Quantifying Qualitative Features

Many features of interest in remote sensing imagery, such as roads, rivers, clouds, trees, and buildings can have high spectral, structural, and textural variability due to variations in reflectance, resolution, intrinsic shape, etc. Nevertheless they have distinctive qualitative properties of their own from the point of human perception. For instance, clouds are typically fluffy or wispy, roads have uniform widths, rivers are rarely straight, and buildings are rectilinear. The efficient quantification of such qualitative structural signatures is important for automatically recognizing and labeling features in imagery. In this paper we demonstrate the value of constrained Delaunay triangulations (CDT) of discretely sampled shape contours for obtaining quantifiers of qualitative characteristics of certain features. These quantifiers are efficient to compute, and fairly robust to partial occlusions, resolution limitations, and noise. This has applications to the automated analysis and understanding of airborne and terrestrial imagery in classifying structures such as, clouds, forests, rivers, cities, etc.