🔭 astrophysics

Radio Galaxies detection and characterization using deep learning techniques

This paper introduces YOLO-Chars, a two-stage deep-learning framework utilizing customized YOLO models for detection and a dedicated network for characterizing radio galaxy properties, which demonstrates competitive performance on the SKA SDC1 benchmark to address the data analysis challenges of future radio telescopes.

Original authors: Sanjay Khatik, Rohit Sharma, Pankaj Jain

Published 2026-08-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sanjay Khatik, Rohit Sharma, Pankaj Jain

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The universe speaks a language of radio waves, invisible to the human eye but rich with information about the cosmos. For decades, astronomers have used giant dishes to listen to these signals, mapping out the distribution of galaxies and the energetic events that shape them. However, the next generation of these instruments promises to generate data on a scale that is almost impossible for humans to process using traditional methods. The sheer volume of images and the complexity of the sources within them—ranging from tiny, point-like stars to vast, sprawling galaxies—create a bottleneck. To make sense of this deluge, scientists are turning to artificial intelligence, specifically a type of machine learning that excels at recognizing patterns in images. The goal is not just to count these cosmic objects, but to measure their physical properties with the same precision a human expert would, but at a speed and scale that only a computer can manage.

In a recent study, researchers developed a new system called YOLO-Chars to tackle this challenge. They tested their approach on a simulated dataset designed to mimic the output of the Square Kilometre Array, a future radio telescope that will be the most sensitive in the world. The simulation included a crowded field of view containing nearly 200,000 radio sources, ranging from compact points of light to extended, fuzzy shapes. The researchers faced a difficult problem: a single computer model often struggles to see both the tiny details of a small source and the broad structure of a large one at the same time. To solve this, they built a two-stage framework that separates the job of finding a source from the job of describing it.

The first stage acts as a wide-eyed scout. It uses two specialized detection models working in tandem. One model is tuned to spot small, compact sources, while the other looks for medium and large, extended structures. By splitting the task, the system avoids the confusion that arises when trying to detect everything with a single set of rules. Once a source is located, the system draws a box around it. This box is then passed to the second stage, a dedicated network designed to characterize the object. Instead of just saying "there is a galaxy here," this second network measures specific physical properties: how bright the source is, how big it appears in the sky, and the angle at which it is oriented. This separation allows the team to optimize each part of the process independently, making the whole system more flexible and accurate.

The results of the simulation were promising. The system successfully identified the vast majority of the radio sources in the test field, achieving a level of accuracy that compares favorably with other leading methods. It managed to recover the physical properties of these sources with a high degree of precision, particularly for those that were bright enough to be clearly distinguished from the background noise. The researchers found that the system worked best when it was allowed to be slightly more inclusive, accepting a few extra detections to ensure it didn't miss any faint objects. When they tightened the rules to be more selective, the system became extremely pure, missing very few false alarms, but it also missed some real sources. This trade-off is a common reality in astronomy, where the goal is often to balance the desire to find everything against the need to be certain about what has been found.

Despite these successes, the authors are careful to note the limitations of their work. The system was trained entirely on simulated data, which is a clean and idealized version of reality. Real radio images from actual telescopes contain imperfections, calibration errors, and complex noise patterns that the simulation did not include. Consequently, while the system performs well in this controlled environment, it may need further adjustments before it can be deployed on real observations from telescopes like the Square Kilometre Array. Additionally, the current version of the tool focuses on measuring the basic shape and brightness of sources. It does not yet attempt to determine the specific type of galaxy or the complex physical processes driving its radio emission, tasks that would require even more sophisticated analysis.

The study demonstrates that a modular approach, where different parts of the problem are solved by different specialized tools, is a viable path forward for the future of radio astronomy. By breaking the massive task of surveying the sky into manageable pieces—finding the sources, then measuring them—the researchers have created a pipeline that can scale to handle the enormous data volumes expected in the coming years. This work does not solve every problem in the field, nor does it claim to replace human intuition entirely. Instead, it offers a robust, automated framework that can handle the initial heavy lifting of data analysis, allowing astronomers to focus on the deeper scientific questions hidden within the radio waves. As the next generation of telescopes comes online, tools like this will be essential for turning the flood of data into a clear picture of our universe.

Technical Summary: YOLO-Chars for Radio Galaxy Detection and Characterization

Problem Statement
The next generation of radio telescopes, particularly the Square Kilometre Array (SKA), will generate data volumes that exceed the processing capabilities of traditional statistical source-finding methods. Current challenges include the need for automated, scalable techniques to detect and characterize radio sources across diverse morphologies (from compact to extended) and varying signal-to-noise ratios (SNR) in crowded fields. While Deep Learning (DL) has shown promise in radio astronomy, many existing approaches either lack scalability for large surveys or rely on classical fitting methods for parameter estimation after detection, limiting integration and flexibility.

Methodology
The authors propose YOLO-Chars (YOLO-based Detection and Characterisation of Radio Sources), a decoupled, two-stage deep-learning framework designed for the SKA Science Data Challenge 1 (SDC1) dataset (560 MHz, 1000-hour integration). The framework separates source detection from source characterization to allow independent optimization and modularity.

Data Preprocessing:
- The study utilizes the SKA SDC1 560 MHz simulated map.
- To avoid amplifying noise in outer regions, primary beam correction is not applied to the images; instead, the apparent flux ( $f_{appr}$ ) is calculated by interpolating beam values at source centroids and multiplying by the intrinsic flux.
- Training data is filtered to include sources with $f_{appr} \ge 1.6 \times 10^{-6}$ Jy and sufficient surface brightness ( $> 5 \times 10^{-8}$ Jy pixel $^{-2}$ ).
- Ground-truth bounding boxes (gt-bboxes) are generated by converting source parameters ( $b_{maj}, b_{min}, \phi$ ) into pixel coordinates, applying a scaling factor of 2.5 to encompass full brightness, and regularizing unresolved sources.
Stage 1: Multi-Scale Detection:
To address the wide dynamic range of source sizes, two dedicated single-stage detectors based on a YOLOv3 architecture are employed:
- Single-Scale Detection (SSD): Optimized for compact sources ( $0.1'' < b_{maj} < 6''$ ). It uses a $64 \times 64$ grid ( $\Delta = 8$ pixels) and 9 anchors per cell.
- Multi-Scale Detection (MSD): Targets medium and extended sources ( $4'' < b_{maj} < 128''$ ). It utilizes two coarser grid scales: $32 \times 32$ ( $\Delta = 16$ pixels) for medium sources and $16 \times 16$ ( $\Delta = 32$ pixels) for large sources.
- Architecture: Both models employ residual blocks and downsampling blocks. The SSD model has ~7.2M parameters, and the MSD has ~6.4M.
- Loss Function: A custom loss combines Complete Intersection over Union (CIoU) for bounding box regression and Focal Loss for object confidence, addressing class imbalance between sources and background.
Stage 2: Multi-Scale Characterisation (MSC):
- A dedicated, lightweight CNN (~0.58M parameters) estimates physical parameters for detected sources.
- Input: Source cutouts extracted from the detection bounding boxes are resized to standard dimensions (ranging from $16 \times 16$ to $128 \times 128$ pixels) via zero-padding and normalized using a logarithmic scale.
- Output: The network directly regresses five parameters: apparent flux ( $f_{appr}$ ), major axis ( $b_{maj}$ ), minor axis ( $b_{min}$ ), and position angle (encoded as $\sin \phi$ and $\cos \phi$ ).
- Loss: Mean Squared Error (MSE) is used for regression.
Catalogue Construction:
- Predictions from SSD and MSD are generated on overlapping grid patches covering the full map.
- A non-maximum suppression (NMS) strategy is applied. Crucially, a specific merging protocol uses $4 \times 4$ pixel centred boxes to resolve duplicates between the different detector outputs (SSD vs. MSD) without discarding valid detections due to differing confidence thresholds or spatial extents.

Key Contributions

Decoupled Framework: The paper introduces a modular architecture where detection and characterization are optimized independently, offering flexibility for future extensions.
Multi-Scale Strategy: By employing separate detectors for compact and extended sources (SSD and MSD), the framework effectively handles the extreme dynamic range of source sizes found in SKA simulations, a challenge for single-model approaches.
Direct Parameter Estimation: Unlike methods that rely on classical fitting post-detection, YOLO-Chars uses a dedicated DL network to directly regress physical parameters, integrating the characterization step into the deep learning pipeline.
Robust Merging: The development of a specific NMS-based merging procedure using centred square boxes to combine detections from multi-scale detectors while preserving source integrity.

Results
Evaluated on the SKA SDC1 560 MHz benchmark:

Detection Performance: The combined multi-scale configuration achieved an overall score ( $F_{sco}$ ) of 483,875 with a catalogue purity of 93.57% and an average sub-score ( $\bar{w}$ ) of 0.8181.
Comparison: These results are competitive with, and in some metrics (specifically $\bar{w}$ and overall score) superior to, other teams' submissions on the same benchmark, including MINERVA (YOLO-CIANNA) and JLRAT2.
Characterization Accuracy: For matched sources, the median fractional errors decrease with increasing SNR. For high-SNR compact sources ( $>15\sigma$ ), the median errors are approximately 11.6% for flux, 29.3% for $b_{maj}$ , and 35.5% for $b_{min}$ . Position angle errors are higher for compact sources due to resolution limits.
Completeness vs. Purity: The study demonstrates a trade-off; increasing the confidence threshold improves purity (up to 99.53%) but significantly reduces source recovery (completeness).

Significance and Claims
The authors claim that YOLO-Chars demonstrates the potential of a scalable, modular deep-learning framework for next-generation radio continuum surveys. The work highlights that:

Source detection and characterization can be effectively treated as separate problems, allowing for specialized network architectures.
Multi-scale detection strategies are essential for handling the diverse morphologies of radio galaxies in crowded fields.
The framework provides a practical approach for automated analysis of massive radio datasets, achieving competitive performance on the SKA SDC1 benchmark.

The paper explicitly notes limitations, including training exclusively on idealized simulations (lacking real-world calibration artifacts), the bias against faint/barely resolved sources due to flux thresholds, and the lack of physical association logic for multi-component sources. The authors state that generalization to real observations requires domain adaptation and that future work will explore hybrid approaches and more advanced architectural variants.

More like this