| This thesis presents a model-data fusion study to derive regional-scale (∼104 km2) CO2 flux estimates for summer 2004 in the northeast United States and southern Quebec, using an end-to-end treatment that goes from strategizing observations to assimilating extensive data into a receptor-oriented modeling framework. Surface fluxes are specified using the Vegetation Photosynthesis and Respiration Model (VPRM), a simple, readily optimized biosphere model driven by satellite data, AmeriFlux eddy covariance measurements, and meteorological fields. The surface flux model is coupled to a Lagrangian atmospheric adjoint model, the Stochastic Time-Inverted Lagrangian Transport Model (STILT), which links point observations to upwind sources with high spatio-temporal resolution. Concentration data for assimilation comes from the CO2 Budget and Regional Airborne Maine (COBRA-Maine) airborne campaign---which is described fully---and the NOAA-ESRL tall tower at Argyle, ME, as well as an ad-hoc regional network of surface observation stations. The variety of independent constraints provided by each input demonstrates the need for large amounts of data, shows the importance of both spatial and temporal coverage, and emphasizes the complementarity of tower and airborne observations. Although the dataset is dense and regionally representative, the surface source function is relatively insensitive to Bayesian optimization, providing an important counterexample to the current working paradigm of CO2 data assimilation studies. Errors in transport and tracer boundary conditions and in representation of the atmospheric boundary layer contribute to variance at the surface which is large enough to limit the effectiveness of atmospheric data assimilation for constraining surface fluxes. |
