Expert Systems Help Extract Information from New Geospatial Data
By Lawrie E. Jordan, III

What a difference a decade makes. Just 10 years ago, the fledgling geotechnologies industry was scrambling to find new sources of spatial data to satisfy the growing needs of geographic information systems (GIS). Consequently, we spent much of the '90s designing advanced imaging satellites, refining digitization techniques, and creating new digital data sets.
      As a result of these efforts, the industry now faces almost the opposite challenge. Spatial data users can now log onto the Internet and obtain a staggering array of off-the-shelf satellite images, digital elevation models (DEMs), pre-packaged vector files, and other geographic data-all in digital formats.
      As the variety of spatial data has increased, so has the richness of the information they offer. Specifically, the one-meter imagery acquired by Space Imaging's (Thornton, Colo.) recently launched IKONOS satellite, and those planned by Orbital Imaging Corp. (Dulles, Va.) and EarthWatch Inc. (Longmont, Colo.), contain a degree of geographic intelligence and detail never before available to us.
      While many people feel it may not be possible to have too much spatial data, the task of sorting through and extracting meaningful information from this multitude of raster and vector information is a serious challenge. In fact, adequately exploiting high-resolution imagery and the other new data sets may be our primary challenge in the next decade.
      Fortunately, expert systems technology is available to handle this immense volume and breadth of geospatial data. When applied to the geospatial environment, expert system technology can process numerous content-rich datasets to perform complex land-use/land-cover analysis and express the results as a thematic raster image map.
      In recent years, expert systems have been used extensively in diagnostic medicine and many other academic and information technology applications. In 1995, ERDAS Inc. began research with support from NASA to incorporate expert system technology into commercial image processing software for the purpose of better exploitation of geospatial data, especially high-resolution satellite imagery.
     Today, expert system technology is a featured component of the recently released ERDAS IMAGINE 8.4 geographic imaging software. The new capability, called the IMAGINE Expert Classifier, is fully integrated into the package, working with and complementing the spatial operators traditionally used to extract meaningful information from raw data.

Introducing Expert Systems
An expert system is a computer application that solves a specific problem or makes a decision based on a series of rules, conditions or hypotheses defined by an expert in a given field. This set of rules is the knowledge base that is often called a decision tree, because schematically it resembles a tree in which the rules are presented as questions whose answers send the analysis down new branches containing additional queries.
      The basic idea behind expert systems is to mimic the human ability to combine knowledge with reason and make conclusions from inferences. In geotechnical applications, this takes automated image processing a huge leap beyond traditional statistical pixel analysis and into the realm of interpreting images based on the context and spatial relationships of features.
      "Many factors can be taken into account beyond spectral values-size, shape, shadow, tone, pattern-all the things we key on in visual interpretation," says Greg Koeln, vice president of Environmental & GIS Services at Earth Satellite Corp. in Rockville, Md., an ERDAS IMAGINE 8.4 beta site.
      Because the expert system can generate conclusions based on so many different types of information, ERDAS designed the IMAGINE Expert Classifier to accept inputs of multiple data sets, such as DEMs, land cover maps, soil layers and GIS vectors, as well as satellite images.
      Using urban classification as an example, the IMAGINE Expert Classifier might prompt a telecommunications planner to input an IKONOS image and building height DEM to differentiate certain roofing materials and roadway surface that appear spectrally similar. The system would identify a linear feature with that spectral value as a road because of its shape. Two other similar features with geometric shapes might be distinguished as a building rooftop and a parking lot based on the DEM information.
      In addition, the user can weight certain variables with confidence values that can be altered during the analysis.
      "Confidence values let us make one data set more important," says Jay Kost, a USGS scientist and IMAGINE Expert Classifier beta tester working on the National Elevation Database program at EROS Data Center. "For instance, soil data is often the most important in land cover mapping, while elevation is sometimes key to differentiating tree species in forestry."
      Following the format of other expert systems, the IMAGINE Expert Classifier is actually two programs in one. The first enables the expert to create the knowledge base by building the decision tree with a simple drag-and-drop graphical user interface. This tree diagram contains the sequence of rules, conditions and variables the expert would consider during a manual analysis of the same data.
      The second part of the program is the Wizard interface, which allows non-experts to apply the knowledge base to their own data. It prompts the users to enter required data sets, automatically applies IMAGINE processing functions as necessary, and takes the analysis through the decision tree to a conclusion.

Putting the Expert System to Work
ERDAS distributed beta copies of ERDAS IMAGINE 8.4 for testing in a variety of disciplines-forestry, mapping, geology, environment, urban planning and many others. EarthSat immediately put the IMAGINE Expert Classifier to work in solving a land-cover/land-use issue it often faces in mapping projects for telecommunications clients.
      EarthSat typically uses traditional statistical classifiers to identify land cover in Landsat Thematic Mapper (TM) images and was looking for a practical method of subdividing certain classes. Specifically, the company wanted to differentiate high- and low-density urban areas.
      "Urban areas are relatively easy to identify spectrally in a Landsat image," says EarthSat's Koeln, "but spectral signature alone can't distinguish density."
      Koeln knew that image texture held the answer to his query. Texture is essentially the frequency of spectral change within a small region, and urban areas tend to have high degrees of texture because so many features-houses, yards, roads and parks-result in different spectral responses in a relatively small area. Higher urban density yields greater image texture.
      They set up the IMAGINE Expert Classifier to require only the Landsat image as an input. Next, they created a statistical model with ERDAS IMAGINE's Model Maker, a unique graphical tool for modeling spatial data, which calculated texture throughout the image in 7x7 pixel matrices. In the expert system, spectral thresholds were established to isolate urban areas.
      Based on the combination of spectral and texture data, the IMAGINE Expert Classifier successfully differentiated urban density. Setting up the rule base took only a couple of hours.
      "We gave the texture values a higher confidence than the spectral data in determining results," said Koeln. "Greater texture and higher spectral response was high-density urban, and lower texture and lower spectral was categorized as low density."
      Koeln expects the expert system technology to offer important advantages when applied to the new one-meter satellite imagery because so many small ground features will affect, and possibly confuse, spectral signatures. In such cases, the ability to analyze texture, shape, size and many other characteristics besides spectral response will be a necessity in accurately classifying land use and land cover.
      From that respect, many beta testers commented that the IMAGINE Expert Classifier is really much more than an image classifier, at least in the traditional sense. It is actually a decision-support tool that can analyze diverse spatial data sets, none of which must be a satellite image.
      A practical example of this is a knowledge base ERDAS created for one of its military clients during the testing of the IMAGINE Expert Classifier. The objective was to build a decision system that would enable military commanders to determine whether specific vehicles were capable of crossing a given terrain.
      The knowledge base was created with performance specifications and tolerances of various Army vehicles, such as the maximum slope they can climb and the surface conditions needed for proper traction. With inputs of soil type maps, DEMs and current weather information for a specific area, the IMAGINE Expert Classifier produces a map of traversable terrain for each vehicle.
      Another capability in the IMAGINE Expert Classifier, called the Pathway Cursor, has received high marks from clients. It allows the user to select any pixel in the output image and view the exact path taken by the decision tree to arrive at that classification.
     "The cursor is a useful tool in helping the user go back and see what rules were used in the classification," said USGS's Jay Kost. "This allows the user to refine the knowledge base."
     The pathway cursor also helps users learn from the knowledge base and take even greater advantage of the expertise incorporated into the system by the expert who created it. The knowledge base can be passed along to thousands of other non-expert users who can reliably repeat the procedure and perform an extremely complex level of spatial analysis. This portability of knowledge has led ERDAS to believe that expert system technology is a key component of the future of geographic image processing.

Back