A solar system object discovery method with convolutional neural networks for future surveys

We present a deep learning method that reliably detects solar system objects (SSOs) in a data-agnostic way.

As a test case, we analyzed CFHT MegaCam observations after adding hundreds of artificial Kuiper Belt objects (KBOs) to an image time series. These images are then sub-divided into 64×64-pixel sub-image pairs (each sub-image is of the same sky location but taken at a different time) and label these sub-images to indicate the presence (or lack) of an artificial source and the location, and magnitude of the source. The labelled sub-images are then fed into ImageNet algorithms to train classifiers and regressors separately.

A trained classifier made from the MobileNet algorithm is able to retrieve (classify a positive image as a positive detection) 91% of sub-images with a moving object, and the regressor predicts the location of the objects to within 1.5 pixels of mean absolute error for bright sources. We found that contamination from chip/detector flaws (bad columns, stellar spikes, etc) in the image lowers the recall from 100% even for the brightest sources.

Using special test cases we determine that our CNN model does not learn from the background or shapes of the moving objects. The detection is, rather, triggered on the change in position of sources in the sub-images without knowledge of the intrinsic shapes of sources, making this approach a data-agonistic model (the user does not need to parameterize the image quality or other features of the image).

With the classifier-filtered sub-images and their regression-predicted location sky coordinates, each detection source is grouped with nearby sources to form tracklets. These tracklets are then analyzed to reveal moving object candidates. After maximum likelihood estimation analysis on time of observation and predicted positions of KBOs, only a handful of KBO candidates (a few 100 per square degree) remain and these candidates are then examined manually.

The flux limit of our CNN approach achieves the same limit of detection as the classical pipeline used by the Outer Solar System Origins Survey (OSSOS) but does not depend on specific knowledge of the shapes of sources in the image, while the OSSOS pipeline makes use of convolution kernels tuned to the expected stellar profiles.

Our deep-learning discovery process for SSOs is highly efficient at object detection. Given the data-agnostic nature of such machine learning-based models, we anticipate that the use of deep CNN for source detection will be of high utility for future SSO discovery pipelines.