Tsuki: YOLOv8-Trained Optical Debris Detection
Neural network trained on telescope debris imagery identifies orbital objects with sub-centimeter detection capability. YOLOv8 architecture accelerates optical tracking workflows.
What is Tsuki?
Tsuki (月 - Japanese for "moon") is a YOLOv8-trained object detection system for optical telescope debris tracking. Trained on weeks of telescope imagery showing debris against star backgrounds. Automates detection that traditionally required manual analyst review. Sub-centimeter debris visible through matched filtering integration.
Optical Debris Tracking Challenge
Optical telescopes photograph the sky to track satellites and debris. Objects appear as streaks or points against star backgrounds. Traditional processing requires analysts to manually identify debris tracks, distinguish them from stars, cosmic rays, and image artifacts, then extract astrometric measurements for orbit determination.
Manual analysis is time-consuming and doesn't scale to large survey telescopes generating thousands of images per night. Automated detection algorithms based on fixed thresholds and morphology filters produce high false positive rates, requiring significant analyst time filtering results.
Machine learning object detection trained on real telescope imagery learns to distinguish debris from false positives more reliably than rule-based algorithms. Tsuki automates the detection pipeline, allowing analysts to focus on orbit determination and conjunction assessment rather than image processing.
YOLOv8 Architecture
YOLO (You Only Look Once) is a convolutional neural network architecture for real-time object detection. Unlike classification networks that scan images with sliding windows, YOLO processes the entire image in a single pass, predicting bounding boxes and class probabilities simultaneously. Version 8 improves accuracy and speed over earlier YOLO variants.
For debris detection, YOLOv8 takes telescope images as input and outputs bounding boxes around detected debris objects with confidence scores. The network learns features distinguishing debris streaks from stars (point sources vs. elongated tracks), satellite glints (brightness patterns), cosmic rays (morphology), and hot pixels (spatial distribution).
Training required weeks of GPU time processing labeled telescope imagery. Training data included debris at various magnitudes, track lengths, orientations, and background star densities. Data augmentation (rotation, brightness scaling, noise injection) improved generalization to new observing conditions.
Training Process
Data Collection & Labeling
Gathered telescope imagery spanning multiple observation campaigns with known debris tracks. Analysts manually labeled debris bounding boxes and classified object types. Training set included 50,000+ labeled images with 200,000+ debris detections across magnitude range 14-20.
Network Training
Initialized YOLOv8 with COCO pre-trained weights then fine-tuned on debris dataset. Training used Adam optimizer with cosine annealing learning rate schedule. Batch size 32, input resolution 640×640 pixels. Trained for 300 epochs over 2 weeks on 4×A100 GPUs. Validation loss plateaued at epoch 280.
Performance Validation
Tested on held-out telescope imagery not used in training. Achieved 94% recall at 95% precision for magnitude 16-18 objects. Dimmer objects (magnitude 19-20) showed 78% recall due to low signal-to-noise. False positive rate 0.2 per image, significantly lower than rule-based methods (5-10 false positives per image).
Sub-Centimeter Detection Capability
Sub-centimeter debris detection requires combining optical detection with matched filtering signal processing. Optical telescopes can't resolve centimeter-scale objects at orbital distances directly, but debris photometric brightness and spectral signatures allow size estimation.
Tsuki identifies candidate detections from telescope data. Matched filtering then analyzes temporal brightness variations, spectral characteristics, and photometric patterns to constrain object size. Objects with sufficient reflectivity and favorable lighting geometry produce detectable signals even at sub-centimeter scale.
Not all sub-centimeter debris is detectable. Detection depends on albedo (surface reflectivity), phase angle (sun-object-observer geometry), and distance. Small debris at GEO (36,000 km) is much harder to detect than same-size debris in LEO (500 km) due to inverse-square law signal strength decrease. Sub-centimeter capability applies to favorable LEO cases with high albedo.
Integration with Tracking Systems
Tsuki runs on incoming telescope imagery in real-time, flagging debris detections within seconds of image capture. Detected objects feed astrometric reduction pipelines that extract right ascension and declination measurements. These observations integrate with orbit determination to update object catalog.
New detections not matching existing catalog entries trigger follow-up observations to confirm discovery and establish preliminary orbits. Newly cataloged objects enter conjunction screening for collision risk assessment.
Optical tracking complements radar coverage by detecting objects in orbits or size ranges where radar sensitivity is limited. Tsuki's automation enables processing larger volumes of telescope imagery than manual workflows, expanding catalog coverage.
Detection Method Comparison
| Method | Automation | False Positives/Image | Recall (mag 16-18) |
|---|---|---|---|
| Manual Analysis | None | 0 (by definition) | ~99% (slow) |
| Rule-Based Thresholding | Full automation | 5-10 | 85% |
| Tsuki (YOLOv8) | Full automation | 0.2 | 94% |
Frequently Asked Questions
What is YOLOv8?
YOLOv8 (You Only Look Once version 8) is a convolutional neural network for real-time object detection. It processes entire images in one pass, predicting object locations and classes simultaneously. Version 8 improves speed and accuracy over earlier YOLO variants through architectural refinements.
How long did training take?
Training ran for 300 epochs over 2 weeks using 4×NVIDIA A100 GPUs. Data preparation (labeling, augmentation) required additional weeks. Total project timeline from data collection to deployed model: approximately 2 months including validation testing.
Can Tsuki detect debris during daytime?
No. Optical telescopes require dark sky to detect faint orbital objects. Most optical tracking occurs during twilight and night. Radar systems provide daytime coverage. Combined radar and optical tracking provides 24/7 surveillance when systems coordinate.
What is magnitude in optical astronomy?
Magnitude measures object brightness. Lower numbers are brighter. Magnitude 0 is bright star visible easily. Magnitude 6 is dimmest star visible to naked eye. Satellites range magnitude 4-10 depending on size and distance. Small debris at LEO: magnitude 16-20. Each magnitude is 2.5× brightness difference.
How does Tsuki handle different telescopes?
Training data included imagery from multiple telescope systems with varying optics, detectors, and image scales. Data augmentation simulated different noise characteristics. Network generalizes reasonably to new telescopes but benefits from fine-tuning on telescope-specific imagery for optimal performance.
Can sub-centimeter debris really be detected optically?
Sub-centimeter detection requires favorable conditions: high albedo (reflective surface), optimal phase angle (sun-object-observer geometry), short distance (LEO not GEO), and integration with matched filtering for signal enhancement. Not all sub-cm debris is detectable, but capability exists for favorable cases.
What happens to false positive detections?
False positives (objects flagged as debris that aren't) proceed to astrometric reduction. If positions don't match catalog objects and follow-up observations fail to confirm, they're discarded. 0.2 false positives per image is manageable. Traditional methods produce 5-10× more requiring significantly more analyst time.
Why is it called Tsuki?
Tsuki (月) means "moon" in Japanese. The name reflects optical tracking tradition of observing celestial objects and acknowledges international collaboration in space surveillance. Short, memorable names facilitate communication between engineering teams and operations staff.
Learn More
Explore LIGO-validated matched filtering for signal enhancement or learn about orbit determination from observations.