INTEGRATING REMOTE SENSING DATA WITH GEMS TO IMPROVE SIMULATION OF CARBON DYNAMICS

Quantifying the spatial and temporal dynamics of ecosystem carbon stocks and fluxes has been a major challenge, and there is an urgent need to improve the accuracy of quantification through combining model simulations with various observations, especially remotely sensed data. Specifically, we need modeling tools that can: (1) adequately detect and remove model structure deficiency, (2) adaptively quantify the behavior of model parameters (temporal and spatial changes), (3) optimally combine model simulations with various observations from field and remote sensing sources, and (4) rigorously assess model output uncertainty. We developed a model-data fusion system to provide mathematical framework and software infrastructure that satisfied these needs. The model-data fusion system consists of the General Ensemble Biogeochemical Modeling System (GEMS), data from various resources, and data assimilation techniques (e.g., Smoothed Ensemble Kalman Filter (SEnKF)). We presented two studies to show the applications of GEMS-SEnKF at plot and regional scales. We first applied GEMS-SEnKF to assimilate eddy covariance measurements at two different FLUXNET sites into GEMS. One of the sites was a cropland located in Nebraska, experienced corn-soybean rotations. The other site was a mature black spruce forest in the Delta Junction of Alaska. The simulation results suggested that GEMS-SEnKF (1) successfully detected inter-species differences, seasonal variations, and biases of the key parameters (e.g., potential production rate and potential decomposition rate) in GEMS, (2) substantially reduced uncertainty of state variables stemmed from errors of parameters, input, and structure. At the regional scale, we applied GEMS-SEnKF to simulate regional carbon sequestration capacity in a federal land and compare the trends with those in surrounding non-federal lands. Results indicated that assimilation of remotely sensed data can dramatically improve the capability of model parameterization to correctly represent the spatial heterogeneity of land surface processes and model parameters.