Knowledge-guided machine learning can improve carbon cycle quantification in agroecosystems.

Area of Science:

• Earth System Science
• Agricultural Science
• Machine Learning

Background:

• Quantifying agroecosystem carbon cycles is vital for climate change mitigation and sustainable food production.
• Existing models (process-based and data-driven) face uncertainties due to complex processes and limited observational data.
• Accurate quantification is needed at scales relevant for decision-making.

Purpose of the Study:

• To develop and validate a Knowledge-Guided Machine Learning (KGML) framework for improved carbon cycle quantification in agroecosystems.
• To address prediction uncertainties inherent in conventional modeling approaches.
• To enhance the spatial resolution of soil organic carbon change assessments.

Main Methods:

• Integration of a process-based model's knowledge with high-resolution remote sensing data and machine learning (ML).
• Development of a hybrid KGML framework.
• Application and testing of the KGML framework in the U.S. Corn Belt.

Main Results:

• The KGML framework demonstrated superior performance over conventional process-based and black-box ML models.
• KGML achieved significantly higher spatial detail in quantifying soil organic carbon changes (86% more than conventional methods).
• The study provides a protocol for enhancing KGML and generalizing hybrid modeling approaches.

Conclusions:

• KGML offers a powerful approach to reduce uncertainties in agroecosystem carbon cycle modeling.
• High-resolution assessments of soil organic carbon dynamics are achievable with KGML.
• The developed framework can be generalized for predicting complex Earth system dynamics.