Using images as a data source
for a GIS: Two different approaches Michael Simmons Vice President International Tydac Technologies Inc. Bangkok, Thailand It has unfortunately become conventional wisdom in GIS that digital maps are expressed only as vectors. If this limiting would, prior to use, either have to be converted to vector format or be used only to amend an existing vector map file. Fortunately for those GIS users who use image data, approaches are available which are not restricted to vector formats. These new methods open some very efficient and effective possibilities for the use of images in GIS. General approach to image use by GIS: The image is raw data; that is before processing it has neither positional information nor does it convey information about the earth's surface or sub-surface. These tasks are performed by an operator with the help of an image analysis system. The resulting processed image is geometrically corrected, positioned accurately on the Earth's surface with reference to a specified projection, and through the use of appropriate classification techniques the data may convey information on vegetation cover, geology, land use, water depth, etc. These classified, geometrically corrected images are maps. Less and less is the photographic classified image product seen as an end in itself. These photographic products are now used primarily for visual interpretation of images prior to digital input as map data to a GIS. In areas in which map data is scarce and even in areas in which rapid change is occurring but which are otherwise well mapped, these images may be a most significant source of map data. In extreme cases imagery may provide the only source of map data. For maximum benefit images should always be used with conventional maps, usually produced from aerial photographs and ground surveys. These maps are converted to digital format through vector based digitising techniques. Data may already have been digitized for another purpose, and need to be transferred from a GIS different from that selected for the project under consideration. To summarize: image analysis has come to assume a role of data input, and sometimes-primary data input, to GIS. The traditional photographic image products are now of secondary importance to digital image products. Geo-positioning of images has become of crucial importance, otherwise the classified images can not be co-registered with other digital map products in a GIS. Conventional vector map to Image Integration The representation of maps as vectors and the representation of images as rasters have presented difficulties for those interested in integration of maps and images. An obvious first step is to super impose vector line files over an image. This, of course will only have meaning if there is a relationship between the two data sets. A justification often made for use of imagery is to amend older maps with more current information. Examples cited include new residential streets and changes in forest vegetation. A new map is made by adding new vectors to the previous map file and perhaps by also adding new entries to the associated database. The approaches developed to achieve this objective of map amendment are of two main types Image analysis systems have added "bit" mapping functions which allow a raster "map' to be constructed by the operator using a mouse to trace a one pixel wide white line around the boundary of regtons on the image. GIS, on the other hand, have used the image as a backdrop or reference over which the operator can trace a vector, also using a mouse. More recent improvements permit previously constructed vector files to be displayed and the vectors "dragged" to new positions reflecting the data displayed on the image. The result of both approaches is creation of a new or amended vector file, which can be introduced to a GIS as a new map. Limitations of the vector approach The vector modification approach works well with simple images. However consider typical classified image which purports to show, for example, land use. There are likely to be several thousand polygons even if as few as only five land use classes are identified. Using the vector approach each polygon must be examined separately and the operator must make a decision on boundary definition for each area independently. Nor can the class of each polygon be automatically assigned. The operator must subsequently edit the vector file to add polygon identification and create a database describing each polygon. This can become an extraordinarily tedious process, which frequently takes longer than conventional cartographic methods. The problems of the vector approach are:
Consider the possibilities if the user is free of the necessity to amena maps created in vector format. The main objective is conversion of data obtained from imagery into useful maps. A raster based GIS allows direct use of raster images as maps without the intermediate step of creation of vector maps. Without the intermediate step of creation of vector maps. A GIS based on a raster data model gives users a previously unavailable flexibility in integrating imagery and all other digital maps. First the geometrically corrected, classified image is imported to the GIS as a map. The cell size selected for the raster map should if possible be identical to that of the pixels of which the image is composed. Image classes can be converted directly to map legend items. As classified images usually represent only a single theme this approach is sufficient, but if necessary a database can also be appended containing multiple fields of data for each class. All existing digital maps of the study area can also be contained in a library of raster maps. It is then possible to directly compare, using overly techniques, the image and any of the maps in this library. Second, if the objective is amendment of an older for example land use map with more current information contained in an image it is possible to use GIS functions to automate this procedure. A two map, or matrix, overlay can be used to identify all locations in which the classes on the existing map and the image are different. Some changes are logically impossible and these can be identified so that the existing map class prevails over the image class. Some images still contain noisy data even after filtering; these areas can be eliminated by specifying size thresholds below which areas will not be recognized. The remaining actual change areas, which will be used to amend the land use map, can then be preserved as a new and separate map and\or integrated with the old map. Third, if vector output is required, the raster maps can be converted, with or without an associated database, as a last step, to vector format. Quadtrees for more efficient integration of images and maps SPANS is a raster based GIS, in which the implementation uses the quadtree data model. The quadtree approach is a data reduction technique which has reduced the size of raster image files by up to ONE HUNDRED times, and typically will reduce raster files of thematic maps by 3-10 times. The quadtree data structure allows the ease and speed of multiple maps overly associated with the raster approach. But it avoids the main disadvantage of rasters which is the very large files created when high resolution is required. At a Quadtree level of 15, images of any size up to in excess of 32,000 by 32,000 can be converted. Any pixel size can be specified and this dimension can be set to the same size can be specified and this dimension can be set to the same size as the smallest quadtree cell. Any number of classes can ocur on the image, also to a maximum of 32000. Images processed on any image analysis system can be read by SPANS. Ideally users would process images on any image analysis system which they prefer and then imaport a geometrically corrected, classified image maps into the GIS. The approach does require as a minimum that imagery has been corretly positioned on the Earth's surface and is expressed in a specified projection. Identification of training areas is crucial to accureate interpretation of imagery. In SPANS the operator can use various map analysis functions or screen drawing techniques to isolate training areas from existing maps. These areas can be transferred to an image analysis system as either vector or raster files. The advantage of using a GIS for this step is that the precise nature of conditions on the ground is known. SPANS is able to freely convert between vector, raster and quadtree file types. By allowing the operator this freedom, map images can be properly and fully utilized as data input to GIS. Spatial analysis functions can be used to assist in the interpretation of the map image; the need for manual intervention is minimised and full advantage is taken of the digital products available in both raster and vector modes Summary Vector and raster data structures in GIS have produced significantly different approaches to the use of images, perhaps much more than is commonly realised. The image, once geometrically corrected, is a primary geographic data source, Image analysis systems are also required to classify, and position the image, the two essential steps in the conversion of an image to a map. Vector based GIS, which are widely used for cartographic purposes, use the image map as a back drop or visual reference for amendment of digital vector maps. Raster based GIS offer much greater versatility in the use of image maps. The classified image can be directly imported as a map for use with other digital maps of the same area. The Quaadtree implementation of raster GIS produces any advantages particularly in file size reduction, very fast operating times for multiple map overlay operations and almost instantaneous data base queries. For cartographic production rasters can be transformed to vector format. |