LEAST SQUARES PREDICTION

The basic principles of least squares prediction using both multiquadric functions and covariance functions are covered briefly. Similarities and dissimilarities of the multiquadric and covariance methods are discussed. Multiquadric kernels are based on geometric and physical considerations rather than stochastic processes as is the case of covariance kernels. Thus the procedure of determining and fitting empirical covariances to select an analytical covariance function is unnecessary in multiquadric analysis. Comparative test results involving topographic features in Hawaii are given. A similar type test involving gravity anomalies in Iowa is also discussed. In general multiquadric kernels were found to be superior to covariance kernels in accuracy and computational efficiency for these applications. Topography, gravity anomalies, and other phenomena are not stationary in the sense of stationary random functions, which is at the heart of the justification for least squares prediction with covariance functions. This is probably true of many of the phenomena of interest in photogrammetry and remote sension. Least squares prediction with multiquadric functions is applied in a demonstration of its capability with respect to image processing and analysis.