Multiscale Segmentation of Elevation Images Using a Mixture-of-Experts Framework

Abstract

Estimating and segmenting topography by fusing data acquired over multiple resolutions has been extensively studied over the years. The standard multiscale Kalman smoother estimator embedded with a single stochastic model parameterized using power-spectral matching methods has been found to give suboptimal performance in estimating nonstationary topographic variations. The modeling is based on the fact that topography and other geophysical phenomena exhibit a 1/f property. Although acceptable for data sets over large areas, this approximation is found to be poor for data sets over small areas. This letter employs multiple models regulated by a mixture-of-experts network to adaptively fuse the estimates. Alternate to power spectral methods, a fractal-based approach is used to segment data and parameterize the multiple models for better performance. Sensitivity and performance analyses are also performed on the parameters of the estimators, and ideal selection criteria are proposed.