Lattice auto-associative memories induced multivariate morphology for hyperspectral image spectral-spatial classification

Abstract

The simultaneous use of spatial and spectral information for the classification-based analysis of hyperspectral images improves classification results and image segmentation quality. In this paper pixel spectra are individually classified by conventional support vector machines (SVM). The result of an innovative watershed transformation is used to postprocess the SVM result, imposing the homogeneity of the watershed region, i.e. classification disagreements inside the watershed region are solved by majority voting. This paper introduces several approaches to define reduced supervised orderings based on the recall distance of lattice auto-associative memories (LAAM). The automatic unsupervised selection of the foreground/background training sets from the hyperspectral image data is performed by the use of endmember induction algorithms (EIA). The proposed approach is compared with a recent state-of-the-art spectral-spatial approach.