ABSTRACT
The most common techniques of satellite image processing, viz., principal component and parallelepiped analyses and ascendant hierarchic, mobile centres and maximum likelihood methods of classification, etc. are all based on textural analysis oi images. The le Voisinage et l'Organisation des Informations Spatialisees par Informatique et Numerisation model enables us to delineate, through the neighbourhood parameter, boundaries of spatial units whose heterogeneity is defined by a pattern identified by a composition vector. In almost all cases, a classification is obtained in which each class comprises several map units that are not compact. However, if the window is too large, smaller objects will disappear and will be integrated with a more complex reference zone. The composition vector is a textural measure giving frequency distribution of various classes in a given neighbourhood. Nuclei of the next iteration are redefined from the statistics of pixels grouped in each nucleus. More nuclei can be added in each iteration and then a new iteration started.
