Chad Mello, United States Air Force Academy; Matthew Mendoza, United States Air Force Academy; Leonardo Camacho, United States Air Force Academy; Damien Eberhardt, United States Air Force Academy
Keywords: machine learning, deep learning, transfer learning, geosynchronous, satellite, spectra,
Abstract:
A recent study utilizing data collected from the Falcon Telescope Network (FTN) for space situational awareness has demonstrated the potential of non-resolved optical signatures for identifying geosynchronous orbit satellites (GEOs) using traditional approaches like Principal Component Analysis (PCA) and Support Vector Machines (SVM). While these methods can be effective, especially for relatively small datasets, they face limitations such as sensitivity to data variance and dependence on manually selected features, which can hinder scalability and accuracy. To better cope with these challenges, we propose leveraging modern machine learning techniques, specifically fine-tuned Convolutional Neural Networks (CNNs) pretrained over large spectral datasets. Our approach may offer significant advantages, including enhanced feature extraction and robustness to spectral variances.
In our experiments, we developed and utilized a hybrid ResNet-Inception deep learning model. We trained one instance against a 600,000-plus star spectra dataset from the Sloan Digital Sky Survey (SDSS) and another instance on satellite spectra generated by a custom Conditional Variational Autoencoder (CVAE). We then fine-tuned these instances using augmented FTN satellite data to better capture the unique characteristics of optical-based satellite spectra, aiming to improve classification accuracy and efficiency. By automatically learning features directly from the data, our model is able to identify complex patterns in satellite spectra that traditional methods might miss. Additionally, the use of generative models and transfer learning allows us to leverage prelearned features from vast datasets, reducing the requirement for extensive labeled satellite data and addressing the challenges associated with small sample sizes typical in space situational awareness research. The results demonstrate that our proposed method achieves over 83% accuracy across 20 GEO classes using only 86 complete samples, showing the potential of our models as more data is collected via FTN. This study underscores the potential of advanced neural network architectures in enhancing the accuracy and efficiency of GEO classification, contributing to the greater goal of improved situational awareness in space.
Date of Conference: September 17-20, 2024
Track: Satellite Characterization