Solar activity detection and prediction using image processing and machine learning techniques

The objective of the research in this dissertation is to develop the methods for automatic detection and prediction of solar activities, including prominence eruptions, emerging flux regions and solar flares. Image processing and machine learning techniques are applied in this study. These methods can be used for automatic observation of solar activities and prediction of space weather that may have great influence on the near earth environment. The research presented in this dissertation covers the following topics: i) automatic detection of prominence eruptions (PBs), ii) automatic detection of emerging flux regions (EFRs), and iii) automatic prediction of solar flares. In detection of prominence eruptions, an automated method is developed by combining image processing and pattern recognition techniques. Consecutive Hu solar images are used as the input. The image processing techniques, including image transformation, segmentation and morphological operations are used to extract the limb objects and measure the associated properties. The pattern recognition techniques, such as Support Vector Machine (SVM), are applied to classify all the objects and generate a list of identified the PBs as the output. In detection of emerging flux regions, an automatic detection method is developed by using multi-scale circular harmonic filters, Kalman filter and SVM. The method takes a sequence of consecutive Michelson Doppler Imager (MDI) magnetograms as the input. The multi-scale circular harmonic filters are applied to detect bipolar regions from the solar disk surface and these regions are traced by Kalman filter until their disappearance. Finally, a SVM classifier is applied to distinguish EFRs from the other regions based on statistical properties. In solar flare prediction, it is modeled as a conditional density estimation (CDE) problem. A novel method is proposed to solve the CDE problem using kernel-based nonlinear regression and moment-based density function reconstruction techniques. This method involves two main steps. In the first step, kernel-based nonlinear regression techniques are applied to predict the conditional moments of the target variable, such as flare peak intensity or flare index. In the second step, the condition density function is reconstructed based on the estimated moments. The method is compared with the traditional double-kernel density estimator, and the experimental results show that it yields the comparable performance of the double-kernel density estimator. The most important merit of this new method is that it can handle high dimensional data effectively, while the double-kernel density estimator has confined to the bivariate case due to the difficulty of determining optimal bandwidths. The method can be used to predict the conditional density function of either flare peak intensity or flare index, which shows that our method is of practical significance in automated flare forecasting