In this ESA project, we will explore whether machine-earning based emulators are capable of not only reproducing European carbon fluxes from the JULES land surface model but going beyond this and providing a means to derive a novel observation-driven dataset of GPP, built on the existing process-level understanding within the model.

placeholder image 1

Question 1

How well can machine learning methods emulate physical process-based land surface models, focused over Europe?

placeholder image 2

Question 2

Can explainable AI techniques provide new insights into process understanding when combining land surface models and Earth Observation data?

Question 3

Are the learnt relationships between the modelled inputs and outputs consistent with those from Earth Observation data?

image In order to answer the above questions, we will undertake the following activities:

  • We will produce land surface model simulations from JULES over Europe for a range of terrestrial essential climate variables.
  • We will develop, train and evaluate machine learning models against the simulated land surface parameters, providing the capability to successfully emulate the complex physical process-based models.
  • These emulators will be used to investigate the complex emergent relationships and feedbacks inherent in such simulations to gain an increased understanding of the underlying Earth System processes.
  • We will use these emulators to test whether data from satellite-based essential climate variables (e.g. ESA-CCI) are consistent with the relationships learnt from the land surface models.
  • We will produce an Emulated-GPP (gross primary productivity) data product based on EO data, using the relationships learnt from the land surface model.