Constraining carbon budgets at a regional scale: fusing forest inventory data with a cohort-based biosphere model

Abstract:

Forest inventories play an essential role in carbon monitoring and REDD+, however they provide a sparse picture of the carbon cycle at a regional scale. Terrestrial Biosphere Models (TBMs) provide a complete picture of the carbon cycle, but efforts at combining inventory data with models have focused primarily on model calibration and purely model-based regional-scale carbon estimation, which ignore observed disturbances, management, and spatiotemporal variability in forest. Our approach is based on assimilating inventory observations in a size- and age-structured model, the Ecosystem Demography model (ED2). Assumptions of large homogenous areas in ecological models result in loss of details that hinder incorporation of observations. We address how to assimilate inventory data with model predictions in a practical way that is readily extensible to the simultaneous fusion of remote sensing and eddy covariance along with inventories.

We updated ED2 predictions on forest growth with Forest Inventory and Analysis program (FIA) data. Data assimilation method was the Ensemble Adjustment Kalman Filter (EAKF) as implemented in Data Assimilation Research Testbed (DART) workflow. The study area is a 1° by 1° grid with the Willow Creek Ameriflux tower in Wisconsin at center. ED2 groups individual trees in cohorts so it captures the landscape-scale heterogeneity. Although this approach speeds up computations, it is not practical to estimate each FIA plot within a chosen area. We classified and averaged data for different plots according to their biomass based on number and size of trees within a plot, focusing on biomass changes over a measurement period. We separately calculated the average diameter at breast height (dbh) and stem density for plants over 5 cm for measured and modeled plots within a biomass class for different Plant Functional Types (PFTs).

The results showed EAKF successfully adjusting the predicted changes in biomass according to observations. Variation in stem density and dbh within the same biomass class was low for plots with mid to high biomass and thus the state estimates fit better with all observed values within that biomass class, while plots with low biomass show high variation. Therefore applying a single correction factor for all plots did not match the updated plots with the observed plots.