STRATEGIES FOR MULTIVARIATE IMAGE REGRESSION

We present multivariate image regression (MIR) as a set of typically problem-dependent strategies for image decomposition guided by the nature of the Y variable and/or training data set delineation in the (X, Y) image domains. Regression techniques common in chemometrics may be applied also to the image regimen (in this paper we treat mainly two-dimensional images). We present applications of both IMPCR and IMPLS-DISCRIM in an effort to delineate the various possibilities for image regression. IMPCR builds directly on our earlier bilinear multivariate image analysis projection approach, while IMPLS-DISCRIM is trained on scene space binary classification masking with subsequent off-screen partial least squares analysis; the results are back-projected as images in the original scene space. Regression may either be carried out for modelling purposes and/or for subsequent prediction purposes. In the image domain this duality is accompanied by several optional training data set delineations in the scene space and/or in the spectral domain. We try to cover as complete a survey as possible of typical, representative regression problem types. We illustrate some of these MIR strategies with an MR-imaging example as well as a simple didactic MIR calibration from analytical chemistry.