A comparative study of multispectral image classifiers: Applications to classification problems with high-dimensional data and high-overlapping spectral signatures. Simulation of high-dimensional images.

摘要: Classi cation of high-dimensional images is of the almost interest in Remote Sensing applications. Storage space, and specially the computational e ort required for classifying this kind of images are the main drawbacks in practice. Moreover, it is well known that a number of spectral classi ers may not be useful-even not valid-in practice when the training sets are high-overlapping in the representation space.In this paper, we compare a large number of spectral classi ers for classifying high-dimensional images. We selected a number of di erent classi ers, both parametric and nonparametric classi ers, in order to made an in depth comparative study. We also propose the use of di erent spectral classi cations as initial classi cations to be used by a contextual classi er (ICM, in this case) in order to obtain some interesting combinations of spectral-contextual classiers for Remote Sensing image classi cation with an acceptable trade-o between the accuracy of the nal classi cation and the computational e ort required. We shall use two synthetical image databases, consisting of high-dimensional images, to test the performance of the classi ers. These datasets have been created from scratch by using a procedure proposed by the authors.