Automated Plasma Region Classification and Boundary Layer Identification Using Machine Learning

Jiye Wang, Xuan Liu, Fei Dai, Rui Zheng, Yilong Han, Yan Wang, Andi Liu, Xinhua Wei, L. Q. Zhang, Hui Li, Chi Wang, Tieyan Wang, J. L. Burch, W. BaumjohannRemote 2025 · DOI

TL;DR

This paper introduces a novel machine learning model, the Wavelet Decision Tree Classifier (WDTC), that automates the identification of plasma regions and dynamic boundary layers (like the bow shock and magnetopause) in space weather data. By combining wavelet analysis for feature extraction with a decision tree for classification, the WDTC achieves high accuracy and provides a valuable tool for understanding complex space plasma interactions and for training future AI models.

What Problem Does This Paper Address?

Understanding the complex interactions between the solar wind and Earth's magnetosphere is crucial for space weather prediction. A key challenge in this field is accurately identifying distinct plasma regions (solar wind, magnetosheath, magnetosphere) and, more importantly, the dynamic boundary layers (bow shock, magnetopause) that separate them. These boundaries are where critical energy transfer and particle acceleration occur. Traditional methods often struggle with the dynamic and sometimes ambiguous nature of these transitions, leading to difficulties in automated classification and analysis, especially with the vast amounts of data generated by modern space missions like MMS.

Methods

The core innovation is the Wavelet Decision Tree Classifier (WDTC). This model takes a two-pronged approach:

  1. Wavelet Feature Extraction: Instead of using raw satellite data directly, the WDTC first applies a discrete wavelet transform (specifically, the db1 wavelet) to key plasma parameters: magnetic field strength (B), plasma density (n), velocity (V), and temperature (T). This process decomposes the signals into different frequency bands and extracts two features for each parameter: wavelet energy (representing signal amplitude strength) and wavelet entropy (representing signal complexity and irregularity). This results in an 8-dimensional feature vector.
  2. Decision Tree Classification: These wavelet-derived features are then fed into a decision tree classifier. The decision tree uses an entropy-based splitting criterion to recursively partition the data, effectively learning the most discriminative features for classifying the plasma into five categories: solar wind (SW), bow shock (BS), magnetosheath (MSH), magnetopause (MP), and magnetosphere (MSP).

The authors chose decision trees over neural networks for their superior interpretability and computational efficiency, especially given their moderate dataset size.

Datasets

The study utilizes data from the Magnetospheric Multiscale (MMS) mission. Specifically, it uses data collected from September 15 to December 31, 2015, focusing on dayside observations. The dataset includes magnetic field and plasma measurements (ion density, velocity, and temperature) at a uniform time resolution of 0.125 seconds, achieved through careful data alignment and interpolation. A manually labeled dataset of 500 samples (100 for each of the five categories) was created for training and validation.

Key Insights and Ideas

  • Wavelet Features are Key: The study highlights that wavelet-derived features (energy and entropy) are more effective than raw data for capturing the complex dynamics of plasma regions and boundary layers. These features encode information about signal strength and complexity across different time scales.
  • "n_Entropy" is Highly Informative: Feature importance analysis revealed that ion density entropy (n_Entropy) was the most significant feature for classification, followed by temperature wavelet energy (T_Wavelet_Energy). This suggests that the complexity of ion density fluctuations plays a crucial role in distinguishing plasma regions.
  • Unified Approach is Superior: The WDTC's integrated framework for both region classification and boundary layer identification is presented as a significant advantage over two-step methods (like GMM), which can be prone to error propagation and suboptimal coordination.
  • Wavelet Transform Outperforms FFT: A comparative analysis showed that the wavelet transform generally achieved better performance (higher accuracy, recall, and F1 scores) than the Fast Fourier Transform (FFT) for classifying plasma regions, especially in the magnetosheath and bow shock regions.

Key Results

  • The WDTC model achieved high classification performance across various plasma regions, with an average accuracy of 96.4% (±1.2% standard deviation) over ten experiments.
  • Boundary layer detection was particularly strong, with F1 scores of 0.95 for the bow shock and 0.94 for the magnetopause.
  • The model successfully processed 711 boundary layer crossings from the MMS dataset, identifying 295 bow shock events and 416 magnetopause crossings.
  • The magnetopause region achieved an exceptional AUC of 1.00 in ROC analysis, indicating perfect classification.
  • The WDTC model demonstrated robustness, with a Matthews correlation coefficient (MCC) of 0.96, indicating strong performance even with potential class imbalance.
  • The model was able to distinguish between "typical" and "ordinary" samples based on wave/turbulence activity, providing a more nuanced understanding.

Limitations and Future Work

  • Transient Event Challenges: The WDTC, like other methods, faces challenges in classifying specific transient plasma bubble events where their wavelet-transformed signatures closely resemble those of boundary layers. This can lead to occasional misclassifications.
  • Future Enhancements: The authors suggest integrating additional in situ measurements (beyond the four used) to further improve boundary layer identification. They also propose extending the WDTC model to analyze atmospheric environments, which also exhibit significant wave activity.
  • Data Completeness: The study notes that excluding data-gap-affected cases significantly enhances the WDTC's accuracy, implying that data quality is crucial for optimal performance.

Summary updated

Abstract

The accurate classification of plasma regions is a critical challenge in space science, with identifying dynamic boundary layers (BLs) being particularly complex. This study introduces a novel wavelet-decision tree classifier (WDTC) designed to automate BL detection. Unlike conventional machine learning methods that rely on raw satellite measurements, the WDTC utilizes processed parameters derived from wavelet analysis as inputs to the decision tree algorithm. For each in situ measurement, including magnetic field strength (B), plasma density (n), velocity (V), and temperature (T), the wavelet analysis generates two features: wavelet energy and wavelet entropy. This results in a total of eight input parameters (two for each of the four in situ measurements) for the decision tree. By incorporating these distinctive wavelet-derived features, the WDTC enhances its ability to accurately and efficiently identify BLs within complex plasma environments. The model was applied to data from the Magnetospheric Multiscale (MMS) mission, focusing on the dayside region, and successfully differentiated between the solar wind, bow shock, magnetosheath, magnetopause, and magnetosphere. From September 15 to December 31, 2015, the WDTC identified 711 BL crossings, including 295 bow shock events and 416 magnetopause crossings. Beyond its scientific applications, the WDTC provides high-quality training datasets and a reliable data labeling tool, contributing to neural network training efforts.
BibTeX
@inproceedings{wang2025automated,
  author = {Jiye Wang and Xuan Liu and Fei Dai and Rui Zheng and Yilong Han and Yan Wang and Andi Liu and Xinhua Wei and L. Q. Zhang and Hui Li and Chi Wang and Tieyan Wang and J. L. Burch and W. Baumjohann},
  title = {{Automated Plasma Region Classification and Boundary Layer Identification Using Machine Learning}},
  booktitle = {Remote},
  year = {2025},
  doi = {10.3390/rs17091565},
}
Figure