Global long-term hourly 9 km terrestrial water-energy-carbon fluxes with physics-informed machine learning

In this study, we generate a global, long-term, hourly terrestrial water-energy-carbon fluxes, using model simulations, in-situ measurements, and physics-informed machine learning. The soil-plant model, STEMMUS-SCOPE was deployed to simulate land surface fluxes over 170 Fluxnet sites (STEMMUS – Simultaneous Transfer of Energy, Mass, and Momentum in Unsaturated Soil; SCOPE - Soil Canopy Observation, Photochemistry and Energy fluxes radiative transfer model). The model input and output data were then used as training data-pairs to develop the STEMMUS-SCOPE emulator using multivariate random forests regression algorithm. Here, physics-informed machine learning refers to the fact that the emulator was trained and constrained by the physical consistency represented by the soil-plant model. We compared the physics-informed emulator (RF_S-S) with the one trained using only Fluxnet in-situ measurements (RF_in-situ), and found that the land surface fluxes predicted by RF_S-S are less scattered than that by RF_in-situ.

We estimate six variables simultaneously: net radiation, latent heat flux, sensible heat flux, gross primary productivity, solar-induced fluorescence in 685 nm and 740 nm. Results show that RF_S-S can estimate fluxes with Pearson Correlation Coefficient score (r-score) higher than 0.97 for these six variables. The testing result using independent stations (not included for developing emulators) show a r-score higher than 0.94. The feature importance shows that incoming shortwave radiation, surface soil moisture, and leaf area index are top predictor variables that determine the prediction performance. We further explore the performance of RF_S-S in predicting soil heat flux, root zone soil moisture, and leaf water potential, which assist the understanding of ecosystems’ drought responses to climate change.