Abstract:

[Objective] To address the inefficiency and subjectivity of traditional manual identification for medium-scale traveling ionospheric disturbances (MSTIDs), this study proposes an automatic deep learning-based framework for MSTID detection and parameter extraction from OI airglow images based on the data collected by the all-sky airglow imager at the Meridian Project’s Xinglong station. [Methods] The framework adopts a three-stage architecture: (1) A custom 10-layer convolutional neural network (CNN) is designed to classify raw airglow images, filtering out environmental interference (e.g., clouds, twilight overexposure) and retaining suitable observational data captured at starring night; (2) A Faster Region-based Convolutional Neural Network (Faster R-CNN) model, trained via transfer learning and virtual data augmentation, achieves precise localization of MSTID waveforms; (3) Edge detection and linear fitting algorithms are developed to extract propagation parameters such as direction, velocity, and wavelength. Innovatively, a hybrid Poisson-Gaussian noise model and a waveform simulation function are proposed to generate synthetic training data, enhancing model robustness against limited real-world samples. [Results] Evaluated on the test dataset, the framework demonstrates high performance: the CNN classifier attains 96.9% accuracy in identifying clear-sky conditions; the Faster R-CNN detector achieves an average Intersection-over-Union (IoU) of >75% for wavefront localization. The proposed system significantly improves the automation and objectivity of MSTID analysis, enabling efficient large-scale statistical studies of ionospheric disturbances.

Key words: medium-scale traveling ionospheric disturbances, CNN, Faster R-CNN, all-sky airglow imager, synthetic image