Putting the “Space” in Hyperspace: Investigating Hyperdimensional Computing for Space Applications

Earth observation (EO) is currently a major application area of satellite operations. As more powerful imagers become more accessible and see increased use, these EO images will increase in size, which also increases the computational complexity of classifying them. Using neural networks for this task has limitations due to the small memory and compute capability possessed by embedded platforms. Instead, this research investigates a shift in machine-learning paradigms from neural networks to hyperdimensional computing (HDC). HDC uses very large vectors to represent and draw relations from data. HDC often has a lower latency, power usage, and memory footprint than neural networks. Using the EuroSAT dataset, this research achieved > 1.4× speedup and energy efficiency using an HDC model over a convolutional neural network. Though, this improvement came at the cost of 4% lower accuracy. These results indicate HDC is well suited for machine-learning tasks in space.