Introduction
Monitoring solar activity—including sunspots, flares, and filaments—is a critical component of protecting Earth's electronic infrastructure from space weather threats. While deep learning models, particularly Convolutional Neural Networks (CNNs), represent the current state-of-the-art in solar feature classification, they often operate as "black boxes." This lack of interpretability makes it difficult to understand the underlying logic of a classification. Furthermore, standard computer vision models can be sensitive to image orientation or global lighting variations.
This research explores an alternative approach using Topological Data Analysis (TDA). By applying algebraic topology to study the "shape" of solar data, the authors propose a classification method that is inherently robust to rotation and brightness shifts while providing mathematically explainable features.
Methodology: Persistent Homology
The core of the proposed method is Persistent Homology, a tool used to extract topological features from data across multiple scales. The pipeline involves several key steps:
- Sublevel Set Filtration: Grayscale solar images are converted into a sequence of binary maps using increasing brightness thresholds ($T$). As the threshold rises, more pixels are activated, causing the topological structure of the image to evolve.
- Feature Tracking: The algorithm tracks the "birth" and "death" of two types of topological features:
- 0-dimensional (0D): Connected components, representing blobs of solar activity.
- 1-dimensional (1D): "Holes" in the data.
- Vectorization and Classification: These birth and death values are plotted on a persistence diagram and transformed into "persistence curves." These vectors are then fed into a Support Vector Machine (SVM) for final classification.
A key insight of this work is that sunspots effectively function as topological holes in the brightness field of the solar surface. TDA is uniquely suited to detect these voids regardless of their specific geometric deformation.
Dataset and Experimental Setup
The study utilized a dataset curated by Armstrong and Fletcher (2019), comprising 13,159 Hydrogen-alpha ($H\alpha$) images from the Hinode/Solar Optical Telescope (SOT) 2006 mission. The images are categorized into five distinct classes: Sunspots, flares, filaments, prominences, and "quiet sun" (baseline).
The evaluation was conducted using a standard 80-20 train-test split, with 1,318 images reserved for testing.
Results and Discussion
The TDA-based approach demonstrated high performance and significant robustness:
- Accuracy: The base TDA method achieved 97.6% accuracy. By optimizing the model to include both 0D and 1D homology data along with image inverses (flipping brightness values), the accuracy reached 98.8%.
- Interpretability: Unlike CNNs, which may reach >99.9% accuracy but offer little insight into their decision-making process, the TDA features are physically meaningful (e.g., the number and persistence of dark voids).
- Invariance: Because TDA focuses on the relationships between pixels (connectedness and enclosure) rather than absolute spatial coordinates, the features are naturally invariant under rotation and global brightness shifts.
Conclusion and Future Work
While there is a slight accuracy trade-off compared to the most advanced deep learning frameworks, TDA offers a more interpretable and computationally efficient alternative for solar feature classification.
Future research aims to extend these topological methods to time-series data. By analyzing video sequences of the Sun, researchers hope to move from classification to prediction, potentially identifying the topological precursors of solar flares to provide earlier warnings for space weather events.
Summary updated