Résumé
Due to the complexity of the tropical terrestrial environment present in Pacific islands and the lack of ground data, remote sensing could offer an appropriate tool for obtaining a better understanding and knowledge of the key parameters necessary to many environmental applications. Moreover, recent sensors provide high spatial resolution and good temporal periodicity which is suitable for the study of tropical environments. The potentiality of an oriented-object technique for land-cover mapping will be illustrated in this paper. Unlike traditional pixel-based classification, this technique is based on object-use topology and shape features for the differentiation of target classes. It offers a complex "knowledge base" about classes which can be directly formulated in classification rule sets. The first step applied to images is "segmentation" which enables the improvement of classification accuracy as compared with that achievable using only individual spectral signature pixels. In fact, indices based on spectral, spatial and textural or structural parameters are explored in order to reduce the confusion between classes. The results from the segmentation are then used to produce a classification of objects. The oriented-object classification technique is carried out on a section of Efate Island (Vanuatu republic) using images acquired in 2007/2008 by Formosat-2 sensor. Finally, the accuracy of the oriented-object classification is established with the help of ground control points.