Improvements on coronal hole detection in SDO/AIA images using supervised classification (1506.06623v1)

Abstract: We demonstrate the use of machine learning algorithms in combination with segmentation techniques in order to distinguish coronal holes and filaments in SDO/AIA EUV images of the Sun. Based on two coronal hole detection techniques (intensity-based thresholding, SPoCA), we prepared data sets of manually labeled coronal hole and filament channel regions present on the Sun during the time range 2011 - 2013. By mapping the extracted regions from EUV observations onto HMI line-of-sight magnetograms we also include their magnetic characteristics. We computed shape measures from the segmented binary maps as well as first order and second order texture statistics from the segmented regions in the EUV images and magnetograms. These attributes were used for data mining investigations to identify the most performant rule to differentiate between coronal holes and filament channels. We applied several classifiers, namely Support Vector Machine, Linear Support Vector Machine, Decision Tree, and Random Forest and found that all classification rules achieve good results in general, with linear SVM providing the best performances (with a true skill statistic of ~0.90). Additional information from magnetic field data systematically improves the performance across all four classifiers for the SPoCA detection. Since the calculation is inexpensive in computing time, this approach is well suited for applications on real-time data. This study demonstrates how a machine learning approach may help improve upon an unsupervised feature extraction method.

• The paper introduces a supervised classification method that leverages image intensity, texture, and magnetic field data to accurately identify coronal holes.
• It employs both intensity-based thresholding and SPoCA segmentation techniques to effectively distinguish coronal holes from filament channels.
• Advanced classifiers, especially Linear SVM, achieved higher True Skill Statistic scores, indicating strong potential for real-time solar wind forecasting.

Improvements on Coronal Hole Detection in SDO/AIA Images Using Supervised Classification

Introduction

The paper discusses the implementation of machine learning algorithms combined with segmentation techniques to distinguish coronal holes and filament channels in SDO/AIA EUV images of the Sun. The goal is to improve upon traditional unsupervised feature extraction methods used in the identification of coronal holes, which are vital for understanding space weather phenomena. The focus is on two segmentation techniques: intensity-based thresholding and SPoCA.

Data Preparation

Two datasets were prepared using segmentation techniques applied to SDO/AIA 19.3 nm images over a period spanning 2011 to 2013. Coronal holes, discerned by their low intensity, were manually labeled by cross-referencing with Kanzelhöhe Observatory Hα filtergrams, which clearly show filament channels, thereby aiding in accurate classification.

Figure 1: Analysis of the detected low intensity regions. (a) AIA 19.3 nm image showing a filament channel; (b) Detected filament channel from EUV image; (c) Spatial relationship of pixel values in the channel.

Feature Extraction

Attributes such as shape measures, magnetic flux properties, and first and second order image statistics were extracted from the SDO/AIA images and corresponding HMI line-of-sight magnetograms. Texture features were derived using co-occurrence matrices to characterize spatial arrangements of pixel values, providing a comprehensive dataset for classification.

Figure 2: Representation of the co-occurrence matrix $C(i,j)$ calculated from the filament channel.

Classification Techniques

Four classifiers were evaluated: Linear SVM, SVM, Decision Tree, and Random Forest. The classification aimed to separate coronal holes from filament channels effectively, with the datasets segmented by both intensity thresholding and the SPoCA algorithm. Performance was quantified using the True Skill Statistic (TSS).

Results

The results showed that both SVM and Linear SVM classifiers achieved superior performance, with Linear SVM slightly outperforming others in terms of median TSS. Importantly, integrating magnetic field data systematically enhanced classification accuracy, particularly when using SPoCA detection maps.

Figure 3: Density plot of TPR versus FPR from intensity-based segmentation using all attributes.

Figure 4: Density plot of TPR versus FPR using the SPoCA algorithm.

Figure 5: Box plots of TSS for different classifiers and segmentation techniques.

Figure 6: ROC curves for SVM and Linear SVM classifiers with calculated AUC.

Discussion and Conclusion

The integration of machine learning with image segmentation techniques markedly improves the identification and classification of coronal holes in solar images. Key findings indicate that using both image intensity and magnetic field data enhances classifier performance, facilitating the discrimination of coronal holes from darker filament channels. The Linear SVM classifier, in particular, provides a practical, computationally efficient solution for real-time applications, with potential implementation in existing solar wind forecasting tools. Future work aims to refine these methodologies and explore their application across a broader set of solar imaging data. The paper represents a significant step in the automation of solar feature classification, leveraging advanced data-driven techniques to improve our understanding of solar dynamics and space weather forecasting.