HR: 0830h
AN: G11A-0246 [PDF]
TI: Multiscale Data Fusion Regulated by a Mixture-of-Experts Network
AU: * Slatton, K C
EM: slatton@ece.ufl.edu
AF: University of Florida, Engineering Building, Rm 459
PO Box 116130, Gainesville, FL 32611 United States
AB:
Laser altimetry (LIDAR) and interferometric synthetic aperture radar (InSAR) have emerged as important tools for remotely
sensing topography at fine and medium scales, respectively. Strip-map InSAR provides large coverage areas, but at spatial
resolutions that are often insufficient for many applications. Conversely, LIDAR provides higher resolution, but covering
large areas can be impractical. Slatton, et al. (2001) demonstrated that digital elevation models (DEMs) derived from LIDAR
and InSAR data could be fused to provide large coverage areas, while maintaining high resolution locally. A multiscale
Kalman smoother (MKS) employing a fractional Brownian motion stochastic model allowed the estimation of fused elevations with
uncertainty measures at every pixel. However, the standard MKS algorithm with a single stochastic model does not
incorporate spatial variations in the elevation statistics. For example, rough undulating terrain yields an elevation
surface with a shorter correlation length than flat smooth terrain. In this work, multiscale Kalman filters are defined in a
multiple-model configuration that accommodates local variations in elevation statistics. Stochastic model realizations for
long and short correlation length surfaces are blended together with a simple Mixture-of-Experts (ME) network. Implementing
classical multiple-model approaches, such as a Magill filter bank, on multiscale data structures would require that a
particular model be selected for every node in the quadtree. The selection of the best model at a parent node becomes
potentially problematic if different models were selected as best at the children nodes. The need to explicitly map
different stochastic models to the quadtree nodes of a multiscale estimator is obviated in the ME approach because the
relative weighting of the individual Kalman estimates is automatically determined based on the innovation sequences to
provide an adaptive estimate of the elevations.
DE: 1294 Instruments and techniques
DE: 6974 Signal processing
SC: Geodesy [G]
MN: 2003 Fall Meeting