A multivariate numeric image can be seen as a 3-way data table: two dimensions of this table are of spatial nature whereas the other characterizes the constitutive univariate images. The process of labeling consists in assigning a qualitative group to each pixel of the original multivariate image. A supervised learning method, stepwise discriminant analysis (SDA) was compared with two unsupervised methods, simple C-Means clustering (CMC) and Fuzzy C-Means clustering (FCMC). As illustrative example, the methods were applied on multivariate images of sections of maize kernels obtained by fluorescence imaging. CMC requires the utilization of a function assessing the distance between some representative patterns and the pixel vectors. The relative interest of Euclidean distance and Mahalanobis distance was investigated. The best results were obtained by using CMC and simple Euclidean distance. In these conditions, it was possible to identify, with no a priori knowledge, the main tissues of maize (coats, horny endosperm, starchy endosperm and germ).
