Edge detection and processing of remotely sensed digital images

The traditional edge extraction process has five basic stages: smoothing, edge detection, thresholding, thinning and linking. All of these stages require different algorithms to carry out their functions, which have typically required artificial limits or constraints set by “heuristics”. This research has designed and implemented an automated technique for edge extraction that has consistent logic linking the various stages of detection and formation, in which artificial limits have been avoided. A wide range of filters and detectors were evaluated in both the spatial and frequency domains, as well as operators using both rate-of-change and orientation criteria. For smoothing, a minimum-variance filter produced the most accurate and reliable results. The best method of edge detection was identified as one of the derivative filters. For thinning and linking, the definition of an edge was most accurate if the orientation of the edge was the prime consideration. These methods produce results which efficiently extract edges for both boundaries and linear features, are relatively tolerant of noise in the data, and utilize a logical set of rules that can be adapted as the data dictate.