Technical Article

Automated detection and classification of solar radio bursts in CALLISTO spectrograms using deep-learning YOLOv5 model and ensemble methods

¹ Solar-Terrestrial Centre of Excellence – Royal Observatory of Belgium, Avenue Circulaire 3 1180, Brussels, Belgium
² Center for Mathematical Plasma Astrophysics, Department of Mathematics, University of Leuven, KULeuven, Belgium
³ Istituto ricerche solari Aldo e Cele Daccò (IRSOL), Faculty of Informatics, Università della Svizzera italiana (USI), CH-6605, Locarno, Switzerland

^* Corresponding author: This email address is being protected from spambots. You need JavaScript enabled to view it.

Received: 4 December 2025
Accepted: 1 April 2026

Abstract

Context. Solar radio bursts in the meter and decameter range wavelengths are indicators of eruptive events in the solar corona. They are routinely monitored by the global Compound Astronomical Low-cost Low-frequency Instrument for Spectroscopy and Transportable Observatory (CALLISTO) network. The development of automated detection and classification tools remains difficult due to the diversity of instrumentation background and limited datasets where bursts have been identified and labeled. Aims. This work evaluates the performance of a deep-learning object detection model, You Only Look Once (YOLO) version 5, which identifies and localizes features in images using bounding boxes. In addition, we combined multiple of these trained models using ensemble methods to improve the automated detection and classification of Type II, III, IV, and Group of Type III solar radio bursts across the e-CALLISTO network. Methods. A dataset of 1108 annotated spectrograms from 49 instruments was used to study the effect of image resolution, data augmentation, and class definition. Ensemble strategies, including hard voting, soft voting, and Weighted Box Fusion, were applied to combine the results from several models into a final detection. Results. Moderate image resolution of 640 × 640 pixels preserved burst morphology while limiting noise amplification. Data augmentation improved generalization across different telescopes, and grouping closely related radio burst categories reduced false detections, although it also increased the number of missed events. Combining data augmentation with category merging provided a balance between optimal precision and recall. Combining the predictions of multiple trained models through ensemble methods further improved overall performance. The best configuration, based on the Weighted Box Fusion technique, achieved the highest mean F1 score of 0.738, exceeding the performance of any single model. Type III bursts remained the most challenging to detect, mainly due to annotation ambiguities and similarity to background noise. Conclusions. Using deep learning combined with ensemble methods improves the automated detection of solar radio bursts compared to single-model approaches, with the Weighted Box Fusion ensemble achieving the highest F1 score. The main challenge remains the ambiguity in labeling bursts, especially for Type III bursts and closely related events, suggesting that more consistent annotations and refined class definitions could further improve model performance.