Hyperspectral sensing and AI pave new path for monitoring soil carbon

Just how much carbon is in the soil? That’s a tough question to answer at large spatial scales, but understanding soil organic carbon at regional, national, or global scales could help scientists predict overall soil health, crop productivity, and even worldwide carbon cycles. 

Classically, researchers collect soil samples in the field and haul them back to the lab, where they analyze the material to determine its makeup. But that’s time- and labor-intensive, costly, and only provides insights on specific locations. 

In a recent study, University of Illinois researchers show new machine-learning methods based on laboratory soil hyperspectral data could supply equally accurate estimates of soil organic carbon. Their study provides a foundation to use airborne and satellite hyperspectral sensing to monitor surface soil organic carbon across large areas.

“Soil organic carbon is a very important component for soil health, as well as for cropland productivity,” says lead study author Sheng Wang, research assistant professor in the Agroecosystem Sustainability Center (ASC) and the Department of Natural Resources and Environmental Sciences (NRES) at U of I. “We did a comprehensive evaluation of machine learning algorithms with a very intensive national soil laboratory spectral database to quantify soil organic carbon.” 

Wang and his collaborators leveraged a public soil spectral library from the USDA Natural Resources Conservation Service containing more than 37,500 field-collected records and representing all soil types around the U.S. Like every substance, soil reflects light in unique spectral bands which scientists can interpret to determine chemical makeup.

“Spectra are data-rich fingerprints of soil properties; we’re talking thousands of points for each sample,”  says Andrew Margenot, assistant professor in the Department of Crop Sciences and co-author on the study. “You can get carbon content by scanning an unknown sample and applying a statistical method that’s been used for decades, but here, we tried to screen across pretty much every potential modeling method,” he adds.

“We knew some of these models worked, but the novelty is the scale and that we tried the full gamut of machine learning algorithms.”

Kaiyu Guan, principal investigator, ASC founding director, and associate professor at NRES, says, “This work established the foundation for using hyperspectral and multispectral remote sensing technology to measure soil carbon properties at the soil surface level. This could enable scaling to possibly everywhere.”

After selecting the best algorithm based on the soil library, the researchers put it to the test with simulated airborne and spaceborne hyperspectral data. As expected, their model accounted for the “noise” inherent in surface spectral imagery, returning a highly accurate and large-scale view of soil organic carbon. 

“NASA and other institutions have new or forthcoming hyperspectral satellite missions, and it’s very exciting to know we will be ready to leverage new AI technology to predict important soil properties with spectral data coming back from these missions,” Wang says.

Chenhui Zhang, an undergraduate student studying computer science at Illinois, also worked on the project as part of an internship with the National Center for Supercomputing Applications’ Students Pushing Innovation (SPIN) program.

“Hyperspectral data can provide rich information on soil properties. Recent advances in machine learning saved us from the nuisance of constructing hand-crafted features while providing high predictive performance for soil carbon,” Zhang says. “As a leading university in computer sciences and agriculture, U of I gives a great opportunity to explore interdisciplinary sciences on AI and agriculture. I feel really excited about that.”

The article is published in Remote Sensing of Environment.