Focus Area
Statistical Data Fusion for Earth Remote Sensing
Since 2007 we have been developing and applying methods for combining heterogeneous Level 2 remote sensing data to create fused products that minimize uncertainty. Our technical formalism relies on spatial statistical models. These models postulate the existence of an underlying true spatial field which exibits spatial dependence; i.e., knowing the value of the physical variable at one location reduces the uncertainty about its value at nearby locations.
Remote sensing data are noisy, spatially aggregated observations of the true field. Data from two different instruments have different measurement errors, aggregate over different ground footprints (called the spatial "support" of the observations), and generally have different sampling characteristics. In other words, they view the same "truth" differently. Given reliable uncertainty information on the inputs, data fusion uses a spatial statistical model to infer the true spatial field. Our methodology weights the inputs using their uncertainties, and capitalizes on spatial dependence to infer the complete field and report accompanying uncertainties. Inferences can be made even where there are not observations, although the output uncertainties will be higher.
We have performed data fusion in a variety of contexts. Currently, we are working to create fused temperature and water vapor data products for the Atmospheric Infrared Sounder (AIRS) and the Cross-track Infrared Scanner (CrIS). Links are also provided to publications on prior data fusion work and work by our outside collaborators using NASA data.