94% Precision: YOLO11‑Based Detection of Near‑Earth Object and Satellite Streaks

Background

Near‑Earth objects (NEOs) are defined as asteroids with perihelion < 1.3 AU, making them a focus of planetary‑defence monitoring. Modern wide‑field surveys produce thousands of images per night, and the increasing number of artificial satellites and debris creates linear streaks that obscure faint moving targets.

Dataset

Real observations were obtained at the La Sagra Observatory (MPC code L98) with a Celestron C14 + Fastar f/2.1 telescope and an SBIG ST‑10X CCD. The instrument provides a pixel scale of ~4.12″ px⁻¹, a field of ~74.9′ × 50.5′, and images of 1092 × 736 px with exposure times from 8 s to 120 s (limiting magnitude 19–20). After flat‑ and dark‑field correction, 2055 images were collected and 765 linear streaks were manually annotated; streak lengths range from 8.5 px to 1161 px (average ≈ 203 px). To augment long‑streak examples, 280 synthetic streaks were generated and injected into real frames. Synthetic streaks span five brightness levels, include multi‑satellite cases (≈ 10 % of images), have a minimum length of 269 px, and use point‑spread‑function and Fourier‑based detector modeling to reproduce real imaging characteristics.

Pre‑processing and Splitting

All images were normalized and converted to PNG. The dataset was partitioned 70 % training / 20 % validation / 10 % test while preserving class ratios. Frames from each night were aligned to a common reference; regions without overlap were discarded.

Detection Module

The core detector is YOLO11‑OBB, a single‑stage network for rotated objects that directly predicts oriented bounding boxes (OBB). This eliminates post‑processing rotation steps and matches the elongated, tilted geometry of satellite or asteroid streaks.

Processing Pipeline

Convert FITS to normalized PNG.

Run YOLO11‑OBB to obtain candidate OBBs with confidence scores.

Cross‑match candidates with the Gaia star catalogue; discard detections near bright stars to suppress diffraction‑spike false positives.

Geometric refinement: analyse the photometric profile along the OBB’s principal axis, extend the box to the true streak endpoints, and cluster corner points to determine precise endpoints and centre.

Frame‑to‑frame association: compute pixel velocity and direction to link detections across consecutive frames into continuous trajectories.

Convert refined detections to the Minor Planet Center (MPC) standard format, cross‑match with satellite ephemerides, assign a confidence score (using a two‑component Gaussian model), and store the records in a SQLite database.

Evaluation

On an independent test set of 273 images (input resolution 640 px, confidence threshold 0.25, IoU 0.45), the model achieved 94 % precision and 97 % recall, detecting 107 of 110 real streaks. Catalogue filtering removed 77 % of bright‑star false positives. Geometric refinement corrected short OBBs by extending them based on photometric profiles.

Results

Compared with manual inspection, the system provides substantial gains in efficiency, repeatability, and sensitivity, enabling large‑scale statistics of space‑object contamination and automated archival pipelines. Detection results are directly integrated into a normalized database, ready for scientific analysis.

Reference: StreakMind: AI detection and analysis of satellite streaks in astronomical images with automated database integration , arXiv:2605.03429.

Code example

本文
约2500字
，建议阅读
5
分钟
西班牙皇家海军学院天文观测站构建的 StreakMind 系统，能够自动识别天文图像中由卫星或小行星拖出的线性轨迹，提取轨迹的长度、位置和方向，为后续的天体测量和数据库入库提供标准化输出。其在独立测试集上，模型对短、中、长拖影均表现可靠，整体精确率达 94%、召回率 97%，110 条真实拖影中成功检出 107 条。