Computer Model Emulation for High-Dimensional Functional Output from OCO-2 Remote Sensing

• Activity Number: 348 - Applications: Gaussian Process and Computer Experiments
• Type: Contributed
• Date/Time: Tuesday, July 30, 2019 : 10:30 AM to 12:20 PM
• Sponsor: Section on Physical and Engineering Sciences
• Author(s): Anirban Mondal* and Pulong Ma and Jonathan Hobbs and Emily Lei Kang and Alex Konomi and Joon Jin Song
• Companies: Case Western Reserve University and Duke University and Jet Propulsion Laboratory and University of Cincinnati and University of Cincinnati and Baylor University
• Keywords: Emulator; Nearest Neighbor Gaussian Process; Remote Sensing; Functional Principal Component; Active Subspace

Abstract

NASA's Orbiting Carbon Observatory-2 (OCO-2) collects tens of thousands of observations of reflected sunlight daily, and the mission's retrieval algorithm processes these indirect measurements into estimates of atmospheric CO2 and other states. The physical forward model describing the mathematical relationship between the atmospheric state and the observed radiances is developed by the OCO-2 science team in the form of an expensive computer simulation. The multiple runs of this expensive computer simulation model makes the retrieval algorithm computationally very expensive. Here we focus on the emulator approach where a statistical representation of the forward model is built based on some simulation runs of the forward model. Once the emulator is built it can be used to predict the observed radiance output for a given set of atmospheric state vectors almost instantaneously. We use functional principal component analysis to reduce the dimension of the functional radiance output and active subspace approach to reduce the dimension input state vector. Nearest neighbor Gaussian process is used to fit the emulator and its performance is compared with a physical surrogate model.