Chapter 4: Clustering and Classification

Title: Partial unmixing of multi or hyperspectral images using ICA and fuzzy clustering techniques: Application to vegetation mapping on vineyards

Author(s): Saeid Homayouni, Jean-Pierre Da Costa, Christian Germain, Olivier Lavialle, Gilbert Grenier

Address: LAPS, UMR 5131 CNRS, Universit Bordeaux 1 – ENSEIRB – ENITAB, 351 Cours de la Libration, F33405 Talence cedex, France | LAPS, UMR 5131 CNRS, Universit Bordeaux 1 – ENSEIRB – ENITAB, 351 Cours de la Libration, F33405 Talence cedex, France | LAPS, UMR 5131 CNRS, Universit Bordeaux 1 – ENSEIRB – ENITAB, 351 Cours de la Libration, F33405 Talence cedex, France | LAPS, UMR 5131 CNRS, Universit Bordeaux 1 – ENSEIRB – ENITAB, 351 Cours de la Libration, F33405 Talence cedex, France | LAPS, UMR 5131 CNRS, Universit Bordeaux 1 – ENSEIRB – ENITAB, 351 Cours de la Libration, F33405 Talence cedex, France

Reference: Atlantic Europe Conference on Remote Imaging and Spectroscopy pp. 115 - 122

Abstract/Summary: The context of this paper is agricultural remote sensing imagery. Its objective is to propose a blind framework for the partial unmixing of multi or hyperspectral images. This framework relies on a linear mixing model with physical constraints related to the positivity and the dependence of the mixing coefficients. It hinges on three steps. The first step is a non-constrained source separation operation using Independent Component Analysis. Then, a vine-related source is selected by comparing the various sources to a rough vegetation index map. Finally, a fuzzy clustering algorithm is applied to the vine-like source to provide a soft classification of pixels into vine and non-vine pixels, i.e., a vine abundance map. The entire framework is successfully applied to hyperspectral CASI data acquired on vineyards. Results are provided which allows discussing the relevance and the capabilities of the approach.

Order a copy of this article