On a texture-context feature extraction algorithm for remotely sensed imagery

An image data set of 54 scenes was obtained from 1/8" by 1/8" areas on a set of 1:20,000 scale photography. The scenes which consisted of 6 samples from each of the nine categories scrub, orchard, heavily wooded, urban, suburban, lake, swamp, marsh, and railroad yard was analyzed manually and automatically. For the automatic analysis, a set of features measuring the spatial dependence of the grey tones of neighboring resolution cells was defined. On the basis of these features and a simple decision rule which assumed that the features were independent and uniformly distributed on identification accuracy of 70% was achieved by training of 53 samples and assigning an identification to the 54th sample and repeating the experiment 54 times. This identification accuracy must be compared with the average 81% correct identification which five photointerpreters achieved with the same scenes, although the 81% correct identification is the accuracy achieved when they used the 9" × 9" photograph to interpret from. Note that the photograph is data of considerably higher resolution having much more context information on it than the small digitized 1/8" × 1/8" area the automatic analysis had available.