Real-world soil carbon prediction using spectral data and machine learning

Overview

This solution delivers a machine learning-based approach to measuring soil organic carbon (SOC) using proximal soil spectral sensing. The system predicts SOC content from VIS-NIR-SWIR-MIR spectral data with demonstrated accuracy (RMSE 0.80%, R² 0.85 across 2,461 samples spanning SOC values from 0.11% to 15.92%). A key innovation is a deployment protocol that anticipates prediction failure by assessing whether incoming spectral samples are sufficiently represented within the model's existing training data. This makes the technology particularly valuable for regenerative agriculture, no-till farming systems, and carbon credit verification, where reliable, scalable soil carbon assessment is critical.

Technical specifications

Core capabilities:

• Predicts soil organic carbon content from VIS-NIR-SWIR-MIR spectral data with R² 0.85 precision
• Uses memory-based learning to compare incoming spectra against existing model spectra
• Employs similarity criteria during both calibration/validation and deployment phases
• Builds multiple specialized local models rather than relying on a single universal model
• Incorporates auxiliary data such as land cover to improve model accuracy
• Designed for proximal soil sensing applications in diverse geographic regions

Validation dataset characteristics:

• 2,461 soil samples from three distinct physiographic regions of Brazil
• Wide range of chemical and physical properties reflecting varied geology, land use, and climate
• Future expansion planned through acquisition of additional MIR soil spectra

Technology readiness level

The technology has been validated on a substantial dataset of 2,461 samples with strong predictive performance metrics already established. The current stage focuses on establishing a robust deployment workflow that ensures prediction reliability when the model encounters new, unseen spectral data. The approach addresses a known limitation in soil spectral ML models: their ability to generalize depends on how well incoming samples are represented in training data. The similarity-based protocol for deployment validation is the next development milestone, with the existing diverse Brazilian dataset providing a strong foundation for testing this workflow.