CVAE-GAN & STFT-CNN for low resolution geomagnetic data reconstruction in Pi2 pulsation identification

Identifying short-period irregular (Pi2) geomagnetic pulsations is crucial for space weather monitoring, as these signals often mark substorm activity that can disrupt satellites, navigation and power systems. An accurate Pi2 detection commonly requires high-resolution (1-s) geomagnetic data to capture key signal features. However, long-term high-resolution records are often limited—especially in older archives due to storage and operational constraints. This study proposes a hybrid Conditional Variational Autoencoder–Generative Adversarial Network (cVAE-GAN) framework to reconstruct 1-s geomagnetic data from widely available 1-min samples. We evaluate the reconstructed signals using a Short-Time Fourier Transform–based Convolutional Neural Network (STFT-CNN) for Pi2 pulsation identification. Experiments are conducted using H-component data from six INTERMAGNET stations (2012–2019), selected based on the Wp index. The proposed method yields higher fidelity reconstructions (MSE: 0.1814, RMSE: 0.4258) compared to baseline traditional models and improves Pi2 classification accuracy to 91.38 % surpassing both original 1-s data (83.64 %) and downsampled 1-min data (54.94 %). Importantly, our results show that the model can reconstruct high-resolution features that support improved Pi2 detection from low-resolution data. However, the framework generates statistically plausible, high-resolution signals conditioned on the low-resolution input and guided by events labels identified from high-resolution(1-s) data. While this approach allows the model to recover features that improve automated Pi2 detection, it does not physically restore frequency components that are irretrievably lost due to downsampling. Instead, the reconstructed signals should be interpreted as model-informed surrogates, shaped by both the observed trends and the high-resolution-derived event labels, rather than literal physical recovery of all original signal content. This demonstrates the potential of generative models for analyzing historical low-resolution data and advancing automated space weather event detection while acknowledging clear limitations in physical signal reconstruction.

Text 123342.pdf - Published Version Restricted to Repository staff only Download (2MB) | Request a copy Official URL or Download Paper: https://www.sciencedirect.com/science/article/pii/...

Actions (login required)

View Item