Performance evaluation of multispectral analysis for surface material classification

Previous work has shown the feasibility of merging surface material information derived from moderate resolution multispectral imagery with estimates of height based upon stereo matching in high resolution panchromatic imagery. The goal is to use surface material information, normally highly correlated with object location in complex urban scenes, as a source of information for small scale mapping of man-made structures such as buildings and roads, as well as natural features such as soil, vegetation, and water. The fusion of height estimates with surface material estimates provides a unique synthetic three dimensional dataset that is not directly available in any airborne imaging sensor. The focus of this paper is to present a performance evaluation of two classification techniques, gaussian maximum likelihood and differential radial basis function, on the task of surface material analysis. In order to carry out this evaluation the authors have created several highly detailed ground truth segmentations based upon manual analysis of the multispectral imagery, as well as by inspection of panchromatic imagery acquired over the same area. Tools built for the generation and validation of the ground-truth surface material map are also discussed