Star-forming clump detection in nearby galaxies using Faster R-CNN and $ugrizy$ imaging data from CLAUDS and HSC-SSP

Technical Summary: Star-forming clump detection in nearby galaxies using Faster R-CNN and ugrizy imaging data from CLAUDS and HSC-SSP

Problem Statement
Giant Star-Forming Clumps (GSFCs) are kpc-scale regions of enhanced star formation with stellar masses between $10^7$ and $10^9 M_\odot$. While commonly observed in high-redshift galaxies ($z \sim 2$), they are rarely detected in low-redshift ($z \lesssim 0.5$) galaxy analogues. Previous studies have relied on small samples, often using visual identification by experts or source extraction algorithms on contrast-enhanced images. These methods are limited in scale and automation. The authors note that while Deep Learning (DL) has been applied to larger samples, robust measurements of physical properties for large samples of low-redshift clumps remain missing due to the scarcity of observed clumpy galaxies at these redshifts.

Methodology
The authors developed a novel object detection model based on the Faster Region-Based Convolutional Neural Network (Faster R-CNN or FRCNN) framework, specifically adapted for multi-band astronomical imaging.

• Data Sources: The study utilizes imaging data from the Hyper Suprime-Cam Subaru Strategic Program (HSC-SSP) and the CFHT Large Area U-band Deep Survey (CLAUDS) across four deep fields. The dataset comprises a sample of $\sim$710,000 galaxies, with a subset of $\sim$14,000 galaxies possessing six-band photometry ($ugrizy$) and the remainder having five-band photometry ($grizy$).
• Model Architecture:
  • Backbone: The model adopts the Zoobot foundation DL-model (version 2.0, based on ResNet50) as the feature extraction backbone. This represents one of the first applications of Zoobot in a downstream object detection task.
  • Input Adaptation: The standard ResNet50 input layer (designed for 3-channel RGB images) was modified to accept 5-channel ($grizy$) and 6-channel ($ugrizy$) inputs. Pretrained weights were adapted by copying weights from the original $g$, $r$, and $i$ channels to the new $u$, $z$, and $y$ channels.
  • Classification Scheme: The detection head was expanded to a six-class scheme to distinguish between: (1) clumps, (2) 'odd' clumps (artifacts/anomalies), (3) foreground stars, (4) foreground/background galaxies, (5) bulges, and (6) generic background. This multi-class approach aims to reduce contamination in the final clump sample.
• Training Data: A dataset of 3,198 galaxies was manually annotated using a "correct-a-machine" approach. An initial model trained on Galaxy Zoo: Clump Scout (GZCS) data provided pre-detections, which were then corrected by volunteers via a Zooniverse interface. This process included correcting bounding boxes and reclassifying false positives (e.g., stars or bulges misidentified as clumps).
• Postprocessing: Raw model outputs underwent Non-Maximum Suppression (NMS) with an IoU threshold of 0.2. Bounding boxes exceeding a physical size of 7.30 kpc were removed, as were those fully containing smaller boxes or located outside the galaxy segmentation mask or near the central bulge. Flux peaks within the remaining bounding boxes were extracted to define precise clump candidate positions.

Key Contributions

1. Multi-band Object Detection: The authors successfully extended the Faster R-CNN architecture to process simultaneous multi-band imaging data (up to six channels), demonstrating that utilizing the full spectral information improves detection performance compared to standard RGB or limited-band inputs.
2. Zoobot Integration: The paper demonstrates the viability of using the Zoobot foundation model as a feature extractor for object detection, leveraging its pretraining on real astronomical data to handle domain-specific features with a relatively small training set ($\sim$3,200 images).
3. Comprehensive Contaminant Handling: By training on a multi-class scheme, the model explicitly learns to identify and separate contaminants (stars, background galaxies, bulges) from true star-forming clumps, a significant improvement over binary classification schemes.
4. Large-Scale Catalogue: The application of these models to the HSC-SSP and CLAUDS datasets resulted in the detection of approximately 1.5 million clump candidates across $\sim$700,000 galaxies, providing a public catalogue for future analysis.

Results
The model's performance was validated using a large set of real galaxies into which simulated star-forming clumps were injected. The simulations covered a range of physical properties (stellar mass, age, metallicity, dust attenuation) and were forward-modeled to match the survey's seeing and depth.

• Completeness and Purity: For simulated clumps brighter than the $5\sigma$ point-source depth limits of the instruments:
  • The 6-channel ($ugrizy$) model achieved a detection completeness of $\gtrsim 0.9$ and a purity of $\gtrsim 0.8$.
  • The 5-channel ($grizy$) model showed slightly lower completeness (dropping by $\sim$2–3% relative to the 6-channel model) but maintained similar purity ($\gtrsim 0.8$).
• Impact of U-band Data: The inclusion of the $u$-band data (from CLAUDS) significantly improved the detection of clumps with very recent star formation, which are often fainter in redder bands. The 6-channel model detected $\sim$17% more clump candidates than the 5-channel model in the overlapping sample, primarily those bright in the NUV.
• Performance Metrics: The models achieved high recall (near 1.0 at low objectness thresholds) and precision improvements when filtering by objectness scores. The $F_1$ scores indicated that fine-tuning the backbone weights generally yielded better performance than transfer learning with fixed weights, though transfer learning showed later signs of overfitting.
• Limitations: Completeness drops for clumps with low contrast against the galaxy background ($u_{cl} - u_{bkg} \gtrsim -1$) or those located very close to the galaxy center ($d_{cl} \lesssim 0.5 r_{eff}$), where they are outshone by the bulge.

Significance and Claims
The authors claim that their DL-based approach enables the automated detection of star-forming clumps in large samples of low-redshift galaxies, a task previously hindered by the limitations of manual inspection and traditional algorithms. By achieving high completeness and purity on simulated data, the study suggests that the detection method is robust and that the primary limiting factor for future studies is the sensitivity of the observing instruments rather than the detection algorithm itself.

The paper modestly concludes that while the 5-channel model is sufficient for large-scale surveys lacking $u$-band data, the 6-channel model offers superior performance for identifying the youngest star-forming regions. The authors emphasize that the resulting catalogue requires careful application of selection criteria (e.g., excluding edge-on galaxies, central clumps, and applying mass completeness limits) for scientific analysis. They note that while the detections are robust photometrically, spectroscopic confirmation (e.g., via H$\alpha$ emission) is necessary to definitively confirm the star-forming nature of the candidates.