Sequential Data Assimilation of Canopy Attributes from Multispectral Bi-directional Reflectance
Adam
Wolf, Carnegie Institution Dept of Global Ecology, adamwolf@stanford.edu
(Presenting)
Joe
Berry, Carnegie Institution Dept of Global Ecology, joeberry@stanford.edu
(Presenting)
Greg
Asner, Carnegie Institution Dept of Global Ecology, gpa@stanford.edu
Monitoring NDVI over the annual cycle provides boundary conditions for modeling the dynamics of the carbon cycle. A strategy for possibly improving the quality of this input is to use the satellite observations in assimilation mode with a model that predicts reflectance based on surface properties and seasonal phenology. Each new observation could, thus, provide information to update the model of surface properties or identify change in those properties. Our approach is to use a model to simulate bi-directional reflectance for an ensemble of realizations of canopy architectures. These are then used in the data assimilation research test-bed, DART developed at NCAR, to assimilate canopy parameters using MODIS observations. To test our approach we make use of four sites in Hawaii Volcanoes National Park that have been carefully characterized. Statistics of canopy parameters and the model DISORD were used to develop ensembles that describe the ecological parameter space. We found that the numerical calculation of the reflectance for each ensemble member at each time-step for each pixel would be prohibitively slow, so we developed a neural network (NN) approximation of the model, trained it on a large number of ensemble realizations by DISORD for each band and for all possible sun/view geometries. This NN approximation was approximately 1000x faster than the numerical model and ~98% accurate with non-significant bias. The NN model was then used in DART to assimilate MODIS bidirectional reflectance in 7 spectral bands using the unique geometry of each observation to estimate the time-varying canopy attributes at each of the four sites.
NASA Carbon Cycle & Ecosystems Active Awards Represented by this Poster: