Plasma bubbles in the ionosphere over equatorial and low-latitude regions are a common nighttime plasma density depletion structure that can cause severe scintillation interference on high-frequency communications, making them an important target of space weather monitoring. Traditional detection methods rely on statistical thresholds derived from temporal fluctuations in ion density; however, plasma bubbles exhibit significant variability in spatiotemporal scales (ranging from kilometers to hundreds of kilometers), and their occurrence frequency is strongly modulated by solar activity and seasonal factors. Consequently, fixed-threshold methods struggle to accommodate this diversity, with miss rates reaching up to 40% during periods of low solar activity. Although deep learning has demonstrated powerful capabilities in automatically extracting and recognizing complex waveform features in fields such as medical image segmentation and signal peak detection, its systematic application to ionospheric plasma bubble detection—and thus addressing the adaptability deficiencies of traditional threshold methods—remains underexplored.
In a recent study published in Space: Science & Technology, a research team from Huazhong Agricultural University proposed an automatic plasma bubble detection model based on deep learning, termed IncResUnet. Using ion density data from the FORMOSAT-1 satellite spanning 1999 to 2004, combined with expert annotations, the study established a dataset containing 13,675 plasma bubble events and defined a new identification criterion: density deviation from background trends exceeding twice the standard deviation, amplitude ≥ 0.2, duration ≥ 5 seconds, and transverse scale ≤ 500 km. Built upon the U-Net architecture, the model incorporates Inception-residual modules, which employ parallel multi-scale convolution kernels to capture density variation features across different temporal spans, while residual connections alleviate the gradient vanishing problem. Experimental results demonstrate that IncResUnet achieves an F1-score of 0.914, recall of 0.958, and precision of 0.874 on the FORMOSAT-1 test set, significantly outperforming the traditional threshold method as well as various U-Net variants such as U-Net++ and Attention U-Net. When the model trained on FORMOSAT-1 data was directly applied to C/NOFS satellite data, it achieved an F1-score of 0.804, which was further improved to 0.859 after a small amount of incremental training, with a recall rate as high as 98.8%. Applied to COSMIC-2 data, the model successfully detected all 36 events. This study represents the first systematic application of deep learning methods to automatic plasma bubble detection, validating the feasibility of learning density depletion waveform features from time-series data, and provides an efficient automated detection tool for space weather monitoring and ionospheric physics research.
First, this paper focuses on the engineering requirements for ionospheric plasma bubble detection and the limitations of traditional methods. Plasma bubbles are a common nighttime plasma density depletion structure over equatorial and low-latitude regions, capable of causing severe scintillation interference on high-frequency communications, and thus represent an important target of space weather monitoring. The ion trap sensor aboard the FORMOSAT-1 satellite provides high-resolution ion density data. Traditional detection methods apply linear detrending to 10-second data segments and compute density fluctuation values, identifying a plasma bubble event when the fluctuation exceeds a fixed threshold. However, plasma bubbles exhibit significant variability in spatiotemporal scales, ranging from kilometers to hundreds of kilometers, and their occurrence frequency is strongly modulated by solar activity, season, and local time. The fixed-threshold approach suffers from a miss rate of up to 40% during periods of low solar activity. As illustrated in Fig. 1, plasma bubbles manifest as sudden density drop features in ion density time series, which bear similarities to signal peak detection problems, thereby providing a rationale for introducing deep learning methods.
Second, the paper presents a complete technical pipeline for automatic plasma bubble detection. Fig. 2 illustrates the full workflow, ranging from data annotation to model output and post-processing. The research team first established a dataset containing 13,675 plasma bubble events, based on ion density data from the FORMOSAT-1 satellite spanning 1999 to 2004, in conjunction with detection results from traditional methods and expert manual annotations. As shown in Fig. 3, the proposed IncResUnet model incorporates Inception-residual modules into the U-Net architecture. The encoder comprises five groups of convolutional modules and Inception-residual modules, progressively extracting multi-scale features through four max-pooling operations; the decoder restores the original resolution through four upsampling operations. The Inception-residual modules employ parallel one-dimensional convolution kernels of multiple sizes to capture density variation features across different temporal spans, and residual connections concatenate the processed features with the original input, effectively alleviating the vanishing gradient problem. Training adopts a time-split strategy, with data from 2001 to 2003 used for training, 1999 for validation, and 2000 and 2004 for testing, while random undersampling is applied to balance the positive and negative sample ratios. After model output, a dual-threshold refinement strategy is applied, in which events with duration shorter than 5 seconds are filtered out as noise, and adjacent events separated by less than 1 minute are automatically merged into complete large-scale plasma bubble structures.
Finally, the paper comprehensively validates the effectiveness of the IncResUnet model through a series of experiments. Table 1 compares the detection performance of the traditional method, ResNet-18, LSTM, U-Net, and various U-Net variants. On the 2000 test set, IncResUnet achieves an F1-score of 0.914, a recall of 0.958, and a precision of 0.874; on the 2004 test set, it attains an F1-score of 0.914, a recall of 0.924, and a precision of 0.905, outperforming all comparison models in terms of recall. Fig. 4 presents four typical detection examples, where blue and orange denote the detection results of the proposed method and the baseline method, respectively. It can be observed that the traditional method suffers from missed detections and false alarms, whereas IncResUnet accurately identifies the onset and offset boundaries of plasma bubbles. In the robustness experiments summarized in Table 2, IncResUnet achieves the highest F1-scores (92.5% and 89.0%) across different data partitions. In the cross-satellite generalization validation, Table 3 shows that the model trained on FORMOSAT-1 data, when directly applied to C/NOFS data, yields an F1-score of 0.804, which can be improved to 0.859 after incremental training, with a recall rate as high as 98.8%; when applied to COSMIC-2 data, all 36 events are successfully detected, achieving a recall of 100%. The study further reveals that the generalization performance discrepancy across different satellite datasets primarily stems from the degree of data distribution shift. This work represents the first systematic application of deep learning methods to ionospheric plasma bubble detection, demonstrating the feasibility of learning density depletion waveform features from time-series data, and provides an efficient automated detection tool for ionospheric physics research and space weather monitoring.
