Photo voltaic radio bursts and their superb spectral buildings include key bodily data associated to plasma instabilities, high-energy particle acceleration, and different necessary processes. They due to this fact function an necessary observational means for learning photo voltaic exercise and house climate. With the rising quantity of photo voltaic radio spectrogram statement knowledge, deep learning-based recognition and detection of photo voltaic radio bursts have develop into a key analysis route. Nevertheless, most research on meter-wave photo voltaic radio spectrogram recognition and detection depend on proprietary datasets, and publicly out there datasets stay scarce. To handle this challenge and facilitate using public datasets for deep studying mannequin validation, thereby selling the event of automated detection strategies for meter-wave photo voltaic radio spectrogram, we suggest a brand new photo voltaic radio spectrogram dataset. In our examine, we additional examine the dataset building course of, its utilization, and the coaching efficiency achieved based mostly on this dataset.
The dataset is publicly out there at: http://62.234.23.17/SRData/SRSD/.
This examine constructs a public dataset based mostly on Learmonth Observatory in Australia and the meter-wavelength observing system of the Chashan Photo voltaic radio Observatory (CSO) of Shandong College. To satisfy the wants of various researchers, we divide the dataset into three elements. The primary is the photo voltaic radio spectrogram dataset A (SRSD-A), which is designed for the automated classification and identification of spectrogram statement knowledge into three classes: burst, non-burst, and irregular knowledge. 。The second is the photo voltaic radio burst dataset B (SRBDB-B), which can be utilized for the localization and detection of 5 varieties of meter-wave photo voltaic radio bursts, particularly sort II, sort III, sort IIIs, sort IV, and sort V bursts. The third is the photo voltaic radio burst dataset C (SRBD-C), which is meant for the localization and detection of three varieties of photo voltaic radio bursts: sort II, sort III, and sort IIIs. Primarily based on the above datasets, a number of deep studying fashions are skilled to confirm their effectiveness. The skilled fashions may be utilized to the popularity and detection of photo voltaic radio spectrograms. Determine 1 reveals the YOLO11 community mannequin.
Determine 1: Schematic diagram of the YOLO11 mannequin structure.
We validated Dataset A utilizing a number of deep studying architectures, together with ResNet, ConvNeXt, and MobileViT. The classification accuracy exceeded 99%, demonstrating glorious classification efficiency. Primarily based on the burst detection datasets, we skilled the YOLO11 mannequin on Dataset B and Dataset C. Determine 2 reveals the coaching course of for Dataset B. As may be seen from the determine, the general coaching course of was comparatively steady. As well as, the skilled mannequin was validated on a extra complicated dataset, attaining a precision of 0.702 and a recall of 0.645, which demonstrates its good generalization functionality. Nevertheless, some misclassified samples nonetheless stay. Noise within the spectrograms and the complexity of burst morphologies are nonetheless challenges that have to be addressed. Growing the quantity of information would assist additional enhance the coaching accuracy.
Determine 2: YOLO11 mannequin coaching efficiency in Dataset-B
Conclusion
The outcomes present that the dataset proposed on this examine can successfully help the popularity and detection of photo voltaic radio observational knowledge. This supplies a unified dataset for experimental analysis on photo voltaic radio burst detection, thereby supporting research on photo voltaic radio knowledge processing and burst detection, whereas additionally selling the combination of radio astronomy and synthetic intelligence. In future work, we plan to additional optimize and broaden the dataset by incorporating extra knowledge sources, whereas additionally bettering the objectivity of the annotation course of and enhancing the applicability of the dataset. These datasets are anticipated to enhance the detection functionality for future burst occasions and advance associated analysis.
Primarily based on a not too long ago printed article: Yan Liu , Hongqiang Track , Fabao Yan and Yanrui Su, Dataset for Recognition and Detection Primarily based on Photo voltaic Radio Spectrogram Information, Analysis in Astronomy and Astrophysics,26:037001 (2026), DOI: https://doi.org/10.1088/1674-4527/ae2b5a
