The Deep Neural Network based Photometry and Astrometry Framework for Wide Field Small Aperture Telescopes

Wide field small aperture telescopes (WFSATs) are preferable observation instruments for time domain astronomy, because they could obtain images of celestial objects with high cadence in a cost-effective way. An automatic data processing algorithm which could detect celestial objects and obtain their positions and magnitudes from observed images is important for further scientific research. In this paper, we extend the ability of a deep neural network based astronomical target detection algorithm to make it suitable for photometry and astrometry, by adding two new branches. Because the photometry and astrometry neural network are data-driven regression algorithms, limited training data with limited diversity would introduce the epistemic uncertainty to final regression results. Therefore, we further investigate the epistemic uncertainty of our algorithm and have found that differences of background noises and differences of point spread functions between the training data and the real would introduce uncertainties to final measurements. To reduce this effect, we propose to use transfer learning strategy to train the neural network with real data. The algorithm proposed in this paper could obtain types, positions and magnitudes of celestial objects with high accuracy and cost around 0.125 second to process an image, regardless of its size. The algorithm could be integrated into data processing pipelines of WFSATs to increase their response speed and detection ability to time-domain astronomical events.