Abstract
This paper presents an improved method for two-dimensional sky background subtraction in the Large Sky Area Multi-Object Fiber Spectroscopic Telescope (LAMOST). By utilizing Locality-constrained Linear Coding (LLC) and a piece-wise optimization strategy, the researchers accurately estimate background noise within celestial spectra. The method leverages correlation coefficients to select optimal base vectors from dedicated sky fibers, significantly improving the signal-to-noise ratio for astronomical observations.
Introduction
LAMOST is a powerful tool for astronomical surveys, utilizing 4000 fibers to capture spectra simultaneously. However, these observations are inevitably contaminated by the "sky background"—noise from the Earth's atmosphere and light pollution. Accurate sky subtraction is critical for obtaining clean spectra of distant celestial objects. In the LAMOST configuration, each CCD manages 250 fibers, with approximately 40 dedicated to monitoring the sky background. The challenge lies in accurately predicting the sky signal for the remaining 210 object fibers based on these sparse samples.
Methodology
The researchers propose a two-step modeling approach combining Locality-constrained Linear Coding (LLC) with a piece-wise optimization strategy.
Locality-constrained Linear Coding (LLC)
LLC is employed to represent the spectra of object fibers as a weighted combination of the 40 dedicated sky fibers, which serve as a "base vector set." Unlike simple averaging, LLC emphasizes local similarities, ensuring that the model is constructed from the most relevant sky samples. A key finding of the study is that the correlation coefficient is a superior metric for selecting these base vectors compared to Euclidean distance.
Piece-wise Optimization
To address the limitations of global modeling, the authors implemented a piece-wise strategy. Each spectrum is divided into four segments, and the LLC parameters are optimized independently for each part. This allows the model to adapt to localized variations in the sky signal that a global fit might overlook.
Results and Discussion
The proposed method achieved high fitting precision, with a Mean Square Error (MSE) consistently below 0.05. Visual analysis confirms that the piece-wise LLC model effectively reconstructs the ground truth sky background with minimal residuals.
Key insights from the study include:
- Metric Superiority: Correlation coefficients provide a more robust basis for vector selection than geometric distance.
- Optimal Weighting: Using a subset of the most correlated fibers (K-nearest neighbors) rather than all 40 base vectors improves modeling accuracy.
- Granular Fitting: The transition to piece-wise modeling was essential for capturing fine-grained spectral details that global models missed.
Conclusion
This work demonstrates the feasibility of using LLC and piece-wise optimization for two-dimensional sky background modeling in LAMOST. Future efforts will focus on exploring additional similarity measures, applying the method to real-time object observations, and optimizing the underlying algorithms for greater speed and precision.
Summary updated