Selection of optimal textural features for maximum likelihood image classification.

摘要

The classification or segmentation of images into land cover or object class is one of the fundamental remote sensing/image processing tasks. The classification techniques have reached a high level of maturity in the forms of statistical pattern recognition techniques applied to multispectral images. According to this approach each pixel has a spectral vector associated with it and pixels are segmented into the class they most closely resemble spectrally. While these techniques are reasonably successful they do not take advantage of the brightness patterns within a class or at object boundaries nor are they readily applicable to monochrome images or highly correlated multispectral images (e.g., true color images). To overcome these limitations several investigators have suggested the use of image derived features as additional factors for use in image classification. Robert (1989) has identified over forty textural features that have been suggested by various authors as useful for scene classification. Regrettably it is very compute intensive to generate all these features for every pixel in an image and to then use them in a classifier. Schott et al. (1988) developed a technique for selecting a small subset of spectral bands from a large set based on criteria intended to optimize maximum likelihood classification. Salvaggio et al. (1990) have implemented code to generate 46 image derived textural features and applied a two-step feature reduction and optimization process based largely on the band selection process of Schott et al. (1988). The current effort drew on this proof-of-concept work mentioned above and improved on several limitations noted in the earlier work. Specifically, the initial feature set reduction algorithm of Robert (1989) is shown to be deficient and an improved method implemented. 18 refs., 62 figs., 13 tabs.