A Journey to the Edge of the Solar System with an AI navigator
Date
2023-12-19
Authors
Lee, Aram
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
I present a deep learning method of searching for solar system objects (SSOs) in wide-field survey imaging data including trans-Neptunian objects (TNOs). Artificially generated sources are added to mosaic images taken with the Canada-France-Hawaii telescope (CFHT) MegaCam instrument to create the convolutional neural network (CNN) training set. The CFHT MegaCam data images are a time series of observations, and the location of the artificial SSO changes between images, in a way that is consistent with a heliocentric Keplarian orbit. The imaging characteristics of the artificial sources were found to be highly similar to those of real SSOs, with rates of sky motion consistent with TNOs. My deep learning approach is based on the detection of moving sources within 64×64-pixel sub-image pairs extracted from the time series of large-format mosaic astronomical imaging data. Each image pair extracted from the training images has been labelled with the presence or absence of a moving source, along with the source location and brightness measured in magnitudes. The labelled sub-images were fed into ImageNet algorithms to train classification models and regression models separately. The algorithm assigns a model-dependent probability that a particular sub-image contains an SSO. The probability threshold required to assert that an SSO has been detected is set based on the evaluation of retrieval and precision of the model and the requirements of the experiment. This thesis evaluates the capabilities of the range of deep learning models and determines which one is most effective in the detection of artificial SSOs. The MobileNet model was selected as the most efficient for this problem space. A trained classification model derived from the MobileNet model retrieved 91% of sub-images with a moving source with a 90% precision on test data sets. A separate regression model then predicted the location of the moving source with a mean absolute error of ±1.5 pixels for sources with SNR > 17 (m_r < 23 in my data set). Although the retrieval rate is high, due to the scarcity of real SSOs in imaging data, the precision achieved (90% of false positives rejected) results in a substantial number of false positives. Further data processing on the candidate list is required to improve the purity of the result.
To improve sample purity, I investigated two post-processing approaches:
• With the classification-filtered sub-images and their regression-measured
locations in sky coordinates, each detected source was grouped with nearby
detected sources as SSOs exhibit nearly linear sky motion for the duration of
the observed time series. Any group of linear source tracks, detected in at
least 1/3rd of the images, was considered a candidate detection. This
approach achieves an effective detection limit (more than 50% of artificial
sources in the data are detected) at SNR=7.2, and the source purity of the
sample was greater than 99% in this case. However, the required
combinatorics of this approach (NxN comparison) make it computationally
slow, and the high SNR required for detection resulted in very few ‘real’
candidates being proposed.
• I also investigate a ‘scoring’ approach for candidate selection. My CNN
classification model output is a model-dependent probability that a particular
sub-image contains a moving source. Each sub-image was given a score
derived by scaling the classification model probability assigned to that
sub-image. A sub-image was then determined to hold a candidate object if its
score exceeded a given threshold (determined by the desired purity of the
sample). With this approach, I achieved an effective detection limit (50% of
artificial sources in the data are detected) at SNR=3.4 and discovered a
number of real SSOs within the test data set. Visual inspection of 1800
scoring-based candidates revealed approximately 200 visibly bright real (not
from the artificial source list) SSO candidates.
I tested trained models on test sets from different sky regions and found that our
models did not learn from the backgrounds or shapes of TNOs, but rather detected
the motion of TNOs. I found that deep-learning object detection algorithms can aid
in the discovery of TNOs and SSOs. When combined with a scoring approach, my
algorithm provides a capability that is similar to that achieved with more classical
approaches without making assumptions of motion rates of the SSOs and without
requiring any substantive data engineering. The CNN approach to SSO detection is
very promising and should be pursued in the development of future SSO discovery
software pipelines.
Description
Keywords
Astronomy, Observational Astronomy, Minor Planets, Trans-Neptunian Objects, Kuiper Belt Objects, Detection, Machine Learning