Improved data-driven ecosystem carbon fluxes under moisture stress through synergistic Earth observations

Abstract

The integration of global land surface remote sensing and in situ measured ecosystem carbon fluxes through machine learning approaches offers a unique data-driven perspective to diagnose the carbon cycle. Earth Observation (EO) data sets from different parts of the electromagnetic spectrum contain specific information on the land surface status, but also on the structural and physiological vegetation conditions. Each EO-derived land surface variable alone has a limited scope, addresses only individual aspects of the complex system, and can be confounded by other factors. Here we use the new generation statistical flux upscaling framework Fluxcom-X to analyse the individual and synergistic contributions of different EO data sets to site-level terrestrial carbon fluxes in tailored cross-validation experiments. Several distinct data streams are explored as predictor variables: land surface temperature (LST) from both polar orbiters and geostationary satellites (MODIS and SEVIRI), far-red SIF from GOME2, multi-spectral vegetation optical depth (VOD) from different sources (ku-band climate archive from Moesinger et al.(2020) and L-band from SMOS), and soil moisture (SM) from ESA CCI. Each predictor variable undergoes a dedicated pre-processing in terms of quality checks and gap-filling. Beyond their overall added value in prediction skill, we are interested in the impacts of the EO predictors on different scales of carbon flux variability (e.g. diurnal, seasonal, seasonal anomalies, inter-annual, and between sites), specifically during situations of unusual water scarcity and surplus. We also compute SHAP values to understand how the machine learning model uses the EO information. Additionally, a second line of analysis addresses the role of acquisition properties for the accuracy of the estimates.

The first results for the predictor variable MODIS LST show that the inclusion of MODIS LST improves GPP estimates on all time scales. The model strongly profits from LST as surrogate for moisture availability during dry anomalies, and for light availability during wet anomalies. Regarding the impact of acquisition properties of MODIS, we find that the variability in viewing geometry and overpass time does not affect the accuracy of simulated site-level GPP. However, failing to account for the clear-sky bias in availability of MODIS LST will result in a substantial decrease in accuracy, especially for overcast days.

Further experiments will include SEVIRI LST, SIF, VOD, as well as soil moisture, and we will analyse their role in the data-driven simulations of carbon fluxes. The lessons learned from the site-level cross-validation experiments will guide the production of gridded estimates of gross and net carbon fluxes for Europe and the globe.