Resident Space Object Classification from Light Curves with Deep Learning

In this study, a novel data-driven pipeline is suggested for classifying resident space objects (RSO) from photometric light curve data. The latter was extracted from the MiniMega-TORTORA (MMT) monitoring system. MMT provides light curves corresponding to a collection of RSO, including inactive satellites, rocket bodies, and space debris. As a manner to classify space objects, a subset of light curves corresponding to several of them is chosen and divided into training and testing. For each 1D feature, the corresponding time series was decomposed by means of Empirical Mode Decomposition (EMD) in order to extract information about the most dominant frequency bands in that feature. The different time series components provided by EMD are then stacked together into a matrix and provided as input to a Convolutional with Attention (CoAtNet) model yielding a 92.0% top-1 accuracy on the testing set. Tests with an imbalanced dataset and with different input sizes were conducted to evaluate the robustness of the model. Experimental findings show that the approach is a robust method for classifying space objects while requiring few input data.