Spatial Inference for Distributed Remote Sensing Data

Monday, 15 December 2014: 2:10 PM
Amy J Braverman1, Matthias Katzfuss2 and Hai Nguyen1, (1)NASA Jet Propulsion Laboratory, Pasadena, CA, United States, (2)Texas A&M University, Statistics, College Station, TX, United States
Remote sensing data are inherently spatial, and a substantial portion of their value for scientific analyses derives from the information they can provide about spatially dependent processes. Geophysical variables such as atmopsheric temperature, cloud properties, humidity, aerosols and carbon dioxide all exhibit spatial patterns, and satellite observations can help us learn about the physical mechanisms driving them. However, remote sensing observations are often noisy and incomplete, so inferring properties of true geophysical fields from them requires some care. These data can also be massive, which is both a blessing and a curse: using more data drives uncertainties down, but also drives costs up, particularly when data are stored on different computers or in different physical locations. In this talk I will discuss a methodology for spatial inference on massive, distributed data sets that does not require moving large volumes of data. The idea is based on a combination of ideas including modeling spatial covariance structures with low-rank covariance matrices, and distributed estimation in sensor or wireless networks.