High-performance embedded system for onboard object detection in hyperspectral images

Hyperspectral imaging instruments can collect hundreds of spectral bands for the same area on the Earth surface which enables extracting extended fine information of the area of interest. Due to limited bandwidth from the satellite to Earth ground station, hyperspectral image classification and object detection on-board have gained significant interest due to their importance in the analysis and interpretation of the content of the image opening new applications based on hyperspectral images. The processing and analysis of images onboard reduces the data volume to be transmitted to the ground station and allows real-time decisions to be made onboard. In this paper, a high-performance embedded system is considered to run target detection in hyperspectral images. The work considers several deep-learning models and their deployment in a Jetson Orin Nano to run object detection. It takes into account the tradeoff between throughput, model complexity, and accuracy. The model is quantized to 8-bits to explore the utilization of INT8 operations of the Jetson Orin GPU, to achieve real-time performance.