Multispectral image characterization by partial generalized covariance

Abstract. A general method is presented for the assessment of data attribute variability, which plays an important role in initial screening of multiand high-dimensional data sets. Instead of the commonly used second centralized moment, known as variance, the proposed method allows a mathematically rigorous characterization of attribute sensitivity given not only Euclidean distances but partial data comparisons by general similarity measures. Depending on the choice of measure different spectral features get highlighted by attribute assessment, this way creating new image segmentation aspects, as shown in a comparison of Euclidean distance, Pearson correlation and γ-divergence applied to multi-spectral images.