DEEP LEARNING-BASED SUNSPOT SEGMENTATION AND CLASSIFICATION USING GRADIENT FEATURES FOR MCINTOSH MORPHOLOGY IDENTIFICATION IN SOLAR IMAGERY
Luthfi Naufal(a*,b), Dhani Herdiwijaya(a), Dwi Irwanto(b), Fargiza Abdan Malikul Mulki(a)

a) Astronomy Research Grup, Bandung Institute of Technology
Jalan Ganesha 10, Bandung 40132, Indonesia
*Luthfi.naufal23[at]gmail.com
b) Computational Science Study Program, Bandung Institute of Technology
Jalan Ganesha 10, Bandung 40132, Indonesia

Abstract

The development of an automated method for detecting and classifying sunspots in solar images serves as the main focus of this study, with a machine learning-based approach as its foundation. The segmentation of umbra and penumbra regions is performed using deep learning architectures such as U-Net and DeepLabV3+, which are known for their effectiveness in recognizing complex morphological patterns. These architectures are considered mature models and have proven to be reliable in various image segmentation studies, making them a stable and dependable choice compared to several newer architectures that are still in the exploratory or developmental phase. In addition to segmentation, the system is also equipped with the capability to classify each sunspot based on the McIntosh morphological classification scheme, which is widely used in solar activity monitoring.
The proposed method does not rely on conventional threshold-based approaches- instead, it incorporates additional features in the form of photon intensity derivatives (gradients) as input to the model. These features are extracted from multi-spectral data to capture more precise spatial variations, enabling the model to more accurately distinguish boundaries between the umbra, penumbra, and photosphere. The dataset used consists of 4096x4096 pixel solar images accompanied by semi-manual annotations in the form of three-class masks, as well as numerical data associated with each pixel.
The outcomes of this research include pixel-level segmentation maps (labeled as background, penumbra, and umbra) and McIntosh class predictions for each sunspot. Evaluation of the model^s performance demonstrates improved accuracy in both segmentation and classification compared to baseline approaches. With the ability to automatically and accurately identify and classify sunspots, this system is expected to support real-time solar activity monitoring and contribute to the study of solar dynamics and space weather.

Keywords: Sunspot- Deep Learning- McIntosh Classification- GPU- Parallel Computing- High-Performance Computing

Topic: Instrumentation in Astronomy