Toward Automatic Detection of Pi2 Magnetic Pulsation Using Machine Learning

Magnetic pulsations of type Pi2 are a well-established category of Ultra Low Frequency (ULF) waves, characterized by irregularly damped oscillations with periods ranging from 40 to 150 seconds (6.7–25 mHz). Nowadays, it is well known that Pi2 occurs at the onset of geomagnetic substorms, is considered an outstanding research topic in space physics, and is a link between ionospheric and magnetospheric processes. The discontinuation of the conventional index previously employed to detect Pi2 pulsations has driven this study to propose an innovative detection method leveraging machine learning (ML). This paper introduces a novel ML-based classification framework that utilizes geomagnetic field data for Pi2 pulsation detection. A comprehensive analysis of various linear, ensemble, and non-linear ML models was conducted, employing hyperparameter optimization to identify the optimal model with high classification performance and minimal computational overhead during testing. Model robustness was assessed using multiple evaluation metrics, including accuracy, F1-score, kappa score, execution time, precision-recall curves, ROC curves, learning curves, and confusion matrices. The proposed gradient boost (GB) classifier demonstrated superior performance, achieving 98.21% accuracy in distinguishing Pi2 pulsations. This detection system offers a reliable and efficient tool for monitoring Pi2 pulsations in the nighttime, contributing to advancements in space weather analysis and substorm detection.