Deep Neural Networks for Detecting Asteroids in the ATLAS Data Pipeline

The ”Asteroid Terrestrial-impact Last Alert System” (ATLAS) currently uses a two stage binary classifier to filter moving astronomical objects from electronic and optical artifacts, in order to detect asteroids that may eventually pass close to, or impact, Earth. Any detections that pass ATLAS’s filter are examined by human analysts. The results of the current filter contains more false positives than true positives, so that the majority of detections are classified as bogus by the analysts. These bogus tracklets cause unnecessary work for the analysts, and increase the time it takes to classify a detection as a real near Earth object, potentially decreasing warning time for a collision. In order to reduce this unnecessary effort, we extend the current classifier to incorporate dynamic motion data. We develop two engineered features which are combined with the output of the original classifier as the input features of a deep neural network. This network generates the probability of a detected object being designated real (i.e. an actual, moving, astronomical object), as opposed to being classified as bogus (i.e. one of the vast majority of false detections resulting from optical or noise artifacts). The new classifier decreases false positives by 59%, while maintaining a low false negative rate at virtually zero.