Compact and Extended Radio Galaxy Classification

For my first PhD project, I looked into how successful deep neural networks are in classifying radio galaxy sources according to the number of radio emission components they have. The data available were images from the now archived Radio Galaxy Zoo.

The figure shows a snapshot of the now archived Radio Galaxy Zoo project. Source: https://radio.galaxyzoo.org

Simple radio morphology

Shown to the right are a couple of examples of radio sources that are relatively simple in morphology - both at radio and infrared wavelengths with the radio contours overlaid.

More complex radio morphology

A couple of more complex examples are shown to the left. There are more radio components observed, and without additional information, it is not clear whether the different radio components belong to the same source.

The two above sets of examples are from the Radio Galaxy Zoo Data Release 1.

We initially had over 200,000 single-channel images of typical size (132,132) pixels, individual galaxy sources to work with, but no labels were provided. Therefore if we wanted to look at supervised learning approaches to help classify the sources, we needed find a way to generate labels on which to train our deep neural networks on.

It is important to note that given radio data only and without the use of data at different wavelengths to perform cross-identification of sources, it is not possible to definitively classify radio galaxies.

To make use of the available information in providing labels, we decided to use a program called the Python Blob Detector and Source Finder. (PyBDSF - https://www.astron.nl/citt/pybdsf/)

The program decomposes radio interferometry images into sources using a set of Gaussians, shapelets, or wavelets. For our work, we made the assumption that each image consists of a single radio source, therefore used PyBDSF to count the number of components attributed to a particular source.

Out of over 200,000 images available, PyBDSF was successfully able to process 175,454. The remainder of the images were not able to be processed as the majority of them contained blank pixels.

PyBDSF Gaussian breakdown

Shown to the right is the breakdown of a source into individual Gaussians using PyBDSF. By eye, the source looks to only have one component, but PyBDSF detects an extra two small components of radio emission to the bottom right. Although these components might look as if they are background noise, PyBDSF works out the overall background noise by analysing the pixel statistics in the entire image and calculating a threshold above which it is decided that the emission is true signal rather than noise. Being able to successfully make this distinction is an important problem in radio astronomy.

We utilised the successfully processed images by labelling them with the number of components they contained.

Original image on top left, the remainder show the augmented images

Given that the performance of convolutional neural network architectures is greatly improved with the use of an increasing number of images, we generated more images with the use of image augmentation, which involved the label-preserving transformations of horizontal and vertical flipping, rotation and translation.

Two-class problem

First, we wanted to see whether it would be possible to train a deep neural network to distinguish between two classes consisting of compact sources and multiple-component extended sources.

The first and second row show a small sample of compact and multiple-component sources respectively. They make up the two classes of images for our two-class classification problem.

To get to an optimised convolutional network architecture, we experimented with simple setups consisting of smaller numbers of nodes and filters, and a single convolutional layer and dense layer, and gradually increased the complexity and adjusted the parameters to see if the training and validation accuracies would improve. It is important to keep track of both and make sure they have similar losses, to make sure the network isn't overfitting.

Performance of two different architectures on the two-class problem

The figure to the right shows the performance of a couple of different architectures on the two-class problem. The use of two fully connected dense layers only results in higher losses (therefore worsened classification results). The addition of two convolutional layers along with the dense layers results in lower losses.

Best-performing architecture

Further experimentation with the architectures revealed that a 3-convolutional and 2 dense layer (3 conv + 2 dense) architecture performed the best overall.

3-conv and 2 dense layer losses on the 2-class problem

The figure to the right shows the performance of the 3 conv + 2 dense layer architecture on the two-class problem. We see that it performs slightly better in terms of training and validation losses compared to the previous architectures explored.

Four-class problem

Sample of images for the four-class problem

Next, we wanted to see how the 3 conv + 2 dense architecture would perform on a 4-class problem of classifying between compact sources, 1, 2, and >=3 -component extended sources as shown in the first, second, third and fourth rows respectively.

We can see that the network still learns to distinguish between the four classes, as shown in the right plot, however the loss is a bit higher compared to what was observed for the two-class problem. This is unsurprising as using more classes increases the potential overlap in appearance between morphologies.

Testing performance on Data Release 1 (DR1) from the Radio Galaxy Zoo

In Data Release 1 from the Radio Galaxy Zoo, citizen scientists were tasked with identifying the components that belonged to particular radio sources. As such, the labels they provided were the 'number of components' and the 'number of peaks' observed for each radio source.

In order to test our 3-conv +2 dense layer architecture on such a problem we had to adjust our labelling scheme to be as follows: Sources with one peak and one component were lumped together as compact/single-component extended sources (there was no way to make the distinction based on the data provided). Sources with two peaks and two components were 2-component extended sources, and those with >=3 components and peaks were multiple-component extended sources.

Although the test classification accuracy for such a setup was > 94%, we found that the high accuracy was influenced by having high-numbers of sources in the Compact/single-component extended class, which had the simplest morphology. The performance declined with increasing numbers of components.

It becomes increasingly difficult to correctly characterise sources having an increasing number of radio components.

This is likely due to the use of pooling, which results in some information loss, while reducing the number of parameters in a network, As such, the way the features relate to each other on a global scale tends to degrade.

A way to address this is to use Capsule Networks, which aim to preserve the hierarchical relationships between features in images.

Conclusions

We found it was possible to design a simple convolutional network architecture to distinguish radio sources based on the number of components. The network was trained on the two classes of opposites consisting of compact and multiple-component extended sources. It was then trained on the 4-class problem of distinguishing between compact, single-, two-, and multiple-extended component sources. When this trained network was applied to the images from Data Release 1 from the Radio Galaxy Zoo, it achieved high classification metrics overall.

However, upon closer inspection we found this was because the majority of sources were in the class with the simplest morphology, and increasingly complex morphologies tended to be more poorly classified. This effect is most likely due to the use of pooling layers. A way to address the issue is to use a different kind of deep learning network that preserves the way the features relate to each other, which will be addressed in the next blog post.

Below is a link to a video about the current work.