Nguyen, Hai; Cressie, Noel; and Braverman, Amy, Multivariate spatial data fusion for very large remote sensing datasets, National Institute for Applied Statistics Research Australia, University of Wollongong, Working Paper 13-16, 2016, 31.
Global maps of carbon dioxide (CO 2) mole fraction (in units of parts per million) in the lower atmosphere are important tools for climate research since they can help identify sources and sinks of CO 2. No satellite instrument currently provides estimates of the lower-atmosphere CO 2, though inferences are possible using data from existing instruments. Two remote sensing instruments, the Orbiting Carbon Observatory 2 (OCO-2) and the Greenhouse gases Observing SATellite (GOSAT), both observe column-averaged CO 2. These data are then used as inputs into flux inversion, which combines a transport model, a priori atmospheric information, and satellite-derived column-averaged CO 2 to produce estimates of sources and sinks.
Here, we demonstrate a method for improving inferences for column-averaged CO 2 using OCO-2 and GOSAT. Both instruments produce estimates of CO 2 concentration, called profiles, at 20 different pressure levels. Operationally, each profile estimate is then convolved into a single estimate of column-averaged CO 2 using a pressure weighting function. However, CO 2 may be more efficiently estimated by making optimal estimates of the vector-valued CO 2 profiles and applying the pressure weighting function afterwards. These estimates will be more efficient if there is multivariate dependence between CO 2 values in the profile. In this article, we describe a methodology that uses a modified Spatial Random Effects model to account for the multivariate nature of the data fusion of OCO-2 and GOSAT. We show that multivariate fusion of the profiles has improved mean squared error relative to scalar fusion of the column-averaged CO 2 values from OCO-2 and GOSAT. The computations scale linearly with the number of data points, making it suitable for the typically massive remote sensing datasets. Furthermore, the methodology properly accounts for differences in instrument footprint, measurement-error characteristics, and data coverages.