Real-time hyperspectral data transmission for UAV-based acquisition platforms

Abstract

Hyperspectral sensors that are mounted in unmanned aerial vehicles (UAVs) offer many benefits for different remote sensing applications by combining the capacity of acquiring a high amount of information that allows for distinguishing or identifying different materials, and the flexibility of the UAVs for planning different kind of flying missions. However, further developments are still needed to take advantage of the combination of these technologies for applications that require a supervised or semi-supervised process, such as defense, surveillance, or search and rescue missions. The main reason is that, in these scenarios, the acquired data typically need to be rapidly transferred to a ground station where it can be processed and/or visualized in real-time by an operator for taking decisions on the fly. This is a very challenging task due to the high acquisition data rate of the hyperspectral sensors and the limited transmission bandwidth. This research focuses on providing a working solution to the described problem by rapidly compressing the acquired hyperspectral data prior to its transmission to the ground station. It has been tested using two different NVIDIA boards as on-board computers, the Jetson Xavier NX and the Jetson Nano. The Lossy Compression Algorithm for Hyperspectral Image Systems (HyperLCA) has been used for compressing the acquired data. The entire process, including the data compression and transmission, has been optimized and parallelized at different levels, while also using the Low Power Graphics Processing Units (LPGPUs) embedded in the Jetson boards. Finally, several tests have been carried out to evaluate the overall performance of the proposed design. The obtained results demonstrate the achievement of real-time performance when using the Jetson Xavier NX for all the configurations that could potentially be used during a real mission. However, when using the Jetson Nano, real-time performance has only been achieved when using the less restrictive configurations, which leaves room for further improvements and optimizations in order to reduce the computational burden of the overall design and increase its efficiency.