Threshold Selection for Line Detection Algorithms

Local operators are used to detect linear features in Landsat data for an area where such features are typically less than 1 pixel in width. A modification to existing techniques is made which gives improved results. The choice of threshold for the detector algorithm is investigated by considering line and background intensity, linewidths, and line alignment with respect to the pixel boundaries and spatial autocorrelation in the data set. It is found that results will always contain a significant amount of noise. Two postprocessing techniques are described which reduce this noise by utilizing contextual information in the data.