Machine Learning-Augmented Model-Data Fusion for Soil Organic Carbon Prediction

Overview

This research proposes embedding machine learning (ML) algorithms into model-data fusion frameworks to enhance the accuracy and precision of soil organic carbon (SOC) predictions in agricultural soils. The approach integrates process-based crop models with ML techniques and observational data to create a composite view of the carbon cycle. By constraining easily observable carbon pools such as plant biomass and leaf area index (LAI) using a Bayesian model-data fusion framework, the methodology addresses unexplained patterns in prediction residuals and identifies missing processes in current crop models. The result is a scalable, harmonized framework for soil carbon monitoring, reporting, and verification with applications in sustainable agriculture, climate-smart farming, and carbon market verification.

Technical specifications

Key features:

• Model-data fusion pipeline based on the APSIM process-based crop model and Ensemble Kalman Filter (EnKF) for assimilating observations
• Machine learning augmentation using algorithms such as deep learning and gradient boosting machines (GBM) to model and predict residuals from APSIM SOC estimates
• Satellite-derived constraints including LAI and plant biomass observations used to constrain the carbon cycle during the analysis step
• Multi-scale validation demonstrated across approximately 500 sites and 3 plant functional types across the United States
• Uncertainty quantification providing national-scale uncertainty accounting for the terrestrial carbon cycle
• Scalable framework designed for harmonizing multiple data constraints to support operational soil carbon monitoring

Technology readiness level

The foundational model-data fusion approach has been validated at national scale across the United States, with results reconciled against independent data products and demonstrated across approximately 500 sites. A manuscript documenting this work is currently under review. The ML augmentation component builds upon an existing model-data fusion pipeline, with a manuscript in preparation. The next phase involves training ML models on APSIM drivers (soil, climate, parameters) to predict prior residuals, then nudging APSIM SOC estimates before passing them through the EnKF analysis step. This represents an advanced research-stage technology moving toward operational deployment for carbon monitoring, reporting, and verification applications.