A Deep Learning Approach for Removing Multi-Source Transient Interference in Satellite Magnetic Field Measurement

Magnetic field measurements are essential for space science missions but are often contaminated by transient stray fields from spacecraft subsystems such as electrical and control units. Traditional mitigation approaches-including strict magnetic cleanliness programs, deployable long booms, and dual-sensor gradient systems-suffer from inherent limitations, as cleanliness programs and long booms impose high cost and system complexity. To overcome these challenges, we propose Multi-Source Adaptive Gradiometry (MSAG), an enhanced gradiometry technique that integrates a neural network-based interference classification framework. The trained network identifies interference types and applies adaptive correction coefficients, enabling accurate multi-source disturbance correction without requiring manual segmentation of long-duration data. We validate MSAG using realistic synthetic data, generated by superimposing key transient interference types-modeled from SMILE ground tests-onto actual THEMIS satellite measurements, and test it through 220 Monte Carlo simulations. MSAG reduces the median RMSE from 0.457 nT to 0.014 nT and achieves a median correlation coefficient of 0.999994 with the ground truth. This improved accuracy could alleviate constraints on magnetic cleanliness and boom length in future missions, highlighting the advantage of MSAG over conventional methods and underscoring the potential of combining machine learning with gradiometry for high-fidelity magnetic field recovery.