Digital Twin-Ready Earth Observation: Operationalizing GeoML for Agricultural CO2 Flux Monitoring at Field Scale

Operationalizing Earth Observation (EO)-based Machine Learning (ML) algorithms (or GeoML) for ingestion in environmental Digital Twins remains a challenging task due to the complexities associated with balancing real-time inference with cost, data, and infrastructure requirements. In the field of GHG monitoring, most GeoML models of land use CO2 fluxes remain at the proof-of-concept stage, limiting their use in policy and land management for net-zero goals. In this study, we develop and demonstrate a Digital Twin-ready framework to operationalize a pre-trained Random Forest model that estimates the Net Ecosystem Exchange of CO2 (NEE) from drained peatlands into a biweekly, field-scale CO2 flux monitoring system using EO and weather data. The system achieves an average response time of 6.12 s, retains 98% accuracy of the underlying model, and predicts the NEE of CO2 with an R2 of 0.76 and NRMSE of 8%. It is characterized by hybrid data ingestion (combining non-time-critical and real-time retrieval), automated biweekly data updates, efficient storage, and a user-friendly front-end. The underlying framework, which is part of an operational Digital Twin under the UK Research & Innovation AI for Net Zero project consortium, is built using open source tools for data access and processing (including the Copernicus Data Space Ecosystem OpenEO API and Open-Meteo API), automation (Jenkins), and GUI development (Leaflet, NiceGIU, etc.). The applicability of the system is demonstrated through running real-world use-cases relevant to farmers and policymakers concerned with the management of arable peatlands in England. Overall, the lightweight, modular framework presented here integrates seamlessly into Digital Twins and is easily adaptable to other GeoMLs, providing a practical foundation for operational use in environmental monitoring and decision-making.