RSTD: Residual Spatiotemporal Diffusion Model for the Dynamic Prediction of On-Orbit Spacecrafts From Spaceborne Image Sequences

Abstract

The spatiotemporal prediction of on-orbit satellites is crucial for intention understanding and ensuring the successful completion of missions. Current spatiotemporal prediction methods primarily use convolutional neural networks (CNNs) and recurrent neural networks (RNNs) to process sequential observation images to predict future states. However, these methods often result in poor prediction performance due to the network’s inherent limited ability to express complex details. In this work, a residual spatiotemporal diffusion (RSTD) model is proposed to learn the spatial and temporal characteristics of targets from spaceborne imaging sequences. Leveraging a historical database, the model utilizes the patterns of image feature changes to assist in predicting target shape variations during the next observation period. A spatiotemporal perception module, capable of capturing long-term dependencies, is incorporated into the denoising process, thereby endowing it with forecasting capabilities. Furthermore, by incorporating a residual dual-stream structure, the model separates the prediction of the target’s overall shape and dynamic changes, thus overcoming the issue of overly smooth predicted images. Comprehensive experiments demonstrate that the proposed method achieves a peak signal-to-noise ratio (PSNR) of about 35 dB during uniform and uniformly variable motion. It also outperforms existing methods in subsequent feature extraction and attitude estimation, supporting the spatiotemporal attitude prediction of on-orbit satellites.