随着新射电望远镜设施数据的增长，用于射电星系形态分类的机器学习方法正得到越来越多的应用。然而，尽管广泛使用的基于卷积神经网络（CNNs）的深度学习模型对图像内的平移具有等变性，CNNs和大多数其他机器学习方法对欧几里得平面的其他等距变换（如旋转和反射）却不具有等变性。近期的工作试图通过使用G-steerable CNNs来解决这一问题，该网络被设计为对二维欧几里得E(2)变换的特定子集具有等变性。尽管这种方法提高了模型性能，但其计算成本是一个公认的缺点。在此，我们考虑使用直接提取的E(2)-等变特征来进行射电星系的分类。具体而言，我们研究了Minkowski functionals (MFs)、Haralick features (HFs)和elliptical Fourier descriptors (EFDs)的使用。我们表明，尽管这些特征在准确率方面不如CNNs，但它们能够为射电星系的分类提供信息，且所需的计算运行时间约少50倍。我们证明MFs提供的信息量最大，EFDs最小，并表明将所有三种特征组合使用仅能带来渐进的性能提升，我们认为这是由于特征集之间存在信息重叠。
英文摘要：
With the growth of data from new radio telescope facilities, machine-learning approaches to the morphological classification of radio galaxies are increasingly being utilised. However, while widely employed deep-learning models using convolutional neural networks (CNNs) are equivariant to translations within images, neither CNNs nor most other machine-learning approaches are equivariant to additional isometries of the Euclidean plane, such as rotations and reflections. Recent work has attempted to address this by using G-steerable CNNs, designed to be equivariant to a specified subset of 2-dimensional Euclidean, E(2), transformations. Although this approach improved model performance, the computational costs were a recognised drawback. Here we consider the use of directly extracted E(2)-equivariant features for the classification of radio galaxies. Specifically, we investigate the use of Minkowski functionals (MFs), Haralick features (HFs) and elliptical Fourier descriptors (EFDs). We show that, while these features do not perform equivalently well to CNNs in terms of accuracy, they are able to inform the classification of radio galaxies, requiring ~50 times less computational runtime. We demonstrate that MFs are the most informative, EFDs the least informative, and show that combinations of all three result in only incrementally improved performance, which we suggest is due to information overlap between feature sets.
Weight sharing in convolutional neural networks (CNNs) ensures that their feature maps will be translation-equivariant. However, although conventional convolutions are equivariant to translation, they are not equivariant to other isometries of the input image data, such as rotation and reflection. For the classification of astronomical objects such as radio galaxies, which are expected statistically to be globally orientation invariant, this lack of dihedral equivariance means that a conventional CNN must learn explicitly to classify all rotated versions of a particular type of object individu…
In this work we introduce group-equivariant self-attention models to address the problem of explainable radio galaxy classification in astronomy. We evaluate various orders of both cyclic and dihedral equivariance, and show that including equivariance as a prior both reduces the number of epochs required to fit the data and results in improved performance. We highlight the benefits of equivariance when using self-attention as an explainable model and illustrate how equivariant models statistically attend the same features in their classifications as human astronomers.
A technique is described that is used to improve the detection of radio-frequency interference in astronomical radio observatories. It is applied on a two-dimensional interference mask after regular detection in the time-frequency domain with existing techniques. The scale-invariant rank (SIR) operator is defined, which is a one-dimensional mathematical morphology technique that can be used to find adjacent intervals in the time or frequency domain that are likely to be affected by RFI. The technique might also be applicable in other areas in which morphological scale-invariant behaviour is de…
Morphologically classifying radio sources in continuum images with the SKA has the potential to address some of the key questions in cosmology and galaxy evolution. In particular, we may use different classes of radio sources as independent tracers of the dark-matter density field, and thus overcome cosmic variance in measuring large-scale structure, while on the galaxy evolution side we could measure the mechanical feedback from FRII and FRI jets. This work makes use of a \texttt{MeqTrees}-based simulations framework to forecast the ability of the SKA to recover true source morphologies at hi…
Transformer-Based Source Detection and Morphological Classification in LOFAR Deep-Field Continuum Images
University of Oxford（牛津大学）；Instituto de Astrofísica de Canarias（加那利天体物理研究所）；Universidad de La Laguna（拉古纳大学）；University of Malta（马耳他大学）；Anqing Normal University（安庆师范大学）；University of Science and Technology of China（中国科学技术大学）
Technical University of Košice（科希策技术大学）；South African Radio Astronomy Observatory（南非射电天文台）
Machine learning techniques have been increasingly used in astronomical applications and have proven to successfully classify objects in image data with high accuracy. The current work uses archival data from the Faint Images of the Radio Sky at Twenty Centimeters (FIRST) to classify radio galaxies into four classes: Fanaroff-Riley Class I (FRI), Fanaroff-Riley Class II (FRII), Bent-Tailed (BENT), and Compact (COMPT). The model presented in this work is based on Convolutional Neural Networks (CNNs). The proposed architecture comprises three parallel blocks of convolutional layers combined and …
Extended radio sources are an important minority population in modern deep radio surveys, because they enable detailed investigation of the physics governing radio-emitting regions such as active galaxies and their environments. Cross-identification of radio sources with optical host galaxies is challenging for this extended population, due to their morphological complexity and multiple potential counterparts. In the first data release of the Low-frequency array (LOFAR) Two-metre Sky Survey (LoTSS DR1) the automated likelihood ratio for compact sources was supplemented by a citizen science vis…
Next-generation radio surveys will yield an unprecedented amount of data, warranting analysis by use of machine learning techniques. Convolutional neural networks are the deep learning technique that has proven to be the most successful in classifying image data. Capsule networks are a more recently developed technique that use capsules comprised of groups of neurons, that describe properties of an image including the relative spatial locations of features. The current work explores the performance of different capsule network architectures against simpler convolutional neural network architec…
We demonstrate the use of Haralick features for the automated classification of radio galaxies. The set of thirteen Haralick features represent an extremely compact non-parametric representation of image texture, and are calculated directly from imagery using the Grey Level Co-occurrence Matrix (GLCM). The GLCM is an encoding of the relationship between the intensity of neighbouring pixels in an image. Using 10,000 sources detected in the first data release of the LOFAR Two-metre Sky Survey (LoTSS), we demonstrate that Haralick features are highly efficient, rotationally invariant descriptors …
We present the application of deep machine learning technique to classify radio images of extended sources on a morphological basis using convolutional neural networks. In this study, we have taken the case of Fanaroff-Riley (FR) class of radio galaxies as well as radio galaxies with bent-tailed morphology. We have used archival data from the Very Large Array (VLA) - Faint Images of the Radio Sky at Twenty Centimeters (FIRST) survey and existing visually classified samples available in literature to train a neural network for morphological classification of these categories of radio sources. O…
Deep Learning for Morphological Identification of Extended Radio Galaxies using Weak Labels
CSIRO Space & Astronomy（澳大利亚联邦科学与工业研究组织空间与天文研究所）；CSIRO Data61（澳大利亚联邦科学与工业研究组织Data61研究所）；Western Sydney University（西悉尼大学）；International Centre for Radio Astronomy Research (ICRAR)（国际射电天文学研究中心）；The University of Western Australia（西澳大利亚大学）；Thüringer Landessternwarte（图林根州立天文台）；Universidad de Guanajuato（瓜纳华托大学）
With the advent of large scale surveys the manual analysis and classification of individual radio source morphologies is rendered impossible as existing approaches do not scale. The analysis of complex morphological features in the spatial domain is a particularly important task. Here we discuss the challenges of transferring crowdsourced labels obtained from the Radio Galaxy Zoo project and introduce a proper transfer mechanism via quantile random forest regression. By using parallelized rotation and flipping invariant Kohonen-maps, image cubes of Radio Galaxy Zoo selected galaxies formed fro…
Novel techniques are indispensable to process the flood of data from the new generation of radio telescopes. In particular, the classification of astronomical sources in images is challenging. Morphological classification of radio galaxies could be automated with deep learning models that require large sets of labelled training data. Here, we demonstrate the use of generative models, specifically Wasserstein GANs (wGAN), to generate artificial data for different classes of radio galaxies. Subsequently, we augment the training data with images from our wGAN. We find that a simple fully-connecte…
Advances on the classification of radio image cubes
Bernoulli Institute for Mathematics, Computer Science and Artificial Intelligence（伯努利数学、计算机科学与人工智能研究所）；University of Groningen（格罗宁根大学）；University of Stellenbosch（斯泰伦博斯大学）；ASTRON（荷兰射电天文研究所）
