Using Maps to Automate the Classification of Remotely-Sensed Imagery

The accurate classification of remotely-sensed imagery usually requires some form of ground truth data. Maps are potentially a valuable source of ground truth but have several problems (e.g., they are usually outdated, features are generalized, and thematic categories in the map often do not correspond to distinct clusters or segments in the imagery). We describe several methods for using maps to automate the classification of remotely-sensed data, specifically Landsat Thematic Mapper imagery. In each, map data are coregistered to all or a part of the image to be classified. A probability model relating spectral clusters derived from the imagery to thematic categories contained in the map is then estimated. This model is computed globally and adjusted locally based on context. By computing the probability model over a large area (e.g., the full Landsat scene) general relationships between spectral categories and clusters are captured even though there are differences between the image and the map. Then, by adjusting and applying the model locally, new features can be extracted from the image that are not contained in the map and, in certain cases, different classes can be assigned to the same cluster in different parts of the image based on context. Experimental results are presented for several Landsat scenes. Several of the methods produced results that were more accurate than the map. We show that these methods are able to enhance the spatial detail of features contained in the map, identify new features not present in the map, and fill in areas in which map coverage does not exist.