Towards a Digital Twin of the Carbon Cycle in Europe

<div> <p><span data-contrast="auto">In recent years, Digital Twins of the Environment are emerging as useful tools for decision support to help address specific, complex environmental problems. We understand a Digital Twin as a digital representation of a physical system that is data-driven, has some predictive capabilities, and can provide decision support for stakeholders. With the increase in computer power, the development of advanced machine-learning techniques, and the ever-increasing wealth of Earth Observation data available, Digital Twins have a lot of potential to help solve or manage important environmental challenges, such as climate change.</span><span data-ccp-props="{"201341983":0,"335551550":6,"335551620":6,"335559739":160,"335559740":259}"> </span></p> </div> <div> <p><span data-contrast="auto">Here we present our recent work undertaken as part of our ESA EO4Society project IMITATE (‘Introducing Machine learning Into Targeted Analysis of Terrestrial Ecosystems’), where we developed the foundations of a Digital Twin of the carbon cycle over Europe using a model-data fusion. Our approach is based on the development of machine-learning emulators of the land surface model JULES (Joint UK Land Environment Simulator), which is the terrestrial component of the UK Earth System Model. For this project, we have focused on a single parameter/process within JULES (Gross Primary Productivity, GPP) and a limited geographical domain (Europe).</span><span data-ccp-props="{"201341983":0,"335551550":6,"335551620":6,"335559739":160,"335559740":259}"> </span></p> </div> <div> <p><span data-contrast="auto">These emulators have several advantages. They are really fast and much less computationally intensive, which means we can run simulations in milliseconds. They can also be used without expert knowledge of model or HPC facilities, embedded within web-applications, etc. In addition, we can take advantage of explainable AI methods to better understand the relationship between the inputs and the outputs. Finally, by driving the emulators (which have been trained on JULES data and thus learned its physical basis) with Earth Observation data, we produce a new GPP dataset, which is based on the physics from JULES and constrained by observations.</span><span data-ccp-props="{"201341983":0,"335551550":6,"335551620":6,"335559739":160,"335559740":259}"> </span></p> </div> <div> <p><span data-contrast="auto">Our emulators match JULES remarkably well, with an average R</span><sup><span data-contrast="auto">2</span></sup><span data-contrast="auto"> better than 0.98. In addition, using these emulators with input EO data such as Land Surface Temperature and FAPAR, the emulated GPP is comparable to other products (MODIS and Sen4GPP), and has good agreement with observations from some FLUXNET stations. Thus, we present here a novel approach with very promising results, and there are still many possibilities for improvements, further developments and new applications, which we will be exploring in future.</span><span data-ccp-props="{"201341983":0,"335551550":6,"335551620":6,"335559739":160,"335559740":259}"> </span></p> </div>