Land Use Classification using Structural Features
After a thorough survey on land use classification, we begin with a set of measures based on straight lines for Ikonos images in this chapter. Subsequent detailed analyses (counting and classifying dwellings, for example) can then be confined to developed areas in the following chapters. Straight line structures will be more prevalent and more organized in developed areas than in wilderness or rural areas. However, for our measures we only need this assumption to hold locally. Four of our most promising measures (based on length and contrast) do not depend heavily on this assumption. On the other hand, our remaining measures (orientation, line spacing, and periodicity) depend on this assumption heavily. As expected, this later group could not perform as well as the length and contrast measures experimentally. Our objective at this stage is the (rough) classification of the image into regions of little or no development (wilderness or rural) and developed regions (urban or residential). We applied Bayes, Parzen window, and nearest neighbor (NN) classifiers to label each image region. Initially, we defined a two-class problem to discriminate “urban” and “not urban” regions and obtained excellent results (roughly 87% correct classification). Although there has been extensive work on land use classification, no structural approaches to this problem have been reported. Our approach, being totally based on straight lines, offers the first such solution, to our knowledge. This approach shows very promising results in extensive testing over a wide variety of land development patterns.