Solar System Object Detection in Time Series Data Using Synthetically Trained Neural Networks

For the a-posteriori detection of solar system objects in time series extracted from image data, a number of preprocessing, clustering and orbital fit algorithms have been proposed. However, for the largest datasets these methods prove to be computationally expensive or even unfeasible. With our work, we explore if and how artificial neural networks can help us with this task and which network architectures are particularly suited to distinguish between stationary and moving objects in observational time series. After defining the training and validation data, selecting five neural network architectures that fit the problem and extensive training of the networks on synthetically generated data, we evaluate all architectures in terms of their classification performance, both on synthetic data and actual observations from the Minor Planet Center's database. The recently proposed InceptionTime architecture shows detection sensitivities of up to 99% with specificities on a similar level on both synthetic data and actual observations. Combined with the option for GPU offloading provided by popular machine learning libraries, we conclude that InceptionTime networks are promising additions to traditional a-posteriori moving object detection pipelines.