MULTI-CLASSIFIEURS DES IMAGES SATELLITAIRES.

Abstract

In remote sensing, clustering, also called unsupervised classification, is an important task that aims to partition a given image in a multispectral space into a number of spectral classes (clusters), when in situ information is not available. Among the many existing clustering algorithms, the most commonly used are K-means, ISODATA, FCM (Fuzzy C-Means), SOM (Self Organizing Map) and more recently K-Harmonic Means. However, with the increase in the amount of remotely sensed data and its heterogeneity, it becomes difficult to obtain relevant clustering results using a single clustering algorithm. Moreover, each algorithm mentioned above requires a number of parameters and the most important of them is the number of clusters, which the user has to define a priori. To cope with these shortcomings, the Multiple Classifier System (MCS) is also known as ensemble clustering , is the consensus of different clustering algorithms can provide the best partition with high accuracy and consequently overcome limitations of traditional approaches based on single classifiers. The MCS involves two stages : the partitions generation and the partitions combination. In this thesis, we investigate the potential advantages of this technique in the unsupervised land cover classification by using various kinds of data : Synthetic data, composite data and remotely sensed data. The first stage of the MCS is assumed by four clustering algorithms, the well-known k-means algorithm, the k-harmonic means algorithm (KHM), Bisecting K-means (BKM) and the self-organizing map (SOM). The best clustering is obtained according to WB index. The relabeling and the voting methods are used in the second stage. Experimental results obtained by the MCS outperform the results of the individual clustering.