Chapter 4: Clustering and Classification

Title: Optimisation of image parameters for landscape cartography with data from multi/hyperspectral images and spectral library

Author(s): Francois Tavin, Audrey Roman, Pierre Gouton, Sandrine Mathieu

Address: LE2I, UMR-CNRS 5158, University of Burgundy, Aile Sciences de l|Ingenieur BP 47870, 21078 DIJON, France | LE2I, UMR-CNRS 5158, University of Burgundy, Aile Sciences de l|Ingenieur BP 47870, 21078 DIJON, France | LE2I, UMR-CNRS 5158, University of Burgundy, Aile Sciences de l|Ingenieur BP 47870, 21078 DIJON, France | Alcatel Alenia Space, BUOS/PG/I 100 Bd du Midi, BP 99 06156 Cannes la Bocca Cedex, France

Reference: Atlantic Europe Conference on Remote Imaging and Spectroscopy pp. 110 - 114

Abstract/Summary: Classification of multi/hyperspectral remote sensing images is a wide field of investigation. Many articles discussed methods, supervised or not, in order to obtain reliable classification for given images in several domains. Our approach takes place before the classification process itself. Our goal is then to estimate the influence of image parameters (spatial, spectral, radiometrical resolutions, SNR) on classification processing. Obviously the results can be highly correlated with the application domain. We choose to focus on the geosciences thematic, for which the classification is still an issue. Three supervised classification methods have been selected: Spectral Angle Mapper, Support Vector Machines and Artificial Neural Networks. A huge library of spectra measured on the ground have been gathered in order to proceed our supervised classifications. Using several images (real or simulated by degradation) of the same place at different resolutions, we analyse the results of our three classifications versus ground truth to estimate the influence of image parameters on these three classification methods. This article aims to expose our methodology.

Order a copy of this article