Deep Learning for Astronomical Imaging: A CNN-GRU Approach on GMU Observatory Data

This paper proposes a machine learning framework that automates the filtering of astronomical imaging data from the George Mason University Observatory. Traditionally, researchers must use manual inspection of science FITS frames to filter out “bad” (faulty) frames–those with streaking, telescope jitter, or motion blur before continuing to calibrate and generate light curves [4]. However, it is a long and error-prone process. We developed a two-stream convolutional neural network and gated recurrent unit (CNN-GRU) model that classifies frames as “good” or “bad”, using both spatial and temporal features. FITS frames were normalized and resized for computational efficiency. We labeled multiple large sets of TOI observations for an entire night, as well as partially sampled sets of TOIs to train on unique data. Initial results show that the model is able to accurately classify problematic frames. Accordingly, this work demonstrates that utilizing deep learning can accelerate telescope data analysis pipelines by automating manual filtering.