We examine deep learning's effectiveness for TNO detection in the CLASSY (Classical and Large-A Solar SYstem) survey, a CFHT Large Program in 2022–2024. In recent years, progress in TNO population science has been driven by large-scale surveys. As we approach the era of the Legacy Survey of Space and Time (LSST) and the resulting petabyte-scale avalanche of data, new tools are needed to successfully unpack the signal from noise. Deep learning can build useful representations directly from the data without a lot of hand-engineering, making it an ideal addition to a survey's analysis pipeline. The steep luminosity function means TNOs are often faint and in the background noise; isolating the region of interest in observations reduces the load for computing the predicted orbital track.
Successfully training neural networks requires excellent labelled datasets. Here, we create composite images of the nightly CFHT MegaCam observations, leveraging the TNO sky motion in consecutive observations to extract a linear series of points (“tracklets”). These tracklet form the basis of our training dataset of ~75,000 images, which was composed of tracklets from both real and planted sources as well as images without tracklets. Several custom convolutional neural network-based models were trained to identify and locate these tracklets in the CLASSY data. The predictions from each of these networks were then combined to boost the predictive power and reduce the false negatives. Our CLASSY datasets and network models will provide a blueprint for including deep learning in the moving object detection pipelines for other surveys, such as the next-generation LSST.