essential for the development of mechanistic forest dynamics models, and for estimating carbon storage and cycling. However, identifying agents of tree mortality is difficult and imprecise. Although lightning kills thousands of trees each year and is an important agent of mortality in some forests, the frequency and distribution of lightning‐caused tree death remain unknown for most forests. Moreover, because all evidence regarding the effects of lightning on trees is necessarily anecdotal and post hoc, rigorous tests of hypotheses regarding the ecological effects of lightning are impossible. We developed a combined electronic sensor/camera‐based system for the location and characterization of lightning strikes to the forest canopy in near real time and tested the system in the forest of Barro Colorado Island, Panama. Cameras mounted on towers provided continuous video recordings of the forest canopy that were analyzed to determine the locations of lightning strikes. We used a preliminary version of this system to record and locate 18 lightning strikes to the forest over a 3‐year period. Data from field surveys of known lightning strike locations (obtained from the camera system) enabled us to develop a protocol for reliable, ground‐based identification of suspected lightning damage to tropical trees. In all cases, lightning damage was relatively inconspicuous; it would have been overlooked by ground‐based observers having no knowledge of the event. We identified three types of evidence that can be used to consistently identify lightning strike damage in tropical forests: (1) localized and directionally biased branch mortality associated with flashover among tree and sapling crowns, (2) mortality of lianas or saplings near lianas, and (3) scorched or wilting epiphytic and hemiepiphytic plants. The longitudinal trunk scars that are typical of lightning‐damaged temperate trees were never observed in this study. Given the prevalence of communications towers worldwide, the lightning detection system described here could be implemented in diverse forest types. Data from multiple systems would provide an outstanding opportunity for comparative research on the ecological effects of lightning. Such comparative data are increasingly important given expected increases in lightning frequency with climatic change.


| INTRODUCTION
Humans have a long history of fear and fascination with lightning (Andrews, Cooper, Darveniza, & Mackerras, 1992;Botley, 1951;Bouquegneau & Rakov, 2010;Franklin, 1769), and scientists have been exploring the physics and biological effects of lightning for more than a century (e.g., Anonymous, 1898;Stone, 1903; also see Rakov & Uman, 2003). However, the specific role of lightning as an ecological disturbance remains one of its least-studied aspects. This is a particularly important knowledge gap in the tropics, where lightning frequency is relatively high, but the ecological effects of lightning strikes have never been quantified at large spatial scales in real time.
This gap in our understanding of the ecological importance of lightning mainly is attributed to two logistical obstacles. First, the spatial unpredictability and temporal unpredictability of lightning constrain documentation of the immediate ecological effects of lightning strikes to serendipitous, unreplicated observations (e.g., Furtado, 1935;Orville, 1968;Tutin, White, & Mackanga-Missandzou, 1996).
Second, whereas ground-based and satellite-based lightning flash detection systems generate accurate flash frequency data at regional scales (Albrecht, Goodman, Buechler, Blakeslee, & Christian, 2016;Boccippio, Cummins, Christian, & Goodman, 2001), they do not provide sufficient spatial accuracy to locate tree damage caused by individual flashes (Mäkelä, Mäkelä, Haapalainen, & Porjo, 2016). Here, we describe field-based methods that collectively enable the accurate quantification of lightning strike frequency, lightning characteristics, and lightning-caused tree mortality at the km 2 scale in near real time (i.e., within hours of a storm event). We focus on tropical forests, but the methods could be applied to any terrestrial ecosystem.
Discussions of lightning-caused ecological disturbance invariably focus on trees. Indeed, lightning strikes thousands of trees worldwide each year (Taylor, 1974) and is an important agent of tree mortality in some forests . Ecologists commonly assess lightning damage to trees via post hoc surveys (e.g., of lightning scars on trunks; Taylor, 1964Taylor, , 1965. However, this approach is inherently biased because lightning effects on trees are extremely variable-from catastrophic trunk shattering (Fernando, Mäkelä, & Cooray, 2010;Stone, 1916;Taylor, 1974) to no obvious damage (e.g., Orville, 1968).
Thus, a considerable fraction of lightning-caused tree damage and mortality likely either goes unnoticed or ultimately is not attributed to lightning. Finally, field-based forest surveys often are conducted at intervals >1 year (commonly 5 years), which greatly limits the accuracy of lightning damage assessments, especially given that lightning scars can become unrecognizable over time due to localized healing, decomposition, or secondary infections (SPY and EMG, pers. obs.).
Historically, post hoc lightning damage surveys have been most effective in temperate pine forests, in part because lightning damage to coniferous trees commonly appears as a conspicuous longitudinal stripe on the trunk Outcalt, 2008;Wadsworth, 1943). By contrast, the best available landscape-scale data for tropical forests are limited to surveys of conspicuous lightning gaps (Anderson, 1964;Brünig, 1964;Magnusson, Lima, & de Lima, 1996). These large group mortality events presumably result from the most intense lightning flashes (although this has never been verified) and represent a small fraction of the total number of strikes in a forest. Collectively, these limitations suggest that post hoc surveys significantly underestimate lightning-caused disturbance in forests, especially in the tropics.
Ecological assessments of lightning disturbance also are limited by a lack of information about the strike itself. Individual lightning flashes can have positive or negative polarity and differ substantially in intensity (measured as peak current) and duration (Rakov & Uman, 2003).
The return strokes (the familiar, visible portion of a lightning discharge) of cloud to ground (CG) and ground to cloud (GC) flashes cause injuries to trees. Many flashes have multiple return strokes, each lasting just tens of microseconds, but some "continuing current" (CC) strokes persist for hundreds of milliseconds and likely initiate forest fires (Bitzer, 2017;Fuquay, Taylor, Hawe, & Schmid, 1972). It is possible to measure the polarity, intensity, duration, and multistroke nature of a lightning flash (Bitzer et al., 2013). However, to our knowledge, only one study has attempted to associate such characteristics with the direct effects of lightning on trees (Mäkelä, Karvinen, Porjo, Mäkelä, & Tuomi, 2009); no similar studies exist for tropical forests.
The only solution to the problems summarized above is to locate lightning strikes and to measure their characteristics as they happen.
As noted above, satellite-based optical detection and land-based networks of electronic sensors can provide regional flash distribution data in near real time (Cummins & Murphy, 2009), but their limited spatial accuracy makes locating lightning damage difficult or impossible (Mäkelä et al., 2016). The methods described here overcome this problem and establish a basis for accurate quantification of lightning strikes at the landscape scale. This information, in combination with field-based assessment of tree condition, and knowledge of tree traits (e.g.,  and lightning discharge characteristics, can provide comprehensive assessment of the ecological effects of lightning in forests.
Accurately quantifying lightning-caused disturbance is important because the frequency of CG lightning is expected to increase with climatic change (e.g., Romps, Seeley, Vollaro, & Molinari, 2014;Williams, 2005). Specifically, models predict that for each 1°C increase in average surface temperature (or each doubling of atmospheric CO 2 concentration), CG lightning as a fraction of total lightning will increase by at least 10% (Price & Rind, 1994a, 1994bWilliams, 1992Williams, , 2005. In the wet tropics, this will likely occur via increased storm intensity, prolonged interstorm intervals, increased drying between rain events (Price, 2009), and increased lightning-initiated fires (which currently are very rare in lowland rainforests). Models further suggest that smoke from agriculture and lightning fires will increase storm intensity and thus lightning frequency, often at locations very distant from the source fire (Cochrane, 2003;Goldammer & Price, 1998;Price, 2009;Price & Rind, 1994b). There is growing evidence that such changes are already happening (Norris et al., 2016). Thus, accurately measuring lightning-caused tree mortality will be relevant to predicting future forest dynamics and structure under a changing climate.
The principal objective of this study was to establish a lightning monitoring network that would enable us to determine the locations of lightning strikes on a forest-wide scale (ca. 20 km 2 ), in near real time, and with high spatial accuracy (ca. 10 m). Two secondary objectives were to (1) demonstrate how ecologically relevant lightning flash characteristics (intensity, polarity, duration, and number of return strokes) can be quantified and (2) generate a standardized list of indicators that could be used to reliably identify lightning damage in the field post hoc.

Field work for this project was conducted in central Panama on Barro
Colorado Island (hereafter, BCI; 9.152°N, 79.846°W) and Gigante Peninsula (hereafter, Gigante; 9.128°N, 79.856°W). BCI is a 15-km 2 island administered by the Smithsonian Tropical Research Institute (STRI) and is one of the best-studied patches of tropical forest on earth (Croat, 1978;Leigh, Rand, & Windsor, 1996). The forests on BCI and Gigante are categorized as seasonally moist, with a well-defined wet season spanning May to December. Much of the rainfall in the area comes from storms associated with tropical low-pressure waves.
Many of the storms produce frequent lightning; BCI receives ca. 40 flashes km −2 year −1 and peak flash rates occur between mid-July and mid-August (from 1995 to 2012 satellite data; Christian et al., 2003). Roughly 25% of those flashes are CG lightning (Boccippio et al., 2001;Price & Rind, 1993) and thus are potentially damaging to trees. Consequently, BCI currently receives ca. 150 CG strikes per year. Some tropical lightning hotspots (e.g., the Congo, the Colombian Chocó, and Lake Maracaibo; Albrecht et al., 2016) have lightning flash frequencies ca. twice that of BCI. We focused on Panama (rather than a lightning hotspot) because it uniquely offers high lightning frequency in combination with political stability, excellent infrastructure, established long-term forest research plots (Hubbell & Foster, 1983), and easy access.

| Video camera network
Measuring CG lightning frequency, distribution, and damage to individual trees requires continuous monitoring of large areas of forest canopy at resolution of <20 m, which is impossible with traditional lightning quantification methods (Mäkelä et al., 2016;Stall, Cummins, Krider, & Cramer, 2009). An alternative approach-placing a large number of electronic sensors in a forest-would provide adequate spatial resolution, but also would be logistically difficult and very costly. We established a video-based lightning monitoring system on BCI in 2014 to overcome these problems. The system consisted of video surveillance cameras mounted on a series of preexisting guyed towers overlooking the forest canopy (Figures 1 and 2 Each video camera was fitted with a 6 mm f/1.2 lens and a 77 mm 3.0 neutral density filter and housed within a metal protective box ( Figure 1 and Table 1). The cameras operated at an interlaced shutter speed of 1/30 s with apertures ranging from f/4 to f/8, depending on local light conditions and camera orientation. Once installed, each camera recorded digital videos continuously to a DVR fitted with a 32 Gb SD memory card located in a weatherproof box at the base of each tower ( Figure 2 and Table 1). The camera and recording components on each tower were powered by a 12-V deep-cycle marine battery charged by two or more solar panels wired in a parallel circuit to generate up to 90 W under full sun exposure ( Figure 2). The preliminary system established in 2014 provided visual coverage of ca. 50% of the island by one camera and 15% of the island by two cameras (Figure 3).
The system was expanded to four cameras in 2015 and 2016 to provide approximately the same amount of spatial coverage (Figure 4), while improving directional diversity and removing the logistical constraint of frequent travel by boat to Gigante. Each camera recorded video for up to 8 days before the SD card approached capacity. The F I G U R E 1 A surveillance camera mounted on one of the towers used in the video network for lightning monitoring. A 3.0 neutral density filter is attached to the front of the camera, and both the power supply and video feed cables are fitted with in-line surge protection. The complete setup is protected within a metal box, which is open for the photograph F I G U R E 2 The solar panels (left) and electronic equipment (right) used to support individual cameras in the video network. The panels collectively provided up to 90 W of power to recharge a deepcycle 12-V battery (not shown) during the day, while also powering the camera, GPS antenna, text overlay unit, and DVR. These latter components were then powered by the battery through the night amount of data recorded per minute varied based on individual camera settings, so the SD cards were removed from the DVRs and replaced with blank cards every 3 or 4 days to avoid data loss.
We used frame-by-frame analyses of the videos from each camera to identify lightning strikes. These analyses were conducted automatically via an IDL program that compared the brightness of each frame to the previous frame on a pixel-by-pixel basis (Supporting Information).
Specifically, if more than one hundred pixels were brighter by a particular threshold (typically, ten bytes), then the frame was marked as potentially containing a lightning strike. Frames with brightness exceeding the threshold were saved and subsequently examined by the authors to verify the presence of lightning flashes. The SD cards collectively contained hundreds of Gb of digital video even after just 3 days, so we expedited the video processing by targeting time intervals surrounding known storm events. We identified these time intervals based on our own observations of electrical storms on BCI, and by polling other researchers residing at the station. We also used Panama Canal Authority (ACP) weather radar images to verify the timing of storm events. The radar images were captured automatically from the ACP website (http://www.pancanal.com/eng/radar/main.html) via a short Unix shell script. The program recorded the radar image every 5 min, and the subsequent image stack was aggregated, using IDL, into an MPEG-4 video loop spanning the previous 24 hr at 06:00 every day (Video S1, Supporting Information). Upon identifying the time frame associated with a specific storm event, we extracted the corresponding videos from the SD cards for frame-by-frame processing as described above.
We determined the approximate locations of lightning strikes by comparing images of the same flash recorded by two or more cameras mounted on different towers (Figures 5 and 6). The azimuth and field of view of each camera provided a reliable estimate of the strike location, which was then mapped on Google Earth (Figure 7) and verified by T A B L E 1 Sources of equipment used in the project The field of view of each of two cameras used in the first iteration of the camera network (2014). Circles = approximate locations of the two strikes shown in Figure 5 F I G U R E 4 The field of view of each of four cameras used in the second iteration of the camera network (2015)(2016). Under this arrangement, ca. 15% (225 ha) of BCI is viewed by at least two cameras hiking to the site. This approach enabled us to locate strike sites within a few days of the event. The successful implementation of the camerabased system provided the logistical foundation for the development of the more sophisticated and efficient ELS system described below.

| Electronic network
In addition to the video monitoring system described above, we will deploy two advanced electronic lightning sensors (ELSs)  the simulation showed that, once installed, the ELS system will yield <12 m error in strike locations over 95% of the study site ( Figure 9).

| Field assessment
We located, identified, and assessed the condition of struck trees and lianas in each strike location following established guidelines (USDA 1999). We recorded up to 25 different pieces of information about each strike location and sketched the spatial distribution of damage.
When necessary, we used the single-rope technique to climb nearby undamaged trees (Perry, 1978) to verify ground-based observations.
Trees not on established forest research plots were measured (as diameter at breast height; dbh) and identified from leaf samples. Each strike location was surveyed as soon as possible after the event, and again 3-4 months later to document mortality and the production of coarse woody debris (CWD). We quantified CWD volume at each strike location by tallying all dead trees >10 cm dbh, crown dieback, and dead liana stems >2 cm dbh. The amount of CWD attributable to crown dieback was estimated based on general tree architecture and associated patterns of allometry in crown area (Bohlman & O'Brien, 2006;Chave et al., 2014;Malhi, Baldocchi, & Jarvis, 1999;Montgomery & Chazdon, 2001;O'Brien, Hubbell, Spiro, Condit, & Foster, 1995). Although rough, these approximations provided an overall estimate of CWD production for each strike location, which will provide a basis for comparison with ongoing CWD inventories in the forest dynamics plots on BCI.
To generate a list of diagnostic cues for the post hoc identification of lightning damage, we qualitatively compared patterns of vegetation damage in strike sites with haphazardly selected control trees F I G U R E 5 Still video frames of two lightning strikes to BCI trees on 14 August 2014, 16:14 local time. The top images were captured by the camera on Gigante, and the bottom images are from the camera on BCI. The two images on the left are of the first flash (Example 1 in the text), and those on the right are of the second flash (Example 2). The field of view of each camera and the approximate strike locations for these two flashes are shown in Figure 3. All images were cropped for clarity (>10 cm dbh) located in patches of forest with no known recent lightning strikes. In total, we surveyed 5,000 control canopy and subcanopy trees distributed across BCI and Gigante. These surveys focused on flashover damage as the primary diagnostic cue (described below), because it was the only cue observable at every known strike site.

| Recorded lightning flashes
As of October 2016, the camera-based monitoring system recorded 91 lightning strikes to the BCI forest and surrounding mainland (e.g.,   Figures 6 and 7).

| Field assessment
Based on surveys of 20 strike sites, we generated a list of standardized diagnostic cues that can be used to identify lightning damage post hoc in tropical forests (Table 2). We also used observations from multiple visits to each site to approximate the time period when each lightning diagnostic cue was typically observable (Table 2). Finally, we measured the frequency of occurrence for each diagnostic cue (Table 2) using data only from strikes that were captured on two or more cameras.
The most consistent post hoc diagnostic cue of a lightning strike event was flashover damage (i.e., directionally biased vegetation mortality caused when electric current jumped from one branch to another across an air gap; Table 2). Whereas tree trunks exhibited no conspicuous damage such as lightning scars (Taylor, 1965) Comparison with dead and damaged vegetation in control surveys showed that the diagnostic characteristics described above are reliable for ground-based lightning damage identification. Specifically, of the 5,000 haphazardly selected canopy and subcanopy trees surveyed, damage resembling flashover among neighboring trees and other plants was exceedingly rare; we never observed evidence resembling flashover damage among more than three individual plants.
We therefore suggest that post hoc field identification of lightning in the absence of camera images requires at least four individual plants that exhibit flashover damage, preferably in combination with another diagnostic cue described above.

| DISCUSSION
Here, we describe a hybrid electronic sensor/camera-based system for locating lightning-caused disturbance in near real time, and we provide a recipe for the reliable post hoc identification of lightning damage to tropical trees, lianas, and other forest components. In concert with each other, these two methods enable the accurate quantification of lightning-caused tree damage and mortality in forests at large spatial scales. Such data will provide an unprecedented opportunity to assess the role of lightning as an ecological disturbance at multiple levels of biological organization (Table 3). The monitoring system described here could be established in any forest with a canopy that can be viewed from at least two emergent points (e.g., towers, buildings, F I G U R E 1 1 A Sterculia apetala tree crown with an Arrabidaea sp. liana 14 months before (top image, June 2014) and 3 days after (bottom image, August 2015) the lightning strike described as Example 2 in the text. The tree crown and the liana were badly damaged, but were still living 2.5 years poststrike. No evidence of this strike or the significant crown damage was visible from the ground F I G U R E 1 2 Ground-based image of the focal tree (Ficus obtusifolia) in the lightning strike described as Example 3 in the text. This photograph was taken 8 months after the strike. Note that many vascular epiphytes and epiphytic ferns are alive in the tree crown. By 12 months poststrike, all branches had fallen such that only the main trunk remained standing or exposed ridges). Given recent growth in abundance of communication towers in forests worldwide, this is fast becoming a relatively minor logistical hurdle.
Preliminary data from this system indicate that lightning is an important agent of tropical tree mortality; indeed, lightning damages and kills more trees in the BCI forest than we anticipated at the start of the project. Most importantly, that mortality is not obviously due to lightning and is unlikely to be attributed to lightning during normal forest surveys without prior knowledge of the event. As a case in point, Example 3 above was included in the 2015 survey of the 50 ha forest dynamics plot on BCI (Hubbell & Foster, 1983), but only the focal tree was dead at the time of the survey and the agent of mortality was not identified. Given rapid changes at this site over the 12 months poststrike, a 2016 plot survey likely would have recorded the agent of mortality for all trees killed at this site as blowdown.
Thus, it is clear that lightning damage is easily overlooked or misidentified in practice.
T A B L E 2 Diagnostic characteristics of lightning-caused tree damage in central Panama. N = number of known strike sites recorded on two or more cameras adjusted based on the presence of requisite organisms for each characteristic; Frequency = fraction of known strike sites exhibiting a given characteristic; First Observed = range of time postflash that each characteristic first becomes clearly evident; Persistence = approximate duration postflash that each characteristic remains observable used around the world to study forest ecology (e.g., Parker, Smith, & Hogan, 1992) likely have similar effects. The study site for this project has both high lightning frequency and a relatively high concentration of metal towers extending >10 m above the surrounding canopy.
Our observations are too preliminary to evaluate the effects of these towers on the frequency of lightning damage to nearby trees, but the monitoring system described in this study is generating data that eventually will resolve this issue.

| CONCLUSIONS
Here, we demonstrate a relatively simple and inexpensive system that provides accurate lightning strike location and flash characteristic information in real time (within minutes to hours) and over large spatial scales. Consequently, this system provides an unprecedented opportunity to track the effects of both conspicuous and inconspicuous lightning damage, thus enabling structured and unbiased tests of hypotheses related to the ecological effects of lightning for the first time.
Moreover, by pairing this system with a well-studied forest, such as a forest dynamics plot, it is possible to conduct long-term and detailed investigations using historic information about monitored individual trees. Because the ecological effects of lightning are largely unstudied, we expect that increased use of this and similar systems will lead to the initiation of many unanticipated avenues for future research.
T A B L E 3 Example research questions that could be answered at three different levels of organization given sufficient lightning strike location data at the landscape scale Individual & Population 1. Are some tree species more attractive to lightning than others?
2. Do some trees have traits that resist the damaging effects of lightning (Gora et al. unpub.)?
4. Is there a relationship between tree condition at the time of a strike and the amount of damage that occurs?
5. Does lightning-caused mortality mediate competition among trees?
Community 1. What is the role of lightning disturbance in gap-phase forest dynamics (Brokaw, 1985)?
2. Do lightning strikes modify the structure of local soil microbial communities?