A Convolutional Neural Network–Based Approach for Detecting Solar System Objects in Wide-field Imaging
We present a deep learning method that utilizes convolutional neural networks (CNNs) to discover trans-Neptunian objects (TNOs) in wide-field survey imaging data. Our CNNs were trained using artificial sources planted into a time series of 44 205 s CFHT MegaCam large-format mosaic images. We extract...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IOP Publishing
2025-01-01
|
| Series: | The Planetary Science Journal |
| Subjects: | |
| Online Access: | https://doi.org/10.3847/PSJ/add409 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | We present a deep learning method that utilizes convolutional neural networks (CNNs) to discover trans-Neptunian objects (TNOs) in wide-field survey imaging data. Our CNNs were trained using artificial sources planted into a time series of 44 205 s CFHT MegaCam large-format mosaic images. We extracted 64 × 64 pixel subimages and labeled each subimage pair with the presence or absence of a moving source and the source’s location and magnitude. Our MobileNet-derived classification model achieved 91% recall with 90% precision on test data. A separate regression model predicted the source locations with a mean absolute error of ±1.5 pixels for sources with a signal-to-noise ratio (SNR) of 16.5 or higher. By grouping sources based on linear sky motion, we achieved ∼40% detection limit for sources with an SNR ≈ 7 in individual frames. We also examined a scoring approach to reduce the false positives. In this approach, images are scored based on the probability values from the classification model. With the scoring approach, we achieved a detection limit of SNR ≈ 3 in individual frames. We detected approximately 200 solar system object (SSO) candidates ( sim 5 moving rates consistent with TNOs) in our 1 square degree of imaging. We tested the trained models on images from different sky regions, confirming that the models learned from the motion of sources rather than from the backgrounds or shapes of sources. We demonstrate that deep learning object detection algorithms can aid in TNO and SSO discovery, supporting future discovery pipeline development. |
|---|---|
| ISSN: | 2632-3338 |