An adaptive parameter optimization method for astronomical image alignment using Bayesian optimization. I. A hierarchical search strategy for FWHM and SNR

arXiv:2609.07023 · astro-ph.IM · Submitted 2026-09-07 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Astrophysics. The day in astrophysics: observations, models and what the data forces.

Vera: We're looking at the paper: "An adaptive parameter optimization method for astronomical image alignment using Bayesian optimization. I. A hierarchical search strategy for FWHM and SNR".

Jocelyn: Here's the gist of the paper. The paper presents an adaptive parameter optimization method for astronomical image alignment using Bayesian optimization to automatically select optimal Full Width at Half Maximum (FWHM) and Signal-to-Noise Ratio (SNR) thresholds, which is crucial for creating robust, automated data processing pipelines The Problem Addressed Traditional workflows rely on manual trial-error processes for selecting key parameters like FWHM and SNR thresholds, which are often inefficient and lack objectivity Inappropriate settings can lead to two failure modes: (1) Overly strict parameters result in too few detected sources to establish a reliable geometric transformation; (2) Overly lenient parameters introduce excessive noise-induced spurious sources or cosmic rays, which clutter the source list and confound the matching algorithm, eventually leading to alignment failure or reduced precision The traditional methods of parameter adjustment such as manual trial and error are not only inefficient but also highly subjective, making it difficult to ensure that the parameters found are the global optimal solution The core of the adaptive framework reframes parameter selection as a formal optimization problem using Bayesian Optimization (BO) This involves defining a search space for FWHM and SNR, which are selected because they have direct physical significance in quantifying stellar profile size and detection sensitivity The hierarchical search strategy is designed to balance computational speed with optimization accuracy through two phases Phase I consists of a Global Coarse Search performed on downsampled images to rapidly explore the parameter space, leveraging scale-space theory to preserve the overall topology of the objective function's response surface Phase II transitions to a Local Fine-tuning Search operating on original, full-resolution images, focusing the search around the best parameters found in Phase I for meticulous refinement The Optimization Logic The optimization problem is defined by maximizing the number of successfully matched source pairs, denoted as Nmatches This metric is chosen because it serves as a direct and intuitive measure of alignment quality, and it synergizes perfectly with relaxing morphological filters like sharpness and roundness The objective function is formulated as the negative of the number of matched pairs to align with standard minimization conventions, seeking to find the parameter set p that minimizes f(p) = −Nmatches(p) The Bayesian Optimization engine uses a Gaussian Process (GP) as a probabilistic surrogate model to approximate this expensive objective function, providing both a predicted mean µt(p) and an uncertainty measure σ2t(p) Results and Validation Experimental results using data from the 1-meter telescope at the Yunnan Astronomical Observatory demonstrate that the HBO framework effectively handles varying seeing conditions and background noise levels The method rapidly converges to a robust set of alignment parameters, achieving sub-pixel accuracy and significantly improving the automation level and success rate of the alignment process For instance, when a fixed parameter set failed due to poor seeing, HBO autonomously recognized the blurred state of the stars and identified optimal parameters p∗ = 12.83 pixels for FWHM The method achieved a minimum astrometric RMS residual of 0.0874 pixels in the Moshup sequence Furthermore, the total optimization time for this dataset was recorded at 47.19 seconds, with parameter stabilization typically observed within approximately 25 iterations Scientific Significance The framework proves that the choice of objective function—maximizing Nmatches—is scientifically sound because a larger number of matched stars provides a more "overdetermined" geometric solution, which suppresses individual errors The hierarchical strategy ensures that the final parameters are optimized for the highest-fidelity data, and inter-night stability tests confirm that HBO maintains a 100% success rate across diverse conditions despite physical variability This adaptive strategy demonstrates the potential to mitigate high failure rates associated with static parameter configurations, offering a more flexible approach for future automated astronomical data processing pipelines The framework's ability to "re-sense" the state of the night ensures that the scientific integrity of long-term baselines is preserved without human intervention Future Work Future work will focus on refining the objective function and exploring additional morphological constraints to improve the reliability of source extraction Additionally, more sophisticated probabilistic models, including Gaussian Cox Processes, may be necessary to better account for non-Gaussian background noise Extensive testing on larger and more diverse datasets remains essential to determine the generalizability and long-term reliability of this method for large-scale astronomical surveys The scope of this study is limited to a finite set of observational scenarios and targets, but the feasibility of the proposed adaptive approach is demonstrated References Bancelin, D., Colas, F., Thuillot, W., Hestroffer, D., & Assafin, M.

Vera: 2012, A&A, 544, A15 Bernstein, G.

Jocelyn: M., Trilling, D.

Vera: E., Allen, R.

Paper summary: Vera: So we're looking at this paper by Yongjie Zhang and Yigong Zhang, titled "An adaptive parameter optimization method for astronomical image alignment using Bayesian optimization. I. A hierarchical search strategy for FWHM and SNR". It’s tackling the problem of how astronomers choose those key settings like Full Width at Half Maximum, or FWHM, and Signal-to-Noise Ratio thresholds manually in traditional image processing workflows.

Jocelyn: Right, so the core claim here is that they use Bayesian optimization to automatically select those optimal FWHM and SNR values for aligning astronomical images. They’re proposing a system that avoids the old trial and error approach which is inefficient and subjective.

Subrahmanyan: That makes sense from a theory standpoint because you can see how setting these parameters incorrectly leads to two failure modes, either too strict or too lenient detection, which messes up the geometric transformation. This paper seems to be trying to build a more objective way to handle that uncertainty.

Vera: Exactly. The paper argues that instead of guessing the right FWHM and SNR, they treat it as a formal optimization problem using Bayesian optimization, which is a way of finding the best settings by intelligently searching the parameter space. They are focusing on defining this search space as FWHM and SNR thresholds because those parameters have direct physical meaning in quantifying stellar profile size and detection sensitivity.

Jocelyn: And they’re using a hierarchical search strategy, which means they aren't just searching blindly all at once; they break it down into two phases: a global coarse search on downsampled images, and then a local fine-tuning search on the full-resolution images based on what the first phase found.

Subrahmanyan: That sounds like a clever way to balance computational speed with accuracy because the initial phase helps explore the whole parameter space quickly using scale-space theory to keep track of how things change in terms of image resolution, and then you refine it precisely. The objective function they're maximizing is simply the number of successfully matched source pairs, which they call Nmatches.

Vera: And that metric, Nmatches, is what they chose because a higher number directly measures alignment quality and pairs well with relaxing morphological filters like sharpness and roundness. The optimization engine uses a Gaussian Process to model this expensive objective function, giving them both a predicted mean of the best settings and an uncertainty measure for those settings.

Jocelyn: That uncertainty part is important because it tells you how sure the algorithm is about its chosen parameters, which helps guide the search process as it refines things. The results they show come from data from the one-meter telescope at Yunnan Astronomical Observatory where they test this framework against varying seeing conditions and background noise levels <ref:2609.07023#pg1>.

Subrahmanyan: It’s interesting to hear that when a fixed parameter set failed because of poor seeing, the HBO framework autonomously recognized the blurred state of the stars and identified optimal parameters p star equals twelve point eight three pixels for FWHM. That suggests it can adapt its strategy based on what it sees in the data, which is a big deal.

Paper summary: Vera: They achieved a minimum astrometric root mean square residual of zero point zero eight seven four pixels in the Moshup sequence, and they also recorded that the total optimization time for this dataset was about forty-seven point one nine seconds, with parameter stabilization usually happening within about twenty-five iterations. Those are concrete numbers showing how it performs under real observational conditions.

Jocelyn: So what this means for us is that we can move away from those slow, manual tuning processes and have a more automated pipeline that finds robust alignment parameters much faster and with higher precision. It suggests the method is quite stable across different nights or conditions because they did inter-night stability tests showing a one hundred percent success rate despite physical variability.

Subrahmanyan: From a cosmological perspective, if we can automate these fundamental alignment steps reliably, it means the long-term baselines for our astronomical surveys will be much more consistent and accurate without constant human intervention to babysit the parameter choices. It’s about preserving scientific integrity over time because you get that adaptive strategy to "re-sense" the state of the night.

Vera: And they’ve shown that maximizing Nmatches is scientifically sound because a larger number of matched stars gives you a more "overdetermined" geometric solution, which helps suppress individual errors in the alignment. The hierarchical strategy ensures those final parameters are optimized for the highest fidelity data, and inter-night stability confirms that this adaptive approach can handle diverse conditions without failing.

Jocelyn: They’ve also noted some limitations in their future work, mentioning that refining the objective function and exploring more morphological constraints could improve reliability, and they suggest using more sophisticated models like Gaussian Cox Processes to better account for non-Gaussian background noise. Plus, they admit their scope is limited to a finite set of observational scenarios and targets right now.

Subrahmanyan: That’s fair; moving from a specific test case to large-scale surveys is the next big challenge for any method like this. The feasibility of using Bayesian optimization here, though, shows that adaptive methods have real potential for automating complex data processing pipelines in astronomy.

Vera: So essentially, this paper presents an adaptive parameter optimization method for astronomical image alignment using Bayesian optimization with a hierarchical search strategy to automatically select optimal FWHM and SNR thresholds for image alignment. It moves the process from manual guesswork to a data-driven formal optimization problem where the goal is maximizing matched source pairs.

Jocelyn: And it shows that this system can adapt its search, moving from a fast global exploration of downsampled images to a detailed local fine-tuning on full images, achieving sub-pixel accuracy with minimal runtime.

Subrahmanyan: The implication is that for future large astronomical surveys, having a flexible approach that automatically adjusts parameters based on the current night's conditions could significantly reduce failure rates and increase the overall success rate of automated data processing pipelines.

Conclusion: Vera: So we've been diving into this paper by Zhang and Zhang, which is all about using Bayesian optimization to automatically pick the right Full Width at Half Maximum and Signal-to-Noise Ratio thresholds for aligning astronomical images.

Jocelyn: Right, so we're looking at how they built a system that doesn't rely on guessing those settings manually when you’re trying to align data from telescopes.

Subrahmanyan: The core idea is that this process turns parameter selection into a formal optimization problem, which means it uses math to find the best settings instead of just trial and error.

Vera: They use a hierarchical search strategy, breaking the search into two parts: a big coarse search on lower-resolution images, and then a fine tuning part on the original high-resolution images.

Jocelyn: That hierarchy is clever because it lets them explore the whole range quickly and then zoom in for precision where it matters most.

Subrahmanyan: And they’re maximizing the number of successfully matched source pairs, which is their way of measuring how good the alignment actually is.

Vera: The results show that this method can handle seeing conditions and background noise levels really well, even when fixed settings would have failed.

Jocelyn: They got sub-pixel accuracy on some sequences, and they only needed about twenty iterations to stabilize the parameters.

Subrahmanyan: It proves that this adaptive strategy can maintain a one hundred percent success rate across different nights of observation without needing human input for every single setting.

Vera: The main implication is that we can build these automated data processing pipelines much more robustly and consistently than before.

Jocelyn: It’s about making sure the long-term baselines for astronomical surveys stay accurate over years, not just one night.

Subrahmanyan: But they also flag that their current test is on a specific set of targets, so we need to see if this generalizable for much larger surveys out there.

Yongjie Zhang, Yigong Zhang, ★Zhenjun Zhang, Xiangming Cheng, Lei Xiong, Xiaoguang Yu, Jie Su, 会Cheng Wang, 大Guo

School of Information Engineering, Kunming University School of Computer of Science and Information Engineering Anyang Institute of Technology Yunnan Observatories Chinese Academy of Sciences Fujian Polytechnic Normal University College of Big Data Yunnan Agricultural University

astro-ph.IM

Submitted: 2026-09-07

Updated: 2026-09-07

Comments: 10 pages,6 figures

Journal ref: A&A 714, A352 (2026)

DOI: 10.1051/0004-6361/202660377

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 84/100

The gist: The paper presents an adaptive parameter optimization method for astronomical image alignment using Bayesian optimization to automatically select optimal Full Width at Half Maximum (FWHM) and…

Key concepts

Bayesian Optimization (BO)
A mathematical technique used to find the best settings for a function where evaluating the function is expensive. BO uses a probabilistic model called a Gaussian Process to intelligently explore the search space by balancing exploration (trying new areas) and exploitation (refining promising areas). This allows it to efficiently find optimal FWHM and SNR values without testing every possibility.
Hierarchical Search Strategy
A two-phase approach for parameter selection. Phase I performs a global coarse search on downsampled images to quickly explore the entire parameter space. Phase II then conducts a local fine-tuning search on the original, high-resolution images, focusing precisely around the best settings discovered in the first phase for meticulous refinement.
Objective Function (Nmatches)
The metric used to judge alignment quality is maximizing Nmatches, which is the total number of successfully matched source pairs. This metric is chosen because a higher count provides a more 'overdetermined' geometric solution, making the resulting alignment more stable and less sensitive to individual errors caused by noise or blurring.
Gaussian Process (GP)
A probabilistic surrogate model used within Bayesian Optimization. The GP approximates the complex objective function (Nmatches) using a mathematical distribution that provides both a predicted mean ($μ$) of the best outcome and an uncertainty measure ($ς$). This allows the optimization engine to decide where to search next, prioritizing areas with high potential gains while accounting for prediction errors.

Terminology

Summary

The paper presents an adaptive parameter optimization method for astronomical image alignment using Bayesian optimization to automatically select optimal Full Width at Half Maximum (FWHM) and Signal-to-Noise Ratio (SNR) thresholds, which is crucial for creating robust, automated data processing pipelines

The Problem Addressed

Traditional workflows rely on manual trial-error processes for selecting key parameters like FWHM and SNR thresholds, which are often inefficient and lack objectivity Inappropriate settings can lead to two failure modes: (1) Overly strict parameters result in too few detected sources to establish a reliable geometric transformation; (2) Overly lenient parameters introduce excessive noise-induced spurious sources or cosmic rays, which clutter the source list and confound the matching algorithm, eventually leading to alignment failure or reduced precision The traditional methods of parameter adjustment such as manual trial and error are not only inefficient but also highly subjective, making it difficult to ensure that the parameters found are the global optimal solution

Methodology Overview

The core of the adaptive framework reframes parameter selection as a formal optimization problem using Bayesian Optimization (BO) This involves defining a search space for FWHM and SNR, which are selected because they have direct physical significance in quantifying stellar profile size and detection sensitivity The hierarchical search strategy is designed to balance computational speed with optimization accuracy through two phases Phase I consists of a Global Coarse Search performed on downsampled images to rapidly explore the parameter space, leveraging scale-space theory to preserve the overall topology of the objective function's response surface Phase II transitions to a Local Fine-tuning Search operating on original, full-resolution images, focusing the search around the best parameters found in Phase I for meticulous refinement

The Optimization Logic

The optimization problem is defined by maximizing the number of successfully matched source pairs, denoted as Nmatches This metric is chosen because it serves as a direct and intuitive measure of alignment quality, and it synergizes perfectly with relaxing morphological filters like sharpness and roundness The objective function is formulated as the negative of the number of matched pairs to align with standard minimization conventions, seeking to find the parameter set p that minimizes f(p) = −Nmatches(p) The Bayesian Optimization engine uses a Gaussian Process (GP) as a probabilistic surrogate model to approximate this expensive objective function, providing both a predicted mean µt(p) and an uncertainty measure σ2t(p)

Results and Validation

Experimental results using data from the 1-meter telescope at the Yunnan Astronomical Observatory demonstrate that the HBO framework effectively handles varying seeing conditions and background noise levels The method rapidly converges to a robust set of alignment parameters, achieving sub-pixel accuracy and significantly improving the automation level and success rate of the alignment process For instance, when a fixed parameter set failed due to poor seeing, HBO autonomously recognized the blurred state of the stars and identified optimal parameters p∗ = 12.83 pixels for FWHM The method achieved a minimum astrometric RMS residual of 0. p. 80874 pixels in the Moshup sequence Furthermore, the total optimization time for this dataset was recorded at 47. p. 819 seconds, with parameter stabilization typically observed within approximately 25 iterations

Scientific Significance

The framework proves that the choice of objective function—maximizing Nmatches—is scientifically sound because a larger number of matched stars provides a more overdetermined geometric solution, which suppresses individual errors The hierarchical strategy ensures that the final parameters are optimized for the highest-fidelity data, and inter-night stability tests confirm that HBO maintains a 100% success rate across diverse conditions despite physical variability This adaptive strategy demonstrates the potential to mitigate high failure rates associated with static parameter configurations, offering a more flexible approach for future automated astronomical data processing pipelines The framework's ability to re-sense the state of the night ensures that the scientific integrity of long-term baselines is preserved without human intervention

Future Work

Future work will focus on refining the objective function and exploring additional morphological constraints to improve the reliability of source extraction Additionally, more sophisticated probabilistic models, including Gaussian Cox Processes, may be necessary to better account for non-Gaussian background noise Extensive testing on larger and more diverse datasets remains essential to determine the generalizability and long-term reliability of this method for large-scale astronomical surveys The scope of this study is limited to a finite set of observational scenarios and targets, but the feasibility of the proposed adaptive approach is demonstrated

References

Bancelin, D., Colas, F. p. 8, Thuillot, W. p. 8, Hestroffer, D. p. 8, & Assafin, M. p. 8 2012, A&A, 544, A15

Bernstein, G. M., Trilling, D. p. 8 E., Allen, R. p. 8 L., et al. 2004, AJ, 128, 1364

Beroiz, M., Cabral, J. p. 8 B., & Sanchez, B. p. 8 2020, Astronomy and Computing, 32, 100384

Chen, G., Xue, B., Li, X., Cao, J., & Yin, J. p. 8 2025, AJ, 169, 146

Chi, H. & Wang, F. 2025, ApJ, 986, 20

Chun-hui, B., Jiang-hui, J. p. 8, Dong-jie, T. p. 8, et al. 2025, Chinese Astronomy and Astrophysics, 49, 241

Goodeve, J. L. 2025, AJ, 169, 151

Hong, J. & Dickerson, J. p. 1 A. 2000, Journal of Guidance, Control, and Dynamics, 23, 728

Li, C.-W., Peng, Q.-Y., et al. 2020, Acta Astronomica Sinica, 61, 91

Lindeberg, T. 1994, Journal of applied statistics, 21, 225

Liu, T., Quan, Y., Su, Y., et al. 2025, Nature Astronomy, 9, 608

Lowe, D. G. 2004, International journal of computer vision, 60, 91

Malbet, F., Boehm, C. p. 9, Krone-Martins, A. p. 9, et al. 2021, Experimental Astronomy, 51, 845

Meurisse, A. & Carpenter, J. p. 0 2020, Planet. p. 9 Space Sci., 182, 104853

Shao, M., Nemati, B. p. 9 p. 9, Zhai, C., et al. 2014, ApJ, 782, 1

Snoek, J., Larochelle, H. p. 9, & Adams, R. p. 9 P. 2012, Advances in Neural Information Processing Systems, 25

Stetson, P. B. 1987, PASP, 99, 191

Tyson, J. A., Guhathakurta, P. p. 9, Bernstein, G. p. 0 M., & Hut, P. p. 9 1992, in American Astronomical Society Meeting Abstracts, Vol. p. 9 181, American Astronomical Society Meeting Abstracts, 06. p. 910

Wang, B., Zhao, H., & Li, B. 2017, Acta Astronomica Sinica, 58, 8

Yang, C.-T. 2019, Acta Astronomica Sinica, 60, 104

Zackay, B., Ofek, E. O., & Gal-Yam, A. p. 9 p. 9 2016, ApJ, 830, 27

Zhai, C., Shao, M., Nemati, B., et al. 2014, ApJ, 792, 60

Zhai, C., Shao, M., Saini, N. p. 9 S., et al. 2018, AJ, 156, 65

Zhai, C., Ye, Q., Shao, M., et al. 2020, PASP, 132, 064502

Zhang, Y., Wang, J., Su, J., Cheng, X. p. 9 p. 9, & Zhang, Z. p. 0 p. 0 2021, AJ, 162, 250

Zackay, B., Ofek, E. O., & Gal-Yam, A. 2016, ApJ, 830, 27

Zhang, Y., Wang, J., Su, J., Cheng, X., & Zhang, Z. 2021, AJ, 162, 250

Zhai C. Shao M Nemati B et al. p. 9 2014 ApJ 792 60

Zhai C Shao M Saini N S et al. 2018 AJ 156 65

Zhai C Ye Q Shao M et al. 2020 PASP 132 064502

Zhang Y Wang J Su J Cheng X & Zhang Z.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Adaptive Parameter Optimization for Image Alignment: The improved system can automatically determine optimal alignment parameters by framing parameter selection as an optimization problem and using Bayesian Optimization to solve it, moving beyond inefficient trial-error processes.

  2. Hierarchical Search Strategy Implementation: The system can efficiently search the parameter space by performing a two-phase approach: a Global Coarse Search on downsampled images to identify promising regions, followed by a Local Fine-tuning Search on full-resolution images for subpixel accuracy.

  3. Adaptive SNR Thresholding: The system can dynamically adjust its sensitivity based on the specific dataset quality by calculating an adaptive SNRmax from the brightest non-saturated source in real time, preventing the optimizer from exploring irrelevant high SNR values.

  4. Objective Function Maximization: The system will maximize Nmatches, defined as the number of successfully matched source pairs, which is a direct and intuitive measure of alignment quality that synergizes with RANSAC-based geometric matching to ensure a robust solution.

  5. Uncertainty-Guided Search Decisions: The system utilizes the Expected Improvement acquisition function, which balances exploration and exploitation by quantifying the expected improvement over the best value found so far using both the GP predictive mean and variance.

  6. Robustness Against Seeing Mismatch: The improved system can maintain a 100% success rate across diverse conditions, even on poor seeing nights, by adaptively re-tuning the detection kernel to match the actual atmospheric seeing of each specific epoch.

  7. Computational Efficiency for Large Surveys: The hierarchical strategy allows the framework to reach parameter stabilization quickly after an initial exploration phase, with total optimization times recorded at 47.19 seconds for a test dataset, making it feasible for real-time sky survey pipelines.

Related papers