A LSTM Based Spacecraft Recognition Method for the Large LEO Constellation

It is a significant issue that spacecraft recognition entirely depends on orbital data without the consideration of multi-source image or radar data. Even more, it becomes a challenging task under the situation of large Low-Earth-Orbit (LEO) constellation spacecraft growing rapidly. This paper proposes a method for efficiently improving spacecraft recognition accuracy by training a Long Short-Term Memory (LSTM) neural network. Rather than directly taking on-orbit measured data as training samples, the paper designed a LEO constellation orbit data generator based on the Latin cubic sampling. It can avoid poor generalization of the trained neural network, because of uneven distribution of measured on-orbit data. The 2-layer LSTM spacecraft recognition framework obtains an excellent advantage of 97.56% recognition accuracy and better convergence rate, compared with the CNN framework and hybrid neural network of the CNN and LSTM. Moreover, the trained network needs as low as 7 sets of orbital data to identify the object, which greatly reduces the need for orbit determination data.