Detection of short-term peaks in VLF/LF signals around the time of gamma-ray burst detections using deep learning

Gamma-ray bursts (GRBs) are recognized as the most energetic and luminous astrophysical phenomena whose high-energy photons can ionize atmospheric particles and thus affect electrical conductivity. Consequently, electromagnetic waves such as Very Low and Low Frequency (VLF/LF) radio signals used for the lower ionosphere monitoring can be used to detect disturbances, i.e. the impact of GRBs on the area in which these signals propagate. This paper addresses application of deep learning for automatic detection the short-term peaks in VLF/LF signals around the time of satellite registration of GRBs. We utilized a refined hybrid architecture combining Convolutional Neural Networks (CNN) for automatic feature extraction and Long Short-Term Memory network (LSTM) for modeling sequential dependencies in the multi-channel time-series data. We used derived, peak-based representations of the raw ionospheric waveforms and trained the proposed neural network architecture on such pre-processed data representations. The experiments with the sample of 54 short-lasting GRBs revealed limitations due to extreme class imbalance; however, a refined approach utilizing manual adjustment of positive-class weights and optimization of the classification threshold achieved the best detection performance. The final CNN-LSTM model demonstrated strong performance metrics, confirming the viability of CNN-LSTM approaches in this task.