Title: Multivariate Data Fusion and Uncertainty Quantification for Remote Sensing
Presenting Author: Amy Braverman
Organization: Jet Propulsion Laboratory

Abstract:
In this talk we discuss our approach to multivarite data fusion: the process of combining multivariate observations from multiple remote sensing instruments to create an optimal (unbiased and minimum variance) estimate of a multivariate geophysical field. The field's estimate is accompanied by a companion uncertainty field. The instruments observe over the same spatio-temporal window, but not necessarily at the same times or locations, or with the same footprints or measurment errors. Our Multivariate Spatio-Temporal Data Fusion algorithm (MVSTDF) exploits spatial, temporal, and inter-variable dependence in the data to drive the uncertainty of the fused product to a minimum. The method is based on a statistical model that relates the observed data to the underlying true geophysical process, and inference is made on the parameters of that model after accounting for the data in each time period. This is accomplished with a Kalman smoother. We demonstrate using synthetic, high-resolution data generated by a simulation system we built for that purpose. The simulation system also uses a statistical model of the geophyscial process, but that model is different than the one that is fit in data fusion, and the simulation model parameters are estimated from coarse resolution physical model output rather then from observations. We simulate and then fuse data that are proxies for atmospheric profiles of carbon dioxide like those that will be produced by OCO-2, and we evaluate MVSTDF by comparing the fused estimtes to the synthetic true fields. Finally, we conclude with a discussion of the computational and visualization challenges of modeling very high-dimensional, spatio-temporal processes using massive remote-sensing data sets.