ABSTRACT
Land cover classi†cations based on ensemble classi†ers have been performed successfully during the recent years and seem particularly interesting for multisource and high-dimensional data sets. Contrary to standard classi†ers, which are based on one single classi†er to obtain a decision, ensemble methods to train several classi†ers and combine their results through a voting process. Many ensemble classi†ers have been proposed [1,2] and two strategies exist to construct an ensemble: (1) a combination of different classi†er algorithms and (2) a combination of variants of the same classi†er. Benediktsson and Kanellopoulos [3], for example, used a neural network and a statistical classi†er for combining synthetic aperture radar (SAR) and multispectral data. The two data sets undergo a separate classi†cation and the individual outputs are combined by decision fusion, which can be de†ned as a strategy for combining information from different data sources. Waske and Benediktsson [4] use a decision fusion strategy that is based on support vector machines (SVMs) for the classi†cation of multitemporal SAR data and multispectral imagery.
