Evaluation of EUCLID IC/CG classification performance based on ground-truth data

Abstract

This work compares the classification accuracy (CA) of two algorithms applied to data from the EUCLID lightning location system (lLS). As CA we call the accuracy of a LLS to correctly distinguish between cloud-to-ground (CG) and intra cloud (IC) discharges. The ground-truth data, used for this evaluation, was taken from optical and electric field data measured in various regions in Austria (2012 and 2015) and France (2014). The data set contains CG and IC discharges of positive and negative polarity. The data set was split up into further sub-categories as long as the number of data was still sufficient to give reasonable results. For a coarse overview of the algorithm performances, the total CA was first calculated for each year and country for both polarities. Furthermore, for the class of CG discharges, the CA of first return strokes, the CA of subsequent return strokes (with and without respect to the polarity) as well as the CA of IC events with respect to their polarity was evaluated. Specifically the subdivision into classes of amplitudes of the peak currents for different events can give further insight to the performance of the algorithms. For that reason the total CA, the CA of negative and positive events and CG and IC was analyzed. The evaluation shows that in combination with the new sensor data format LS the new algorithm exhibits an improvement of 2% at the CA. In combination with the old data format IMPACT, the new algorithm performs worse than the old one. In total, IC events have been classified much better by the new algorithm, irrespective of the sensor data format. CG discharges on the other hand show a worse CA throughout all years for the new algorithm.