IMPLEMENTATION OF CONVOLUTIONAL NEURAL NETWORK MODEL ON HMI AND AIA IMAGES FOR M AND X CLASS SOLAR FLARE DETECTION

• Main Author: Lee, Sylvie
• Format: Final Project
• Language: Indonesia
• Subjects: Astronomi
• Online Access: https://digilib.itb.ac.id/gdl/view/81221
• Institution: Institut Teknologi Bandung

<p align="justify">In recent years, interest in Solar data processing has increased rapidly. This data processing aims to monitor and understand Solar activity, including solar flares. Monitoring and understanding solar flare occurrences are important because these events determine space weather, which impacts infrastructure and technology used in communication, satellite navigation, and space missions. Manual monitoring of solar flares by humans cannot be fully relied upon. This is because detection and identification by humans are highly dependent on the observer's abilities and are time-consuming. Therefore, the development of automatic solar flare detection without human intervention is a challenge in this regard. The focus of automatic solar flare detection lies in developing efficient feature-based classification. Various automatic solar flare detection techniques that have been developed and implemented mostly use deep learning approaches. In this final project research, the author employs image processing techniques to make the analysis process more efficient by applying a Convolutional Neural Network (CNN) approach directly to raw pixels. The input data used in this system are daily solar flare activity data from Space Weather Live and raw image data from the Solar Dynamics Observatory (SDO; Pesnell et al., 2012). There are two algorithms that the author applies in this final project research: 1) The first system is designed to recognize solar flares through the analysis of active region data extracted from full-disk magnetogram images (HMI). 2) The second system is designed to recognize solar flares through the analysis of ultraviolet intensity data extracted from full-disk ultraviolet intensity images (AIA). From this research, the author found that a CNN model with a ResNet-50 architecture using AIA data from 2020-2023 provided the best prediction performance among other architecture models. Additionally, the author also found that the choice between using HMI or AIA data and augmentation variations showed no correlation with model performance, but this depends on the architecture influencing the model's performance. Furthermore, the author also found that an increase in the amount of data improves the model's generalization ability.<p align="justify">