Slitless Spectroscopy Source Detection Using YOLO Deep Neural Network

TL;DR

The paper presents a YOLOv8l-based detector for identifying zeroth-order images and first-order spectral lines in CSST slitless spectroscopy data, addressing the lack of direct photometric-slitless pairs in CSST. It trains on 1,560 simulated slitless images from Cycle 6 and Cycle 9, using 640x640 tiles to handle crowded regions and achieve robust detection. The model attains strong validation metrics (overall mAP 0.868; lines 0.910; zeroth-order 0.825) and demonstrates reliable detection across brightness ranges, with processing times around 9 ms per image. The work discusses challenges such as CR contamination and tile-boundary effects, and outlines integration of these detections into a future spectral-extraction pipeline for CSST data.

Abstract

Slitless spectroscopy eliminates the need for slits, allowing light to pass directly through a prism or grism to generate a spectral dispersion image that encompasses all celestial objects within a specified area. This technique enables highly efficient spectral acquisition. However, when processing CSST slitless spectroscopy data, the unique design of its focal plane introduces a challenge: photometric and slitless spectroscopic images do not have a one-to-one correspondence. As a result, it becomes essential to first identify and count the sources in the slitless spectroscopic images before extracting spectra. To address this challenge, we employed the You Only Look Once (YOLO) object detection algorithm to develop a model for detecting targets in slitless spectroscopy images. This model was trained on 1,560 simulated CSST slitless spectroscopic images. These simulations were generated from the CSST Cycle 6 and Cycle 9 main survey data products, representing the Galactic and nearby galaxy regions and the high galactic latitude regions, respectively. On the validation set, the model achieved a precision of 88.6% and recall of 90.4% for spectral lines, and 87.0% and 80.8% for zeroth-order images. In testing, it maintained a detection rate >80% for targets brighter than 21 mag (medium-density regions) and 20 mag (low-density regions) in the Galactic and nearby galaxies regions, and >70% for targets brighter than 18 mag in high galactic latitude regions.

Figures (11)

• Figure 1: The simulated CSST slitless spectroscopy images, showing 1/16 of the full-size image: (a) the low-density region without cosmic rays; (b) the low-density region with cosmic rays; (c) the medium-density region without cosmic rays; (d) the medium-density region with cosmic rays; (e) the high-Galactic latitude region. The full slitless spectroscopy image has a size of approximately 9k $\times$ 9k pixels. For improved detection performance, each image was divided into 16 equal sections by resampling them to 640 $\times$ 640 pixels. Compared to the simulated images of the Milky Way and nearby galaxy regions, the high-Galactic latitude field contains a larger number of galaxies.
• Figure 2: The normalized distribution of datasets in the GU, GV, and GI bands for high galactic latitude regions, as well as for low-density and medium-density regions of the Galactic and nearby galaxies, where the vertical axis "Density" represents the proportion of sources within that magnitude interval relative to the total number. For the Milky Way and nearby galaxy regions, the simulated images use g-band magnitudes, whereas for the high-Galactic latitude region, the slitless spectroscopy images use white-light magnitudes corresponding to each chip. The distributions for the training, validation, and test sets within each chip and region are generally consistent, ensuring balanced training performance.
• Figure 3: The completeness of visual annotation. The completeness of manual annotation is lowest in the high galactic latitude regions, primarily because these regions contain a large number of galaxies, which are extended sources. At a given magnitude, galaxies appear visually fainter than point sources, reducing their detectability. Moreover, the completeness varies across different bands due to differences in SNRs. In particular, the GU band images exhibit relatively low SNR, which limited the number of sources that could be reliably identified through visual inspection.
• Figure 4: The precision-recall curve of the source detection model, evaluated on the validation dataset. The model achieves a mean Average Precision (mAP@0.5) of 0.868 across all classes, with the "line" class (spectral lines) performing best at 0.910 precision and the "dot" class (zero-order image) at 0.825.
• Figure 5: Schematic diagram of slitless spectroscopy image detection: (a) Illustration of target detection in a low-density celestial region of the Galactic and nearby galaxies, (b) Illustration of target detection in a medium-density celestial region of the Galactic and nearby galaxies, (c) Illustration of target detection in a high galactic latitude region. The figure shows the final CSST slitless spectral simulation image, assembled by stitching together 16 sub-images after object detection. Panel (d), (e), and (f) are local images magnified 16 times from panel (a), (b), and (c) respectively, to better demonstrate the target detection effect.