A Modified K-Means Method to Evaluate Location Accuracy of Lightning Networks Based on Video Data

Abstract:

The evaluation of a lightning location network should contain two basic information, the detection efficiency, which shows the percentage of total lightning actually observed, and the location accuracy, which is related to the uncertainty about the point of contact of the ground discharge. Direct and indirect methods are used for the assessment of location accuracy. Direct methods are based on accurate knowledge of the discharge location (e.g., tower or artificially triggered lightning), while indirect methods are based on other methods where location is not precisely known (camera analysis, network self-evaluation, comparison between networks, etc.).

In this work, we proposed a new method to evaluate the lightning location accuracy of BrasilDAT network using on a modified K-means method, using high-speed videos as base data. We refined a camera-based indirect method, previously used in the literature. The great advantage of this method is that it can be used to benchmark location accuracy anywhere without the need for special conditions. The lightning data were observed for three summer seasons (2012, 2013 and 2014), with high-speed cameras that filmed the phenomena at up to 2500 frames per second, synchronized by GPS and compared to data provided by BrasilDAT.

We developed a program that calculated the centroid of locations provided by BrasilDAT for those discharges that hit the same ground contact point. Confirmation that a group of discharges on the same ground contact point was given by the video data. The program still iteratively eliminated from the centroid calculation the locations whose error exceeded a certain threshold. An optimized centroid was calculated after a few iterations. Because this centroid represents the optimal location of the discharge, the error is measured between the location given by the network and the centroid.

To test the precision of the method, the location accuracy was measured for three CG flashes, and seven upward lightning flashes producing M-components and subsequent return strokes, with precise known locations. Those observations occurred in 2012, 100 km away from the rest of the data analyzed, so the results on the location errors should be similar.

More details on the method are presented on the paper, as well as other evaluations of BrasilDAT network.