Toward an Interpretable CNN Model for the Classification of Lightning‐Produced VLF/LF Signals.

An interpretable convolutional neural network model is proposed for the classification of very low frequency and low frequency lightning electric field waveforms. This model adopts multi‐scale convolutional kernels and shortcut connections to enhance the ability of lightning waveform classification. Based on the data recorded from five provinces in China, the proposed model achieves an accuracy of 98.56% for a four‐type classification task including return strokes, the intra‐cloud lightning, preliminary breakdown, and narrow bipolar events. The proposed model is validated with another open‐source data set from Argentina with an accuracy of 98.45%, which shows good robustness. To ensure the classification, the features learned by the model are visualized. The class activation mapping (CAM) method is adopted to visualize the class‐specific contribution of different waveform parts by using the feature maps of the final convolutional layer. It is highlighted by the CAM method that the proposed model focuses on waveform parts that align with those areas of interests identified by human experts. The high‐contribution waveform parts are furtherly analyzed, which indicate that the proposed model possesses the capability to associate waveform features with the corresponding lightning discharge processes. Plain Language Summary: Electromagnetic waveforms in very low frequency and low frequency bands are usually used to detect and locate lightning activities. Traditional waveform classification methods have difficulties in distinguishing multiple types of lightning waveforms. Although machine learning models have great potential in multi‐type waveform classification tasks, these models rely on the features proposed by human experts and cannot capture the features of different scales in lightning waveforms. To this end, this paper proposes an improved convolution neural network model, which incorporates modifications to the model structure to better suit the lightning waveform classification task. The data set for model training comes from five provinces in China and contains different meteorological conditions. The proposed model achieves a classification accuracy of 98.56% on this data set and 98.45% on an open‐source data set from Argentina. Meanwhile, the classification process is interpretable by visualizing the convolution outputs. The analysis of the visualization results shows that the high performance of the proposed model is reliable for its ability to focus on waveform parts that align with areas of interests identified by human experts. A closer inspection of these waveform parts suggest that the proposed model possesses the capability to associate waveform features with the corresponding lightning discharge processes. Key Points: The proposed model achieves an accuracy of 98.56% for a four‐type lightning waveform classification task and shows good robustnessThe model is interpretable by visualizing the contribution of different waveform parts to the classification resultsThe proposed model aligns with human‐expert classifications and suggests potential to link waveform features with lightning processes [ABSTRACT FROM AUTHOR]