Skip to main content
Top

2022 | Book

Deep Learning in Solar Astronomy

Authors: Prof. Long Xu, Prof. Yihua Yan, Dr. Xin Huang

Publisher: Springer Nature Singapore

Book Series : SpringerBriefs in Computer Science

share
SHARE
insite
SEARCH

About this book

The volume of data being collected in solar astronomy has exponentially increased over the past decade and we will be entering the age of petabyte solar data. Deep learning has been an invaluable tool exploited to efficiently extract key information from the massive solar observation data, to solve the tasks of data archiving/classification, object detection and recognition.

Astronomical study starts with imaging from recorded raw data, followed by image processing, such as image reconstruction, inpainting and generation, to enhance imaging quality. We study deep learning for solar image processing. First, image deconvolution is investigated for synthesis aperture imaging. Second, image inpainting is explored to repair over-saturated solar image due to light intensity beyond threshold of optical lens. Third, image translation among UV/EUV observation of the chromosphere/corona, Ha observation of the chromosphere and magnetogram of the photosphere is realized by using GAN, exhibiting powerful image domain transfer ability among multiple wavebands and different observation devices. It can compensate the lack of observation time or waveband. In addition, time series model, e.g., LSTM, is exploited to forecast solar burst and solar activity indices.

This book presents a comprehensive overview of the deep learning applications in solar astronomy. It is suitable for the students and young researchers who are major in astronomy and computer science, especially interdisciplinary research of them.

Table of Contents

Frontmatter
Chapter 1. Introduction
Abstract
Deep leaning has been developing very fast in recent years due to big data, high-performance computing and the breakthrough of neural network training techniques. It has been particularly successful in computer vision, machine translation, speech recognition and natural language processing. Modern astronomy concerns a big data challenge owning to high-resolution, high-precision and high-cadence telescopes. The big data presents a great challenge to data processing, statistical analysis and scientific discovery. Therefore, it is highly demanded to develop artificial intelligent algorithms to process big data aromatically, further discover complex relationship and mine knowledge hidden in massive data. As the best representative of artificial intelligence, a bunch of classical models have been developed for processing single image, video, speech and natural language. Among of them, convolutional neural network has been verified most efficient for processing image. To process time series input, like video, recurrent neural network, e.g., long short-term memory (LSTM), was developed, which was widely known for forecasting the future, e.g., event occurrence, physical parameter prediction. An overview of artificial intelligence, deep learning and astronomical big data is presented in this chapter, as the background of this book.
Long Xu, Yihua Yan, Xin Huang
Chapter 2. Classical Deep Learning Models
Abstract
Deep learning has achieved a big success in computer vision, NLP, audio processing and machine translation. Accordingly, there have been a bunch of classical deep learning models designed for these tasks. In this chapter, convolutional neural network (CNN), LSTM, autoencoder (AE) and GAN are discussed briefly. These models are most efficient for processing image, time series (e.g., video, NLP) and image generation respectively, as the foundation of our proposed models in this book. Recently, more advanced deep learning models/principles have emerged, such as attention (e.g., non-local, squeeze and excitation (SE), global context (GC), and most popular transformer), graph convolution network (GCN), self-supervised learning and contrastive learning. They can further boost model performance, extend application filed and break the limits of lack of labelled data, noise data and etc.
Long Xu, Yihua Yan, Xin Huang
Chapter 3. Deep Learning in Solar Image Classification Tasks
Abstract
The exponential increasing of data being collected in astronomy has raised a big data challenge. Mining valuable information timely and efficiently from massive raw data is highly demanded. Even simple binary classification of collected raw data is of great importance, reducing the burden of the following data processing. Inspired by the success of image classification with deep learning, we investigated solar radio spectrum classification using deep learning, including the premier deep belief network (DBN), the most popular convolutional neural network (CNN) and long short-term memory (LSTM) network. For model training, a database of solar spectrum was established and published to the public. As far as we know, it is the first one in the world. The database contains 8816 spectrums with different image patterns which represent different solar radio mechanisms. Then, each spectrum was given a label by the invited experts of solar radio astronomy.
Long Xu, Yihua Yan, Xin Huang
Chapter 4. Deep Learning in Solar Object Detection Tasks
Abstract
Solar observation provides us abundant solar images containing plentiful information about solar activities. Especially, solar instruments onboard satellite continuously record high-resolution and high-cadence full-disk solar images. These images are used for solar activity forecasting and statistical analysis. Usually, it is required to mine key information from full-disk images firstly. Then, over extracted information, one can establish classification, recognition or forecasting models by using machine learning or deep learning. In a full-disk solar image, active region, filament, coronal hole and sunspot are the objects carrying major information about solar activities. In computer vision, object detection is one of the most classical tasks, which has been well investigated. In this chapter, we present two examples of object detection from solar image, by using well pre-trained deep learning models in computer vision.
Long Xu, Yihua Yan, Xin Huang
Chapter 5. Deep Learning in Solar Image Generation Tasks
Abstract
It has been witnessed that deep learning has been applied to classification in previous chapters. In fact, deep learning also demonstrated great ability of image generation which is more challenging than classification. In this chapter, several applications of deep learning in solar image enhancement, reconstruction and processing are presented, including image deconvolution of solar radioheliograph, desaturation of solar imaging, generating magnetogram, image super-resolution. These tasks are all concerned with image generation, by employing generative neural networks. As a representative of generative networks, GAN was widely exploited in image generation tasks. It can generate high fidelity and photo-realistic content mainly owning to an adversarial loss.
Long Xu, Yihua Yan, Xin Huang
Chapter 6. Deep Learning in Solar Forecasting Tasks
Abstract
Besides classification and generation, deep learning is also applicable to time series analysis. Unlike CNN which accepts singe image input, RNN is specifically designed for handling time series input, e.g., video sequence, natural language processing. As the best representative of RNN, LSTM has been widely exploited in various of time series analysis, achieving big success. In this chapter, it is applied to solar activity/event forecasting and solar radiation index prediction. As one of the most violent solar eruptions, solar flare is the main driving source of catastrophic space weather, so forecasting of solar flare is of great importance. The solar radio flux of 10.7 cm is a typical index for measuring global solar activity. It is a typical indicator of long-term space weather.
Long Xu, Yihua Yan, Xin Huang
Metadata
Title
Deep Learning in Solar Astronomy
Authors
Prof. Long Xu
Prof. Yihua Yan
Dr. Xin Huang
Copyright Year
2022
Publisher
Springer Nature Singapore
Electronic ISBN
978-981-19-2746-1
Print ISBN
978-981-19-2745-4
DOI
https://doi.org/10.1007/978-981-19-2746-1

Premium Partner