machine learning

Constellations of small sensors called cube-sats are working together in space to image the entire land surface of our planet every day at resolutions fine enough to resolve individual plants. These dense image time series provide the combination of high temporal frequency and fine spatial resolution to map individual trees and quantify population dynamics throughout whole continents. Dealing with these data requires statistical methods that can handle variable detection and missing observations. The Kellner lab is developing this framework and applying it to tree populations across thousands of miles of Amazonian forest.

Extracting the most important information from cube-sat time series requires the ability to identify and monitor objects, like individual plants, in contrast to pixel-based summaries. This new capability for object-based remote sensing at regional scales is fundamentally transformative. Developing the computational resources for object-based remote sensing at planetary scale requires advanced machine-learning and cloud-based computing to align millions of satellite images, and to identify and track objects through time. The promise of these high-density time series is the ability to determine how forests are changing over time with the granular resolution of individual plants, to test biological hypotheses across regions and continents, and to interpret high-resolution image time series from a biological point of view.

Funding sources: National Science Foundation, Brown Data Science Initiative