I work on spatial-prediction and mapping in forests and agriculture using high-resolution remote sensing from the air and space.
Supported by
- NASA
- National Science Foundation
- USDA
- U.S. Department of Defense
Outside the university, I’ve contributed to the founding and development of a privately funded company focused on remote sensing in agriculture, including venture-capital fundraising, and I’ve worked in partnership with state programs in Colorado and North Dakota.
Global Ecosystem Dynamics Investigation
We focus on the development of statistical methods in support of the NASA Global Ecosystem Dynamics Investigation (GEDI).
GEDI is a waveform lidar sensor on the International Space Station. The sensor generates measurements of vertical canopy structure from recorded laser waveforms. Our work produces the statistical models that convert calibrated geolocated laser waveforms into units of aboveground biomass density. Output from these models is publicly available as the GEDI level-4A data product. This work is supported by NASA in collaboration with the University of Maryland, College Park.
Strengthening GEDI algorithms through improved stratification and quality filtering
Quality filtering and stratification are two major challenges faced by spaceborne lidar. Although the mission has collected tens of billions of individual waveforms, most of them don’t pass quality filters. This happens for a variety of reasons. Some measurements are acquired under leaf-off conditions, and cloud-cover or atmospheric haze can degrade measurement quality.
Stratification refers to the necessity of applying waveform-processing and aboveground biomass density algorithms within categories dictated by geography and vegetation cover. Currently the mission uses geographic distinctions that roughly correspond to continents, and vegetation cover is based on categories derived from the NASA MODIS instrument at 1 kilometer resolution.
In work supported by NASA in collaboration with Boston University, we are investigating the application of high-resolution stratification derived from the NASA Landsat program. The Landsat program generates land-surface measurements worldwide at 30 meter resolution. Higher-resolution images allow us to more precisely control quality filtering and prediction than is currently possible at 1 kilometer.
Development of machine-learning models for GEDI aboveground biomass density
The current approach to statistical model development for GEDI aboveground biomass density is based on 13 linear models applied to 35 combinations of vegetation cover and geographic world region. Each model uses 1–4 predictor variables. The simplicity of this approach makes the models easy to troubleshoot and limits the potential for artifacts. But it introduces complexity in design decisions, including which transformations are applied, which input variables are considered, and how the models should be stratified.
With support from NASA, we are working to apply machine learning to GEDI aboveground biomass density. The objective is to produce a single, worldwide algorithm that does not require arbitrary transformation or explicit stratification, that is easy to train using off-the-shelf software, and that performs at least as well as current linear models.
Digital Soil Mapping
I serve as the Chief Scientist for Perennial, a private, venture-backed company founded by three former Brown University undergraduates with support from the Nelson Center for Entrepreneurship.
The company’s mission is to make regenerative agriculture measurable, investable, and scalable by delivering registry-compliant measurement of soil carbon and ecosystem outcomes on agricultural land. The core technology is a digital soil mapping platform driven by high-resolution satellite remote sensing and machine learning. I worked closely with the founders – Jack Roswell ’20, David Schurman ’20, and Alex Zhuk ’20 – when they were undergraduates at Brown, and work with the company through the Brown University Policy on Outside Professional Activities for Faculty.
Together we have published peer-reviewed academic journal articles demonstrating the potential of digital soil mapping in agriculture, generated new intellectual property and patent filings, developed a protocol for private-sector carbon accounting, employed dozens of people, and raised $35 million in venture capital. The company is supported by investments led by Temasek, a global investment company headquartered in Singapore, and Bloomberg, a global leader in business and financial information.
Drone remote sensing
The quantitative improvement represented by low-altitude drone flight is significant, but the most important advance is conceptual.
Low-altitude drone flight can generate images at scales clearly aligned with biological processes, like metabolism, natural selection, and resource allocation within and among individual plants. New measurements from low-altitude drones open the door to seeing things that have been beyond our grasp, and to properties related to organismal condition, like leaf chemistry, canopy temperature and solar induced fluorescence.
We have completed drone flights at sites in the Czech Republic, Switzerland, and the Atlantic lowlands of Costa Rica with very high density lidar sensors that can produce thousands of measurements per square meter, and with spectral sensors that are able to measure solar induced fluorescence. This research is being done in collaboration with Oregon State University, and has been funded by Brown University, NASA, and the National Science Foundation.
Field work in Costa Rica
Tropical rain forest canopy height
Forest blowdown in the Czech Republic
A tropical rain forest canopy
Lidar scan of a European beech
Heavy-lift drone in Costa Rica
Join the lab
Undergraduates, graduate students and postdocs are encouraged to join the lab.
If you’re interested, email James R. Kellner explaining why you want to join. Include a description of your background, experience, and goals.
