🔭 astrophysics

Stellar Density Classification and Regression for CSST Multi-color Imaging Using Deep Learning

This paper proposes a hierarchical two-stage deep learning framework that classifies CSST imaging data into six stellar density categories and performs regression for bright star prediction to optimize source extraction and calibration across the telescope's vast dynamic range of stellar densities.

Original authors: Jinzhi Lai, Man I Lam, Jianjun Chen, Xin Zhang, Hao Tian, Xiaohan Chen, Jialu Nie, Ming Yang, Chao Liu

Published 2026-05-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jinzhi Lai, Man I Lam, Jianjun Chen, Xin Zhang, Hao Tian, Xiaohan Chen, Jialu Nie, Ming Yang, Chao Liu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the Chinese Space Station Survey Telescope (CSST) as a giant, ultra-high-definition camera floating in space, tasked with taking a massive panoramic photo of the entire universe. The problem is that the universe isn't uniform. Some parts of the sky are like empty deserts with just a few stars (voids), while other parts are like the most crowded city square on Earth, packed with millions of stars and galaxies (like the center of our Milky Way).

The paper describes a new "smart assistant" built using deep learning (a type of artificial intelligence) to help the telescope's computer system handle these wildly different environments without getting confused.

Here is how the system works, broken down into simple concepts:

The Problem: One Size Does Not Fit All

If you try to use the same photo-editing software to fix a picture of an empty desert and a picture of a crowded concert, you'll run into trouble.

In the empty desert: Standard tools might get confused by background noise, thinking a speck of dust is a star.
In the crowded concert: Standard tools might try to count individual people but fail because everyone is standing too close together, blending into a single blob.

The CSST telescope faces this exact issue. It needs to count stars and measure their positions accurately, whether it's looking at an empty patch of sky or the dense center of our galaxy.

The Solution: A Two-Stage "Traffic Cop" System

The authors created a two-step AI system that acts like a smart traffic cop, deciding exactly how to process each photo based on how crowded it is.

Step 1: The "Crowd Detector" (Classification)

First, the system looks at the image and asks a simple question: "How crowded is this?"

It uses a neural network called ResNet-34 (think of it as a highly trained expert who can instantly recognize patterns).
It sorts every image into one of six categories, ranging from "Empty Field" to "Super-Crowded Galactic Center."
The Result: This step is incredibly accurate (98.83% correct). It acts as a gatekeeper. If the image is empty, it sends it down one path. If it's a super-crowded city of stars, it sends it down a different, more specialized path designed to untangle the mess.

Step 2: The "Star Counter" (Regression)

Once the system knows the general crowd level, it needs to know exactly how many bright stars are in the picture. These bright stars are crucial because they act like streetlights for navigation, helping the telescope know exactly where it is pointing.

It uses a slightly deeper, more powerful AI called ResNet-50.
Instead of just sorting the image, this AI predicts the exact number of bright stars.
The Result: It predicts the count with high precision. If the real number is 100, the AI's guess is usually within about 21% of that number. This is good enough to ensure the telescope's navigation is accurate.

Why This Matters

Before this system, scientists had to use rigid rules. They might have had to manually decide which parts of the sky to study or use slow, complicated math that often failed in crowded areas.

This new AI system is like having a smart, automated switchboard:

It instantly recognizes the environment.
It routes the data to the right tool (a simple tool for empty space, a complex tool for crowded space).
It ensures that the final data is clean and accurate, whether the telescope is looking at the edge of the universe or the heart of our galaxy.

The Bottom Line

The paper demonstrates that by splitting the job into "sorting" (classification) and "counting" (regression), the telescope can handle the extreme variety of the universe. It prevents the computer from getting overwhelmed in crowded areas and ensures it doesn't miss details in empty areas. This makes the data from the Chinese Space Station Survey Telescope much more reliable for future scientific discoveries.

Technical Summary: Stellar Density Classification and Regression for CSST Multi-color Imaging Using Deep Learning

Problem Statement
The Chinese Space Station Survey Telescope (CSST) is designed to map the universe across an unprecedented dynamic range of stellar densities, spanning from extragalactic voids (a few sources) to the crowded Galactic center (>10⁵ sources per detector). Processing such heterogeneous data with a single, general-purpose source extraction pipeline introduces significant systematic uncertainties. Standard algorithms, such as SExtractor, suffer from poor accuracy and increased astrometric uncertainty in crowded fields due to source blending and background noise dominated by unresolved sources. Conversely, specialized iterative tools like DAOPHOT are computationally prohibitive and unstable in sparse regions. The CSST's continuous survey mode inevitably traverses these varying density regimes, necessitating an adaptive approach to ensure photometric and astrometric fidelity.

Methodology
To address these challenges, the authors propose a hierarchical, two-stage deep learning framework designed for adaptive data reduction. The system decouples density characterization from source extraction, allowing downstream algorithms to be optimally matched to the specific stellar density environment.

Data Preparation:
- Simulation: The study utilizes high-fidelity pixel-level mock data generated by the CSST simulation software, covering a 25 deg² sky region and various representative fields (including the Galactic Center, Fornax, and Sculptor dwarf spheroidals).
- Preprocessing: Original 9232×9216 pixel images are processed using a mask-aware standardization pipeline. Invalid pixels (NaNs, infinities) are zero-filled, and valid pixels are standardized. Images are resized to 224×224 pixels, creating a two-channel tensor (image + binary mask) to preserve spatial information regarding data voids.
- Augmentation: Training sets are expanded fourfold using horizontal flipping, discrete rotations, and random color jittering to enhance diversity and reduce overfitting.
Stage 1: Density Classification (ResNet-34):
- Objective: To categorize images into six discrete density classes (ranging from 0 to >10⁵ stars per image).
- Architecture: A ResNet-34 model is employed, modified to accept dual-channel inputs. It utilizes residual connections to mitigate vanishing gradients and global average pooling to reduce overfitting.
- Training: The model is trained using Cross-Entropy loss and the AdamW optimizer with gradient clipping and early stopping.
- Selection: Among tested architectures (VGGNet-16, DenseNet-121, AlexNet, Baseline CNN), ResNet-34 was selected for its optimal balance of high accuracy (98.83%) and computational efficiency (2.77 ms/image), avoiding the memory bottlenecks of DenseNet and the lower accuracy of lighter models.
Stage 2: Bright Star Regression (ResNet-50):
- Objective: To predict the precise count of bright stars (<23.5 mag) within a field, a critical parameter for astrometric calibration.
- Architecture: A deeper ResNet-50 model, pre-trained on ImageNet and fine-tuned, serves as the regression engine. The classification head is replaced with a custom regression head (512-256-128-1 neurons) with dropout and batch normalization.
- Training: A staged training approach is used, progressively unfreezing layers. A custom combined loss function (integrating MSE, MAE, and Relative MSE) is applied to handle the wide dynamic range of star counts. Labels undergo a logarithmic transformation ( $y' = \log_{10}(1+y)$ ) to stabilize training.

Key Contributions

Hierarchical Framework: The paper introduces a two-stage pipeline that uses classification as a "decision gate" to route data to specialized processing branches, while simultaneously using regression to assess calibration quality.
Adaptive Processing: By decoupling density estimation from source extraction, the model ensures that photometric and astrometric algorithms are dynamically matched to the local stellar density, mitigating systematic errors in both crowded and sparse fields.
Model Optimization: The study rigorously evaluates multiple CNN backbones, demonstrating that ResNet-34 offers the best trade-off between accuracy and throughput for the classification task, while ResNet-50 provides the necessary capacity for precise regression.
Robustness to Dynamic Range: The framework successfully handles a stellar density range spanning five orders of magnitude (0 to >10⁵), a scale that typically breaks standard pipelines.

Results

Classification Performance: The ResNet-34 classifier achieved a global accuracy of 98.83%. It demonstrated exceptional performance at extremes, with 100% precision/recall for empty fields (Class 0) and 99.62% for the most crowded fields (Class 5). Misclassifications were primarily "soft errors" occurring at class boundaries (e.g., near the 1,000-source threshold), which have minimal impact on downstream pipeline routing.
Regression Performance: The ResNet-50 regression model achieved a Mean Absolute Error (MAE) of 0.0824 dex and an RMSE of 0.1333 on the logarithmic scale. This translates to an average prediction error factor of approximately 1.21 (±21%). The model achieved an $R^2$ score of 0.9076, indicating it captures over 90% of the variance in bright star counts.
Efficiency: The inference time is highly efficient, with the classification model processing images in ~2.77 ms and the regression model in ~0.1 ms, making it suitable for real-time integration into the CSST data reduction pipeline.

Significance and Claims
The authors claim that this framework significantly enhances the fidelity and homogeneity of CSST data products by ensuring that the correct reduction algorithms are applied based on the local stellar density. By acting as a critical decision gate, the classification stage prevents standard algorithms from failing in crowded fields and avoids the computational waste of using specialized tools in voids. The regression stage provides a reliable metric for astrometric calibration, allowing the pipeline to flag fields with insufficient reference sources.

The paper positions this work as a foundational module for future large sky surveys that must operate across heterogeneous stellar landscapes. While the current models are trained on simulated data, the authors assert that the framework provides a robust, automated, and computationally efficient foundation that can be fine-tuned with actual CSST commissioning data to bridge the gap between simulation and observation. The approach is presented as a solution to the systematic biases introduced by static, one-size-fits-all survey strategies.