SILO: A Machine Learning Dataset of Synthetic Ground-Based Observations of LEO Satellites

Images of space objects may have their interpretability assessed with a Space-object National Imagery Interpretability Rating Scale (SNIIRS) score. The rules for such scores are specific, but the process of rating a large number of images can be time-consuming even for a skilled analyst. As this scale is subjective and based on interpretability of resolved features, it is also difficult to provide automated SNIIRS assessments with a simple algorithmic procedure. A Convolutional Neural Network (CNN) may be able to solve this problem, but such an effort requires a large labeled dataset of images. In this paper we will describe the effort to use wave-optics simulations to generate a dataset of SNIIRS-scored images of Low Earth Orbit (LEO) satellites observed from a ground-based optical observatory with varied turbulence conditions. This first iteration of the Scored Images of LEO Objects (SILO) dataset is intended to serve as a foundation for deep learning efforts, similar to how MNIST and ImageNet have been foundational datasets in other machine vision domains. This dataset is already being used in numerous machine learning efforts, including those pertaining to using CNNs to perform image interpretability assessment and to produce higher-resolution image recoveries from degraded image sets. In this paper we also describe some of the other potential uses for this dataset.