Detection of Occultation Events by Machine Learning for the Unistellar Network

The Unistellar network is composed of more than 4,600 digital, smart and compact telescopes owned by citizen scientists who used them regularly to conduct scientific observations, including occultations by asteroids. The SETI Institute team predicts around 6 events per week for several areas of the world including Europe, North America, Japan, Oceania, India and soon more. Consequently we are obtaining every week a rising amount of occultation data that needs to be analysed quickly so the citizen scientists can learn quickly if their observations were successful.Typically ~20% of the existing data include a positive detection of an occultation, seen by the observer as the disappearance of a star. The purpose of this work is to demonstrate that a Convolutional Machine Learning algorithm can be used for a fast and trustworthy detection of occultations. Ultimately our goal will be to embed this algorithm on board the telescope computer so the user will know almost immediately whether or not the occultation was successful. This Neural Network has been trained on a large sample of artificial star images (~800,000), composed of 20 different types of light variation it can encounter. This training set has been designed following statistics on the 6 months of observation prior this study from Unistellars database (from October 2020 to March 2021). The test sample is, unlike the training test, composed of 30 real observations, and among them, 15 observations containing an occultation and 15 not containing such event. We also gave the final model a set of noisy observations to understand how it could react with unusable data. Our preliminary test suggests that our algorithm successfully detects the event at the good time during the expected time of the occultation at a rate of 74,3% and has a rate of false detection at 7,9% in a short period of time (typically 1 second for 100 frames on our machine-learning dedicated computer). Machine Learning algorithms represent a solid base for future event detection based on raw data acquired from CMOS/CCD detectors on digital telescopes, like the Unistellar eVscope. We will include more observations in our future analysis and validate the algorithm under wide range of conditions including poor weather, windy environment or cloudy sky.