RGC: a radio AGN classifier based on deep learning. I. A semi-supervised multiclass model for VLA images
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Astrophysics. The day in astrophysics: observations, models and what the data forces.
Vera: We're looking at the paper: "RGC: a radio AGN classifier based on deep learning. I. A semi-supervised multiclass model for VLA images".
Jocelyn: Here's the gist of the paper. Bent radio active galactic nuclei (RAGNs)—wide-angle tails (WATs) and narrow-angle tails (NATs)—trace dense environments in galaxy groups and clusters, yet no multiclass classifier simultaneously separates them from straight Fanaroff–Riley types (sFRI, sFRII) using visually inspected labels and unlabelled data The four-class scheme introduced here enables WAT/NAT-resolved catalogues that can serve as environment probes and progenitor classifications for diffuse cluster radio emission Dataset Construction The study introduces FIRST-2060, a four-class labelled dataset of 2060 RAGNs (sFRI, sFRII, WAT, NAT) constructed from three publicly available catalogues through multi-tier visual inspection This dataset is provided in two preprocessing variants: RL1, which retains spurious sources, and RL2, from which they are removed The labelled dataset RL is constructed by combining selected sources from three catalogues: MiraBest (Porter & Scaife 2023), the Sasmal catalogue (Sasmal et al.
Vera: 2022), and the Lao catalogue (Lao et al.
Jocelyn: 2025) Visual inspection involved six co-authors, with four individuals each assigned to one of the four target classes The classification was based primarily on the opening angle θOA, where sources with θOA ≈ 180◦ were labelled straight, those with 90◦ ≤ θOA ≤ 180◦ were labelled WAT, and those with θOA ≤ 90◦ were labelled NAT Model Architecture and Training The RGC model (M2) integrates the self-supervised framework BYOL (Bootstrap Your Own Latent) with an E(2)-equivariant steerable CNN (E2CNN) encoder, pre-trained on the unlabelled data and fine-tuned on the labelled sets The model operates in two stages: a task-agnostic self-supervised pretraining stage on RU, followed by a task-specific supervised finetuning stage on RL1 and RL2 In the pre-training stage, BYOL learns a representation by training an online network—comprising an encoder fθ, projector gθ, and predictor qθ—against a momentum-updated target network with weights ξ The encoder fθ is E2CNN, which preserves invariance under the Euclidean group E(2) Performance Metrics and Model Ranking The RGC model M2 ranks second overall with a macro-F1 of 0.79 ± 0.02, just 0.01 behind M1 (0.80 ± 0.02) The per-class F1-scores of M2 reflect this competitiveness: on RL1, it achieves sFRI 0.75 ± 0.05, sFRII 0.80 ± 0.03, WAT 0.77 ± 0.02, and NAT 0.83±0.4 The most consistent pattern across all models is the class-dependent difficulty of the task NAT achieves the highest per-class F1-score for every model on both datasets—reaching ∼ 0.84 ± 0.02 for M1 on RL1—while sFRI is the hardest class, with F1-scores ranging from 0.61 ± 0.06 (M6) to 0.76 ± 0.05 (M1, RL2) Morphological Fidelity Analysis Grad-CAM analysis shows that M2 is the only model whose attention contours consistently trace the morphological structure of RAGNs—lobes, jets, and bends—rather than defaulting to compact blobs or diffuse patterns The structural similarity index (SSIM) between the normalized heatmap and the 3σ source mask is highest for M2 and M6 The contrast is most pronounced for bent and extended sources (e.g.
Vera: L0809, L2121), where M2’s contours bend with the source while all other models default to circular or image-filling patterns Astrophysical Relevance The four-class scheme adds a dimension that existing binary or three-class classifiers lack: by resolving bent sources into WATs and NATs rather than treating them as a single undifferentiated class, it enables a first-order separation between the BCGs that typically host WATs (O’Dea & Baum 2023) and the satellite galaxies that typically host NATs (Miley et al.
Paper summary: Vera: So we're looking at this paper now called "RGC: a radio AGN classifier based on deep learning. I. A semi-supervised multiclass model for VLA images." The main idea here is that they built a model to sort radio active galactic nuclei, or RAGNs, into four classes using deep learning.
Jocelyn: Exactly. They’re dealing with bent radio galaxies, which we call wide-angle tails, or WATs and narrow-angle tails, NATs—and they want to separate those from the straight Fanaroff–Riley types, sFRI and sFRII.
Subrahmanyan: That is a big problem for classification because visually inspecting these sources can be hard to do consistently across different surveys or teams. So the goal is to see if a machine can do what people are trying to do manually but faster and more consistently.
Vera: The paper claims they built this four-class labelled dataset, called FIRST-two thousand sixty from three public catalogues using visual inspection by six co-authors <ref:2510.22190#pg1>. They also used a model called RGC one point zero that uses twenty thousand unlabelled sources to help train the main classifier <ref:2510.22190#pg1>.
Jocelyn: It's interesting because they are using a semi-supervised learning approach, which means they use those unlabelled sources alongside the labelled ones to get better results than just training on the labelled data alone.
Subrahmanyan: That’s smart because in real astronomy, we don't always have perfectly labelled data for everything, so leveraging that large pool of unlabelled sources is a way to make the model more robust against those gaps.
Vera: The model they focus on is RGC M2, which uses a framework called BYOL and an E(two)-equivariant steerable CNN encoder <ref:2510.22190#pg1>. That means the network is designed to understand rotation and orientation in space, which is key for these tailed sources.
Jocelyn: And the performance numbers are pretty solid. They say that RGC M2 ranks second overall with a macro F1 score of zero point seven nine plus or minus zero point two just behind model M1 which gets a macro F1 of zero point eight zero plus or minus zero point two.
Paper summary: Subrahmanyan: That ranking tells us that the deep learning approach is performing well when compared to other methods they tested, especially against supervised baselines. But we need to look at what those F1 scores actually mean for the individual classes, not just the overall number.
Vera: Right, because I see that NAT sources get the highest per-class F1 score across all models—reaching around zero point eighty-four plus or minus zero point two for M1 on RL1. That suggests that classifying those narrow-angle tails is where the model is most confident.
Jocelyn: But then you have sFRI sources, which are consistently the hardest class, with F1 scores ranging from a low of zero point sixty-one plus or minus zero point six for M6 to about zero point seventy-six plus or minus zero point five for M1 on RL2.
Subrahmanyan: That's a lot of variability, which shows the difficulty in separating those straight types from the bent ones, even with these sophisticated models. What I wonder is how much that error bar actually matters when you are trying to classify something for environmental studies.
Vera: Well, this leads into what they call morphological fidelity analysis using Grad-CAM attention maps. They found that RGC M2 is the only model where the attention contours consistently trace the actual shape of RAGNs—the lobes, jets, and bends—instead of just picking up random blobs or diffuse patterns.
Jocelyn: That’s a strong point because it means the model isn't just guessing based on color or brightness; it’s actually looking at the structure. The structural similarity index between the heatmap and the actual source mask is highest for M2 and M6.
Subrahmanyan: So if we want to use this to probe environments, we need a model that understands the physical structure—the jets and tails—not just some statistical pattern on a detector image. That makes sense for connecting it to the physics of the galaxy group environment.
Vera: And they are using this capability because they think it matters for real-world applications. The four-class scheme allows them to create catalogues that can act as probes for the environment and as progenitor classifications for diffuse cluster radio emission.
Paper summary: Jocelyn: So what does this mean practically? It means we can start building catalogues that tell us not just where a galaxy is, but what kind of radio emission it might be producing, which helps us understand the fossil plasma supply chain.
Subrahmanyan: That connects directly to the bigger picture—automated classification at survey scale could let us study which types of progenitor sources seed which classes of diffuse radio emission we see in clusters. It gives a way to statistically link them all together.
Vera: And they have a couple of models, M1 and M2, that show complementary strengths. M1 is better at cross-type separation—telling sFRI from WATs—while M2 is better at within-type discrimination—telling two types of tails apart when they are both present.
Jocelyn: So for a survey where you have a huge mix of sources, you might actually want to use an ensemble approach, combining the strengths of both models rather than just sticking with one.
Subrahmanyan: That’s a practical engineering thought about how to deploy this kind of classifier across massive datasets, which is exactly what we need if we are going to scale this up for things like EMU or the SKA.
Vera: And as a final note, they pointed out that RGC M2 is also the best-calibrated model by a substantial margin. It achieves an expected calibration error of zero point seven plus or minus zero point one on RL1 and about zero point eight plus or minus zero point one on RL2.
Jocelyn: So the final message is that if you want to extract quantitative environmental information from classifier internals at survey scale, like bending curvature or jet tail transitions, you need a model like M2 because it actually pays attention to those physical features.
Subrahmanyan: It’s about moving away from just seeing a result and starting to understand the mechanism of how the machine arrived at that classification. That is where the real cosmological insight lies.
Conclusion: Vera: So we've been looking at how this new model, RGC, tackles classifying radio active galactic nuclei based on their shape. Now we're wrapping up with a look at the whole paper title and what this actually means for us in the sky.
Jocelyn: The title is "RGC: a radio AGN classifier based on deep learning I. A semi-supervised multiclass model for VLA images." It tells you right away that this is a computer model, not just another old way of looking at these radio sources.
Subrahmanyan: Specifically, it’s using deep learning to classify them into four classes—sFRI, sFRII, WAT and NAT—and they are using a semi-supervised method because they have both labelled and unlabelled data.
Vera: That semi-supervised part is important because it means the model isn't just guessing from what we explicitly told it; it’s learning from thousands of other radio sources that we didn't label ourselves.
Jocelyn: And what this means for us is that we can start creating catalogues of these tailed sources—the WAT and NAT ones—that are resolved, meaning they aren't just lumped into one big category.
Subrahmanyan: Exactly. If you get a catalogue that separates the straight types from the tails, it becomes a probe for the environment around those galaxies.
Vera: That’s right because those tails trace dense environments in galaxy groups and clusters, which is where we want to study structure formation.
Jocelyn: And they found that this four-class scheme lets us distinguish between the central bright galaxies that usually host wide-angle tails and the satellite galaxies that host narrow-angle tails.
Subrahmanyan: That distinction helps us understand how these different types of galaxies interact within a larger cosmic structure.
Vera: The whole point is to move from just seeing a radio source to understanding what kind of environment it lives in based on its shape.
Jocelyn: It also gives us a way to look at the fossil plasma supply chain, which is how we understand the material feeding those active galactic nuclei.
Subrahmanyan: So this paper sets up a framework where automated classification can help us statistically link the type of central source to the type of diffuse radio emission surrounding it.
Vera: It's a big step toward using machine learning not just for discovery, but for understanding the physical processes happening out in those deep cluster environments.
M. S. Hossain, M. S. H. Shahal, K. M. B. Asad, P. Saikia, A. Khan, F. Akter, A A Ali, M A Amin, D P Guha, M O B Jihad, A Momen, S Sen, and A K M M Rahman
Center for Computational and Data Sciences, Independent University of Bangladesh Center for Astronomy Space Science and Astrophysics Department of Physical Sciences Department of Computer Science and Engineering Department of Astronomy and Physics Yale University Department of Agricultural and Biosystems Engineering North Dakota State University
astro-ph.IM, astro-ph.CO, cs.LG
Submitted: 2025-10-25
Updated: 2026-05-26
Comments: 12 pages, 8 pages appendix, 7 figures, re-submitted to A&A
Journal ref: A&A 714, A5 (2026)
DOI: 10.1051/0004-6361/202557828
Code: https://github.com/inigoval/fixmatch
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 87/100
The gist: Bent radio active galactic nuclei (RAGNs)—wide-angle tails (WATs) and narrow-angle tails (NATs)—trace dense environments in galaxy groups and clusters…
Key concepts
- FIRST-2060
- This is a new, four-class dataset containing 2060 RAGNs. It combines sources from three existing catalogues using careful visual inspection. The four classes are sFRI, sFRII (straight types), WAT (wide-angle tails), and NAT (narrow-angle tails). This dataset allows for testing classifiers on different morphological types.
- RGC Model
- The RGC model is a deep learning classifier that uses a self-supervised method called BYOL combined with an E2CNN encoder. It learns from both labelled and unlabelled data. It is designed to classify RAGNs into the four categories, aiming to be robust even when trained on limited labeled examples.
- E2CNN Encoder
- The E2CNN is a specialized convolutional neural network (CNN) encoder that respects the Euclidean group E(2). This means it is designed to understand and preserve geometric properties of radio sources, such as their shape, orientation, and curvature. It helps the model focus on the physical structure of RAGNs rather than just pixel patterns.
- Grad-CAM Analysis
- Grad-CAM is a technique used to visualize which parts of an image a deep learning model focuses on when making a prediction. The study found that the best model (M2) uses Grad-CAM to consistently highlight the actual morphological features of RAGNs, such as lobes and bends, proving it understands the physical shape.
Terminology
Summary
Bent radio active galactic nuclei (RAGNs)—wide-angle tails (WATs) and narrow-angle tails (NATs)—trace dense environments in galaxy groups and clusters, yet no multiclass classifier simultaneously separates them from straight Fanaroff–Riley types (sFRI, sFRII) using visually inspected labels and unlabelled data The four-class scheme introduced here enables WAT/NAT-resolved catalogues that can serve as environment probes and progenitor classifications for diffuse cluster radio emission
Dataset Construction
The study introduces FIRST-2060, a four-class labelled dataset of 2060 RAGNs (sFRI, sFRII, WAT, NAT) constructed from three publicly available catalogues through multi-tier visual inspection This dataset is provided in two preprocessing variants: RL1, which retains spurious sources, and RL2, from which they are removed The labelled dataset RL is constructed by combining selected sources from three catalogues: MiraBest (Porter & Scaife 2023), the Sasmal catalogue (Sasmal et al. 2022), and the Lao catalogue (Lao et al. p. 3 2025) Visual inspection involved six co-authors, with four individuals each assigned to one of the four target classes The classification was based primarily on the opening angle θOA, where sources with θOA ≈ 180◦ were labelled straight, those with 90◦ ≤ θOA ≤ 180◦ were labelled WAT, and those with θOA ≤ 90◦ were labelled NAT
Model Architecture and Training
The RGC model (M2) integrates the self-supervised framework BYOL (Bootstrap Your Own Latent) with an E(2)-equivariant steerable CNN (E2CNN) encoder, pre-trained on the unlabelled data and fine-tuned on the labelled sets The model operates in two stages: a task-agnostic self-supervised pretraining stage on RU, followed by a task-specific supervised finetuning stage on RL1 and RL2 In the pre-training stage, BYOL learns a representation by training an online network—comprising an encoder fθ, projector gθ, and predictor qθ—against a momentum-updated target network with weights ξ The encoder fθ is E2CNN, which preserves invariance under the Euclidean group E(2)
Performance Metrics and Model Ranking
The RGC model M2 ranks second overall with a macro-F1 of 0.79 ± 0.02, just 0.01 behind M1 (0. p. 280 ± 0.02) The per-class F1-scores of M2 reflect this competitiveness: on RL1, it achieves sFRI 0. p. 675 ± 0.05, sFRII 0. p. 080 ± 0.03, WAT 0.77 ± 0.02, and NAT 0. p. 083±0.4 The most consistent pattern across all models is the class-dependent difficulty of the task NAT achieves the highest per-class F1-score for every model on both datasets—reaching ∼ 0. p. 684 ± 0.02 for M1 on RL1—while sFRI is the hardest class, with F1-scores ranging from 0. p. 661 ± 0.06 (M6) to 0.76 ± 0.05 (M1, RL2)
Morphological Fidelity Analysis
Grad-CAM analysis shows that M2 is the only model whose attention contours consistently trace the morphological structure of RAGNs—lobes, jets, and bends—rather than defaulting to compact blobs or diffuse patterns The structural similarity index (SSIM) between the normalized heatmap and the 3σ source mask is highest for M2 and M6 The contrast is most pronounced for bent and extended sources (e.g. L0809, L2121), where M2’s contours bend with the source while all other models default to circular or image-filling patterns
Astrophysical Relevance
The four-class scheme adds a dimension that existing binary or three-class classifiers lack: by resolving bent sources into WATs and NATs rather than treating them as a single undifferentiated class, it enables a first-order separation between the BCGs that typically host WATs (O’Dea & Baum 2023) and the satellite galaxies that typically host NATs (Miley et al. 1972; Rudnick & Owen 1976) Applied at scale to EMU, LoTSS, and ultimately the SKA, a WAT/NAT-resolved catalogue would provide not just cluster candidates but a morphological proxy for the host galaxy’s dynamical state within its environment The Grad-CAM analysis of Section 6.1 is relevant here: M2 is the only model whose attention contours consistently trace bends, lobes, and extended tails (Appendix C) rather than defaulting to compact central blobs (M6) or diffuse imagefilling patterns (M4) The morphology of radio tails also carries quantitative information about the ICM itself M2 is the best-calibrated model by a substantial margin, achieving the lowest ECE of 0. p. 007 ± 0.01 (RL1) and 0. p. 008 ± 0.01 (RL2) A classifier that attends to the morphological features carrying ICM information—bending curvature, jet–tail transitions, lobe asymmetry—is a necessary condition for any future attempt to extract quantitative environmental information from classifier internals at survey scale The sFRI/sFRII versus WAT/NAT classification is therefore relevant to the fossil plasma supply chain—automated classification at survey scale would enable statistical studies of which progenitor types seed which classes of diffuse emission
Conclusion
The principal findings of this work are as follows: 1. The model ranking M1 > M2 > M3 > M4 > M5 > M6 in macro-F1 and accuracy is consistent across both dataset regimes, RL1 and RL2 2. p. 6 NATs are the most reliably classified class across all models, while sFRI sources are consistently the hardest 3. p. 6 NATs achieve the highest per-class F1-scores across all models in Fig. p. 6 3, despite being less numerous 4. p. 7 Grad-CAM analysis (Section 6. p. 01, Fig. 6) shows that M2 is the only model whose attention contours consistently trace the morphological structure of RAGNs—lobes, jets, and bends—rather than defaulting to compact blobs or diffuse patterns (Appendix C) 5. p. 0 M2 is the best-calibrated model (ECE ≈ 0. p. 707; Fig. 5), substantially ahead of the next-best architecture 6. p. 1 Grad-CAM analysis shows that M2 is the only model whose attention contours consistently trace bends, lobes, and extended tails (Appendix C) rather than defaulting to compact blobs or diffuse patterns (M6) or diffuse imagefilling patterns (M4) 7. p. 0 M1 and M2 are the most robust to class imbalance (Fig. p. 9 7), but error decomposition reveals complementary strengths: M1 excels at cross-type separation, while M2 excels at within-type discrimination The four-class scheme introduced here—resolving bent sources into WATs and NATs rather than a single undifferentiated class—is directly applicable to the ongoing EMU and LoTSS surveys and to the forthcoming SKA Scaling the pre-training to larger unlabelled radio datasets, following Slijepcevic et al. p. 0 (2024), is a natural next step toward building domain-specific foundation models for radio astronomy
Data and Code Availability
The FIRST-2060 dataset, trained model weights, and all code used in this study are publicly available in our GitHub repository: https://github.com/cassaiub/rgc The authors used Claude and Gemini as writing and coding aids to refine portions of the manuscript for clarity and readability
Acknowledgments
Asad, Amin, and Momen acknowledge support from ICT Innovation Fund Grant No. 148 (2020) of the ICT Division, Government of Bangladesh, and from Sponsored Research Grant No. p. 1
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Develop an ensemble approach for survey-scale morphological catalogues by combining M1 (ConvNeXT) and M2 (RGC). This combination leverages M1's strength in cross-type discrimination and M2's strength in within-type discrimination, suggesting
an ensemble approach may offer a practical framework for survey-scale morphological catalogues.
-
Implement a model that utilizes BYOL pre-training on unlabelled radio sources to generate representations that are inherently invariant to orientation, thus encoding the fact that
the morphological class of a radio source is independent of its orientation on the sky.
-
Create an automated system capable of producing four-class catalogues (sFRI, sFRII, WAT, NAT) for environments like EMU and LoTSS. This allows for
a first-order separation between the BCGs that typically host WATs and the satellite galaxies that typically host NATs.
-
Design a classifier whose internal representations are spatially faithful by using an encoder like E2CNN, which is noted as
the only model whose attention contours consistently trace bends, lobes, and extended tails rather than defaulting to compact blobs or diffuse patterns.
Sources
Related papers
- On-sky demonstration of a vector Zernike wavefront sensor in a cascaded adaptive optics system
- Euclid Data Release 1 (DR1): Calibration strategy and observations
- The General Antiparticle Spectrometer (GAPS) Antarctic Balloon Payload
- Detection and excision of radio-frequency interference in radio arrays
- Unified matched-filter and excess-power pipeline for gravitational-wave detector characterization
- Ushering in sub-arcsecond angular resolution observatories in the far-infrared: The case for a double-Fourier interferometer