Neural Network Analysis of Multi-Channel Light Curves for the Characterization of Resident Space Objects

Recent years have seen a ceaseless increase in the number of resident space objects which has led to a crowded orbital environment, raising concerns about the risk of collisions and the long-term sustainability of space operations. To address these challenges, monitoring and characterizing objects in orbit has become essential. Among the available observation methods, telescopes are able to collect valuable data by capturing changes in an object's brightness over time, i.e., providing Light Curves. These measurements can reveal key information about the characteristics and operational status of space objects. Telescopes may operate in different spectral bands contributing to provide a richer database which, however, is harder to manage and process as operating in multiple channels may introduce artifacts to the signals' characteristic due to discontinuities and inconsistencies between the acquisitions from different channels. The work addresses these problems by investigating the use of neural networks applied to real multi-channel light curves obtained from an open-source catalogue in order to classify space objects as active or inactive. The paper involves the study of the signals in the temporal and spectral domain to identify the most suitable approach. The results highlight the potential of this technique to support more effective monitoring of space objects and contribute to safer and more sustainable space operations. In particular, the comparative analysis of hybrid recurrent-convolutional and convolutional architectures demonstrates how different network designs capture complementary aspects of the signals, suggesting the feasibility of ensemble approaches for improved classification accuracy. Furthermore, the study emphasizes the impact of class imbalance and normalization strategies, offering practical insights for real-world applications where inactive objects largely dominate. Ultimately, this research represents a step forward in integrating machine learning into operational space situational awareness, paving the way for more reliable space traffic management solutions.