Hyperspectral Image Segmentation through Evolved Cellular Automata

Abstract

Segmenting multidimensional images, especially hyperspectral ones, remains a challenging topic due to two primary issues. First, most existing methods fail to maintain the multidimensional nature of the data throughout the segmentation process. They typically reduce the hyperspectral information to a two-dimensional representation early on, which results in significant loss of the rich spectral information these images provide. Second, the lack of sufficient and reliable ground truth data makes it difficult to effectively train and fine-tune segmentation and classification algorithms. This paper introduces a novel approach that addresses these challenges by implementing a combined two-step process for segmenting and classifying regions in multidimensional images. The initial step utilizes cellular automata (CA) and their emergent behavior to generate homogeneous regions within the hyperspectral cube, preserving the data’s multidimensional properties. The subsequent step involves using a support vector machine (SVM) to assign labels to these regions for classification purposes.

While the application of cellular automata in hyperspectral image segmentation is not new, previous methods often involve manually designing the rules for the automata, which typically leads to averaging out the spectral information. This paper’s key contribution is the use of evolutionary techniques to automatically generate CA rule sets that yield optimal segmentation results across various scenarios, without any form of dimensionality reduction until the final stage. Moreover, the proposed evolutionary process for generating the CA rules can be carried out using RGB images, allowing the resulting automata to be effectively applied to hyperspectral data. This approach circumvents the issue of limited labeled ground truth images. The method has been validated on both synthetic and real hyperspectral images, achieving highly competitive results compared to other existing methods.