Unsupervised cooperative partitioning approach of hyperspectral images for decision making

Abstract : Hyperspectral and more generally multi-component images are complex images which cannot be successfully partitioned using a single classification method. The existing non-cooperative classification methods, parametric or nonparametric can be categorized into three types: supervised, semi-supervised and unsupervised. Supervised parametric methods require a priori information and also require making hypothesis on the data distribution model. Semi-supervised methods require some a priori knowledge (e.g. number of classes and/or iterations), while unsupervised nonparametric methods do not require any a priori knowledge. In this thesis an unsupervised cooperative and adaptive partitioning system for hyperspectral images is developed, where its originality relies i) on the adaptive nature of the feature extraction ii) on the two-level evaluation and validation process to fuse the results, iii) on not requiring neither training samples nor the number of classes. This system is composed of four modules: The first module, classifies automatically the image pixels into textured and non-textured regions, and then different features of pixels are extracted according to the region types. Texture features are extracted for the pixels belonging to textured regions, and the local mean feature for pixels of non-textured regions. The second module consists of an unsupervised cooperative partitioning of each component, in which pixels of the different region types are classified in parallel via the features extracted previously using optimized versions of Fuzzy C-Means (FCM) and Adaptive Incremental Linde-Buzo-Gray algorithm (AILBG). For each algorithm the number of classes is estimated according to the weighted average dispersion of classes. The third module is the evaluation and conflict management of the intermediate classification results for the same component obtained by the two classifiers. To obtain a final reliable result, a two-level evaluation is used, the first one identifies the pixels classified into the same class by both classifiers and report them directly to the final classification result of one component. In the second level, a genetic algorithm (GA) is used to remove the conflicts between the invalidated remaining pixels. The fourth module is the evaluation and conflict management in the case of a multi-component image. The system handles all the components in parallel; where the above modules are applied on each component independently. The results of the different components are compared, and the adjacent components with highly similar results are grouped within a subset and fused using a GA also. To get the final partitioning result of the multi-component image, the intermediate results of the subsets are evaluated and fused by GA. The system is successfully tested on a large database of synthetic images (mono and multi-component) and also tested on two real applications: classification of invasive plants and pine trees detection.
Complete list of metadatas

Cited literature [129 references]  Display  Hide  Download

Contributor : Abes Star <>
Submitted on : Monday, March 9, 2015 - 5:51:23 AM
Last modification on : Wednesday, September 18, 2019 - 8:53:52 AM
Long-term archiving on : Wednesday, June 10, 2015 - 11:25:12 AM


Version validated by the jury (STAR)


  • HAL Id : tel-01127927, version 1


Akar Taher. Unsupervised cooperative partitioning approach of hyperspectral images for decision making. Signal and Image processing. Université Rennes 1, 2014. English. ⟨NNT : 2014REN1S094⟩. ⟨tel-01127927⟩



Record views


Files downloads