Abstract
The Support Vector Machines (SVM) are increasingly far-reaching in remotely sensed data classification. As supervised classifiers, the SVM output depends on the input pixels, pointing out that training is potentially an important stage for optimizing classification accuracy. The SVM consist in projecting pixels into a high-dimensional feature space and then fitting in a hyperplane, maximizing the distance between the closest vectors and the hyperplane itself. This study aims to locate the pixels acting as support vectors and identify the ecological features they contain in a tropical vegetation context. The analyses focused on a Quickbird-2 image where two vegetation types occur. The physical boundary between classes was delineated in the field, and we used an iterative method to mark and localize the pixels acting as support vectors on the image. Our results highlight that vegetation sampling should focus on ecotones (the transition area between two different and adjacent vegetation classes) in order to minimize the field survey effort and maximize the mapping accuracy.
Acknowledgements
We are grateful to Cédric Lardeux (SAPHIR Team, University of Rennes, France) for implementing the LIBSVM library on IDL and the Urbanism Department of the Government of French Polynesia for providing the Quickbird-2 data.