Remote sensing observations and eddy covariance measurements are both widely used in ecology to improve understanding of biosphere-atmosphere-hydrosphere interactions across scales and in various ecosystems. Continuous measurements from flux towers facilitate exploration of the exchange of carbon dioxide, water and energy between the land surface and the atmosphere at fine temporal and spatial scales, while satellite observations can fill in the large spatial gaps of in-situ measurements and provide long-term temporal continuity. Here we demonstrate a machine learning approach to upscale ecosystem-scale carbon flux estimates to the Southwest (SW United States and NW Mexico) regional scale using remotely sensed and gridded meteorological inputs. Our upscaling method leverages the strengths of both the satellite and flux data, producing spatially and temporally continuous high-resolution estimates of Gross Primary Productivity (GPP). We focus here on water-limited ecosystems, which have been shown to disproportionately impact variability in the global terrestrial carbon sink. Existing upscaled flux products are sparsely informed by water-limited ecosystem measurements. Our machine learning approach was designed specifically for semi-arid ecosystems: with explicit consideration for the impacts of the water balance and drought on carbon dynamics, and validation procedures that assess both interannual and seasonal variability in vegetation carbon uptake.
Results/Conclusions
Our spatially and temporally continuous upscaled GPP product help us understand linkages between the carbon and water cycles in semi-arid ecosystems and informs predictions of vegetation response to future climate conditions. By including a multi-scalar drought index (SPEI; Standardized Precipitation Evapotranspiration Index) at multiple timescales as a predictor in our machine learning models, we captured the response of vegetation to short-term drought, seasonal water availability, and interannual precipitation variability. We found that our 1 km spatial resolution was necessary to accurately quantify drought impacts on carbon uptake in the Southwest due to spatially heterogeneity in vegetation and topography. Our product improves on existing globally upscaled products, which do not generally perform well in semi-arid regions. Our machine-learning approach using moderate-resolution (i.e. 1km) satellite and meteorological inputs combines ground measurements of carbon fluxes and spaceborne estimates of vegetation productivity to produce continuous estimates of GPP through space and time that reflect semi-arid ecosystem dynamics. Machine learning approaches can bridge ground and spaceborne observations, with potential applications to improve estimates of ecosystem processes across spatial and temporal scales.