Ph.D. Dissertation Defense: Azim Ahmadzadeh

Machine Learning of Scientific Events: Detection, Classification, and Segmentation

Azim Ahmadzadeh
Advisor: Dr. Rafal Angryk

Algorithms for detection, segmentation, and classification of objects in images have been widely used in a variety of domains in the past few decades. Medical images such as MRI, CT, and X-rays are perhaps the first types of images that were exploited by these algorithms. With the recent exponential growth in the number of ground-based, air-borne, and space-borne observatories, a new generation of images have become available with two distinguishing features: they are abundant and they are not restricted by any licensing or privacy regulations. Heliophysics, the science of the Sun and the ways it impacts the solar system and the Earth, has heavily benefited from this new generation of image data. However, taking full advantage of this deluge of data with manual analysis is simply infeasible. This calls for rigorous automation of processes such as integration of data as well as detection and classification of events.

This dissertation tackles these tasks in three main parts. Part one presents the automated classification of solar events (active regions and coronal holes), with the main objective being the optimization of a set of image parameters in order to improve the classification of events using classical machine-learning algorithms. In part two, detection and segmentation of a solar event (filaments) is studied. In contrast to the first part, segmentation of these instances is carried out using deep neural networks in which the features are optimized automatically. Part three presents a large image parameter data set as the final product of the optimized parameters obtained in part one. This data set, with an ongoing extraction process, covers AIA data since January 2011 with a cadence of six minutes, resulting in roughly 1 TiB of data per year. A public web API has been designed to make the access to this dataset easier and possibly provide a good source for other interesting problems in this domain, such as tracking of solar events, or even a broader array of problems such as content-based image retrieval.

Committee
Dr. Rafal Angryk (chair)
Dr. Berkay Aydin
Dr. Dustin J. Kempton
Dr. Petrus C. Martens
Dr. Manolis K. Georgoulis (Academy of Athens, Greece)