The future is hyperspectral: Development of platform-agnostic hyperspectral foundational models

Hyperspectral imaging is an emerging technology for Earth-observation that has the potential to revolutionize our ability to measure chemical, physical and biological characteristics of various materials. Hyperspectral images have 100s to 1000s of channels per pixel (compared to three channels in commonly used RGB imagery) and captures visible and “unseen” signals, with wide transformative applications in agriculture, ecology, and environmental science. Currently, data from available hyperspectral platforms are incompatible, with the lack of data harmonization strategies leading to siloed workflows.

This project uses an AI approach to develop the first hyperspectral “foundation” model to enable data harmonization and the generation of standardized data outputs, thereby preventing reinvention of the wheel for each platform, especially as new satellites and affordable airborne sensors come online. A harmonized foundation model allows leveraging data streams from around the world, lowers barriers to entry, increases uptake of this exciting technology, and democratizes standardized scientific and commercial applications.

PRINCIPAL INVESTIGATOR

Philip Townsend, professor of forest and wildlife ecology

CO-PRINCIPAL INVESTIGATORS

Matthias Katzfuss, professor of statistics

Prabu Ravindran, scientist III in the Botany Department

CO-INVESTIGATORS

Sunduz Keles, professor of statistics

Kyle Kovach, postdoctoral research associate in forest and wildlife ecology

Frederic Sala, assistant professor in computer science