Using neural networks to locate edges and linear features in satellite images

Abstract

Edges and linear features are manifestations of discontinuities. In geologic applications of satellite imagery, edges and linear features are used to identify faults, lineaments, or lithology changes. Two techniques for detecting these types of features in satellite imagery are skilled human interpretation and mathematical manipulation (e.g. Fourier analysis). The former approach has the advantage of being able to learn, combined with a high degree of fault tolerance. Disadvantages to this approach are that it is labor intensive (slow) and somewhat arbitrary in its decision process. The latter approach is purely mathematical and, when implemented on a computer, is fast and consistent. The disadvantages of the mathematical approach are the lack of both fault tolerance and learning ability. A third approach, applied here, is to use neural networks (NNs). NNs combine the speed and accuracy of computers with the fault tolerance of human beings. NNs are arrays of highly interconnected, simple processing units which excel at pattern recognition. Because of this ability, they are appropriate tools for recognizing edges and linear features in two-dimensional (2-D) data, such as satellite imagery, aeromagnetic, and gravity data. Image interpretation is comprised of two phases. The syntactic phase identifies image primitives such as lines and edges, and the semantic phase resolves the meaning of the lines and edges. This study focused on the first aspect of locating syntactical information in an image. We determined that NNs are of only limited use for detecting linear features, but they are capable of detecting edges at various scales.