Accurate and rapid quantification of greenhouse gas (GHG) emissions from the agroecosystem is critical for striking a balance between agricultural production and climate change mitigation. Traditional cropping system models, while useful, often fall short due to inadequate representations of physical and biogeochemical processes and uncertainties in numerous model parameters. These shortcomings become particularly pronounced when such models are applied across diverse landscapes with limited observational data. Knowledge-guided machine learning (KGML) is a novel hybrid modeling framework that deeply integrates process-guided models within machine learning models to significantly reduce the data requirements for learning and enhances out-of-sample prediction accuracy. Coupled with inverse modeling techniques, KGML is ideally suited for representing complex system dynamics with numerous missing, unmeasurable, or even unknown processes and variables. This presentation will highlight our recent advancements in hybrid modeling for improved GHG monitoring and predictions in agroecosystems. For instance, we have developed a KGML model to predict the "hot-moment, hot-spots" patterns of carbon dioxide (CO2) and nitrous oxide (N2O) emissions, and used a novel machine learning-based data assimilation approach to significantly reduce prediction uncertainties with remote and in-situ sensing data. Another example includes using KGML to assist the first-of-its-kind partitioning of autotrophic (Ra) and heterotrophic (Rh) respirations at the field-to-regional level, which provides a unique opportunity to investigate the spatial and temporal patterns of how Ra and Rh may respond differently to climate change and different management practices, thereby identify preferred strategies for mitigating climate change. Overall, our studies underscore the high potential of KGML application in complex agroecosystem modeling and offer insights into the development of the next generation of AI-empowered agroecosystem prediction frameworks.