Realizing Machine Learning’s Promise in Geoscience Remote Sensing

As machine learning and pattern recognition methods flourish in commercial applications, systematic efforts aim to introduce these technologies into Earth and planetary science practice. This is especially true in remote sensing, where scientists seek to convert massive, noisy archives into science insight. Such problems seem well suited for machine learning and pattern recognition. As the data science revolution matures, it is worth asking if it has yet made a significant impact on remote sensing. Rather than attempt a comprehensive review, we examine research in imaging spectroscopy, also known as hyperspectral imaging, as a case study of a data-centric remote sensing discipline which would be expected to benefit from machine learning. Our analysis demonstrates that articles using the term ``imaging spectroscopy'' or ``hyperspectral'' come from a handful of independent and self-citing communities. The volume of research has recently accelerated from all groups, especially from those focusing purely on the methodology of machine learning and signal processing. Nevertheless, these are seldom cited by mainstream Earth and planetary scientists. The traditional sciences (including clinical, Earth, and space science) are responsible for all but one of top-ranked journal publications, with fewer than 40% of total articles. The remaining 60% represent thousands of manuscripts and vast person-centuries of work by highly educated researchers, a reservoir of potential that geosciences could better exploit.