Computer-based Landsat image interpretation has neglected the spatial organization of the image in favour of the spectral and temporal organization. A brief survey of techniques that exploit spatial information, including multistage sampling, is given. Semantically-guided region-merging methods have been used successfully but they require sophisticated and expensive list processing facilities. Similar semantic and spatial sensitivity can be introduced by exploiting a pyramidal, hierarchical representation of the image advocated by Kelly, Tanimoto and Levine. The image pyramid is constructed bottom-up with the original image as the base. Each level is a reduced resolution version of the level below, constructed by averaging the signatures of adjacent pixels at the lower level. By classifying pixels at the higher levels one is efficiently classifying semantically uniform regions in the original image. If, however, a region's signature lies in the spectral overlap of two or more classes its subregions will have to be considered for classification. Several refinements of this technique, including the use of semantically-based region splitting and merging techniques at each level of the pyramid, are described.
These techniques are used to classify forest cover types on Vancouver Island in a Landsat image. The results of several initial experiments indicate that, compared to a baseline of a traditional supervised maximum-likelihood classifier, the cost of maintaining the pyramid is balanced by the vast reduction in the number of pixel classifications. The spatial homogeneity or readability of the segmented image, as measured by the number of regions, is improved by a factor of three while the accuracy of the classification is unaffected or slightly improved. When the region splitting and merging techniques are applied at each level of the imaqe pyramid the accuracy and the readability of the final segmentation both increase markedly. It is thereby demonstrated that these pyramidal techniques offer many of the advantages of the semantically-driven region-merging approach in a more flexible and efficient fashion. Indeed the two approaches have been combined to achieve substantial benefits for Landsat image interpretation.