Ryan Walden, Georgia State University; Taslim Dosumnu, Georgia State University; Stuart Jefferies, Georgia State University; Daniel Pimentel-Alarcon, University of Wisconsin-Madison
Keywords: SSA, image reconstruction, deep learning, satellite, classification, CNN, transfer learning, convolutional, neural networks
Abstract:
For the last years, space domain awareness has hinged on high-resolution restoration of satellite imagery that has been severely distorted by the Earth’s turbulent atmosphere. Restoration of this sort is a computationally demanding procedure that typically requires screening and processing thousands of images, and can take hours of CPU time on dedicated servers, impeding real-time analysis.
Restoration, however, is only an intermediate step towards more succinct goals, such as identifying the satellite model, its orientation, health, and intent. To the human eye, image restoration is unavoidable prior to extracting these satellite characteristics, as our brains are not trained to distinguish between distorted images directly. However, the fact that restoration is even possible suggests that all the necessary information to recover the desired characteristics is indeed embedded in the distorted images, only encoded in a different way, due to the light diffraction patterns produced by the atmosphere.
Hence, here we hypothesize that by training a deep convolutional neural network (D-CNN) to interpret distorted images directly, we can recover the desired satellite characteristics (e.g., model and orientation) in one step, with no need for image restoration nor further classification and processing. The key to our approach lies in a mixture of synthetic data, transfer learning, neural architecture search, and multi-task training under a D-CNN framework. Our rationale is that convolutional filters are qualified by design to undertake convolution transformations, like the ones produced by the atmosphere.
Training a network of this sort requires a sufficiently large dataset, rich enough to capture all the spectrum of satellites variability. To this end we introduce a novel synthetic data generation strategy that combines a state-of-the-art 3D graphics engine with an atmospheric distortion pipeline to generate realistic samples of satellite models at various orientation and distances from an observatory. To speed up the training process we use transfer learning, which we implement using a fine-tuning strategy on D-CNN architectures pre-trained on the ILSVRC ImageNet dataset. Compared to traditional, randomly initialized networks, this method has demonstrated faster convergence rates for a wide range of image processing domains, including classification and regression. Additionally, we utilize a Bayesian search strategy to optimize model hyperparameters and determine an ideal architecture from a selection of several D-CNN architectures. This search approach further exploits joint multi-task learning using a combination of satellite classification and orientation labels, potentially converging to an overall higher final accuracy on our test dataset. To show the promise of this approach, we compare our strategy against the state-of-the-art, based on a mix of Lasso and Fourier restoration, followed by specialized classification and recognition algorithms.
Date of Conference: September 15-18, 2020
Track: Machine Learning Applications of SSA