KERNEL-BASED PLS REGRESSION CROSS-VALIDATION AND APPLICATIONS TO SPECTRAL DATA

Multivariate images are very large data structures and any type of reg ression for their analysis is very computer-intensive. Kernel-based pa rtial least squares (PLS) regression, presented in an earlier paper, m akes the calculation phase more rapid and less demanding in computer m emory. The present paper is a direct continuation of the first paper. In this study the kernel PLS algorithm is extended to include cross-va lidation for determination of the optimal model dimensionality. To sho w the applicability of the kernel algorithm, two examples from multiva riate image analysis are used. The first example is an image from an a irborne scanner of size 9 x 512 x 512. It consists of nine images whic h are regressed against a constructed dependent image to test the accu racy of the kernel algorithm when used on large data structures. The s econd example is a satellite image of size 7 x 512 x 512. Several diff erent regression models are presented together with a comparison of th eir predictive capabilities. The regression models are also used as ex amples for showing the use of cross-validation.