Improving image quality of the Solar Disk Imager (SDI) of the Lyman-alpha Solar Telescope (LST) onboard the ASO-S mission

Basic Information

• Paper ID: 2501.00231
• Title: Improving image quality of the Solar Disk Imager (SDI) of the Lyman-alpha Solar Telescope (LST) onboard the ASO-S mission
• Authors: Hui Liu, Hui Li, Sizhong Zou, Kaifan Ji, Zhenyu Jin, Jiahui Shan, Jingwei Li, Guanglu Shi, Yu Huang, Li Feng, Jianchao Xue, Qiao Li, Dechao Song, Ying Li
• Classification: astro-ph.SR (Solar and Stellar Astrophysics), astro-ph.IM (Astronomical Instrumentation and Methods)
• Journal: Research in Astronomy and Astrophysics
• Paper Link: https://arxiv.org/abs/2501.00231

Abstract

In-orbit calibration and performance testing of the Solar Disk Imager (SDI) aboard the Lyman-alpha Solar Telescope (LST) of China's Advanced Space-based Solar Observatory (ASO-S) mission revealed that its spatial resolution falls far short of expectations. This paper develops the SDI Point Spread Function and Image Bivariate Optimization Algorithm (SPIBOA) to improve SDI image quality. This bivariate optimization method ingeniously combines deep learning with optical system modeling. Despite the lack of prior information about true SDI images and optical system functions, the algorithm can effectively estimate the PSF of the SDI imaging system directly from large-scale observational data. Deconvolution correction of observed SDI images using the estimated PSF shows that the corrected spatial resolution improves by more than three times compared to the observed images. Simultaneously, the method significantly reduces inherent noise in observed SDI images. SPIBOA has been successfully integrated into routine SDI data processing, providing important support for data-driven scientific research.

Research Background and Motivation

Problem Background

1. ASO-S Mission Overview: The Advanced Space-based Solar Observatory (ASO-S) is China's first comprehensive space solar observation platform, launched on October 9, 2022, carrying three payloads. The LST is specifically designed for observations of the Lyman-alpha spectral line (121.6 nm).
2. SDI Performance Issues: The SDI was designed with a spatial resolution of 1.2″, but in-orbit testing revealed an actual resolution of only approximately 9.5″, far below design specifications. Images also exhibit anomalous phenomena such as "horseshoe-shaped" bright structures.
3. Scientific Impact: Low resolution severely affects the scientific objectives of LST, including prominence evolution, flare feature analysis, and coronal mass ejection source region studies.

Research Motivation

1. Technical Challenges: Estimating system PSF from observational data alone without prior knowledge of true images and optical system functions
2. Practical Requirements: Developing practical image enhancement algorithms that can be integrated into routine data processing pipelines
3. Methodological Innovation: Exploring new approaches combining deep learning with optical modeling to provide novel methods for astronomical telescope system identification and image quality enhancement

Core Contributions

1. Proposed SPIBOA Algorithm: Developed the SDI Point Spread Function and Image Bivariate Optimization Algorithm, ingeniously combining deep learning with optical system modeling
2. Achieved Significant Performance Improvement: Spatial resolution enhanced by more than 3 times while effectively reducing image noise
3. Successful Engineering Application: Algorithm integrated into SDI routine data processing pipeline, supporting scientific research
4. Methodological Innovation: Provides new insights for astronomical telescope system identification and observed image quality enhancement
5. Provided Application Guidance: Detailed discussion of key factors and considerations in SPIBOA application

Methodology Details

Task Definition

Input: SDI observed image $I_{SDI}$ and corresponding AIA 304Å image $I_{304}$Output: Estimated PSF and enhanced SDI image Constraints: Optical imaging system modeling within the linear convolution domain

Model Architecture

1. Basic Principles

SPIBOA is based on the linear convolution imaging model: $I_{SDI} = I_{ref} \otimes PSF + noise$

where $I_{ref}$ is the ideal SDI image and $PSF$ is the system point spread function.

2. Bivariate Optimization Framework

Core optimization objective: $e = \min_{\Theta_N, wavefront} ||I_{SDI} - I_{ref} \otimes PSF||^2$

where:

• $I_{ref} = N(I_{304}, \Theta_N)$ (ideal SDI image output by CNN)
• $PSF = |F(A \cdot e^{i \cdot wavefront})|^2$ (PSF calculated from wavefront)

3. CNN Architecture Design

• Input Layer: 1→64 convolution kernel, converting input image to 64 feature maps
• Hidden Layers: 72 layers of convolution with width 64
• Output Layer: 64→1 convolution kernel, generating single-channel output image

4. Algorithm Workflow

1. Data Preprocessing: Register SDI and AIA images, suppress noise
2. Initialization: Initialize $I_{ref}$ with AIA 304Å image, initialize wavefront with ground test data
3. Error Computation: $e = ||I_{SDI} - I_{ref} \otimes PSF||^2$
4. Parameter Update: Simultaneously optimize $\Theta_N$ and wavefront parameters using Adam algorithm
5. Convergence Judgment: Output PSF when error falls below threshold, otherwise continue iteration

Technical Innovations

1. Bivariate Joint Optimization: Simultaneously estimate ideal image and system PSF, avoiding uncertainties from separate estimation
2. Physical Constraint Integration: Incorporate optical system design parameters as constraints to improve estimation accuracy
3. Combination of Deep Learning and Optical Modeling: CNN handles nonlinear transformations while optical model ensures physical consistency
4. Unsupervised Learning Strategy: Requires only observational data without true image labels

Experimental Setup

Dataset

• SDI Observational Data: Accumulated large-scale observational data over the past two years
• Training Sample Composition:
  • Solar limb, quiet region, active region, flare region = 1:1:1:2
  • Flare classes ranging from C-class to X-class
• Reference Data: Corresponding AIA 304Å images
• Preprocessing: Image registration, cosmic ray and thermal noise suppression

Evaluation Metrics

• Spatial Resolution: Measured using SIRAF (Spatial Image Resolution Assessment by Fourier Analysis) technique
• Power Spectral Density: Analysis of mid-to-high frequency energy enhancement
• Visual Quality: Comparison of image clarity and contrast before and after correction

Comparison Methods

• Original SDI observed images
• AIA 304Å and 1600Å images as reference standards

Implementation Details

• Optimizer: Adam algorithm
• Image Magnification: 2.73× (accounting for undersampling)
• Window Function: Hanning window to suppress spectral leakage
• Deconvolution: Richardson-Lucy iterative algorithm

Experimental Results

Main Results

1. Spatial Resolution Enhancement

2. Frequency Domain Analysis

• Normalized cutoff frequency before correction < 0.05
• Normalized cutoff frequency after correction ≈ 0.15
• Significant enhancement of mid-to-high frequency energy

Case Studies

1. X1.0 Flare on May 8, 2024

• Full-disk image comparison shows significant enhancement of structural details
• Power spectral density analysis confirms recovery of high-frequency information
• Structural consistency verified through comparison with AIA images

2. PSF Estimation Results

• Estimated PSF exhibits "horseshoe-shaped" structure, consistent with observed phenomena
• Validates accurate modeling of system optical characteristics by the algorithm

Experimental Findings

1. Significant Performance Improvement: All test cases achieve resolution improvement exceeding 3 times
2. Noise Suppression: Effectively reduces inherent image noise
3. Structure Recovery: Blurred bright structures become clear, highly consistent with AIA images
4. Algorithm Robustness: Effective for different types and intensities of solar activity

Related Work

Deep Learning Applications in Astronomical Adaptive Optics

1. Early Applications: Angel et al. (1990) used neural networks for aberration correction on the Multiple Mirror Telescope
2. Modern Developments:
   • Nishizaki et al. (2019): CNN directly estimates Zernike coefficients from single images
   • Niu et al. (2020): Deep learning for interferometric wavefront detection
   • Wang et al. (2021): Atmospheric turbulence wavefront sensing and aberration correction
   • de Bruijne et al. (2022): Extended source deep learning wavefront sensing

Advantages of This Work

1. No Hardware Modification Required: Pure software solution applicable to in-orbit instruments
2. Joint Optimization Strategy: Simultaneous estimation of image and PSF improves accuracy
3. Physical Constraint Integration: Combines optical design parameters to ensure physical validity

Conclusions and Discussion

Main Conclusions

1. SPIBOA successfully addresses SDI spatial resolution deficiency, achieving improvement exceeding 3 times
2. Algorithm successfully integrated into routine data processing pipeline, supporting scientific research
3. Provides new methods for astronomical telescope system identification and image enhancement

Limitations

1. Ringing Artifacts: May exhibit ringing phenomena in high-contrast regions
2. Spatiotemporal Variation: Current version does not account for temporal and spatial PSF variations
3. Error Assessment: Lacks precise quantitative error assessment capability
4. Improvement Variability: Enhancement effects vary across different locations and times

Future Directions

1. Long-term Evaluation: Conduct comprehensive long-term performance evaluation and improvement
2. Extended Applications: Extend to the other two instruments of LST (WST and SCI)
3. Method Optimization: Address spatiotemporal variation and ringing artifact issues
4. Error Modeling: Develop precise error assessment methods

In-depth Evaluation

Strengths

1. Strong Innovation: First to combine deep learning with optical modeling to address space telescope image quality issues
2. High Practical Value: Algorithm successfully engineered and applied, generating actual scientific value
3. Rigorous Methodology: Based on solid optical theory and deep learning techniques
4. Significant Results: Resolution improvement exceeding 3 times has important scientific significance
5. Comprehensive Documentation: Detailed discussion of application key points and considerations

Weaknesses

1. Insufficient Theoretical Analysis: Lacks theoretical guarantees for convergence and uniqueness
2. Limited Verification: Primarily relies on comparison with AIA images, lacks independent verification
3. Computational Complexity: Lacks detailed analysis of algorithm computational complexity and efficiency
4. Generalization Capability: Applicability to other types of telescopes requires further verification

Impact

1. Scientific Value: Significantly enhances scientific output capability of ASO-S/LST
2. Technical Contribution: Provides new insights for space telescope image processing
3. Engineering Significance: Demonstrates feasibility of software methods solving hardware problems
4. Promotion Potential: Methods can be extended to other astronomical instruments

Applicable Scenarios

1. Space Telescopes: Particularly suitable for in-orbit instruments where hardware adjustment is not feasible
2. Ground-based Telescopes: Can supplement traditional adaptive optics systems
3. Historical Data: Can improve quality of existing observational data
4. Other Fields: Principles can be extended to medical imaging, microscopy, and other domains

References

Key references include:

• Gan et al. (2019, 2023): ASO-S mission overview
• Li et al. (2019), Chen et al. (2019, 2024): LST instrument description and performance
• LeCun et al. (2015): Deep learning fundamentals
• Angel et al. (1990): Early applications of deep learning in astronomical adaptive optics
• Nishizaki et al. (2019): Modern CNN wavefront sensing techniques

Overall Assessment: This is a paper of significant practical value and technical innovation that successfully addresses a critical technical challenge of China's first space solar observation platform, providing new solutions for image processing of space astronomical instruments. Despite some theoretical and verification limitations, its successful engineering application and significant performance improvements make it of considerable scientific and technical value.