---

title: >- [Paper Note] STARLINC: Satellite Trail Artifact Removal using Inter-Frame Correlation description: >- [ECCV 2026][segmentation][satellite trail removal] Tackling severe satellite trail contamination in astronomical surveys from LEO mega-constellations, STARLINC introduces an annotation-free deep framework combining synthetic trail simulation, inter-frame SSIM difference maps, and Grad-CAM coarse guidance, achieving 0.485 mIoU for precise pixel-level masking. tags: - ECCV 2026 - segmentation - satellite trail removal - astronomical imagery - weakly supervised learning date: 2026-09-19 content_hash: edfce41454bce775

STARLINC: Satellite Trail Artifact Removal using Inter-Frame Correlation

Conference: ECCV 2026
Paper: ECCV Official
Code: https://github.com/starioKim/STARLINC
Area: Segmentation
Keywords: satellite trail removal, astronomical imagery, inter-frame correlation, weakly supervised localization, synthetic data

TL;DR

STARLINC tackles the surging crisis of low Earth orbit satellite trail contamination in astronomical surveys by leveraging temporal redundancy across repeated exposures to construct SSIM-based differential maps and pairing them with classifier Grad-CAM heatmaps, training a U-Net on synthetic trails to achieve 0.485 mIoU pixel-level masking without requiring any manual pixel-level annotations on real astronomical images.

Background & Motivation

The rapid proliferation of low Earth orbit (LEO) satellite constellations, most notably Starlink, poses an escalating threat to ground- and space-based astronomical surveys. With over 12,000 active LEO satellites in orbit and numbers expanding at roughly 30% annually, satellite streaks are no longer isolated anomalies but ubiquitous contaminants. Recent observations from missions like SPHEREx report contamination in 73.3% of survey exposures, averaging 2.18 streaks per frame. These sunlight-reflecting satellite trails inject artificial photon flux into astronomical images, inducing severe photometric biases in measured stellar magnitudes and propagating into significant inaccuracies in derived physical properties, such as a 3.9% deviation in stellar mass and a 10% bias in stellar age estimates.

Current survey workflows face a crippling dilemma. Modern telescopes generate terabytes of data every night, far exceeding the capacity of manual visual screening. Furthermore, when contaminated exposures are identified, conventional pipelines often discard the entire frame, sacrificing valuable scientific data. At the same time, off-the-shelf computer vision solutions fail to bridge the domain gap. Astronomical images operate in an inherently photon-starved, ultra-low signal-to-noise ratio regime (SNR typically in the single digits, peaking around 16.8, compared to 35–116 in natural image datasets). Satellite trails lack semantic contextual cues and high-contrast boundaries; instead, they appear as faint, razor-thin lines embedded in instrumental and atmospheric noise. Classical line detectors like Hough and Radon transforms, as well as modern deep detectors like DeepLSD, suffer catastrophic performance drops when SNR falls into astronomical regimes. Moreover, existing astronomy-specific tools depend heavily on scarce, tedious pixel-level manual masks and exhibit poor cross-survey generalization.

To address these challenges, the authors recognize two intrinsic physical characteristics of time-domain astronomical surveys: satellite streaks are transient and occupy only a minute fraction of the image area, whereas repeated exposures of the same sky field capture virtually static celestial backgrounds. The core idea is to completely bypass real-image pixel annotations by injecting physically modeled thin synthetic trails for supervision, suppressing persistent stellar structures via local SSIM-based inter-frame differential maps, and incorporating classifier Grad-CAM heatmaps as coarse spatial priors to guide a U-Net in producing fine-grained pixel masks that exclude contaminated pixels while preserving valid observations.

Method

Overall Architecture

STARLINC consists of four interconnected stages: synthetic trail data injection, background suppression via inter-frame differential maps, pseudo-localization via classifier activation heatmaps, and multi-channel composite U-Net segmentation. Taking a temporal sequence of repeated exposures (e.g., 30 frames) of a target sky region as input, it produces a clean, binary pixel-level trail mask for the target frame.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Multi-epoch survey exposure sequence"] --> B["1. Synthetic Trail Generation<br/>Beta distribution models faint thin trails for mask-free training"]
    A --> C["2. Inter-Frame Differential Map<br/>Local window SSIM difference suppresses static celestial stars"]
    C --> D["3. Pseudo-Localization via Heatmaps<br/>Grad-CAM from binary classifier provides coarse spatial trail prior"]
    B -.->|"Offline synthetic supervision"| E
    A --> E["4. Multi-Channel U-Net Masking<br/>Composite input of image + diff map + heatmap yields pixel mask"]
    C --> E
    D --> E
    E --> F["Preserved Scientific Exposure<br/>Selective pixel masking retains clean observation regions"]

Key Designs

1. Synthetic Trail Generation with Skewed Thickness: Bypassing Manual Pixel Supervision

To resolve the dual bottlenecks of absent pixel annotations and extreme class imbalance in real survey data (where only ~20% of frames contain trails), STARLINC introduces a parametric synthetic trail generator. To replicate realistic telescopic line streaks, each synthetic trail is governed by length \(L\), orientation \(\omega\), thickness \(\varepsilon\), and brightness \(m\): - Length \(L \sim \text{clip}(\mathcal{N}(D, \sigma_L^2); \varpi, \varrho)\), with \(D = 0.5 \min(H, W)\); - Orientation \(\omega \sim \mathcal{U}(0, 2\pi)\); - Thickness \(\varepsilon \sim \text{Beta}(a, b)\), parameterized with shape parameters \((a, b) = (2, 6)\); - Brightness \(m \sim \mathcal{U}(m_{\min}, m_{\max})\).

The crucial design insight lies in adopting a Beta distribution biased towards small values for trail thickness \(\varepsilon\). Because satellite streaks under low-SNR survey conditions appear predominantly as faint, single-pixel-wide line segments that are easily drowned out by background fluctuations, heavily sampling thin lines during training forces the segmentation model to specialize in subtle, faint linear artifacts, yielding exact ground-truth masks \(Y_{\text{syn}}\) without human labeling.

2. Local SSIM-Based Inter-Frame Differential Maps: Suppressing Persistent Celestial Backgrounds

Standard pixel-wise subtraction (such as absolute or squared difference) is highly vulnerable to observational brightness variations, atmospheric seeing perturbations, and saturated stellar cores. Exploiting the fact that stars and distant galaxies remain identical across temporal exposures while satellite trails are strictly transient, STARLINC builds an inter-frame differential map using structural similarity within a \(k \times k\) sliding window \(\mathcal{N}_k(m, n)\):

\[M_{i,j}(m, n) = \sum_{(u, v) \in \mathcal{N}_k(m, n)} d(x_i(u, v), x_j(u, v))\]

The distance metric \(d(\cdot, \cdot)\) is formulated as the complement of SSIM to quantify local structural perturbation:

\[d(x_i, x_j) = 1 - \text{SSIM}(x_i, x_j) = 1 - \frac{(2\mu_i \mu_j + C_1)(2\sigma_{ij} + C_2)}{(\mu_i^2 + \mu_j^2 + C_1)(\sigma_i^2 + \sigma_j^2 + C_2)}\]

where \(\mu\) and \(\sigma\) denote local mean and variance, and \(\sigma_{ij}\) is local covariance. By averaging pairwise distances between target frame \(x_i\) and all remaining frames in the temporal group, persistent astronomical sources yield values near zero, while transient satellite trails generate high-response peaks that clearly outline the streak while filtering out stellar cores and background noise.

3. Pseudo-Localization via Classifier Heatmaps: Weakly Supervised Spatial Priors

Even with structural difference maps, residual artifacts from bright variable stars or sensor defects can introduce ambiguity. To provide a coarse spatial prior, STARLINC incorporates an auxiliary binary classifier that predicts whether an image contains a satellite trail (\(\hat{p}_{\text{line}}\)), taking the raw frame and its differential map as input.

Using Grad-CAM, the channel weights \(\varpi_k = \frac{1}{HW}\sum_{i}\sum_{j}\frac{\partial y_c}{\partial f^k(i, j)}\) are computed from the logit \(y_c\) with respect to the last convolutional feature maps \(f^k\), producing a coarse localization heatmap:

\[H_{\text{GC}} = \text{ReLU}\left(\sum_{k=1}^K \varpi_k f^k\right)\]

This activation map provides a global spatial prior indicating the approximate trajectory of the trail across the field, preventing false alarms in distant sky regions without requiring any pixel-level ground-truth training masks.

4. Composite Three-Channel U-Net Masking: End-to-End Fine Boundary Extraction

The final segmentation network ingests a concatenated three-channel tensor combining the original grayscale observation \(I\), the mean inter-frame differential map \(\bar{M}\), and the Grad-CAM heatmap \(H_{\text{GC}}\): \(X = [I, \bar{M}, H_{\text{GC}}] \in \mathbb{R}^{H \times W \times 3}\). A standard U-Net architecture \(f_\theta\) processes this composite input to predict a continuous probability map \(\hat{Y} = f_\theta(X) \in [0, 1]^{H \times W}\). Supervised strictly by synthetic masks \(Y_{\text{syn}}\) during training, the network accurately segments real satellite trails at test time, following their exact geometric boundaries.

Loss & Training

The framework is trained in two decoupled stages. First, a ResNet-based binary classifier is trained on survey images labeled only at the image level (trail-present vs. trail-free) using standard cross-entropy loss. Second, the classifier weights are frozen to generate Grad-CAM heatmaps, and synthetic trails are injected into clean 30-frame temporal groups. The U-Net segmentation network is trained end-to-end using the synthetic three-channel inputs and supervised by the synthetic binary masks. No real pixel-level masks are ever required.

Key Experimental Results

Main Results

Evaluation was conducted on real observations from the 7-Dimensional Telescope (7DT) surveying the Orion Molecular Clouds region (1,140 samples with manually annotated pixel masks via LabelMe). Table 1 compares STARLINC against classical line detectors, general deep detectors, and astronomy-specific models.

Category	Method	mIoU	Dice	Precision	Recall	ROC-AUC	PR-AUC
Classical	Hough [11]	0.142	0.202	0.785	0.293	0.692	0.741
Classical	Radon [30]	0.000	0.001	0.653	0.001	0.670	0.696
Classical	LSD [13]	0.074	0.126	0.077	0.517	0.776	0.714
Learning-based	DeepLSD [29]	0.116	0.192	0.120	0.642	0.867	0.827
Learning-based	DT-LSD [17]	0.240	0.332	0.663	0.507	0.769	0.838
Astronomy-specific	ASTA [34]	0.148	0.217	0.399	0.236	0.730	0.511
Astronomy-specific	MaxiMask [28]	0.155	0.268	0.838	0.159	0.795	0.830
Astronomy-specific	UnetLSD [10]	0.197	0.329	0.303	0.359	0.622	0.399
Proposed	STARLINC	0.485	0.653	0.815	0.545	0.918	0.865

In cross-dataset zero-shot transfer evaluations (Table 2), STARLINC trained purely on 7DT Orion data was directly applied to the Zwicky Transient Facility (ZTF, 4,875 images) and 7DT NGC galaxy field (4,812 images). STARLINC achieved 0.469 mIoU / 0.639 Dice on ZTF (surpassing fine-tuned MaxiMask at 0.365 / 0.535) and 0.471 mIoU / 0.640 Dice on NGC (more than doubling the best fine-tuned baseline DeepLSD at 0.199 / 0.331).

Ablation Study

The ablation experiments quantify the respective contributions of the inter-frame differential map and the heatmap localization prior (Table 3), as well as comparing alternative difference map operators (Table 4).

Table 1: Ablation of key components (corresponding to original Table 3)

Diff Map (§4.2)	Heatmap (§4.3)	mIoU	Dice	Precision	Recall	Note
✗	✗	0.000	0.000	1.000	0.000	Without temporal/spatial priors, U-Net fails completely
✓	✗	0.410	0.582	0.585	0.578	Diff map successfully suppresses stars, establishing strong baseline
✗	✓	0.198	0.331	0.276	0.411	Coarse heatmap alone yields fuzzy masks and background noise
✓	✓	0.485	0.653	0.815	0.545	Full STARLINC with dual complementary priors

Table 2: Comparison of inter-frame differential maps (corresponding to original Table 4)

Differential Operator	Core Mechanism / Characteristic	mIoU (Mean ± Std)
RMSE	Pixel-level root mean square; sensitive to star core residual	0.469 ± 0.014
MAE	Pixel-level absolute difference; retains stellar edges	0.472 ± 0.002
ZOGY [41]	Astronomical subtraction optimized for point sources; suboptimal on lines	0.392 ± 0.016
OIS [3]	Optimal image subtraction; kernel convolution deforms thin lines	0.431 ± 0.033
DIA [8]	Difference image analysis; targeted at variable point sources	0.436 ± 0.008
SSIM (Ours)	Local structural covariance; robustly isolates thin line streaks	0.485 ± 0.012

Key Findings

• Inter-frame correlation is the primary performance driver: Eliminating both auxiliary inputs collapses mIoU to 0.000, while adding the differential map alone jumps mIoU to 0.410. In ultra-low-SNR regimes, static single-frame features cannot separate thin streaks from background noise; temporal redundancy is indispensable.
• SSIM outperforms classic astronomical image subtraction: Classical subtraction algorithms like ZOGY (0.392 mIoU) and OIS (0.431 mIoU) are tailored for point-source PSF matching (e.g., supernovae) and degrade on extended non-Gaussian satellite lines. SSIM captures local structural differences, yielding superior streak isolation (0.485 mIoU).
• Foundation models struggle on raw astronomical images: In Table 5, zero-shot Segment Anything Model (SAM) achieves only 0.358 mIoU / 0.528 Dice, indicating that general-domain visual foundation models cannot reliably detect faint astronomical artifacts without temporal domain cues.

Highlights & Insights

• Annotation-free deep segmentation: By modeling satellite streak physics with a Beta-distributed thickness parameter, STARLINC trains directly on synthetic data while outperforming fully supervised baselines on real data.
• Converting temporal redundancy into spatial clarity: Instead of fighting noise in isolated exposures, the pipeline exploits telescope multi-exposure cadences to eliminate static stars via local window structural dissimilarity.
• Data preservation over image rejection: Replaces the conventional practice of discarding entire contaminated frames with precise pixel-level masking, safeguarding valuable astronomical observation time.

Limitations & Future Work

• Reliance on multi-epoch repeated exposures: STARLINC depends on consecutive overlapping exposures (30 frames in experiments). Its performance degrades under single-epoch, drift-scan, or non-aligned observation cadences where temporal matching is unavailable.
• Diffraction spikes from ultra-bright satellites: Extremely bright satellites cause CCD charge bleeding and telescope spider diffraction spikes that deviate from simple linear geometry. While PSF-aware synthetic rendering improves mIoU to 0.539 (Table 5), it trades off precision (0.748 vs. 0.815).
• Future directions: Integrating satellite orbital telemetry (TLE ephemeris) to provide predictive trajectory priors, and coupling the segmentation mask with astronomical deep inpainting models to reconstruct underlying celestial flux.

Related Work & Insights

• vs MaxiMask [28] & ASTA [34]: MaxiMask relies on extensive human pixel annotations and achieves high precision (0.838) but poor recall (0.159, mIoU 0.155) on unseen domains; STARLINC doubles the overlap performance (0.485 mIoU) without human pixel masks.
• vs DeepLSD [29] & DT-LSD [17]: General-domain line detectors assume high-contrast object boundaries, yielding false alarms and severe precision degradation (0.120 Precision for DeepLSD) under single-digit SNR astronomical noise; STARLINC demonstrates the necessity of physics-inspired astronomical priors.

Rating

• Novelty: ⭐⭐⭐⭐⭐ Elegant integration of astronomical temporal cadences with structural difference maps and Beta-distributed synthetic training.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ Rigorous benchmarking on real 7DT data with manual ground truth, cross-dataset transfer on ZTF and NGC fields, and comprehensive ablations.
• Writing Quality: ⭐⭐⭐⭐⭐ Well-structured narrative connecting astrophysical measurement errors directly to computer vision methodology.
• Value: ⭐⭐⭐⭐⭐ Crucial open-source tool addressing a severe, accelerating crisis in global observational astronomy.

Related Papers

• [ICML 2026] Segment Anything with Robust Uncertainty-Accuracy Correlation
• [ICML 2025] ConText: Driving In-context Learning for Text Removal and Segmentation
• [AAAI 2026] Generalizable Slum Detection from Satellite Imagery with Mixture-of-Experts
• [ICLR 2026] LiFR-Seg: Anytime High-Frame-Rate Segmentation via Event-Guided Propagation
• [ICCV 2025] Object-level Correlation for Few-Shot Segmentation