Detection of Traveling Ionospheric Disturbances Using Computer Vision

The ionosphere, spanning approximately 50–1500 km in altitude, contains free electrons and ions whose distribution strongly influences radio-wave propagation and, consequently, Global Navigation Satellite System (GNSS) measurements. Variations in ionospheric total electron content (TEC) are dominated by diurnal cycles but can also exhibit traveling ionospheric disturbances (TIDs), often generated by acoustic-gravity waves. Recent research has increasingly focused on detecting such disturbances using machine learning and advanced signal processing techniques. Prior studies have applied random forests, gradient-boosted decision trees, convolutional neural networks, and recurrent neural networks to time series to classify co-seismic ionospheric disturbances (CIDs), identify tsunami-related signals, and detect anomalous ionospheric behavior in near-real time. Building on these developments, this study investigates the use of the Mask Region-based Convolutional Neural Network (Mask R-CNN) framework to detect TID signatures in distance–time plots derived from GNSS TEC observations from a ground based station network. This image based representation integrates spatial and temporal information, enabling the identification of characteristic disturbance patterns. By applying the method to datasets from three major earthquakes, we evaluate its effectiveness in identifying TIDs, potentially linked to seismic sources, and assess its suitability for future ionospheric monitoring. For the Tōhoku and Kaikōura data, the model performs well, achieving Area Under the Precision-Recall Curve (PR-AUC) values of 0.87 and 0.83, respectively. However, for the Illapel data, which is characterized by a substantially sparser station network, the model fails, yielding a PR-AUC of 0.06. These results demonstrate that disturbance patterns can be successfully detected using distance–time plots, but also highlight the need for a sufficiently dense station network for reliable identification. In addition, we propose a Radon-transform-based method for estimating the slope of each detected wave pattern and inferring its propagation velocity. We validate this method against manually traced wavefronts on the Kaikōura test set, obtaining a median absolute error of 0.33 km/s across 45 detections.