SPACE OBJECT SHAPE CHARACTERIZATION FROM PHOTOMETRIC DATA USING RECURRENT NEURAL NETWORK

We present an approach that employs machine learning techniques to determine the most probable shape of resident space objects (RSO). Our algorithm focuses mainly on photometric data obtained from optical sensors and analysis of the computer simulation. Aiming at RSO of LEO, an identification model is trained by recurrent neural network (RNN) technique to find the shape among a number of candidate shape models. Firstly, the basic 3D shape model of the space object is established, including cylinders, pyramids, and cuboids and so on. Then, the bidirectional reflectance distribution function (BRDF) model is processed to obtain the photometric time series signals of the basic shape model and the photometric data will be used as training samples. Next, the RNN is trained with the photometric sequential data as input. Finally, new photometric timing signals obtained by optical sensors and analysis of the computer simulation are used as input to the network to identify the object shape. The initial experimental results show that the algorithm presented in this paper could determine the shape of RSO effectively and accurately, and this approach has more adaptive performance and can obtain satisfactory results.