ItemConference Paper

Synthetic Data Augmentation for Robust Solar Flare Classification: A Comparative Analysis of Conditional DCGAN, VAE, and Diffusion Models

Title
Synthetic Data Augmentation for Robust Solar Flare Classification: A Comparative Analysis of Conditional DCGAN, VAE, and Diffusion Models
Creator
Swetha, S.; Prakash, Gladis Sam; Thomas, Jestin; Kalpana, P.; Jose, Teena; Sumathi, P.
Description
Solar flares are extremely dangerous to the ground and space-based resources. Solar flares have to be classified properly and in good time to offer protection to assets in both the environments. Deep Learning-based flares have been divided into 3 classes (C, M and X). The main problems with this kind of classification are that high impact M- or X-class solar flares are extremely rare, and cannot be well sampled, thus leading to a very unbalanced sample. This paper exemplifies a comparative analysis of three models of Conditional Generative Models (cDCGAN models, cVAE models and cDDPM) to produce realistic images of solar flares considering each of the low frequency and high-impact solar flare types. The research question will be how such models can be evaluated in terms of their capacity to create realistic, class-specific images (magnetograms, and EUV) and time-series data which could be used to create class-balanced training samples. The initial experiments make use of the cDCGAN, cVAE, cDDPM architecture and considering the generation of class-conditional solar flare images portray high levels of stability (convergence was stated in less than 600 epochs in the case of the cDCGAN, cVAE models and cDDPM) and the generation of images which could be considered as practically indistinguishable to real life images. The results indicate that cDDPM may be a possible solution to a high-fidelity production of solar features. We measured the efficiency of these models in quantitative terms (popular metrics, like the Frhet Inception Distance, Structural Similarity Index) in a manner that we could determine the best manner of training model based solar flare classification systems using realistic data. This research is aligned with Sustainability Development Goals 9- Industry, Innovation and Infrastructure with focus on verticals 9.1 and 9.5. 2026 IEEE.
Source
Proceedings of 2nd International Conference on Visual Analytics and Data Visualization, ICVADV 2026;pp.896-901
Date
01-01-2026
Publisher
Institute of Electrical and Electronics Engineers Inc.
Subject
cDCGAN; Class Imbalance; Conditional Diffusion Model; Conditional Generative Models; cVAE; Data Augmentation; SDG Goal 9; Solar Flare Classification; Synthetic Data
Coverage
Swetha S., Christ (deemed to be University), Department of Computer Science, Karnataka, Bengaluru, India; Prakash G.S., Christ (deemed to be University), Department of Computer Science, Karnataka, Bengaluru, India; Thomas J., Christ (deemed to be University), Department of Computer Science, Karnataka, Bengaluru, India; Kalpana P., Christ (deemed to be University), Department of Computer Science, Karnataka, Bengaluru, India; Jose T., Christ (deemed to be University), Department of Computer Science, Karnataka, Bengaluru, India; Sumathi P., Vyasa College, Department of Computer Science, Tamil Nadu, Salem, India
Rights
Restricted Access; Hardcopy may be available in the library
Relation
ISBN: 979-833159237-0;
Format
online
Language
English
Type
Conference paper

Citation

Swetha, S.; Prakash, Gladis Sam; Thomas, Jestin; Kalpana, P.; Jose, Teena; Sumathi, P., “Synthetic Data Augmentation for Robust Solar Flare Classification: A Comparative Analysis of Conditional DCGAN, VAE, and Diffusion Models,” CHRIST (Deemed To Be University) Institutional Repository, accessed June 18, 2026, https://archives.christuniversity.in/items/show/26141.

Output Formats