YOLO-CIANNA: Galaxy detection with deep learning in radio data: II. Winning the SKA SDC2 using a generalized 3D-YOLO network

TL;DR

This work generalizes YOLO-CIANNA to 3D hyperspectral HI cubes, introducing a dedicated 3D CNN backbone and a regression-based 3D bounding-box framework guided by a DIoU loss for robust detection and characterization. The method is trained via a bootstrap strategy on the SKA-like SDC2 data, combining the LDEV truth catalog with dynamic augmentation and a refined selection function to improve detection completeness and purity. It achieves state-of-the-art performance on the MAIN cube, with a 9.5% boost over the top SDC2 score, a high detection purity of $92.3\%$, and a $45\%$ increase in confirmed sources, while processing ~1 TB in ~30 minutes on a single GPU. The results demonstrate the viability of 3D CNN detectors for large hyperspectral HI data and chart a path toward applying YOLO-CIANNA to SKA observations and precursors, including transfer learning and deployment considerations for future surveys.

Abstract

As the scientific exploitation of the Square Kilometre Array (SKA) approaches, there is a need for new advanced data analysis and visualization tools capable of processing large high-dimensional datasets. In this study, we aim to generalize the YOLO-CIANNA deep learning source detection and characterization method for 3D hyperspectral HI emission cubes. We present the adaptations we made to the regression-based detection formalism and the construction of an end-to-end 3D convolutional neural network (CNN) backbone. We then describe a processing pipeline for applying the method to simulated 3D HI cubes from the SKA Observatory Science Data Challenge 2 (SDC2) dataset. The YOLO-CIANNA method was originally developed and used by the MINERVA team that won the official SDC2 competition. Despite the public release of the full SDC2 dataset, no published result has yet surpassed MINERVA's top score. In this paper, we present an updated version of our method that improves our challenge score by 9.5%. The resulting catalog exhibits a high detection purity of 92.3%, best-in-class characterization accuracy, and contains 45% more confirmed sources than concurrent classical detection tools. The method is also computationally efficient, processing the full ~1TB SDC2 data cube in 30 min on a single GPU. These state-of-the-art results highlight the effectiveness of 3D CNN-based detectors for processing large hyperspectral data cubes and represent a promising step toward applying YOLO-CIANNA to observational data from SKA and its precursors.

Figures (9)

• Figure 1: 3D view of a $30\times 30\times 70$ subcube centered on a bright source from the MAIN catalog (id. 15 770) using a rescaled intensity. Transparency is inversely proportional to voxel intensity.
• Figure 2: Top: Two-dimensional histogram of LDEV truth source volume as a function of their integrated line flux. The dashed red line represents our selection function, with sources to the right of the line being selected. Bottom: Line flux distribution for the full LDEV catalog and the selected sample.
• Figure 3: Our SDC2 3D CNN backbone architecture. The left column provides layer structural properties, whereas the right column indicates the spatial output dimension starting from a $64 \times 64 \times 256$ input cube. Layers are stacked top to bottom. The layer width scales with the number of filters. Layer colors encode the stride setting: green preserves dimensions, orange reduces the frequency dimension by two, red reduces all dimensions by two, and blue reduces sky dimensions by two. The output grid size and the predicted parameters are shown next to the last layer.
• Figure 4: Evolution of the validation loss subparts (natural, Sect. A.7 in Paper I) during training for our four successive bootstrap models. The first 20 iterations are omitted. Scores computed over the MAIN cube are indicated every 200 iterations after 400.
• Figure 5: (a) Two-dimensional histogram of the MAIN cube detection matches target volume as a function of their target integrated line flux. The dashed red line represents the selection function as in Fig. \ref{['fig:selection_function']}. (b) Histogram of the target line flux for the MAIN truth catalog and the matched sources. False positives are added based on their predicted flux. Completeness is overplotted for relevant bins. (c) Histogram of the predicted line flux for the matched predicted sources and false positives. The MAIN truth catalog is plotted in the background based on the target line flux. Purity is overplotted for relevant bins. (d) Histogram of the smoothed S/N for the MAIN truth catalog and the matched sources. Completeness is overplotted for relevant bins. (e) Identical to d but using the raw non-smoothed S/N.