Their research details an innovative data transformation approach, which integrates deep learning and explainable artificial intelligence with mid-infrared spectroscopy, to predict soil organic carbon (SOC) fractions, a critical aspect of understanding overall SOC and soil health. The researchers’ innovative approach makes predicting SOC fraction outcomes more accurate and cost-effective, making the technology more available to farmers and land managers and better allowing them to assess SOC sequestration potential and monitor soil conditions.

Highlights

  • CLR transformation ensures the sum of estimated fractions equals total SOC.
  • MIR spectroscopy with CNN accurately estimates the SOC fractions compositionally.
  • CNN outperforms cubist for estimating SOC fractions from MIR spectra.
  • SHAP values identify critical functional groups in the SOC fractions across land uses.
MIR spectra of soil samples and statistic.
The total SOC and the SOC fractions in the observations and the estimates with and without CLR.
Feature importance for MAOC fraction derived from the SHAP values of the CNNs.