| Abstract |
Deep learning architectures have revolutionized research in numerous scientific domains and triggered a paradigm shift from traditional machine learning methodologies and feature engineering to architecture design and the so-called "end-to-end" training. While the efficacy of deep learning networks can be strongly attributed to their vigorous capacity of extracting aggregated knowledge, as the size of the available data increases, at the same time they exhibit an underwhelming performance when trained with a limited amount of annotated examples. Our main aim, in this thesis, is to explore the impact of the utilization of state-of-the-art deep learning methodologies, in both cases of data abundance and data deficiency, in two major research topics in the fields of cosmology and remote sensing.
In the first case study, we address the problem of spectroscopic redshift estimation in astronomy. Due to the expansion of the Universe and its statistical homogeneity and isotropy, galaxies recede from each other on average. This movement causes the emitted electromagnetic waves to shift from the blue part of the spectrum to the red part, due to the Doppler effect. This redshift is one of the most important observables in astronomy and cosmology, allowing the measurement of galaxy distances. Several sources of astrophysical and instrumental noise render the estimation process far from trivial, especially in the low signal-to-noise regime of many astrophysical observations. In recent years, new approaches for a reliable and automated methodology of the redshift evaluation have been sought out, in order to minimize our reliance on currently popular techniques that heavily involve human intervention. The fulfillment of this task has evolved into a grave necessity, in conjunction with the insatiable generation of immense amounts of astronomical data, falling into the category of the so-called Big Data. We propose an alternative approach that transforms the issue at hand from a regression problem to a multi-class classification task, opening the field for the deployment of a currently dominating deep learning classifier, commonly known as Deep Convolutional Neural Networks. This approach is extensively evaluated on a spectroscopic dataset of full spectral energy galaxy distributions, modelled after the upcoming Euclid satellite galaxy survey. Experimental analysis on observations of idealistic and realistic conditions demonstrate the potent capabilities of the proposed scheme.
In the second case study, we examine a flourishing research topic in the field of remote sensing, namely land cover classification. Conventional methodologies mainly focus either on the simplified single-label scenario or on pixel-based approaches that cannot efficiently handle high resolution images. On the other hand, the problem of multi-label land cover scene categorization remains, to this day, fairly unexplored. While deep learning and Convolutional Neural Networks have demonstrated an astounding capacity at handling challenging image classification tasks, they significantly underperform when trained on limited in size datasets. To overcome this issue, we propose an online data augmentation technique that can drastically increase the size of a smaller dataset to copious amounts. Our experiments on a multi-label variation of the UC Merced Land Use dataset demonstrates the potential of the proposed methodology, which outperforms the current state-of-the-art by more than 6% in terms of the F-score metric.
|
Deep learning techniques in signal processing
