Dataset of photographed lightning events attaching to and around the Brixton tower, Johannesburg, South Africa for the 2015-2016 thunderstorm season.

This data article describes a dataset of videos of lightning flashes to and around a tall tower (the Brixton tower) in Johannesburg, South Africa. The videos were collected during the 2015-2016 South African thunderstorm season and a total of 3623 .mp4 videos are available in the dataset. Three different cameras were used, two in a similar location and the third at a different location giving a 90 degree perspective. Each video is timestamped and labelled depending on the type of event seen (attachment to the tower, nearby the tower, far from the tower, intracloud etc.). This dataset provides ground-truth, timestamped evidence of lightning events a known location and of differing types and can benefit atmospheric research scientists as well as lightning detection operators, particularly with regards to evaluating detection networks operating in the area. As the dataset contains a significant number of labelled videos, it also of use to pattern or image recognition supervised machine learning techniques and researchers.


a b s t r a c t
This data article describes a dataset of videos of lightning flashes to and around a tall tower (the Brixton tower) in Johannesburg, South Africa. The videos were collected during the 2015-2016 South African thunderstorm season and a total of 3623 .mp4 videos are available in the dataset. Three different cameras were used, two in a similar location and the third at a different location giving a 90 degree perspective. Each video is timestamped and labelled depending on the type of event seen (attachment to the tower, nearby the tower, far from the tower, intracloud etc.). This dataset provides ground-truth, timestamped evidence of lightning events a known location and of differing types and can benefit atmospheric research scientists as well as lightning detection operators, particularly with regards to evaluating detection networks operating in the area. As the dataset contains a significant number of labelled videos, it also of use to pattern or image recognition supervised machine learning techniques and researchers.

Value of the Data
• This dataset provides ground-truth evidence of the occurrence of lightning events or flashes with a known timestamp, allowing the data to be compared with other measurements; • The dataset also provides ground-truth evidence of lightning to a known location (the Brixton tower, Johannesburg, South Africa) as well as evidence that a known location was NOT struck by lightning. Ground-truth evidence of different types of lightning (cloud-to-cloud, upward, downward) is also provided; • This dataset can benefit atmospheric scientists, lightning protection engineers, climatologists and meteorologists, lightning detection network operators, pattern/image recognition and machine learning experts; • This dataset can be time-correlated with other measurements (electric and magnetic-field measurements) made within the area -particularly by lightning location systems -to provide visual evidence of detections and evaluate such systems performance; • The dataset also provides a significantly sized labelled dataset, allowing for the training and testing of supervised machine learning image recognition algorithms;

Data Description
The dataset described in this article involves lightning events in Johannesburg, South Africa -specifically, to and around the Brixton (also known as the Sentech tower), a tall communications tower just outside the Johannesburg city for the 2015-2016 thunderstorm season [1] . The data files describing these events are: 3623 .mp4 videos of lightning events, Table 1 . Additionally, to the timestamp, each video is labelled indicating the type of lightning captured in the video. Each file name follows the format YYYY-MM-DD_HH_MM_SS_label_CamX.mp4. Table 1 shows the description and an example (figure 2a -g) of each label category, as well as the total number videos of this type. Not all Cam 3 files are labelled.
Between Cam 1, 2 and 3 a total of 55 videos of lightning events attaching to the Brixton tower were captured. Table 2 lists these videos and both the upward and downward type are indicated along with the date and timestamp (in both SAST and Universal Coordinated Time UTC). This dataset is used in the associated publication by Hunt et al. [2] .

Experimental Design, Materials, and Methods
Johannesburg is the main economic city in South Africa and is located in the Gauteng province in the North of the country. The city itself has two tall communications towers -the Hillbrow tower and the Brixton/Sentech tower. The Brixton/Sentech tower is frequently struck by lightning and is located at 26.1925 °South and 28.0068 °East.   Fig. 4 shows one of the flashes to Brixton/Sentech tower as captured by all 3 cameras. Camera 1 and 2 have a very similar image with a slight perspective shift, but camera 3 shows a significantly different perspective.

Materials
The cameras used were simple webcams and the motion-triggering was performed using iSpy® motion-capture software. The cameras could operate from 5-30 frames-per-second but typically operated between 5-10 frames-per-second (with an average of 7.25 fps) and therefore had an approximate resolution of 60-100 ms and were able to capture numerous images of a lightning flash. However, the resolution was not enough to distinguish the number of individual strokes that constitute a flash. The camera was time-synchronised to the Network Time Protocol (NTP) server of the University of the Witwatersrand.

Methods
During the season, the cameras were checked for captures everyday. Any video captures that were unrelated to lightning (birds flying past the camera, changes in light or cloud cover etc.) were deleted to maintain space while any evidence of thunderstorm activity was kept. Subsequent to the 2015-2016 thunderstorm season, the videos were manually watched and labelled based on the categories described in table 1 .
The following criteria is to be used when labelling such datasets: • towerupward Upward flashes initiate from tall structures and exhibit branching initiating at the tower tip and splitting into the clouds [3] . Therefore, 2 criteria must be met for this label: 1. Videos where lightning can clearly be seen attaching to a tall structure (ie. the bright channel ends at the top of the tall tower). 2. Branching (multiple channels) must begin at the top of the tower and spread outwards towards the clouds (indicated in fig. 5 ).  • towerdownward : Downward flashes initiate from the clouds, with multiple channels branching from the clouds until one attaches [3] . Therefore, this label has 2 criteria: 1. Videos where lightning can clearly be seen attaching to a tall structure (ie. the bright channel ends at the top of the tall tower). 2. Branching (multiple channels) must begin in the clouds and spread outwards towards the ground, with only one of the branches connecting to the tower (indicated in fig. 5 ). • close : Lightning events that do not attach to the tower are downward events (only tall towers initiate upward events) and have channels that extend from the clouds to the ground. The full channel can be seen and if the events were close-by the tower, this should occupy the whole camera frame height. The criteria for this label is therefore: 1. Videos of lightning events attaching to the ground and NOT the tower. 2. Videos where the channel extends for 90% of the frame height.
• far : Lightning events that do not attach to the tower are downward events (only tall towers initiate upward events) and have channels that extend from the clouds to the ground. The full channel can be seen and if the events were far from the tower, the channel should be short in length. The criteria for this label is therefore: 1. Videos of lightning events attaching to the ground and NOT the tower. 2. Videos where the channel is less than 50% of the frame height. • unclear : The video is sometimes partially saturated due to brightness of the flash (or rain may obscure clarity) and it is not possible to see the channel. Here, it is not known whether the flash was to ground, upward or in the clouds. Therefore, the criteria for this label is: 1. Video captures a bright flash, saturating at least 25% of the frame. 2. No discernible channel can be seen. • unclearbehind : Lightning events that occur behind the camera are not captured, but the light is reflected in the clouds and this can be sufficient to trigger the cameras. The criteria for these is: 1. Videos where no channel is visible.
2. Videos where light reflections can be seen in the clouds. • CC : Intra-cloud lightning events occur where lightning discharges occur between clouds [3] .
In this case, channels can be seen clear but are not vertical and do not extend to the ground. Therefore, the label criteria is: 1 Videos where a discernible channel can be seen, but is not vertical and does not attach to the ground.