Hyperspectral Images Segmentation: A Proposal. Proposition d’une Stratégie de Segmentation D’Images Hyperspectrales

Hyper-Spectral Imaging (HIS) also known as chemical or spectroscopic imaging is an emerging technique that combines imaging and spectroscopy to capture both spectral and spatial information from an object. Hyperspectral images are made up of contiguous wavebands in a given spectral band. These images provide information on the chemical make-up profile of objects, thus allowing the differentiation of objects of the same colour but which possess make-up profile. Yet, whatever the application field, most of the methods devoted to HIS processing conduct data analysis without taking into account spatial information.Pixels are processed individually, as an array of spectral data without any spatial structure. Standard classification approaches are thus widely used (k-means, fuzzy-c-means hierarchical classification...). Linear modelling methods such as Partial Least Square analysis (PLS) or non linear approaches like support vector machine (SVM) are also used at different scales (remote sensing or laboratory applications). However, with the development of high resolution sensors, coupled exploitation of spectral and spatial information to process complex images, would appear to be a very relevant approach. However, few methods are proposed in the literature. The most recent approaches can be broadly classified in two main categories. The first ones are related to a direct extension of individual pixel classification methods using just the spectral dimension (k-means, fuzzy-c-means or FCM, Support Vector Machine or SVM). Spatial dimension is integrated as an additionnal classification parameter (Markov fields with local homogeneity constrainst [5], Support Vector Machine or SVM with spectral and spatial kernels combination [2], geometrically guided fuzzy C-means [3]...). The second ones combine the two fields related to each dimension (spectral and spatial), namely chemometric and image analysis. Various strategies have been attempted. The first one is to rely on chemometrics methods (Principal Component Analysis or PCA, Independant Component Analysis or ICA, Curvilinear Component Analysis...) to reduce the spectral dimension and then to apply standard images processingtechnics on the resulting score images i.e. data projection on a subspace. Another approach is to extend the definition of basic image processing operators to this new dimensionality (morphological operators for example [1, 4]). However, the approaches mentioned above tend to favour only one description either directly or indirectly (spectral or spatial). The purpose of this paper is to propose a hyperspectral processing approach that strikes a better balance in the treatment of both kinds of information.

To achieve this, a generic scheme is proposed to associate more closely the spectral and spatial aspects symmetrically and conjunctively. This method, called butterfly, aims to perform an iterative and a cross analysis of data in the spectral and the spatial domains lead to the segmentation of the hyperspectral image. The strategy is based on two steps: 

- Extraction of a spatial structure (topology) incorporating a spectral structure,

- Extraction of a spectral structure (latent variables) incorporating a spatial structure,

These steps are processed in a successive, iterative and inter-dependent way.

In this article, we will focus solely on specific notions of topology i.e. the notions of connectivity and adjacency. Thus, the first stage deals with the use of commonly used image processing tools (region segmentation algorithms) on a limited number of score images. This makes it relatively easy to process. To carry out the second step, we use chemometric tools to reveal a subspace (latent variables) enabling the characterization of heterogeneity of the obtained image partitions. However, the scheme can be applied on two different ways depending on the region segmentation strategy used i.e. top down approaches (splitting) or bottom-up approaches (merging). We have implemented this scheme by using a split and merge strategy based on the quadtree approach. Each phase is done over successive steps (iterations). At each iteration of the split phase, the data are projected on k1 suitable latent variables. The split of each existing region (partition) into four disjoints quadrants is tested and the one maximising the Wilks Lambda parameter is chosen. At each iteration of the merge phase, the data are projected on k2 suitable latent variables and all the merging of two adjacent regions are tested. The one maximising the Wilks Lambda parameter is chosen.

Lastly, we tested our approach on a simple synthetic image to show more precisely its characteristics and also on two real images at different scales (in field acquisition on crop, remote sensing image of urban zone). The results obtained on real images underline the benefit of the butterfly approach. However, futher work will be undertaken to investigate the influence of various parameters. Moreover, other topology notions and image analysis algorithm could be also investigated.

Résumé

Cet article présente une stratégie de segmentation d’images hyperspectrales liant de façon symétrique et conjointe les aspects spectraux et spatiaux. Pour cela, nous proposons de construire des variables latentes permettant de définir un sous-espace représentant au mieux la topologie de l’image. Dans cet article, nous limiterons cette notion de topologie à la seule appartenance aux régions. Pour ce faire, nous utilisons d’une part les notions de l’analyse discriminante (variance intra, inter) et les propriétés des algorithmes de segmentation en région liées à celles-ci. Le principe générique théorique est exposé puis décliné sous la forme d’un exemple d’implémentation optimisé utilisant un algorithme de segmentation en région type split and merge. Les résultats obtenus sur une image de synthèse puis réelle sont exposés et commentés.