STARLINC: Satellite Trail Artifact Removal using Inter-Frame Correlation

• More than 12,000 active LEO satellites are increasing satellite-trail contamination in astronomical surveys.Source: §1 · p. 1
• Manual inspection cannot scale to terabytes of nightly survey data, and discarding contaminated images loses otherwise valuable observations.Source: §1 · p. 2
• Existing line detectors fail on astronomical data because sparse, low-SNR imagery differs from general-domain training conditions.Source: §1 · p. 2
• STARLINC performs pixel-level segmentation without real astronomical pixel annotations, selectively masking contaminated pixels while preserving valid data.Source: §1 · pp. 2–3
• +102% relative mIoU over the strongest baseline was demonstrated on real 7DT data, with transfer evaluated on ZTF and NGC-region observations.Source: §1 · p. 3

• Satellite trails corrupt stellar photometry, so a single affected datapoint can impair time-domain analyses.Source: §2.1 · pp. 3–4
• Brightness and magnitude errors can propagate into inaccuracies in derived stellar quantities such as mass and age.Source: §2.1 · p. 4
• Astronomical images have SNR values of at most 16.84, versus median SNR values of 35 to 116 in general image datasets.Source: §2.2 · p. 4
• Satellite trails are faint, thin, isolated signals embedded in noise rather than high-contrast contextual edges.Source: §2.2 · p. 5
• DeepLSD performance remains stable above SNR 70 but degrades rapidly near astronomical levels, with sharp metric drops in the single-digit SNR regime.Source: §2.2 · p. 5

• Classical detectors use transforms, gradients, or edge detection to identify line structures from image pixels.Source: §3.1 · p. 5
• Deep learning approaches recast line detection as region coloring, graph inference, or learning in Hough-transform parameter space.Source: §3.2 · p. 5
• STARLINC uses synthetic trail injection, inter-frame differential maps, and classifier heatmaps to guide segmentation without manual masks.Source: Figure 4 · p. 6
• Astronomy-specific methods include boundary tracing, morphological filtering, CNNs, and machine-learning classifiers for satellite trails.Source: §3.3 · p. 6
• Existing trail-segmentation methods require expensive manually examined pixel-level annotations for supervision.Source: §3.3 · p. 6

4.1 Synthetic Trail Generation for Training

• STARLINC addresses missing pixel-level annotations and trail-image imbalance by generating synthetic trails for scalable training.Synthetic trails are injected into clean astronomical images, producing paired images and pixel-level ground-truth masks.Source: §4.1 · p. 7
• Synthetic trails are parameterized by length, orientation, thickness, and brightness, with thickness sampling biased toward thin trails.This emphasizes faint, narrow structures that are difficult to distinguish from low-SNR background noise.Source: §4.1 · p. 7

4.2 Inter-Frame Differential Map for Background Suppression

• Inter-frame differential maps suppress persistent celestial structures and emphasize transient satellite trails across repeated observations.STARLINC aggregates local structural differences within sliding windows and uses SSIM complement distances for robustness to observational noise.Source: §4.2 · p. 8

4.3 Pseudo-localization via Activation Maps

• A trail-presence classifier produces Grad-CAM heatmaps that provide coarse spatial localization cues directly to the segmentation network without pixel-level supervision.The heatmap is formed from weighted final-layer feature maps and ReLU activation.Source: §4.3 · pp. 8–9

4.4 Overall Pipeline for Trail Masking

• The U-Net receives the original image, mean differential map, and Grad-CAM heatmap as a three-channel input and predicts a binary trail mask.Training uses synthetic masks, while inference thresholds the predicted trail-probability map.Source: §4.4 · p. 9

Experimental Settings

• The 7DT Orion Molecular Clouds dataset contains 1,140 samples split into 720 training, 180 validation, and 240 test samples.Temporally adjacent frames construct inter-frame differential maps with window size k = 5.Source: §5.1 · p. 10

Segmentation Evaluation

• Several baselines show discrimination without accurate localization: Hough has precision 0.785, recall 0.293, and mIoU 0.142.Radon and LSD achieve mIoU 0.000 and 0.074, respectively.Source: §5.2 · p. 11
• STARLINC achieves the best localization, with mIoU 0.485 and Dice 0.653, while maintaining precision 0.815 and recall 0.545.These results double the best baseline mIoU.Source: §5.2 · p. 11
• STARLINC obtains ROC-AUC 0.918 and PR-AUC 0.865, with the favorable trade-off persisting across decision thresholds.Figure 5 presents ROC and PR curves for the segmentation model.Source: §5.2 · p. 11, Figure 5 · p. 11

Qualitative Analysis

• Qualitatively, STARLINC produces continuous masks closely following trail geometry, unlike fragmented, dispersed, or absent baseline masks.The comparison includes Hough, DT-LSD, MaxiMask, and STARLINC across astronomical images.Source: §5.3 · pp. 11–12, Figure 6 · p. 12

Cross-Dataset Validation

• STARLINC achieves the best mIoU and Dice on both ZTF and NGC-region datasets without target-dataset fine-tuning.Inter-frame differential maps emphasize transient trails while suppressing persistent celestial structures.Source: §5.4 · pp. 12–13, Table 2 · p. 13

Ablation Study

• Removing both the inter-frame differential map and heatmap yields zero segmentation performance, while their combination performs best.The differential map suppresses persistent celestial structures and the heatmap guides trail localization.Source: §6.1 · p. 13

Differential Map Comparison

• SSIM provides the strongest differential-map cue because it better captures frame-specific structural changes than pixel-wise flux differences.RMSE and MAE are sensitive to brightness fluctuations and residual stellar cores, whereas ZOGY, OIS, and DIA target point-source variability and normalized subtraction.Source: §6.2 · pp. 13–14

Alternative Designs

• Image-only SAM segmentation reaches mIoU 0.358 and Dice 0.528, below STARLINC’s performance.The comparison removes both inter-frame differential maps and classifier heatmaps.Source: §6.3 · p. 14
• PSF-aware rendering raises mIoU from 0.485 to 0.539 and Dice from 0.653 to 0.700, but precision falls from 0.815 to 0.748.The result indicates improved overlap accompanied by reduced precision.Source: §6.3 · pp. 14–15
• Thresholding or extracting lines directly from differential maps yields high recall but very low precision because stellar cores, detector artifacts, and fluctuations also respond strongly.The differential map is therefore used as a segmentation cue rather than a standalone mask generator.Source: §6.3 · p. 15
• IRNet and SEAM underperform STARLINC because their activation maps capture discriminative trail fragments rather than the full thin trail extent.STARLINC uses the heatmap as an auxiliary cue for pixel-level segmentation.Source: §6.3 · p. 15

Conclusion

• STARLINC combines temporal redundancy, inter-frame differential maps, synthetic trail generation, and activation-map pseudo-localization for pixel-level segmentation without real survey masks.The implementation uses a U-Net with three-channel inputs.Source: §7 · p. 15, A Detailed Experiment Settings · p. 19
• Selective masking of contaminated regions avoids discarding entire frames and is presented as a practical approach for large-scale survey artifact handling.The conclusion frames this as a scalable application of the method.Source: §7 · p. 15
• Cross-dataset evaluation applies a model trained on one dataset directly to Serpens and Messier 81 images without additional retraining.These fields differ in background structure and object distribution.Source: B Additional Cross-Dataset… · p. 19
• STARLINC localizes faint and multiple satellite trails while preserving surrounding celestial sources in unseen fields.The qualitative examples include a barely visible trail in Serpens and multiple trails in Messier 81.Source: B Additional Cross-Dataset… · p. 19

Input Modalities

• The original image contains the astronomical scene, the differential map captures temporal inconsistency, and the heatmap supplies classifier-derived localization cues.Together they form the segmentation model’s three-channel input.Source: C Qualitative Results of Inter-Frame… · pp. 19–20, Figure 7 · p. 20
• Trail-containing frames t1, t3, and t4 produce strong differential-map and heatmap responses, whereas trail-free frame t2 shows weak responses.The trail is transient relative to comparison frames, creating localized structural inconsistency.Source: C Qualitative Results of Inter-Frame… · p. 20, Figure 7 · p. 20
• The complementary cues support accurate segmentation by combining temporal anomaly detection with visual trail localization.The original image, differential map, and heatmap are used jointly rather than as isolated representations.Source: C Qualitative Results of Inter-Frame… · p. 20

Qualitative Comparisons

• Classical detectors often trigger on bright celestial structures or galaxy edges and miss faint or partially visible trails.DeepLSD and DT-LSD improve sensitivity in some cases but can remain fragmented or unstable.Source: D Further Qualitative Comparison · p. 21
• STARLINC produces cleaner, more coherent localization while preserving surrounding celestial structures across faint, complex, and variable-brightness trails.Figure 10 broadens qualitative evaluation across trail appearances and background conditions.Source: D Further Qualitative Comparison · pp. 20–21

Classification Results

• STARLINC achieves classification accuracy 0.983, F1 0.959, and ROC-AUC 0.998 with precision and recall both equal to 0.959.Removing synthetic trail generation and the inter-frame differential map degrades classification performance.Source: §E.2 · pp. 21–22

Classification sensitivity

• F1 performance generally improves as classifier training samples increase, particularly from 150 to 450 samples.Source: §E.3 · p. 23
• A 50% synthetic ratio reaches the highest F1 score of 0.959 with 900 training samples.Source: §E.3 · p. 23
• Higher synthetic ratios, including 70% and 90%, degrade classifier performance across most dataset sizes.Source: §E.3 · p. 23

Segmentation sensitivity

• STARLINC achieves 0.504 mIoU even when synthetic trails comprise 95.1% of trail-containing images, comparable to the reference setting.Source: F Synthetic Fraction Sensitivity · p. 24

Fine-tuned baselines

• Fine-tuning DeepLSD, DT-LSD, ASTA, and MaxiMask improves their applicability to 7DT observations, but their performance remains below STARLINC.Source: G Fine-tuned Baseline Results · p. 24

Low-SNR motivation

• As SNR decreases, DeepLSD misses valid line segments and produces fragmented or unstable predictions.Source: H Additional Motivational Results on… · p. 25
• Line-detector accuracy degrades substantially when astronomical signals are weak relative to background noise.Source: H Additional Motivational Results on… · p. 25

Backbone sensitivity

• STARLINC mIoU remains stable from 0.473 to 0.488 across tested U-Net sizes.Source: I Network Size Sensitivity · p. 25

Input bit depth

• The study focuses on detectable faint trails with SNR > 3, where visual or algorithmic labeling is considered meaningful.Source: J Effect of Input Bit Depth · pp. 25–26
• The 16-bit STARLINC variant achieves comparable but slightly lower performance than the main 8-bit setting.Source: J Effect of Input Bit Depth · p. 26
• FITS-native or fully 16-bit processing is treated as an important direction for future work.Source: J Effect of Input Bit Depth · p. 26

• Full-image rejection removes valid astronomical information outside detected trail regions along with the artifact.Source: K Data Preservation through… · p. 26
• Pixel-level removal leaves uncontaminated regions usable as satellite contamination becomes more frequent.Source: K Data Preservation through… · pp. 26–27
• Conventional image rejection retains about 79.58% of data, whereas selective masking preserves approximately 99.95% of valid pixels.Source: K Data Preservation through… · p. 27

Limitations

• Broader validation is limited because many modern astronomical observations remain proprietary during release periods.Source: L Limitations · p. 27
• Future public dataset releases would enable validation across more diverse astronomical observations.Source: L Limitations · p. 27

Additional qualitative results

• STARLINC predictions on Serpens and Messier 81 accurately localize trails while preserving surrounding celestial structures.Source: Figure 9 · p. 28
• The paper compares inter-frame differential maps computed with MAE, RMSE, and SSIM.Source: Figure 11 · p. 31