Histogram correlation in the context of land cover classification from multi-polarized SAR data

Land cover classification is a widely used application of remote sensing data. However, some very basic problems occur when the sophisticated approaches recently developed for the classification of optical remote sensing data are applied to SAR data. All problems can be referred to the “strange” statistical behavior of SAR data. For instance, the multiplicative noise contribution complicates the pixel-wise classification as even pixels belonging to the same class may differ greatly. Furthermore, even if the target shapes are given in advance, the standard classification algorithms often fail because a normal distribution of the underlying image data is required which is approximately true for optical data set, but completely wrong with view to SAR data. One way to adapt SAR data to the needs of these classification algorithms is the use of logarithmic intensities or normalized Kennaugh elements with respect to multi-polarized images. Alternatively, different classification algorithms could be utilized that are completely independent of the underlying distribution [1]. In this contribution, a classification approach based on histogram correlation (often used in Computer Vision) is studied. First, the distribution of training classes in terms of Kennaugh elements and Schmittlet indices is focused. Then, the classes are correlated in order to perform a basic separability analysis. Finally, the sample histograms are correlated with the whole image and the class of the best-fitting sample histogram is accepted for each pixel. Visual inspection underlines that this is a very simple, but remarkably sensitive approach. Numerical validation will follow soon