Space Target Image Sequence Prediction Based on Multilevel Motion Memory Intelligent Learning

Predicting sequences of space target images is a highly complex spatiotemporal forecasting task with crucial applications in space resource management, space situational awareness, space operations, satellite constellation control, and configuration maintenance. Despite some progress made by deep learning approaches in spatiotemporal sequence prediction, existing methods often struggle with long-term temporal dependencies, resulting in spatial inconsistencies or blurring in predicted space target image sequences. To address these challenges and enhance the quality of space target image sequence prediction, this article introduces a novel multilevel motion memory-based intelligent learning (M3IL) algorithm. This algorithm employs a sophisticated memory-based mechanism to accurately capture and predict the dynamic patterns of satellite motion, ensuring spatiotemporal consistency in the predicted sequences. Moreover, a novel improved sigmoid sampling strategy (IS3) is proposed to optimize input data and reduce training difficulties, enhancing the model’s ability to gradually learn complex dependencies. Experimental validation was conducted on the SPARK 2022 Stream-2 dataset, where a comprehensive comparison with several advanced algorithms was made across five critical metrics, GPU memory usage, training time consumption, Friedman test, and the visual quality of predicted images, confirming the applicability and effectiveness of the proposed M3IL algorithm in handling the task of space target image sequence prediction.