Scale and texture in digital image classification
Abstract
This thesis is a theoretical and empirical study on textural properties of digital images. Spatial information exists at a hierarchy of scales and texture is a consequence of the objects in that hierarchy. Within-class texture results from the spatial arrangement of objects at the next finer level in the hierarchy than the informational class. Between-class texture results from spectral differences between adjacent classes and is most obvious near class edges, especially for smooth classes. In rough classes between-class variance may not differ much from within-class variance. Errors in classifications using texture, therefore, are most likely associated with, class edges; however, investigators often avoid edges in evaluating texture or classification.;The window sizes needed to produce a stable texture measure are often large. Experiments with ADAR 1-meter data suggest that windows of 50 to 300 meters are necessary. Small windows are required to minimize edge effects. This is inherently contradictory as windows used to produce stable texture measures also produce a large edge effect.;Experiments with simulated data showed that separability of classes increased when texture was used in addition to spectral information. Separability of texture also improved with larger scale windows. This improvement was over-estimated when pixels were chosen away from class edges. The ADAR data showed that separability of the interiors of classes improved with the addition of texture, but for the class as a whole, the class separability actually fell. Maximum Likelihood classification of the ADAR data demonstrated the effect of edges and multiple scales in reducing the accuracy of classification incorporating texture.