The radiation environment in space and the related radiation exposure is seen as one of the main health detriments for human missions in and beyond low Earth orbit (LEO). In addition to national space agencies sending astronauts to space, the near future is likely to bring numerous commercial endeavours that will facilitate the access to space, and specifically to LEO, for a growing number of people. The components of the cosmic radiation forming the environment in LEO are the galactic cosmic radiation (GCR), inner and outer radiation belt particles and sporadic solar energetic particle events. The steady flux of energetic particles in the galactic cosmic radiation produces a chronic low-dose-rate exposure, which is heavily influenced by several factors including variations during the solar cycle, the Earth’s magnetic field and spacecraft shielding. Investigations of the GCR variations over the course of a solar cycle provide valuable data for exploration mission planning and for the determination of the radiation load received due to the GCR environment. The RAMIS (Radiation Measurement in Space) instrument onboard the DLR Eu:CROPIS satellite, named after its primary payload: Euglena and Combined Regenerative Organic-Food Production in Space, has been measuring the radiation environment in a sun-synchronous polar orbit at 500 km to 600 km altitude since December 2018 and has been providing data on the exposure from GCR, inner and outer radiation belt particles and during numerous solar energetic particle events. Measurements at high latitudes and low geomagnetic shielding provide information on the exposure outside the magnetosphere in near-Earth interplanetary space.
This paper presents a quantitative measure for the intensities of auroras, as well as a new magnetic index which can indicate the probability of auroras locally. Auroras are an important factor in the business of nature tourism in polar areas. Nowcasts and short-term forecasts of auroral activity can be based on near-real-time all sky camera (ASC) and magnetometer data. This paper presents a new parameter (hereafter I) which can indicate quantitatively the brightness of auroras as measured using ASC images at the three brightest aurora wavelengths (428, 558 and 630 nm), as well as a parameter (hereafter the R-index) indicating the variations of magnetic field measured by a magnetometer, which is proposed as a proxy for the aurora. Statistical results of both I and R are presented. It is found that both parameters have similar statistical variations with time of day, season and solar cycle phase, which are also similar to those previously reported for auroral activity. Empirical relations between I and R for each of the three aurora wavelengths are also established. With this relation, combined with previously published formulae for Rayleigh-to-candela conversion, we can build a link between magnetic field measurements and auroras visible for the naked human eye. Based on our data sets, average probabilities of human eye detectable auroras in rural night sky conditions vary in the range of 2-30% at magnetic latitudes 55º-65º.
This paper examines the radiation hazards posed to civil aviation by extreme space weather events, focusing on the impacts of Extreme Solar Particle Events (ESPEs) at flight altitudes. At the same time, aviation benefits from atmospheric and geomagnetic shielding, but significant residual radiation persists, particularly during the so-called Ground Level Enhancements (GLEs), with implications for both aircrew and avionics reliability. Recent high-profile incidents, such as the Airbus A320 grounding due to cosmic-ray-induced Single Event Upsets (SEUs), underscore the growing vulnerability of modern aircraft systems. The study first assessed the spectral models of GLEs and historical ESPEs to establish the total event fluence and peak event flux spectra for a selection of the strongest events. These spectra were then used as inputs to the MAIRE+ model to generate the enhanced in-flight radiation doses and SEU rates above the normal GCR values. The impact of a Carrington-type geomagnetic storm has also been assessed and analyzed. Further, the ongoing efforts to standardize aviation radiation protection are discussed, and a new atmospheric radiation scale is proposed, aiming to enhance operational mitigation and ensure safety in the enhanced space weather environment.
The Earth’s ionospheric dynamics are multi-scale in nature, and the small-scale phenomena have been difficult to measure by in-situ techniques because of low data resolution. In this paper, we make use of the high-resolution Swarm faceplate plasma density data at 16 Hz to develop a set of parameters that can characterize multi-scale ionospheric structures and irregularities along the Swarm orbit. The new data is named as the MUlti-Scale Irregularities produCt (MUSIC) dataset. We present the methods for calculating density gradients over different window sizes, rate of change of density index, power spectral density and the spectral slope at both low and high latitudes. The faceplate plasma data are not continuously available over the years. However, from late 2014 to the end of 2025, the available data from Swarm A amounts to about eight years. Some statistical results from Swarm A are presented. The variations of plasma structures and irregularities are dependent on solar activity, season, local time and geomagnetic activities, and the variations show different patterns between low and high latitudes. For example, the high-latitude ionosphere is characterized by persistent ionospheric structures and irregularities poleward of ±60° magnetic latitude, while the low-latitude ionospheric irregularities are only dominant during 19–01 local time near the magnetic equator. The occurrence of a steep spectral slope at high latitudes shows clear seasonal variations, i.e., it maximizes during local summer and minimizes during local winter in both hemispheres, with the variations attributed to the seasonal variations of the E region conductance in the polar regions. However, the occurrence of a steep spectral slope at low latitudes is only sensible when significant plasma structures and irregularities are present. We further calculate the histogram of spectral slopes at low latitudes when the rate of change of the density index is enhanced. The histogram resembles a Gaussian distribution with an expected value of 1.97. The processed data are available to the wider community. Given its high resolution, this new data product will be useful for the scientific communities that are interested in the magnetosphere-ionosphere-thermosphere coupling and the near-Earth space environment. The MUSIC dataset is now available through the Swarm dissemination server at https://swarm-diss.eo.esa.int/#swarm/Advanced/Plasma_Data/TDS_EFI_MUS_FP.
The Low Frequency Array (LOFAR) is one of the most advanced radio telescopes in the world. When radio waves from a distant astronomical source traverse the ionosphere, structures in this plasma affect the signal. Results are presented from a statistical study using 2,810 hours of observations of Cassiopeia A from a LOFAR station located in the Netherlands (station CS032, located at 52.9o N; 6.9o E) between 28th June 2014 and 27th November 2016. Ionospheric structures were identified in 469 (~17 %) of these observations. A comparison with proxies for geomagnetic activity (the Kp index) and solar activity (the F10.7 cm solar radio flux) showed that geomagnetic or solar effects were not the primary driver of these ionospheric structures. Ionospheric structures were more common in summer and between ~21 LT – 02 LT. These patterns in season and local time showed similarities to the occurrence of lightning strikes. When ionospheric structures were present, the mean number of lightning strikes in a spatial region close to the LOFAR observations (51.9o – 56.5o N; 3.9o – 9.9o E) two hours prior to the LOFAR observations was (70 ± 25) per hour. This was substantially larger than the mean value of (19 ± 5) per hour when the ionospheric structures were absent. This suggests that quasi-upward propagating Atmospheric Gravity Waves (AGWs) launched by thunderstorm activity could be one of the sources of the ionospheric structures. Collectively, these observations show that LOFAR can be used to infer ionospheric signatures of vertical coupling processes in the mid-latitude atmosphere. Such observations have the potential to be used to develop or validate existing Global Circulation Models (GCMs) or new models of vertical coupling within the terrestrial atmosphere.
Motivation and Aims: The main objectives of this paper are twofold: (1) to present the Space Weather Awareness Training Network (SWATNet), an innovative Marie Skłodowska-Curie Training Network designed to provide high-quality doctoral training in space weather, and (2) to highlight the scientific advances achieved within the project. Methods: By combining cutting-edge research with structured international and intersectoral training, SWATNet reached fundamental advances in understanding space weather phenomena and trained a new generation of scientists equipped to tackle modern challenges in both research and industry. Results: SWATNet delivered advancements in modelling solar eruptions and the solar corona, and improved studies of the propagation and transport of Solar Energetic Particles, often utilising interdisciplinary methods. These include machine and deep learning, image processing, and various numerical modelling approaches. The research produced 26 peer-reviewed journal articles, along with various presentations at international conferences. In addition, SWATNet successfully completed its ambitious training programme on practical observatory and industry training.
Measurements of the polarization of the aurora remain challenging but would enable the inference of the magnetic field’s dynamics and the energetics of the impacting charged particles during geomagnetic storms. Here we present a novel three-color imaging instrument for measuring the polarization of the aurora. The instrument concept is based on a commercial micro-patterned linear polarization camera with Red-Green-Blue color filters. This detector type enables a very compact instrument concept where both the spectral selectivity and the polarization measurement is performed in a snapshot fashion by the imaging chip. We implement additional spectral filters to specifically select the wavelengths of the auroral lines and add a rotating half-wave plate to enhance the polarimetric performance. We have developed a first prototype, called PANIC (Polarization of the Aurora in the North Imaging Camera), to proof the measurement principle. The first observations during a field campaign in Skibotn (Norway) exhibit a sub-percent polarimetric sensitivity and accuracy, rendering the instrument concept viable for routine observations. We discuss future upgrades to improve the instrument stability and its capability to discern auroral signals from other light sources.
Solar wind energy is continuously deposited in the magnetosphere-ionosphere-thermosphere system, causing significant modifications primarily in the high latitude ionosphere. These variations are reflected most instantaneously in the ionospheric electron density (Ne) or in the total electron content (TEC). The drivers of ionospheric variability at high latitudes are not yet fully understood. This variability due to solar wind-magnetosphere-ionosphere coupling could be investigated under winter conditions, while ionization from EUV radiation is minimal, and ionization mostly comes from the coupling processes. This study characterizes the contributions of ionospheric drivers to winter TEC variability. We present a quantitative evaluation of the respective impact of the convection and particle precipitation processes on the TEC variability. We use comprehensive datasets of IGS and EISCAT TEC measurements, alongside merging electric field (Em) calculated from solar wind parameters. We apply a lagged correlation method covering the wintertime to assess the temporal and spatial characteristics of ionospheric response. EISCAT UHF Incoherent Scatter Radar campaigns that consist of several days of continuous measurements are used to estimate the ionospheric response time to the solar wind in the E- and F-region separately and to identify the relevant coupling processes. Our results reveal that the highest correlation between IGS TEC and Em is at a lag time of approximate to 2 h. The EISCAT results show distinctions between the E- and F-region ionosphere responses. In the E-region ionosphere, shorter delays of approximate to 71 min are observed. We suggest that the E-region TEC is driven by auroral particle precipitation during substorm processes, and the delay can be attributed to the loading and unloading times of the magnetosphere. In the F-region, the delays are longer with approximate to 101 min, indicating the effect of polar cap plasma convection, because this duration matches well with the duration of quiet time plasma convection across the polar cap. Under certain conditions, where the F-region is driven by dense polar cap patches and associated convection features, the delay in the F-region can be as short as 90 min. We find that the overall TEC response of approximate to 2 h originates mainly due to the F-region processes, where the electron density is modulated strongly by the convection of the plasma.
In this study, we utilise satellite measurements to investigate variations in the magnetosphere. Using 8- to 12-h time windows, we map the time-varying night-side magnetospheric fields up to degree and order 3, capturing lateral offsets from a simple tilted dipole model. We determine the field over the night-side hemisphere during geomagnetic storms and examine how azimuthal symmetry is disrupted at storm maxima and re-established during the storm recovery phase, as exemplified by the St. Patrick's Day storm in 2015. For this event, we demonstrate substantial variations in magnetospheric field intensity throughout the storm, followed by a gradual recovery to background levels over several days. Analysing over a decade of data, we investigate long-term seasonal and solar cycle-related variations alongside the impact of intense geomagnetic storms. We observe that the alignment of the Earth's dipole with the Sun-Earth line plays a key role in shaping the latitudinal geometry of the magnetospheric field. Furthermore, we find that, during quiet conditions, the night-side magnetosphere is characterised by a dipolar field with q10 approximate to 8 q 1 0 approximate to 8 $ q_1<^>0 \approx 8 $ nT. On average, geomagnetic storms induce a tilt of the inducing dipole towards dawn and an enhancement of the equatorial ring current near dusk.
We present the Small Phased Array DEmonstrator (SPADE), a compact phased array spectrograph designed for the monitoring of solar activity in the decameter range, where a majority of bursts of interest for solar physics and space weather occur. Unlike other existing phased arrays, SPADE operations like Sun tracking and spectra productions are made entirely digitally, and rely on the use of open source library software and commercial generic software defined radio receivers. In this paper, we describe the instrument and its associated software and present the first observations of the Sun and the Jovian magnetosphere conducted between September and December 2024. For the solar observations, in particular, we highlight the capacities of the instrument in terms of temporal and frequency resolution, making it capable of addressing science topics like turbulence in the coronal medium and propagation effects affecting radio waves, which are currently only accessible by larger but non-solar dedicated phased arrays.
Aurorae are visual manifestations of space weather and space climate. Their historical records occasionally help us to reconstruct the space weather and space climate in the past, while their records are scattered and forgotten in various archival documents. One of such cases was Alfred Russel Wallace’s travel account. Here, Wallace documented his observations of an unusual optical phenomenon and tentatively associated it with the aurora, but his doubts remain unanswered to this day. This study critically re-examines Wallace’s account of this unusual “aurora”, documenting his background and report. This study located and dated his observation at Muka in Waigiou Island (S00°26', E130°49') on 4 July 1860. The location and date allows us to calculate a magnetic latitude of −11.3°, with the phenomenon reported geomagnetically in the equatorward sky. This combination contradicts the auroral hypothesis for this phenomenon. Although the timing was close to a solar maximum, the contemporaneous geomagnetic measurements also did not show any significant geomagnetic disturbances. Morphologically, the reported phenomenon is more consistent with sprites or transient luminous phenomena; there was a constant succession of faint vertical flashings/flickerings and a great cloud that produced a storm two days after the observation. While we suggest the observation might be explained by sprites or transient luminous events, we cannot completely exclude other possibilities either. As such, Wallace’s concern was correct. This report in turn seems a probable account for sprites or transient luminous phenomena that predates their earliest known documentation in 1885.
Solar energetic particles are of major importance for understanding space weather. While numerous spaceborne instruments measure the particles with energies below several MeV and ground-based measurements address the particles with energies above approximately 1 GeV, the intermediate energies are sparsely covered. The CHerenkov Atmospheric Observation System (CHAOS) is a particle detector for solar energetic particles and galactic cosmic rays that bridges this gap in energy coverage. CHAOS uses multiple solid-state detectors and a bismuth germanium oxide scintillator to enable energy-resolved measurements of different particle species using the dE/dx − dE/dx method. Furthermore, an aerogel Cherenkov detector serves as a velocity threshold detector. The Cherenkov detector allows a clean measurement of protons with kinetic energies of tens of MeV up to 2.1 GeV, free from electron contamination. In this paper, we describe the design of CHAOS and show the first results to demonstrate the instrument’s capabilities. For this, we use measurements obtained during a stratospheric balloon flight as part of the BEXUS 35 campaign as well as measurements conducted at the CERN-EU high-energy Reference Field and compare them with GEANT4 Monte Carlo simulations.
The Digisonde Portable Sounder (DPS) ionosonde at Hainan Station (19.5°N, 109.1°E; magnetic latitude: 11°N) has been monitoring ionospheric conditions since 2002, routinely recording ionospheric plasma profiles, sporadic E layers, and Spread-F structures through ionograms. A Spatial Group-wise Enhanced ConvTransformer (SGE-ConvTransformer) is proposed in this study for spatiotemporal ionospheric prediction at the Hainan station, with emphasis on Spread-F, enabling a one-hour lead time with a 15-minute sampling resolution. The SGE module optimizes semantic feature extraction from the global spatial context, dynamically recalibrating attention to prioritize information-rich regions, such as F-layer traces, over background noise. To further improve visual clarity, a super-resolution Enhanced Deep Super-Resolution (EDSR) module is integrated to sharpen the predicted ionograms. Leveraging DPS ionosonde data from 2002 to 2015, we constructed a large-scale ionogram sequence dataset comprising 36,240 Spread-F instances and 396,931 non-Spread-F instances, which were further categorized into five distinct classes. On the 2016 test set, our model achieved an average Spread-F classification accuracy of 90.05% and a correlation coefficient of 0.8115 for the predicted F-trace. Demonstrating superior robustness under disturbance conditions, the model maintained high performance during six representative geomagnetically disturbed intervals (2023–2024), achieving a classification accuracy of up to 95.69%. Furthermore, the model's generalizability was examined by applying pre-trained weights to data from low-latitude (Brazil, Peru), mid-latitude (Irkutsk), and high-latitude (Zhigansk) stations. Quantitative Spread-F Classification Accuracy (SFCA) metrics at low latitudes and qualitative visual assessments across all regions demonstrate the morphological transferability of our approach across diverse geospatial environments.
Birkeland field-aligned currents (FACs) are associated with solar wind-magnetosphere energy coupling leading to substorms/convection events. FAC intensity enhancements have been studied during geomagnetic storms driven by interplanetary coronal mass ejections and corotating interaction regions. Here we present an in-depth analysis of long-term FAC variation, strongly modulated by solar wind high-speed streams (HSSs) emanated from solar coronal holes. Variations of both dayside and nightside FAC intensities during June 2016 through June 2017 are characterized by prominent periodicities of similar to 14, similar to 26, and similar to 29 days (in descending order of amplitude), highly correlated to the empirically estimated rate of magnetic flux opening at the dayside magnetopause. The periodicities represent the similar to 27-day solar rotation period and its harmonic, resulting from the combined impacts of recurring HSSs emanated from multiple long-lived coronal holes, corotating with the Sun. Modulation of the long-term FAC intensity variation by recurring HSSs is confirmed by statistical regression, cross-wavelet, and coherence analyses. These results may be useful in developing FAC prediction models.
In this study, we analyze ionospheric and thermospheric changes during a composite geomagnetic storm on 4–5 November 2023 with two activity periods. On 4 November, the corotating interaction region (CIR) compression resulted in moderate activity (SYM-H = −60 nT), while on 5 November the arrival of two Coronal Mass Ejections (CMEs) embedded into the ongoing CIR caused intense geomagnetic disturbances (SYM-H = −188 nT). Using global observations of vertical total electron content (VTEC) from the Universitat Politècnica de Catalunya (UPC) Global Ionosphere Maps (GIMs) and the normalized Ionospheric storm scale from UPC GIMs (IsUG) storm-time index, we examine ionospheric anomalies during these events. The IsUG index reveals that both events began with positive VTEC changes, but their delayed responses differed significantly. During the 4 November event, the negative phase did not develop, while the 5 November storm produced severe negative VTEC anomalies. Using complementary satellite observations of thermospheric composition, temperature and density, along with ionosonde-derived large-scale traveling ionospheric disturbances (LSTIDs) velocities, we show that these contrasting responses were caused by differences in thermospheric heating and composition. On 4 November, high-latitude heating was weak, while in the 5 November storm it was significant and led to the development of global disturbance winds and strong O/N2 decreases, causing severe negative VTEC anomalies. Despite the complicated morphology, the IsUG index could track these dynamics in agreement with the underlying physical processes, and can be used for monitoring of complex ionospheric storms.
We report four low-latitude auroral events in 2024 as observed from Hokkaido, Japan (June 28, August 4, September 12, and November 9). These auroral events occurred during moderately intense magnetic storms, with the peak Dst index of approximately −110 nT, accompanied by significant magnetospheric compression. We estimate the altitudes of these red auroras to be > ~500 km, via widespread citizen science efforts. During the red aurora appearances in Japan, the ASYM-H index increased significantly to around 150 nT, which was approximately 1.3–2.0 times larger than the SYM-H peak amplitude for all events, suggesting that the actual storm intensities were underestimated. We further propose that the very dense solar wind of >~30 /cc is a key for causing the majority of extended red auroras during moderately intense magnetic storms, possibly via the stronger-than-usual enhancement of the atmospheric Joule heat in the subauroral latitude.
A method for real-time solar flare detection and characterization using ground-based Very Low Frequency (VLF, 15–45 kHz) data is presented. The D-region, the ionosphere’s lowest region, is monitored by VLF waves propagating in the Earth-Ionosphere waveguide. The D-region electron density increases during sudden surges in X-ray radiation from solar flares. This subsequently enhances HF absorption. By seeking trend changes in VLF phase data, an incremental algorithm finds solar flares. 82.7% of M and X solar flares are detected within one fourth of their rise time. In addition, several VLF transmitters are monitored simultaneously. Combining information from their phase variations leads to an estimation of the Sun’s X-ray flux. Last, propagation models such as LMP or LWPC are combined with the VLF measurements to compute D-region electron density profiles. This method and its implementation in a new Python package are a step towards building a more resilient system for flare detection and alerts. Its reliance on ground-based data alone ensures an easy maintenance and a backup in case a satellite failure. It can provide earlier ionospheric impact detection than operational M5 satellite alerts in many cases, due to shortened data latency.
Context: The geomagnetic storm of 10–11 May 2024, also known as the Mother’s Day storm or Gannon storm, was among the most intense geomagnetic storms in recent decades. Such storms are associated with hazards, including induced currents in power grids, caused by intense ionospheric currents associated with auroral processes. Since the equatorward expansion of the auroral oval is related to the intensity of the storm, such intense events can affect infrastructures that are seldom exposed to such hazards. The May 2024 storm was the first event of this intensity in two decades, and therefore the first that can be studied with our current observational capabilities. Aim: The expansion of the auroral oval was observed by the FRIPON all-sky camera network, originally designed to detect fireballs. With over 150 cameras deployed globally, the network provides a valuable opportunity to map the evolution of the auroral oval’s expansion at relatively low latitudes during the night. In addition, several Low Earth Orbiting satellites with capabilities to observe auroral processes passed over Europe during the night of the storm, allowing us to compare their observations. Methods: In a proof-of-concept approach, we analyse the evolution of the brightness intensity at zenith for each FRIPON camera in Europe. This simple technique enables the tracking of the Southern limit of the visible aurorae over time. The results are compared with ground magnetometer measurements, optical satellite remote-sensing observations of auroral emissions, field-aligned currents from magnetometers on the Swarm satellites and enhancements in ionospheric Total Electron Content (TEC) obtained from Global Navigation Satellite System (GNSS) receiver networks. Results: We observe that the aurorae extend as far south as 46° N geographic (42° N Quasi-Dipole geomagnetic latitude), a significantly lower position than predicted by some models. In addition to this observation, a faint increase in luminosity was detected over the zenith for a large majority of cameras across Southern Europe witnessing a Stable Aurora Red (SAR) Arc, spanning from 37° to 50° in latitude. Conclusions: This work validates the use of the FRIPON network (or other similar networks) for Space Weather research. Complementary to other Space Weather instruments, it may help in studying the dynamics of the auroral oval.
In this study, we investigate the solar and geomagnetic parameters of the 1859 Carrington event using deep learning and empirical relationships. For this, we apply an image translation model, a popular deep learning method based on conditional Generative Adversarial Networks, to the generation of magnetograms from sunspot drawings. We train the model using pairs of sunspot data from Debrecen Photoheliographic Data and their corresponding Solar and Heliospheric Observatory/Michelson Doppler Imager (SOHO/MDI) and Solar Dynamics Observatory/Helioseismic and Magnetic Imager (SDO/HMI) magnetograms from 1996 to 2018, using data from January–July and December of each year for training and data from August and November for validation. To test the model, we compare actual magnetograms with artificial-intelligence-based (AI-based) ones for September and October. Our results show that the unsigned magnetic fluxes of AI-based magnetograms closely match those of the originals. Applying this model to Carrington’s full-disk sunspot drawing of 1 September 1859, we generate an AI-based magnetogram and estimate its unsigned magnetic flux. To estimate solar and geomagnetic parameters, we use the following empirical relationships: magnetic flux and flare peak flux, magnetic flux and coronal mass ejection (CME) speed, CME speed and transit time, CME speed and interplanetary coronal mass ejection (ICME) speed, and ICME speed and the Disturbance Storm Time (Dst) index to obtain upper-limit estimates for an extreme event. We find that the estimated Sun-Earth transit time is 16.7 h, consistent with the historical observations. The corresponding Dst value is about −1313 nT, which is broadly consistent with previous reconstruction-based estimates for the Carrington storm.
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.
