|
This work advances Space Situational Awareness (SSA) by analyzing ground-based hyper/multispectral images of Unresolved Resident Space Objects (URSO). Machine-learning models are constructed for satellite classification using unresolved spectral imagery. The study uses simulated data of observations of nine distinct satellites retrieved from the U.S. Space Force Unified Data Library (UDL). The dataset consists of unresolved hyperspectral imagery of satellites in different poses collected with a slitless spectrograph imager. The slitless spectrograph allows the collection of the spectral signature of the URSO with a single focal plane array in the 630 nm to 980 nm spectral range. In practice, the expectation is to have more unlabeled than labeled samples. Thus, the proposed classifier leverages the concept of Semi-Supervised machine learning. The network architecture leverages image reconstruction via a Convolutional AutoEncoder (CAE) and Multi-layer Perceptron (MLP) for classification. The architecture consists of three submodels: an Encoder, a Decoder (the two traditional components of a CAE), and a separate MLP. The Decoder and MLP use the Encoder’s low-dimensional representation of the samples as their input. The Encoder performs the task of creating the shared latent space via dimensionality reduction of the spectral data. The CAE (Encoder/Decoder) can be trained on labeled and unlabeled data, while the MLP requires labeled data for training. The newly developed classification models achieve a mean validation accuracy of 84% and a mean testing accuracy of 82%, utilizing 10-fold cross-validation. Additionally, the decoder achieves low image reconstruction error with a mean test error of 0.008, measured by Mean Squared Error, over the cross-validation folds. This network architecture demonstrates the ability to map slitless spectral data for accurate identification of RSO despite the challenges of using a limited-size and unbalanced dataset.
|
