Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Detecting streaks in smart telescopes images with Deep Learning

Olivier Parisot, Mahmoud Jaziri

TL;DR

This study tackles the challenge of detecting satellite- and object-induced streaks in raw astronomical images captured with smart telescopes. It evaluates four deep-learning pipelines—ASTRiDE, a ResNet50 classifier with XRAI, a Pix2Pix heatmap proxy, and YOLOv7—on the MILAN Sky Survey dataset collected over a year, including real and synthetic streaks. Results show that streak contamination is relatively rare (around 0.16%), with YOLOv7 offering the best balance of precision and speed, while ASTRiDE excels at filtering non-streaks but yields more false positives. The work provides open datasets for field use and demonstrates practical detection pathways to mitigate streak effects and support Space Domain Awareness.

Abstract

The growing negative impact of the visibility of satellites in the night sky is influencing the practice of astronomy and astrophotograph, both at the amateur and professional levels. The presence of these satellites has the effect of introducing streaks into the images captured during astronomical observation, requiring the application of additional post processing to mitigate the undesirable impact, whether for data loss or cosmetic reasons. In this paper, we show how we test and adapt various Deep Learning approaches to detect streaks in raw astronomical data captured between March 2022 and February 2023 with smart telescopes.

Detecting streaks in smart telescopes images with Deep Learning

TL;DR

This study tackles the challenge of detecting satellite- and object-induced streaks in raw astronomical images captured with smart telescopes. It evaluates four deep-learning pipelines—ASTRiDE, a ResNet50 classifier with XRAI, a Pix2Pix heatmap proxy, and YOLOv7—on the MILAN Sky Survey dataset collected over a year, including real and synthetic streaks. Results show that streak contamination is relatively rare (around 0.16%), with YOLOv7 offering the best balance of precision and speed, while ASTRiDE excels at filtering non-streaks but yields more false positives. The work provides open datasets for field use and demonstrates practical detection pathways to mitigate streak effects and support Space Domain Awareness.

Abstract

The growing negative impact of the visibility of satellites in the night sky is influencing the practice of astronomy and astrophotograph, both at the amateur and professional levels. The presence of these satellites has the effect of introducing streaks into the images captured during astronomical observation, requiring the application of additional post processing to mitigate the undesirable impact, whether for data loss or cosmetic reasons. In this paper, we show how we test and adapt various Deep Learning approaches to detect streaks in raw astronomical data captured between March 2022 and February 2023 with smart telescopes.
Paper Structure (12 sections, 9 figures, 2 tables)

This paper contains 12 sections, 9 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related works
  3. Data acquisition
  4. Method
  5. Detection with ASTRiDE
  6. Detection with a dedicated ResNet50 classifier and XRAI
  7. Fast approximation with a Pix2Pix model
  8. An annotated dataset and a dedicated YOLOv7 model
  9. Discussions
  10. Results
  11. Limitations
  12. Conclusion and perspectives

Figures (9)

  • Figure 1: Observation of NGC6482 by the authors with a smart telescope: a streak has completely damaged the image during acquisition .
  • Figure 2: Tracking problem in the FITS file is stored in NGC457-20220807.zip: it's not a streak (image from data24).
  • Figure 3: Graphical output of ASTRiDE, after the detection of a large streak in a FITS file captured during an observation session of the Barnard 142 Dark nebula (stored in Barnard142-143-20220922.zip, image from data24).
  • Figure 4: At the top, a stretched raw image of Messier 57 (Ring Nebula) with an exposure time of 10 seconds. At the bottom, the XRAI heatmap highlighting the pixels that are considered by the trained Resnet50 model for the classification, by keeping only 10% of the highest XRAI attribution scores (image from data24).
  • Figure 5: Example of a 256 $\times$ 256 output generated with a Pix2Pix model from a patch extracted from a raw image -- with highlighted streaks (image from data24).
  • ...and 4 more figures