In hyperspectral imaging, the volumes of data acquired often reach the gigabyte for a single scene observed. Therefore, the analysis of these data complex physical content must go with a preliminary step of dimensionality reduction. Therefore, the analyses of these data of physical content complex go preliminary with a step of dimensionality reduction. This reduction a two objective, the first is to reduce redundancy and the second facilitates post-treatment (extraction, classification and recognition) and therefore the interpretation of the data. Automatic classification is an important step in the process of knowledge extraction from data. It aims to discover the intrinsic structure of a set of objects by forming groups that share similar characteristics. In this thesis, we focus on dimensionality reduction in the context of unsupervised classification of spectral bands. Different approaches exist, such as those based on projection (linear or nonlinear) of high-dimensional data in a representation subspaces or on the techniques of selection of spectral bands exploiting of the criteria of complementarity-redundant information do not allow to preserve the wealth of information provided by this type of data. 1- We have made a comparative study on the stability and similarity of nonparametric algorithms and unsupervised projection methods and also the selection of spectrales bands used in the dimensionality reduction at different noise levels determined. The tests are conducted on hyperspectral images, classifying them into three categories according to their level of performance to preserve the amount of information. 2- We introduced a new approach of criterion based on the dissimilarity of the attributes spectral and used into a local space on matrices of data; the approach was used for to define a rate of safeguarding of a rare event in a given mathematical transformation. However, we limited his application to the context of the thesis related to the reduction of the size of the data in a hyperspectral image. 3- The comparative studies allowed a first hybrid proposal for an approach for the reduction of the size of an image hyperspectral allowing a better stability: BandClustering with Multidimensional Scaling (MDS). Examples are given to show the originality and the relevance of the hybridization (BanClust\MDS) of the analysis carried out. 4- The trend of hybridization was generalized thereafter by presenting an adaptive hybrid algorithm unsupervised based on Fuzzy logic (Fuzzy C-means), a method of projection as the analysis in principal component (PCA) and a validity index of a classification. The classifications carried out by Fuzzy C- means, make it possible to assign each pixel of an image hyperspectral to all the classes with degrees of membership varying between 0 and 1. This property makes the FCM interesting method for the detection of progressive or between different spectral bands or from spectral heterogeneities. With conventional methods known validity indices of classes, we determined the optimal number of classes of FCM and the fuzzy parameter. We show that this hybridization leads to a relevant reduction rate in hyperspectral imaging. Consequently, this algorithm applied to different samples of hyperspectral data, allows a spectral imagery much more informative, in particular on the level of spectral heterogeneity.
