A data-driven approach for modeling regional evapotranspiration and net ecosystem exchange of carbon

Evapotranspiration (ET) and net ecosystem exchange (NEE) of carbon are two critical variables in terrestrial ecosystems for the understanding of spatiotemporal distributions of water and carbon dynamics. These two variables have been the active subjects in process-based ecosystem models for decades.;In process-based models, the behavior of an ecosystem is represented by a set of functional components interacting with each other. These models commonly use empirical equations and parameters to describe the components and their interactions. However, the mechanisms behind ecosystems are often too complex to be captured explicitly by a set of equations and parameters. Consequently, these models have difficulties in estimating highly dynamic and/or poorly understood ecosystem processes such as ET and NEE of carbon at regional to global scales.;Parallel to the process-based ecosystem modeling efforts are massive data inventories from satellite remote sensing and ground observation networks, the development of statistical machine learning techniques in computer science and statistics, and the accumulation of extensive knowledge gained from ecosystem studies. These have provided the opportunities for the development of data-driven ET and NEE models.;This thesis focused on using data-driven approaches for ET and NEE modeling at regional scales by integrating remote sensing products (land surface temperature, enhanced vegetation index, surface shortwave radiation, and land cover) with AmeriFlux network and domain knowledge into machine learning, thus providing an alternative for the generation of information on the spatial and temporal distributions of ET and NEE at a regional scale.;Evaluations of the data-driven ET model with observations from AmeriFlux network found that it predicted ET with a root mean squared error (RMSE) of 0.63 mm/d (∼40% of the mean observed ET) and an R2 of 0.71, which was promising given the heterogeneity of the AmeriFlux network. When the model was applied to the coterminous U.S., it produced spatiotemporal ET estimates consistent with expected patterns.;The data-driven model for NEE predictions had poor performance with an RMSE of 1.11 gC/m2/d (>100% of the mean observed NEE). This is likely because the remote sensing products included in this thesis research (land surface temperature, enhanced vegetation index, surface shortwave radiation, and land cover) are not adequate to capture small NEE changes. However, when the NEE estimate problem was reformulated into a carbon source and sink identification problem, the data-driven model was able to locate carbon sources and sinks with an accuracy of 79% when evaluated with AmeriFlux network data, which was encouraging. When applied the model to the coterminous U.S. the model generally captured the spatiotemporal distribution of carbon sources and sinks with expected results.;Limitations of this thesis research include: (1) the use of ground-based shortwave radiation for model development and 0.5° shortwave radiation from satellite remote sensing for the coterminous U.S. application is not appropriate and is likely to introduce errors for the continental application; (2) the flux sites included in this thesis may not fully represent the spatiotemporal variation of actual ET/NEE; (3) the use of a 7 km x 7 km region as the spatial representativeness for AmeriFlux site may inadequately represent an area represented by each flux site; and (4) data-driven models are more restricted to predictions with limited spatial and temporal flexibility compared to process-based models because the reliability of data-driven models is highly affected by the representativeness of the input data.;In summary, although there are potentially confounding factors and it is difficult to validate ET and NEE distributions over the coterminous U.S., this thesis research shows that data-driven models generally capture the expected spatiotemporal variations of ET and NEE over the coterminous U.S. The result strongly suggests that data-driven models trained at flux sites can be generalized to larger regions.;Keywords. Evapotranspiration, Net Ecosystem Exchange of Carbon, AmeriFlux Network, Machine Learning, Remote Sensing.