Multispectral edge detection and its application to feature classification, image matching and geographic information systems (GIS).

摘要

The advent of high spatial resolution multispectral and hyperspectral overlapping imagery comes with a need for effective manipulation algorithms for all the bands of data. The two main questions that are tackled in this research are first, on what basis should bands be selected from a hyperspectral cube and second, how can multispectral imagery be used for image matching. The first question is akin to asking which bands contain salient information for the scene being viewed. For this question, a band selection method is proposed that distinguishes scene phenomena based on their spectral response patterns. The successful separation of two scene phenomena is characterized by an edge in one or more bands. Further, a multispectral edge detection method is proposed for image matching and phenomena separation. This method generates edges that are assigned a color which is dependent on the bands participating in the edge formation. By using the scheme on overlapping imagery an image matching strategy is illustrated showing how the edge data in a predefined search space is dramatically reduced by using color based matching. For example, a red edge on one image is matched only with a red edge in the second overlap image. The use of orientation information further reduces the number of potential match candidates. The a priori knowledge of the bands generating an edge between two phenomena aids in topologically tagging abutting areas, allowing for direct entry into a geographic information system (GIS). The entry into the GIS allows for spatial queries, both on single image features as well as on overlapping image features. Also, color edge grouping is illustrated as a method for identifying interesting edge color regions for relational or multiple image matching. This multispectral edge detection method has potential for real time mapping applications, given the scene phenomena response patterns with previously selected bands to distinguish physical ground phenomena.