CSST Slitless Spectra: Target Detection and Classification with YOLO

TL;DR

This study establishes machine learning as a paradigm shift in slitless spectroscopy, unifying detection, classification, and preliminary parameter estimation in a scalable system.

Abstract

Addressing the spatial uncertainty and spectral blending challenges in CSST slitless spectroscopy, we present a deep learning-driven, end-to-end framework based on the You Only Look Once (YOLO) models. This approach directly detects, classifies, and analyzes spectral traces from raw 2D images, bypassing traditional, error-accumulating pipelines. YOLOv5 effectively detects both compact zero-order and extended first-order traces even in highly crowded fields. Building on this, YOLO11 integrates source classification (star/galaxy) and discrete astrophysical parameter estimation (e.g., redshift bins), showcasing complete spectral trace analysis without other manual preprocessing. Our framework processes large images rapidly, learning spectral-spatial features holistically to minimize errors. We achieve high trace detection precision (YOLOv5) and demonstrate successful quasar identification and binned redshift estimation (YOLO11). This study establishes machine learning as a paradigm shift in slitless spectroscopy, unifying detection, classification, and preliminary parameter estimation in a scalable system. Future research will concentrate on direct, continuous prediction of astrophysical parameters from raw spectral traces.

Figures (18)

• Figure 1: This figure illustrates the pixel intensity distribution of a ZScale-normalized slitless spectral image from Simulation A mwl90b20, Chip 01 (GI). The ZScale algorithm, a standard tool in astronomical image processing, optimizes contrast by dynamically adjusting pixel values to highlight critical spectral features while suppressing noise. By stretching the intensity range to emphasize subtle variations, this method enhances the visibility of faint or overlapping spectral traces, enabling clearer identification and analysis of the data’s structural details.
• Figure 2: This figure illustrates the process of splitting the original image into strips. The original image size is 9216$\times$9232 pixels. We divided the image into 10 sub-images, each with a size of 9216$\times$979 pixels. To ensure continuity and avoid losing any information at the edges, we included an overlapping area of 9216$\times$62 pixels between each sub-image. This overlapping region helps maintain the integrity of the data and ensures that all relevant features are captured in the sub-images. By splitting the image in this manner, we can manage the large image size more effectively and facilitate more efficient processing and analysis.
• Figure 3: The 9216$\times$979 pixel strips (Figure \ref{['fig:train_img_0']}) were resized to a square 4544$\times$4544 format, enhancing adaptability to processing pipelines while preserving the structural integrity of key features. This resizing strategy achieves an optimal trade-off: retaining sufficient resolution for precise target detection while minimizing computational overhead. Zero-order (B, yellow) and first-order (A, blue) spectral traces are annotated to aid classification, streamlining analysis within the standardized sub-images. Stars fainter than magnitude 22 are considered as background noise, ensuring focused detection of relevant targets. Each resized sub-image contains approximately 220 objects.
• Figure 4: (a) Ground truth (with a magnitude threshold of 22, targets fainter than the threshold are labeled as background) and (b) predictions from model YOLOv5s (with scores $\ge$ 0.25). Class A, marked by the blue bounding box, represents the first-order slitless spectral image. Class B, marked by the yellow bounding box, corresponds to the zero-order slitless spectral image. The differences between (a) and (b) are highlighted by magenta (false negatives) and cyan (false positives) boxes. Class A is our primary target. When achieving the maximum F1-score of 0.961, the recall and precision values are 0.950 and 0.973, respectively. This indicates that the model is highly effective in identifying true positives while maintaining a low rate of false positives and false negatives. The high precision value of 0.973 reflects the model’s ability to accurately identify objects within Class A, while the recall value of 0.950 demonstrates its proficiency in capturing the majority of relevant instances.
• Figure 5: This figure demonstrates YOLOv5’s capability to detect spectral traces in densely populated star fields, using the Simulation A mwl90b10 as an example. Panel (a) illustrates the ground-truth annotations filtered by a magnitude threshold of 22. Panel (b) displays predictions from the YOLOv5s model trained on the mwl90b20 dataset. Despite significant overlap between bounding boxes—which complicates precise pairing of true and predicted detections for conventional metric evaluation (e.g., precision, recall)—visual inspection reveals that the model accurately identifies the zero-order (compact) and first-order (elongated) spectral images for nearly all bright sources. This qualitative assessment underscores YOLOv5’s robustness in dense stellar environments, confirming its suitability for astronomical tasks requiring reliable detection in crowded fields.