Authors:
Chiara Stuardi、Claudio Gheller、Franco Vazza、Andrea Botteon
Paper:
https://arxiv.org/abs/2408.10871
Radio U-Net: A Convolutional Neural Network to Detect Diffuse Radio Sources in Galaxy Clusters and Beyond
Introduction
The cosmic web, an intricate network of filaments interconnecting galaxy clusters, is a fundamental structure of the universe. The thermal plasma within these filaments and clusters is permeated by weak magnetic fields, which, through processes like shocks and turbulence, accelerate particles to ultra-relativistic energies. These particles emit radio synchrotron radiation, observed as diffuse radio sources in galaxy clusters. Detecting these sources is crucial for understanding cosmic magnetic fields, particle acceleration mechanisms, and the formation of large-scale cosmic structures.
However, the forthcoming generation of radio telescope arrays, such as the Square Kilometre Array (SKA), promises significant advancements in sensitivity and resolution, making manual cataloging methodologies insufficient. Radio interferometric images of diffuse sources present challenges due to noise, artifacts, and embedded radio sources. To address these challenges, the study introduces Radio U-Net, a fully convolutional neural network based on the U-Net architecture, designed to detect faint and extended sources in radio surveys.
Related Work
The detection of diffuse radio sources has historically relied on manual inspection and tailored data reduction techniques. However, the increasing volume of data from modern radio interferometers like LOFAR, MeerKAT, MWA, and ASKAP necessitates automated solutions. Deep learning (DL) has shown promise in various astronomical applications, including the detection and classification of radio galaxies. Previous work, such as COSMODEEP, explored the potential of Convolutional Neural Networks (CNN) in identifying faint signals from diffuse radio sources. However, COSMODEEP’s inability to perform accurate image segmentation limited its effectiveness.
Research Methodology
Radio U-Net is an adaptation of the original U-Net architecture, designed for image segmentation tasks. The network consists of a contractive path for downsampling and an expansive path for upsampling, with feature maps summed at each level to preserve spatial information. The final output layer returns segmented images, identifying the presence of diffuse radio emission.
The network was trained on synthetic radio observations based on cosmological simulations, which included random and systematic perturbations related to the telescope and environment. The training aimed to enable the network to recognize emission structures even below the noise level.
Experimental Design
LOFAR Observations
Radio U-Net was developed to detect diffuse radio sources in the LOFAR Two-meter Sky Survey (LoTSS). The second data release (DR2) of LoTSS covers 27% of the northern sky and consists of images at 6′′ resolution with a median root-mean-square (rms) sensitivity of 83 µJy/beam. The study used a sample of 309 galaxy clusters from the second catalog of Planck Sunyaev Zel’dovich (PSZ2) that lie within the LoTSS-DR2 area.
Synthetic Observations
Synthetic observations were generated using magneto-hydrodynamical (MHD) cosmological simulations produced with the grid code Enzo. The simulation box covered a volume of 1003 Mpc3 with a uniform spatial cell resolution of 41.65 kpc. The emission at 150 MHz from relativistic electrons accelerated by cosmic shocks was computed, assuming diffusive shock acceleration as the primary mechanism.
Training Strategy and Validation
The training set comprised 100 sky images, divided into 192×192 pixel tiles, resulting in 10,000 tiles. A fraction of these tiles (5%) was kept as a validation set, while another 10 images were used as a test set. The training involved transforming the data into a logarithmic scale, normalizing it, and processing it through the network. The categorical cross-entropy loss function was used to compare the segmented images with reference masks, and the RMSprop algorithm optimized the weights and biases.
Results and Analysis
Application to the LoTSS-DR2/PSZ2 Cluster Sample
Radio U-Net was applied to the 309 clusters in the LoTSS-DR2/PSZ2 sample. The network’s performance was evaluated based on the sum probability of all pixels within a circle of radius 2.2R500 around the cluster center. The results showed that Radio U-Net could successfully segment and detect diffuse radio sources, even in low-quality images.
Classification Accuracy
The network achieved a maximum accuracy of 73% with a recall of 83%, indicating that 17% of clusters hosting diffuse radio emission were missed. False positives were primarily due to the presence of extended radio galaxies, suggesting that the network can generalize to different types of diffuse radio emission.
Fine-Tuning and Data Augmentation
Fine-tuning the network with a subset of the LoTSS-DR2/PSZ2 dataset improved recall but not overall accuracy. Data augmentation strategies, such as rotations and flipping, were employed to increase the variety of the training sample, but further improvements were limited by the small size of the observational dataset.
Overall Conclusion
Radio U-Net demonstrates the potential of deep learning for the automated detection and segmentation of diffuse radio sources in large-area surveys. The network’s ability to generalize to different types of diffuse radio emission and its effectiveness in low-quality images highlight its utility in the era of next-generation radio telescopes. Future work will focus on creating more comprehensive simulations, including radio emissions from galaxies, and incorporating multi-wavelength data to enhance the network’s classification capabilities.
The study underscores the transformative potential of deep learning in astronomical research, paving the way for more efficient and automated analysis of vast and complex datasets.
