A Non-Parametric Classification Strategy for Remotely Sensed Images using Both Spectral and Textural Information

A classification strategy which does not require a priori assumptions about the statistical distribution of training pixels in each class is proposed. This method uses an indicator kriging approach in feature space to classify remotely sensed images incorporating both spectral and textural information of bands. Texture information is used only in cases where spectral information is not sufficient to resolve the assignment of the pixel to a class. Application of the proposed methodology on a remotely sensed natural scene shows an improvement in the overall classification accuracy with respect to the case when the scenes are classified by the traditional supervised Gaussian maximum likelihood classification (GMLC) method using either spectral band only or using both spectral and textural bands. A marked improvement in classification accuracy is obtained particularly for the classes for which the GMLC’s assumption of multivariate normal distribution of training pixels in a class fails miserably.