A Knowledge-Guided Machine Learning (KGML) framework is proposed to improve the quantification of carbon cycle dynamics in agroecosystems. This framework integrates knowledge from process-based models, high-resolution remote sensing data, and machine learning techniques. Using the U.S. Corn Belt as a testbed, the KGML framework outperforms conventional process-based and black-box machine learning models in quantifying carbon cycle dynamics. The framework provides high-resolution insights into soil organic carbon changes, revealing 86% more spatial detail than conventional methods. The framework also outlines a protocol for improving KGML through various paths, enabling the development of hybrid models to better predict complex Earth system dynamics. Agroecosystems cover about one-third of Earth's land surface and play a key role in the global terrestrial carbon cycle. Agriculture is a significant source of greenhouse gases, but crops also remove carbon dioxide from the atmosphere. Practices that increase soil organic carbon (SOC) are a low-cost strategy for reducing atmospheric greenhouse gas concentrations. Accurate quantification of carbon fluxes and changes in SOC is essential for identifying appropriate conservation practices. Traditional carbon quantification methods face challenges due to high spatial heterogeneity and seasonality. Process-based models can be computationally prohibitive, while data-driven machine learning models suffer from out-of-sample prediction failure and lack interpretability. KGML combines the strengths of process-based models and machine learning to provide accurate, interpretable, and cost-effective carbon quantification. The KGML-ag-Carbon framework integrates a process-based model with a deep learning model to generate reliable predictions of agricultural carbon fluxes, crop yields, and changes in soil carbon stocks. The framework assimilates in-situ eddy covariance data, regional survey yield data, remotely-sensed gross primary production data, and synthetic data generated by a process-based model. The framework demonstrates high accuracy in predicting carbon fluxes, crop yields, and changes in soil carbon at high spatial and temporal resolution. The framework's performance is evaluated using synthetic and observed data, showing significant improvements in prediction accuracy compared to conventional models. The framework's ability to capture complex carbon flux dynamics and provide interpretable results makes it a valuable tool for carbon budget quantification. The framework's high-resolution outputs provide detailed insights into carbon sequestration and help inform sustainable land management practices. The framework's potential for application in other areas, such as predicting other target variables and simulating carbon dynamics in different ecosystems, is also discussed.A Knowledge-Guided Machine Learning (KGML) framework is proposed to improve the quantification of carbon cycle dynamics in agroecosystems. This framework integrates knowledge from process-based models, high-resolution remote sensing data, and machine learning techniques. Using the U.S. Corn Belt as a testbed, the KGML framework outperforms conventional process-based and black-box machine learning models in quantifying carbon cycle dynamics. The framework provides high-resolution insights into soil organic carbon changes, revealing 86% more spatial detail than conventional methods. The framework also outlines a protocol for improving KGML through various paths, enabling the development of hybrid models to better predict complex Earth system dynamics. Agroecosystems cover about one-third of Earth's land surface and play a key role in the global terrestrial carbon cycle. Agriculture is a significant source of greenhouse gases, but crops also remove carbon dioxide from the atmosphere. Practices that increase soil organic carbon (SOC) are a low-cost strategy for reducing atmospheric greenhouse gas concentrations. Accurate quantification of carbon fluxes and changes in SOC is essential for identifying appropriate conservation practices. Traditional carbon quantification methods face challenges due to high spatial heterogeneity and seasonality. Process-based models can be computationally prohibitive, while data-driven machine learning models suffer from out-of-sample prediction failure and lack interpretability. KGML combines the strengths of process-based models and machine learning to provide accurate, interpretable, and cost-effective carbon quantification. The KGML-ag-Carbon framework integrates a process-based model with a deep learning model to generate reliable predictions of agricultural carbon fluxes, crop yields, and changes in soil carbon stocks. The framework assimilates in-situ eddy covariance data, regional survey yield data, remotely-sensed gross primary production data, and synthetic data generated by a process-based model. The framework demonstrates high accuracy in predicting carbon fluxes, crop yields, and changes in soil carbon at high spatial and temporal resolution. The framework's performance is evaluated using synthetic and observed data, showing significant improvements in prediction accuracy compared to conventional models. The framework's ability to capture complex carbon flux dynamics and provide interpretable results makes it a valuable tool for carbon budget quantification. The framework's high-resolution outputs provide detailed insights into carbon sequestration and help inform sustainable land management practices. The framework's potential for application in other areas, such as predicting other target variables and simulating carbon dynamics in different ecosystems, is also discussed.