HyperspectralViTs: General Hyperspectral Models for On-Board Remote Sensing

Abstract

On-board processing of hyperspectral data with machine learning models would enable an unprecedented amount of autonomy across a wide range of tasks allowing new capabilities, such as early warning systems and automated scheduling across constellations, of satellites. However, current classical methods suffer from high false positive rates and therefore prevent easy automation while previously published deep learning models exhibit prohibitive computational requirements. We propose fast and accurate machine learning architectures, which support end-to-end processing of data with high spectral dimension without relying on handcrafted products or spectral band compression techniques. We create three new large datasets of hyperspectral data containing all relevant spectral bands from the near global sensor EMIT. We evaluate our models on two tasks related to hyperspectral data processing—methane detection and mineral identification. Our models reach a new state-of-the-art performance on the task of methane detection, where we improve the F1 score of previous deep learning models by 27% on a newly created synthetic dataset and by 13% on the previously released large benchmark dataset. Our models generalize from synthetic datasets to data with real methane leak events and boost performance by 6.9% in F1 score in contrast with training models from scratch on the real data. Finally, with our newly proposed architectures, one capture from the EMIT sensor can be processed within 30 s on a realistic proxy of the ION-SCV 004 satellite and in less than 0.64 s on a GPU powered Jetson AGX Xavier board.