Machine Learning-Based VHF Lightning Radiation Sources Identification

Abstract

Identifying the validity of the location result is an important step in lightning radiation source mapping, which can eliminate the interference of noise location results, retain the real radiation source, and obtain a clear and continuous lightning channel development map. The localization methods, such as electromagnetic time reversal and multiple signal classification have high location accuracy, but the validity identification of their location result depends on the subjectively set threshold, which makes it hard to accurately distinguish the location results of weak radiation source and noise. In order to retain the weak radiation sources as much as possible and eliminate the noise interference, this article proposes two machine learning-based validity identification methods, namely, the continuous wavelet transform-based convolutional neural network model (CWT-CNN) and the spatiotemporal clustering algorithm. The CWT-CNN model can learn the time–frequency characteristics of the sliding window data to identify the lightning radiation source in advance and only retain the data containing useful signals. The spatiotemporal clustering algorithm can adaptively adjust the clustering parameters by learning the spatial and temporal distribution properties of the known location results to restore weak radiation sources that were incorrectly eliminated by former criteria. Experiments and analysis show that compared with the previous validity identification methods, the two methods proposed in this article are good at separating location results of weak radiation source from noise points, can obtain more continuous lightning maps without noise interference, and find some additional lightning branches.