© The Authors 2024

1 Introduction

A solar radio burst (SRB) refers to the enhanced radiation flux from the Sun in the radio wavelength band (Gary & Keller 2004; Aschwanden 2004). Spectrum observation is one of the earliest methods employed to study SRBs and a common technique for observing solar radio emissions (McLean & Labrum 1985; Melrose 1980). A radio spectrometer is used to map electromagnetic radiation flux signals into a frequency-time distribution, resulting in a dynamic spectrum (McLean & Labrum 1985). The horizontal axis typically represents time in a dynamic spectrum, and the vertical axis represents frequency. SRBs are classified into types I–V based on their morphological characteristics, including duration, bandwidth, frequency drift rate, and degree of aggregation within the spectrum(Marassi & Monstein 2022). Different radio burst types correspond to different excitation source properties, background conditions, and radiation mechanisms (Aschwanden 2004); among them, radio bursts of types II and III are the most common.

Type II radio bursts are mainly characterized by narrowband signals with a slow frequency drift in the solar radio spectrum (Kahler 1992; Minta et al. 2023). A distinctive feature of type II radio bursts is the presence of fundamental frequencies and harmonic structures (Feng et al. 2018). A comprehensive analysis that includes a detailed comparison of the source size, brightness temperature, fundamental frequency, and harmonic characteristics of type II bursts, along with the investigation of fine structures, holds a unique significance in identifying and understanding the solar eruption-driven shock waves and the acceleration of electrons by the shock waves (Patel et al. 2021). Leveraging high-resolution radio data acquired from various solar radio dynamic spectrometers and imaging instruments operating at different frequency bands enables a more in-depth analysis of type II bursts and their fine structures (Holman & Pesses 1983; Magdalenic et al. 2020). This type of analysis allows the diagnosis of coronal parameters, such as electron number density and magnetic field strength. It facilitates the determination of the velocity and configuration of associated shock waves (Kumari et al. 2019). This allows for a deeper understanding of solar eruptions and radiation processes (Cliver et al. 1999). Nevertheless, numerous unresolved questions persist concerning the origin of fine structures in type II bursts, emphasizing the need for continued research.

Type III solar radio bursts are characterized as rapid drift ranging from several hundred to tens of KHz in dynamic spectra, manifested as an isolated or sequential burst. They are the most common and intense radio burst emissions during solar eruptions (Kane 1981; Dulk et al. 1987; Yoon et al. 2002; Cairns et al. 2018). It is generally accepted that type III radio bursts are generated by plasma emission and appear around the fundamental and second harmonics of local source Langmuir frequency (Lin et al. 1973; Dulk & Suzuki 1980). The study of type III bursts can offer valuable insights into the dynamic processes of solar eruption, for instance, particle acceleration and transport as well as coronal magnetic morphology, etc. (Reid & Ratcliffe 2014; Hamini et al. 2021).

Solar radio observations date back to 1950 when radio spectrometers were first employed by Wild & McCready (1950). Recent advancements have developed observation systems with high temporal, spatial, and frequency resolution capabilities that generate substantial sets of high-quality data (Bastian & Gary 2005). Accumulating large amounts of high-quality data is instrumental in bolstering the reliability of statistical analysis outcomes and propelling research on SRBs. In this context, automated event identification and feature detection methods are pivotal in scientific research because they efficiently handle extensive data volumes. Furthermore, we can optimize the time resolution of solar radio observing instruments based on the predicted results of SRBs. By adjusting the data acquisition time resolution according to the varying predicted data volumes, we have ensured the acquisition of SRB data at a high resolution, while reducing the storage requirements for non-SRB data.

In solar radio physics, Lobzin et al. (2009, 2010) developed a radio burst identification system known as Automated Radio Burst Identification System (ARBIS) to detect type II and type III solar bursts automatically. The underlying principle of ARBIS involves employing Radon and Hough transforms to identify linear structures within the spectrum. The results demonstrated that this image-processing algorithm achieved high recognition rates, while maintaining low false alarm rates for automated identification. Gu et al. (2015) applied principal component analysis to reduce the dimensionality of radio spectra. These authors then utilized support vector machines to classify the reduced data automatically. Zhang et al. (2018) designed an event recognition and analysis system that leverages the Hough transform as its foundation. It can automatically detect type III SRBs in Nançay Decameter Array data and extract vital physical information, including start and stop frequencies and drift lines. Salmane et al. (2018) employed channel removal to mitigate interference signals in the spectrum. They introduced a statistical feature-based classification system by applying signal smoothing to a median filter.

As machine learning techniques progress, advances have been made in applying neural networks to SRB classification and identification methods. For example, Xu et al. (2019) developed a multimodal deep learning network and a long short-term memory network to classify SRBs using the solar radio spectrum obtained using a solar broadband radio spectrometer (SBRS). The field of solar radio has been undergoing an evident rise in the introduction and use of automated recognition and intelligent processing methods.

Regarding the classification and localization of solar eruptions, Hou et al. (2020) investigated enhancements to the original faster region-based convolutional neural network (Faster R-CNN) model. These authors proposed a multiscale detection framework and a multilayer feature fusion training method, which substantially improved the accuracy of small-scale eruption event detection. Scully et al. (2023b,a) achieved the successful automatic detection and classification of type III bursts by employing a coherent deep learning model that combines generative adversarial networks (GANs) with the you only look once (YOLO) algorithm.

These algorithms employ deep convolutional neural networks to extract features from images and subsequently classify and locate targets based on candidate regions. Among these methods, YOLO is the only algorithm capable of achieving high accuracy and real-time detections based on the dataset, even though the other listed methods have also succeeded in object detections. However, its application for SRBs detection is limited, leaving room for improvement in terms of its accuracy. Therefore, this study aims to apply the YOLO algorithm to detect solar radio spectrum bursts, accomplishing detections and localizations, as well as extracting pertinent physical information from these events. YOLOv7 (Wang et al. 2022), which is known for its real-time object detection capabilities, showcases impressive speed and high accuracy. Its anchor-based detection methodology effectively predicts bounding boxes and class probabilities across multiple grids, rendering it well-suited for detecting SRB events.

The primary objective of this study is to enhance the YOLOv7 algorithm for the precise detection of type II and type III bursts. This improved algorithm successfully extracts essential burst information, including the start and end times, start and end frequencies, frequency drift rates, and relative bandwidths. Furthermore, this method exhibits transfer ability and can be effectively applied to process data from various radio observatories. In response to the challenges encountered in SRB detection, this study employs clustering and additional techniques to optimize the enhanced YOLOv7 model. In complex backgrounds, this model has demonstrated the ability to detect SRBs accurately, achieving an average precision rate of 73.5%, a notable improvement over other existing models. The specific elements encompass the generation of anchor boxes tailored for type II and type III bursts and the application of diverse data augmentation techniques to mitigate the challenge of imbalanced samples.

In summary, the primary objective of this study is to realize efficient and precise automatic detection and information acquisition of type II and type III SRBs utilizing the YOLOv7 model. To accomplish this objective, our research progresses methodically through various phases, encompassing dataset construction, model optimization, SRB information extraction, and statistical analyses. Section 2 introduces the dataset used in the algorithm, including data sources, characteristics, and preprocessing methods. Section 3 describes the proposed SRB detection method based on YOLOv7, including model architecture, evaluation metrics, and experimental results. In Sect. 4, we demonstrate the application of the proposed method in detecting SRBs and conducting parameter statistics and analysis, utilizing the Chashan Broadband Solar radio spectrograph at meter-wavelength (CBSm) SRB data. Lastly, Section 5 offers a comprehensive summary of the study. Through the research presented in this paper, we aim to provide new tools and perspectives for radio astronomy and offer strong support for a deeper understanding of SRBs.

2 Data overview

2.1 Establishment of the dataset

This study is primarily focused on type II and type III SRBs (shown in Fig. 1) with the aim to investigate their properties. To ensure our the detection accuracy of our approach, we used data from the Learmonth Solar Observatory, spanning 2001–2022, as the dataset for training the object detection model.

We extracted and curated data related to type II and type III SRB events based on the SRB log information provided by the Learmonth Observatory. Partially displayed images and irrelevant data were manually excluded from establishing an SRB image database and the labeling tool was used to label the binding boxes around type II and type III SRB features in the images. This process involved manual marking and categorizing these burst events to train the object detection model.

Fig. 1 Example of dynamic spectrum data from the Learmonth Observatory, displaying type II and type III SRBs occurring between 25 and 180 MHz.

2.2 Dataset split and statistics

To ensure the fairness and effectiveness of our experiments, we randomly divided the SRB dataset into training, validation, and testing sets. The ratio of the training and validation sets to the testing set was 9:1; the ratio of the training set to the validation set was also 9:1. This division ensures that the model learns from a diverse range of data and can be generalized to new examples.

The dataset comprised 759 pictures, with 107 samples of type II SRBs and 927 samples of type III SRBs. Each sample was labeled with the corresponding bounding box coordinates for accurate object localization. We conducted a statistical analysis of the label distribution in the dataset, and the results are presented in Table 1.

Figure 2 depicts the distribution of the two label types. The graph shows that the label data are widely distributed, with a visible concentration in the upper-middle portion of the image. This distribution is crucial for understanding the spatial characteristics of the dataset. In the graph on the right, the horizontal axis represents the ratio of label width to image width, whereas the vertical axis represents the ratio of label height to image height. This analysis helps us understand the distribution of object sizes within our dataset, essential for making targeted adjustments when setting the anchor boxes.

2.3 Data augmentation

In deep learning models, many samples are required for model training to reduce the occurrence of overfitting. However, the number of radio burst events is limited and the two types of radio burst data are unbalanced. To mitigate these challenges, we employed data augmentation techniques that artificially expand the dataset to enhance the model performance and robustness, after the dataset was divided to prevent increased similarity between the training, validation, and test sets.

The data augmentation strategies employed in this study included Mosaic and Mixup. The Mosaic method combines four images by randomly scaling, cropping, and arranging them together, which enhances the generalization ability of the model and improves its classification performance. The Mixup data augmentation method blends two images by interpolation, further augmenting the dataset and enhancing model robustness. Examples of these two data augmentation methods are shown in Fig. 3. To prevent any increase in similarity among the training, validation, and test sets, data augmentation was performed after the dataset was divided, and only the training set was augmented in this case. To enhance the accuracy of data detections, data augmentation was only enabled in the first 90% of training epochs, while no augmentation was applied in the final 10% of training. This was done to optimize the model convergence and stability during the latter stages of training.

3 YOLOv7-based SRB detection

The YOLOv7 algorithm utilizes a fully convolutional neural network structure to divide an image into grid cells. Within each cell, it predicts multiple bounding boxes along with class probabilities and simultaneously performs a regression for the object position and size. This enables rapid and accurate realtime object detections. In this section, we focus on applying the YOLOv7 algorithm to detect type II and type III solar radio bursts.

3.1 Detection and information extraction of type II and type III radio bursts

The detection framework for solar radio bursts based on YOLOv7 is shown in Fig. 4.

The specific steps are as follows:

• 1.

  To satisfy the input requirements of the YOLO model, the spectrum processed in Sect. 2 is resized by padding zeros. This ensures that the original dimensions of the image remain unchanged, which is crucial for achieving a distortion-free effect before and after resizing.

• 2.

  Appropriate anchor boxes (bounding box priors) of different sizes are generated using clustering algorithms based on all ground-truth boxes in the dataset. These anchor boxes help the model match the target sizes of type II and type III radio burst events with the initial anchor box sizes, facilitating more accurate object detections. Parameters such as the number of training iterations are set and the network is trained while the model parameters are recorded.

• 3.

  The YOLO model is loaded and the object detection is performed. The reprocessed image is input into the YOLO model for object detection. The YOLO model outputs the bounding box information of the objects, including the coordinates of the object’s position, confidence level, and object class.

• 4.

  Post-processing and non-maximal suppression steps are performed. The SIoU NMS method (Chen et al. 2023) is utilized to improve the process of non-maximum suppression, reducing the occurrence rate of multiple detections of the same object and enhancing the accuracy of the model’s output results.

• 5.

  The processed target bounding boxes on the original image are visualized by overlaying bounding box outlines in the spectrum. Additionally, important information, such as the start and end time of the SRBs, as well as the start and end frequencies, are observed.

Fig. 2 Images show the distribution of the labels. In the left graph, the horizontal x-axis represents the ratio of the horizontal coordinate of the label center to the image width, and the vertical y-axis represents the ratio of the vertical coordinate of the label center to the image height. In the graph on the right, the horizontal axis represents the ratio of label width to image width, whereas the vertical axis represents the ratio of label height to image height.

Fig. 3 Examples of data augmentation, (a) enhanced result using the Mosaic method and (b) enhanced result using the Mixup method.

3.2 Experimental environment description

To validate the effectiveness of our method, we trained and tested the YOLOv7 network using the following computer configuration and hyperparameter settings (listed in Table 2).

The experimental environment was configured to provide ample computational resources and robust software support, ensuring the successful completion of our research. We employed a high-performance computing cluster equipped with an NVIDIA A30 Tensor Core GPU and an Intel® Xeon® Gold 6348H CPU @ 2.30 GHz to accelerate the training and testing of the YOLOv7 network. Additionally, we utilized software tools and libraries, including TensorFlow¹ 2.10, and Python 3.8, to facilitate data pre-processing model training, and results analysis.

Fig. 4 Detection framework of the proposed method.

3.3 Evaluation metrics

To evaluate the performance of the model used in this study comprehensively and objectively, we conducted a comprehensive evaluation using a confusion matrix. This tool for assessing the classification performance of a model comprises four key metrics:

True positives (TP): these are cases where the model correctly predicts a positive value, and the actual value is positive. In our context, these are the correct detections for type II and type III solar radio bursts.

False negatives (FN): these cases occur when the model predicts a negative value (no event), but the actual value is positive (an event occurs). In our scenario, these represent missed detections of solar radio bursts.

False positives (FP): these cases occur when the model predicts a positive value (an event), but the actual value is negative (no event). In our context, these are instances in which the model mistakenly identifies an event that did not occur.

True negatives (TN): these are cases where the model correctly predicts a negative value and no event occurred. These are correct rejections of non-events.

These metrics, derived from the confusion matrix, allow us to calculate various performance indicators, including precision, recall, F1-score, and accuracy, which collectively provide a comprehensive evaluation of the effectiveness of the model for detecting SRBs. Precision and recall are fundamental metrics in object detection. They quantify different aspects of model performance.

Precision is a measure of the accuracy of positive predictions. It answers the following query, namely: how many instances there are of the model giving a positive classification and how many were actually positive, out of all the instances. Recall measures the completeness or coverage of positive predictions, answering the question of how many were correctly identified by the model out of all the actual positive instances. These metrics are calculated using the following formulas: $$\text{ Precision }=\frac{\text{TP}}{\text{TP}+\text{FP}},$$(1) $$\text{ Recall }=\frac{\text{TP}}{\text{TP}+\text{FN}}$$(2)

The average precision (AP) is a metric used to evaluate the performance of object detection algorithms in a single class. It is based on the precision-recall (PR) curve, which is generated at various confidence thresholds. The steps for calculating AP are as follows:

• 1.

  Obtain the confidence scores of the model and corresponding predicted labels on the test set;

• 2.

  The confidence threshold gradually decreases within a certain range to generate a series of PR data points;

• 3.

  Plot the PR curve based on these data points.

The mean average precision (mAP) is an aggregate metric that assesses the overall performance of an object detection algorithm across multiple classes. This is the average of the AP values calculated for each class. The mAP provides a single value ranging from 0 to 1, representing the average performance. The mAP is calculated as follows:

• 1.

  Calculate the AP for each class separately;

• 2.

  Average these AP values to obtain the mAP score.

In summary, the mAP serves as a comprehensive measure of the algorithm’s accuracy and comprehensiveness across all classes.

3.4 Experimental results

Table 3 presents the detection results of the network for each class. type III radio bursts exhibit high precision and low FN rates. However, owing to their complex morphology and fewer occurrences than type III bursts, type II bursts exhibit lower precision and recall rates. Both categories exhibit distinct characteristics in terms of the trade-offbetween recall and precision.

To assess the detection accuracy of the model, we employed AP50 as the metric and conducted a comparative analysis with several well-established models, including Faster-RCNN (Ren et al. 2015) and YOLOv5 (Wu et al. 2021). The experimental outcomes and comparisons are presented in Table 3. In summary, a comparative analysis of detection models, including Faster-RCNN, YOLOv5, and our proposed model, reveals distinct performance characteristics. Notably, our model demonstrates competitive accuracy, particularly with respect to detecting type III instances with the highest AP of 79.96%, precision of 80.18%, and recall of 73.74%. These results demonstrate that our model achieved a balanced trade-off between precision and recall, positioning it favorably in comparison to established models such as Faster-RCNN and YOLOv5.

We further evaluate the performance of the model using the PR curve. The PR curve considers the precision and recall of each class as detected by the model. Figure 5 shows the PR curves for the two classes. The x-axis represents recall and the y-axis represents precision. When the curve in the graph is closer to the top right corner, it indicates that as recall increases, the drop in precision is minimal.

4 Application and analysis

To demonstrate the effectiveness of the proposed method in detecting SRBs and extracting eruption information, we apply the YOLOv7 network trained in Sect. 3 to the radio data observed by the Chashan Solar Radio Observatory 12-meter antenna (CBSm). Figure 6 illustrates the detection process flowchart and provides an overview of the key steps in applying the trained network to the radio data. The detection results are then organized and analyzed, including the start and end times, start and end frequencies, frequency drift rates, and relative bandwidths.

4.1 Data preprocessing

The data used in this experiment were sourced from CBSm², a newly established instrument located at the Chashan Solar Observatory (CSO). CBSm’s primary purpose is to monitor the intricate details of SRBs within the metric wave range. The CSO is managed by the Institute of Space Sciences of Shandong University, an institution known for its contributions to space research and solar observations (Shang et al. 2022). The data employed in this study span a frequency range of 90–600 MHz, featuring a frequency resolution of 76.29 kHz and a time cadence of 0.84 ms. CBSm operates routinely from 6:30 am to 6:30 pm, generating an extensive dataset of approximately 1.2 TB daily. Given the sheer volume of data, the utilization of automated algorithms is imperative to efficiently organize and summarize the pertinent phenomena of interest.

Fig. 5 Operation result curves of the model, (a) P-R curve of type II, (b) P-R curve of type II, and (c) the AP values for each class.

4.2 Radio burst detection and information extraction

Based on the example shown in Fig. 7, we input the processed spectrum into the trained YOLO model for detection. After detecting type II and III SRBs, for each burst event, we obtained the burst information in pixels. The start and stop time, as well as the start and stop frequencies of the events, need to be calculated next, based on the labeling boxes.

In the example illustrated in Fig. 8, a cluster of continuous type III burst events is marked on the spectrum. We obtained the coordinates of the bounding box as [262, 378, 504, 483]. The order of the four coordinate values is [y_top , x_left, y_bottom, x_right], denoting the top-left and bottom-right corners of the bounding box on the spectrum. With this coordinate information, we can calculate the duration, Δt, frequency bandwidth, F, start and stop time, t_StoP, t_start, start and stop frequencies, ƒ_stop , ƒ_start, relative bandwidth, ΔF, and frequency drift rate, Δƒ/Δt, of the detected events. The calculation formulas are given in Eqs. (3)–(l0).

The calculation process for burst parameters is as follows: $$t_{\text{start }}=x_{\text{left }}\times t_{\text{resolution }}+t_0,$$(3) $$t_{\text{end }}=x_{\text{right }}\times t_{\text{resolution }}+t_0,$$(4) $$f_{\text{start }}=\left(f_1-y_{\text{top }}\right)\times f_{\text{resolution }},$$(5) $$f_{\text{end }}=\left(f_1-y_{\text{bottom }}\right)\times f_{\text{resolution }},$$(6) $$F=\text{Δ}f=f_{\text{stop }}-f_{\text{start }},$$(7) $$\text{Δ}t=t_{\text{stop }}-t_{\text{start }},$$(8) $$\text{Δ}F=\frac{2F}{f_{\text{start }}+f_{\text{stop }}},$$(9) $$\frac{\text{Δ}f}{\text{Δ}t}=\frac{f_{\text{stop }}-f_{\text{start }}}{t_{\text{stop }}-t_{\text{start }}},$$(10)

In the calculation formula (3)–(10), t_resoiution represents the time resolution of the original data used for plotting; f_resolution represents the frequency resolution (MHz) of the original data used for plotting and t₀ indicates the starting time and ƒ₁ ending frequency of the spectrum, respectively. Therefore, the duration and frequency range of the burst event were obtained by multiplying the width and height of the detection boxes by their corresponding resolutions. In contrast, the calculation of start and stop time and start and stop frequency is related to the resolutions and depends on the starting time and starting frequency of the original data used for the spectrum.

During the data processing phase, the initiation frequency for CBSm’s spectrum is 90 MHz, with a designated cutoff frequency represented as ƒ₁ set at 600 MHz. Each spectrum has a duration of 15 min, leading to a bandwidth of 510 MHz and a temporal coverage of 900 s. In the plotting procedure, the visual representation, excluding coordinate axes and margins, was configured to be 900 × 510 pixels. As a result, the frequency f_resolution is established at 1 MHz and the temporal resolution t_resoiution is maintained at 1 s.

Fig. 6 Flowchart of solar radio spikes identification and feature extraction process.

Fig. 7 Dynamic spectrum example illustrating the type II and type III SRBs between 90 and 600 MHz as observed by CBSm.

Fig. 8 Example of radio burst tagging.

4.3 Burst event analysis

By applying the method described in this study to the CBSm data from December 2022, 107 complete type III burst events were selected for statistical analysis because of the relatively low occurrence rates of type II bursts during this period (See Appendix A for events list).

The first picture of Fig. 9 presents a scatter plot of the frequency variation (in MHz) of the 107 type III burst events arranged chronologically. The blue dots represent the starting frequencies of the events, while the orange dots at the same horizontal coordinate indicate the corresponding ending frequencies. The graph shows that the type III burst events predominantly end at approximately 90-100 MHz and start at approximately 590600 MHz. The start and stop frequencies obtained here may not represent the actual emission frequencies of type III bursts due to the limited frequency range of the CBSm observations and the persistent noise near 90 MHz.

We counted the average frequency drift rate for each event. The second picture of Fig. 9 shows a scatter plot of the frequency drift rates of 107 separate and clear type III bursts. Thus, from the statistics of the mean frequency drift rate, the absolute value of the mean frequency drift rate of type III bursts is determined to be mainly distributed in the range of 0.55-21.52 MHz s^-1. The third picture of Fig. 9 shows the association between bandwidth and event duration. The observation reveals that the majority of type III bursts exhibit a duration of 200 s or less.

Fig. 9 Analysis of type III SRB data. (a) Frequency chart depicting the starts and ends of events. (b) Statistics on the drift rate of the event date. (c) Distribution analysis of frequency bandwidth and duration for type III SRBs

5 Conclusion and summary

The primary goal of this study is to address the challenges related to the detection, classification, and information extraction of SRBs. To achieve this, we employed the YOLOv7 object detection method, which has proven to be a powerful tool in our research.

We initiated our study by training a model using public data from Learmonth, which achieves an impressive mAP accuracy of 73.5%. This accuracy underscores the effectiveness of our approach for detecting and classifying SRB events.

The true innovation of our method lies in its simplicity and versatility compared with conventional approaches used to extract burst feature parameters. By extracting the location coordinates of burst events from the object detection network and performing straightforward calculations, we obtained the essential basic feature parameters. The statistical analysis of these parameters provided valuable insights into the two types of burst events.

Our experiments confirmed that our proposed method simplifies the process of extracting burst feature parameters and offers a more universal and convenient approach compared with traditional methods. Our research has promising aspects in the future. The continuous development of CBSm backend data processing techniques has enabled near real-time data recording and processing. This opens up the possibility of utilizing the realtime software pipeline of YOLO for burst detections in CBSm, which is a significant advancement in near real-time spatial weather monitoring.

Our method is not limited to CBSm, it will also be applied to the Mingantu Spectral Radioheliograph and the Daocheng Solar Radio Telescope in the Chinese Meridian Project II. This innovative project, with its comprehensive data collection and processing capabilities, has enhanced our ability to conduct autonomous scientific research, environmental space forecasting, and interdisciplinary applications. We are confident that we can achieve near-real-time type II and type III SRB analyses by assessing the real-time detection capability of YOLOv7 and the data processing capabilities of high-performance computers.

In conclusion, our research successfully addresses the challenges of the real-time detection and classification of type II and type III SRBs using deep learning models. This achievement holds significant importance in near real-time spatial weather monitoring, space weather research, and radio astronomy. As we continue our research and improve the datasets at hand, we believe that our method will yield even more remarkable results in the future.

Acknowledgements

This research was supported by the grants of the National Natural Science Foundation of China (42374219,42127804) and the Qilu Young Researcher Project of Shandong University. Thanks to the Learmonth Solar Observatory for the observations.

Appendix A Events list

References

All Tables

All Figures

	Fig. 1 Example of dynamic spectrum data from the Learmonth Observatory, displaying type II and type III SRBs occurring between 25 and 180 MHz.
In the text

Fig. 2 Images show the distribution of the labels. In the left graph, the horizontal x-axis represents the ratio of the horizontal coordinate of the label center to the image width, and the vertical y-axis represents the ratio of the vertical coordinate of the label center to the image height. In the graph on the right, the horizontal axis represents the ratio of label width to image width, whereas the vertical axis represents the ratio of label height to image height.

In the text

	Fig. 3 Examples of data augmentation, (a) enhanced result using the Mosaic method and (b) enhanced result using the Mixup method.
In the text

	Fig. 4 Detection framework of the proposed method.
In the text

	Fig. 5 Operation result curves of the model, (a) P-R curve of type II, (b) P-R curve of type II, and (c) the AP values for each class.
In the text

	Fig. 6 Flowchart of solar radio spikes identification and feature extraction process.
In the text

	Fig. 7 Dynamic spectrum example illustrating the type II and type III SRBs between 90 and 600 MHz as observed by CBSm.
In the text

	Fig. 8 Example of radio burst tagging.
In the text

	Fig. 9 Analysis of type III SRB data. (a) Frequency chart depicting the starts and ends of events. (b) Statistics on the drift rate of the event date. (c) Distribution analysis of frequency bandwidth and duration for type III SRBs
In the text