An analysis of the performance of classification methods applied to lightning images

The monitoring of lightning flashes recorded by video cameras can show the evolution of convective cells, allowing researchers to study the behavior of thunderstorms. The main goal of this work is to analyze the performance of four traditional image classification methods applied to lightning images. A system was implemented for this purpose, consisting of two modules: detection module and classification module. The detection module analyzes recorded videos and, using computer vision and image processing techniques, detects if there are significant differences between frames in a sequence, indicating that a lightning flash occurred. Whenever there is a significant difference between two frames, this module stores these images for posterior classification. The images obtained with the detection module were manually separated into five classes: background images, images of vertical lightning, images of horizontal lightning, images of other types of lightning and images of flashes. Background images do not show any evidence of lightning, the images of other lightning are those that have not fit as horizontal or vertical lightning and the flashes images are detected by sudden changes between frames, but do not show necessarily visible lightning. The five classes were then used to analyze the performance of four classification methods: the Nearest Neighbors (NN) algorithm, the K-Means Clustering (K-Means), Artificial Neural Networks (ANN) and Support Vector Machines (SVM). The Nearest Neighbors classifier was tested with the Euclidean and Correlation metrics, varying the number of neighbors from 1 to 4. The backpropagation algorithm was used to train a forward feed ANN composed of two hidden layers, with 20 nodes in the first hidden layer and 10 nodes in the second hidden layer. The SVM classifier used had a Multi Layer Perceptron (MLP) Kernel. The experiments consisted of comparisons among the classification methods using the five classes mentioned above, taken two by two, thus covering all possible binary classification tasks between them: background versus vertical, background versus horizontal, background versus others, background versus flashes, vertical versus horizontal, vertical versus others, vertical versus flashes, horizontal versus others, horizontal versus flashes and other versus flashes. The experiments showed that the Artificial Neural Network is the best classifier for most of the binary classification tasks and the Nearest Neighbors is the second best classifier for these comparisons. The classification task that was easier for all the algorithms is the separation between the background images and the images of flashes, because of the large difference between the images in these classes. On the other hand, the most difficult task is to separate horizontal images from other lightning images, which is difficult because the majority of the image in the class “others” is composed by horizontal lightning. The videos used in the experiments were acquired by seven video cameras installed in Sao Bernardo do Campo, Brazil, that continuously record lightning flashes events during the summer. The cameras were disposed in a 360º loop, recording all data with a time resolution of 33ms. In this period, several convective storms were recorded.