Authored and led by Harvest partner Dr. Kaiyu Guan’s group, along with colleagues at the University of Illinois Urbana-Champaign and University of Nebraska-Lincoln, “Using soil library hyperspectral reflectance and machine learning to predict soil organic carbon: Assessing potential of airborne and spaceborne optical soil sensing” has recently been published in Remote Sensing of Environment. The study utilized a continental-scale soil laboratory spectral library to analyze several machine learning algorithms, along with pre-processed spectra, in order to quantify soil organic carbon concentration. Soil organic carbon (SOC) is essential for determining soil health as it affects both chemical and physical properties of soil, including: water absorption, moisture-holding capacity, nutrient availability, ecosystem services, and global carbon cycles.
The authors note that spectroscopy - the study of the absorption and emission of light - and particularly optical hyperspectral reflectance coupled with machine learning techniques, “can provide rapid, efficient, and cost-effective quantification of SOC.” Optical sensing, which covers visible, near-infrared, and shortwave-infrared light ranges, is flexible enough to be integrated with satellite-based platforms in order to quantify soil organic carbon at scale. Using the coupled soil-vegetation-atmosphere radiative transfer model, the team was able to evaluate the efficacy of this approach in estimating SOC concentration of surface bare soils. These types of evaluations, using simulated air and space data, can provide key insights about which remote sensing data and real-environmental noises are most appropriate for further soil organic carbon studies.
The main objective of the study, as noted by the authors, was to explore the applications of machine learning algorithms to improve understanding and make full use of hyperspectral reflectance in predicting soil organic carbon concentration.
Overall, results demonstrated that fusion data from multispectral satellite missions can effectively be used to monitor global soil carbon, and is particularly applicable in evaluating historical SOC changes. However, the researchers clarify that SOC models should be distinguished according to their distinct spectral signatures, and that shortwave infrared light is critical for surface SOC monitoring using remote hyperspectral sensors.
Read the full study to learn more about the machine learning methodologies used, and for in-depth explanations of the results.