Author ORCID Identifier

0000-0002-1631-5336

Date of Award

5-4-2021

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computer Science

First Advisor

Rafal A. Angryk

Second Advisor

Berkay Aydin

Third Advisor

Dustin J. Kempton

Fourth Advisor

Petrus C. Martens

Fifth Advisor

Manolis K. Georgoulis

Abstract

Classification and segmentation of objects using machine learning algorithms have been widely used in a large variety of scientific domains in the past few decades. With the exponential growth in the number of ground-based, air-borne, and space-borne observatories, Heliophysics has been taking full advantage of such algorithms in many automated tasks, and obtained valuable knowledge by detecting solar events and analyzing the big-picture patterns. Despite the fact that in many cases, the strengths of the general-purpose algorithms seem to be transferable to problems of scientific domains where scientific events are of interest, in practice there are some critical issues which I address in this dissertation. First, I discuss the four main categories of such issues and then in the proceeding chapters I present real-world examples and the different approaches I take for tackling them. In Chapter II, I take a classical path for classification of three solar events; Active Regions, Coronal Holes, and Quiet Suns. I optimize a set of ten image parameters and improve the classification performance by up to 36%. In Chapter III, in contrast, I utilize an automated feature extraction algorithm, i.e., a deep neural network, for detection and segmentation of another solar event, namely solar Filaments. Using an off-the-shelf algorithm, I overcome several of the issues of the existing detection module, while facing an important challenge; lack of an appropriate evaluation metric for verification of the segmentations. In Chapter IV, I introduce a novel metric to provide a more accurate verification especially for salient objects with fine structures. This metric, called Multi-Scale Intersection over Union (MIoU), is a fusion of two concepts; fractal dimension from Geometry, and Intersection over Union (IoU) which is a popular metric for segmentation verification. Through several experiments I examine the advantages of using MIoU over IoU, and I conclude this chapter by a follow-through on the segmentation results of the previously implemented filament detection module.

File Upload Confirmation

1

Share

COinS