Abstract
Effective land management and climate adaptation call for tools that are reliable and easy to use. In this paper, we present the development and evaluation of machine learning emulators trained to reproduce ecosystem model outputs of urban vegetation carbon fluxes: photosynthesis (GPP) and net ecosystem exchange (NEE), on two different time scales, daily and monthly. By training the emulators on a set of JSBACH simulations, they learn to map meteorological drivers to the target fluxes. We demonstrate that our daily emulators achieve a mean R 2 of 0.93, and the monthly emulators a mean R 2 of 0.91, averaged across plant functional types and targets, relative to JSBACH. Thus, simulating monthly mean values provides a computationally efficient alternative to daily outputs, enabling informed decision-making with a negligible reduction in precision. Additionally, we compare the NEE estimates with eddy covariance measurements at an urban site in Helsinki and the GPP estimates with satellite-based solar-induced fluorescence data across Helsinki. Our emulators lay the groundwork for further development of lightweight tools that avoid the setup and computational requirements of the underlying process-based model and can inform planning and monitoring of land use and climate adaptation.
Funding Information
  • ACCC and FAME Flagship programs of the Research Council of Finland (59342)
  • TerraDT project which has received funding from the European Union’s Horizon Europe Framework Programme (101187992)
  • Strategic Research Council (358257)
  • Research Council of Finland (361723)

This publication has 49 references indexed in Scilit: