An Ecohydrologically-informed Machine Learning Approach for Understanding Dryland Carbon Dynamics

Drylands have tightly coupled water and carbon cycles due to persistent water scarcity, making them valuable systems for understanding coupled ecohydrological and biogeochemical processes. In addition, dryland ecosystems contribute significantly to the interannual variability of the terrestrial carbon sink. To better characterize dryland carbon dynamics, we present DryFlux, a machine learning upscaled product based on a dense network of eddy covariance sites in the North American Southwest. This product combines in-situ fluxes with remote sensing and meteorological data to estimate gross primary productivity in drylands while explicitly accounting for water limitation during the model development process. DryFlux outperforms existing products in capturing interannual and seasonal variation in carbon uptake when used globally. We specifically explore how machine learning techniques can accurately upscale fluxes at multiple spatial (1 km and 9 km) and temporal (daily, weekly, monthly) scales to find the best resolution for capturing spatial and temporal heterogeneity in carbon and water fluxes. In addition, we discuss how remotely sensed soil moisture from satellites can help capture biogeochemical 'hot spots' and 'hot moments' in drylands. Our findings can help us better understand dynamic carbon fluxes in drylands, as well as the spatiotemporal resolution needed to resolve water-carbon dynamics in these and other systems. Machine learning methods that explicitly incorporate water limitation in model development can contribute to a more comprehensive understanding of carbon, energy, and water fluxes at multiple scales.