Skip to main content
SearchLogin or Signup

Identification of asteroid streaks in simulated ESA Euclid images

Presentation #211.03D in the session “Asteroid Surveys: Sifting through the Data”.

Published onOct 26, 2020
Identification of asteroid streaks in simulated ESA Euclid images

The ESA Euclid space telescope observes up to 150 000 asteroids as a sideproduct of its primary cosmological mission. Asteroids appear as streaks in the images. Owing to the survey area of 15 000 square degrees and the number of sources, automated methods have to be used for finding them. Euclid is equipped with a visible camera VIS and a near-infrared camera NISP with three filters. The Euclid mission substantially increases the number of asteroids with multi-band photometry that extends to near-infrared.

We tested two methods for finding the asteroid streaks in simulated Euclid images. The first method is StreakDet, a software developed to detect space debris. We optimized the parameters of StreakDet to maximize completeness and developed a post-processing algorithm to improve the purity of the sample of detected sources by removing false-positive detections. The second method is deep learning, i.e., deep artificial neural networks. We developed a custom-built object detection model based on convolutional neural networks.

StreakDet finds 96.9% of the synthetic asteroid streaks with apparent magnitudes brighter than 23 and streak lengths longer than 15 pixels (10 arcsec/h), but this comes at the cost of finding a high number of false positives. The number of false positives can be radically reduced with multi-streak analysis, which utilizes all four dithers obtained by Euclid. StreakDet is a good tool for identifying asteroids in Euclid images, but there is still room for improvement, in particular for finding short (less than 13 pixels, corresponding to 8 arcsec/h) and/or faint streaks (fainter than apparent magnitude 23).

The deep learning model can detect streaks up to a magnitude fainter than StreakDet, and it also works for short streaks. However, because of the gap between synthetic and real image distributions, a deep learning model trained only with simulated training data probably does not work optimally out of the box for the real images. Therefore, StreakDet and deep learning are complementary. After Euclid is launched and real data becomes available, StreakDet can gather non-synthetic training examples for the neural network, which can then detect fainter and shorter streaks than StreakDet can.


No comments here