Taking Artificial Intelligence Into Space Through Objective Selection of Hyperspectral Earth Observation Applications: To bring the “brain” close to the “eyes” of satellite missions

Abstract

Recent advances in remote sensing hyperspectral imaging and artificial intelligence (AI) bring exciting opportunities to various fields of science and industry that can directly benefit from in-orbit data processing. Taking AI into space may accelerate the response to various events, as massively large raw hyperspectral images (HSIs) can be turned into useful information onboard a satellite; hence, the images’ transfer to the ground becomes much faster and offers enormous scalability of AI solutions to areas across the globe. However, there are numerous challenges related to hardware and energy constraints, resource frugality of (deep) machine learning models, availability of ground truth data, and building trust in AI-based solutions. Unbiased, objective, and interpretable selection of an AI application is of paramount importance for emerging missions, as it influences all aspects of satellite design and operation. In this article, we tackle this issue and introduce a quantifiable procedure for objectively assessing potential AI applications considered for onboard deployment. To prove the flexibility of the suggested technique, we utilize the approach to evaluate AI applications for two fundamentally different missions: the Copernicus Hyperspectral Imaging Mission for the Environment (CHIME) [European Union/European Space Agency (ESA)] and the 6U nanosatellite Intuition-1 (KP Labs). We believe that our standardized process may become an important tool for maximizing the outcome of Earth observation (EO) missions through selecting the most relevant onboard AI applications in terms of scientific and industrial outcomes.