GIS data is traditionally collected from digitized maps and other hardcopy spatial and non-spatial sources. Imagery from aerial photographs or satellites is an alternative data source that is more accurate, more up-to-date, and often more affordable. Software packages that extract GIS data from imagery are called photogrammetric packages. Photogrammetric packages, such as SOCET SET, use sophisticated algorithms to extract terrain data (i.e. digital terrain model) and vector (i.e. feature) data from imagery. Much of the data can be collected automatically from imagery, without human intervention. Imagery can be used to create or revise a GIS database. Imagery generally yields much higher accuracy than alternative GIS data sources. Imagery can be collected by flying an airplane over a region, or from satellites such as IKONOS which collect 1 meter resolution imagery of the Earth. GIS data collected from imagery can be three-dimensional, meaning every vertex has a unique Z elevation value. Photogrammetric packages, such as SOCET SET, are available on desktop PC platforms, and are closely integrated with GIS systems, such as Esri. SOCET SET can directly edit GIS data in SDE^TM format without the need to perform data conversions. Using imagery to create or update GIS data yields accurate and up-to-date GIS databases.

The largest cost of managing a GIS database is creating and updating the GIS data. The simple process of converting data from hardcopy format to digital forma accounts for more than 50% of the cost of GIS operations (Stevenson, 1995). Manually generating vector data is error prone and contains inherent inaccuracies. Using imagery, rather than hardcopy maps, helps GIS operations in several key ways:

• Cost
• Accuracy
• Currency (up-to-date)

Costs are reduced because collection from imagery is mostly automatic, and many processes that formerly required a human operator, such as terrain generation, can now be performed by an unattended computer. Accuracy is increased because image-based data is derived from photographs of the actual ground, rather than paper map that can be of dubious origin. Currency is improved because imagery can be collected upon demand, and the data extracted from it can be used to create or revise a GIS database.

All these benefits come from using computer-based photogrammetric systems, such as SOCET SET. Analytical stereo plotters, available for about 25 years, use imagery, but still require manual extraction and significant data conversion efforts. Recent breakthroughs in the field of photogrammetry now offer a computer-based approach that eliminates many conversion steps, as illustrated in figure 1.

When photogrammetry is used, data generation and conversion costs can be reduced by an order of magnitude. For example, the SOCET SET software product can generate 1 million terrain data points per hour. Conversion costs are minimized, because SOCET SET can directly edit the vector data in SDE databases without the need for any file format conversions. Accuracy is ensured, because photogrammetry employs sophisticated algorithms that compute the precise ground position (latitude, longitude, and elevation) of each pixel in the image.

The typical workflow for creating GIS data from imagery is shown in figure 2.

Imagery can come from many sources. Commercial satellites, based on classified government spy technology, are now available for use by the GIS community. IKONOS, QuickBird, and OrbView are three satellites that can or will collect 1-meter resolution imagery. However, the most common source of imagery is still airplanes. Digital cameras, such as those from TerraSource and promised by LH Systems, are becoming more popular, but the vast majority of imagery is film, usually 9" square negatives. Aerial surveying firms, readily located in the Yellow Pages, can be hired to fly the plane, take photographs, and scan the film into digital format.

The image sources supported by SOCET SET are listed in Table 1.

The sensor models use stereoscopic principles, similar to the binocular vision of two-eyed animals, to precisely compute the ground locations of objects in the imagery. Generally, the accuracy of data derived from the imagery is around 0.5 pixels or better. For example, if the image resolution is 6 inches per pixel, then the accuracy can be 3 inches.

Because stereoscopic algorithms are used in photogrammetry, two images are required for all regions that are being processed. If there is only a single image, it is still possible to generate GIS data, but without obtaining the full benefits of photogrammetry, because the 3-dimensional characteristics of the ground cannot be extracted.

After imagery is acquired, the next step is to register the images together. The purpose of registration is to compute very high accuracy camera locations. Registration is optional, because most image sources provide an approximate camera location with the image. However, to maximize accuracy, the camera locations must be estimated even more accurately.

The registration process uses "tie points" (see Figure 4) and control points to estimate the camera locations and orientations more accurate. Tie points are objects in the imagery - single pixels, such as the corner of a building – that are used to connect the imagery. Tie points improve relative (inter-image) accuracy. Control points are surveyed points on the ground (such as benchmarks or survey points) that have a known latitude, longitude, and elevation. Control points are used by the registration process to improve absolute accuracy.

After the imagery is registered, the next step is to generate terrain data. If one is only extracting vector (feature) data from imagery, it is not necessary to generate terrain, but it is still recommended because it can be used during vector extraction to set the elevation values of the vector vertices automatically.

The terrain generation process is fully automated, and can be run without an operator. The process is called Automatic Terrain Extraction (ATE) and is the most sophisticated process in the photogrammetric system. The operator must specify whether the terrain data will be in grid (raster) format or triangular (TIN) format. Grid format is more interchangeable, but triangular format can model the terrain with better fidelity.

ATE uses an adaptive image correlation algorithm to match corresponding points in the stereoscopic images. Image correlation has been used in tens of thousands of data sets and has proved to be a reliable and accurate algorithm. A successful implementation of image correlation largely depends on a set of "correct" parameters which control the image correlation algorithm. The set of parameters should change adaptively according to terrain type, signal power, flying height, X and Y parallax, and image noise level. ATE uses an inference engine to generate the set of image correlation parameters. AATE uses multiple images and multiple bands, and selects the best images for image correlation. ATE compensates for residual Y parallax and performs image epipolar resampling on the fly. ATE also uses an image noise filter to reduce image noise. ATE works well on large and small scale images.

The inference engine generates a set of image correlation parameters by a set of rules based on a set of facts. The set of facts are derived from terrain type, signal power, flying height, X and Y parallax. The set of rules is static and can be applied to any digital photogrammetry project. As ATE progresses, new and more accurate facts are derived and a more suitable set of parameters for image correlation are generated. Different regions with different terrain type, signal power, and X and Y parallax may have different sets of parameters.

ATE uses multiple images and selects the best images for image correlation. Image relief distortions and obscurities due to relief are two of the major problems for image correlation. In regions covered by more than two images, ATE can exploit images which "see" the terrain the best. When the best image pair has not been epipolar resampled, ATE resamples it on the fly. When multiple bands of an image are available such as color images, ATE uses multiple bands instead of a single band for image correlation.

Image noise can be problematic especially with film images scanned at high resolution. This is due to several factors including film grain noise, original image quality, scratches, scanner sensor noise, etc. When working at high resolution, the ATE process uses a noise filter when there is significant image noise.

The most important GIS data that most people extract from imagery is feature data, also called vector data. The SOCET SET photogrammetric system includes a software tool to generate features interactively from the imagery. At first glance, the user interface for feature extraction appears similar to the user interfaces of CAD/CAM or mapping tools, but on closer inspection, there are some significant differences:

• Three-dimensional data extraction
• Computer vision technologies
• Stereoscopic visualization
• Attributes from imagery
• Direct editing of SDE databases
• Automatic terrain draping.

Three-dimensional vectors are vectors with a unique Z elevation value for each vertex. This is one of the key benefits of using imagery, because the stereoscopic characteristics of multiple image views enable the elevation to be computed for every object in the imagery.

Computer vision technologies in the photogrammetric software find edges in the imagery, and convert the edges into feature delineation. This speeds up the process of extracting features immensely.

Stereoscopic visualization is one way of visualizing a GIS database in three-dimensions, allowing the operator to validate the accuracy of the data and update GIS databases. Two images are displayed simultaneously on a CRT, and special polarized glasses are used to provide depth perception. Vector graphics are overlaid on the imagery, again in a stereoscopic manner, and can be visualized in relation to the imagery. Any errors in the XYZ ground coordinates of the vector data will be apparent because the graphical overlays will appear to be above or below the imagery. This is a very fast and efficient way to validate the accuracy and currency of a large GIS database. Called "superimposition", this facility also excels in database update, when features on the imagery and not in the database, are readily identified.

Attributes are facts about a feature, such as its size, purpose, age, condition, parcel number, material, history, or address. In many GIS applications, attributes are just as important as the geometry (location and shape) of the feature. Attributes suffer from the same accuracy and currency problems as the geometric data. Using imagery can ensure that attributes are accurate and up-to-date. SOCET SET automatically computes size attributes (area and length), and other attributes such as condition, material, and purpose can be readily ascertained by simply looking at the imagery. Revisions to a GIS database, for example, to review building sizes for property tax purposes, can be performed quickly and easily with imagery and a photogrammetric software package.

If a terrain database is available, the feature extraction process can automatically use the terrain data to assign an elevation value to every vertex of features that are on the ground.

SOCET SET 3D feature extraction can interface with either the internal 3D topological feature database or Esri’s SDE, which then interfaces with other relational database management systems. The internal 3D feature database is derived from the Vector Product Format (VPF) optimized for efficient feature extraction and supports topology. Features in the internal feature database are always sorted by their ID numbers, which allows efficient data access based on feature ID numbers. There is a built-in spatial index quadtree which supports efficient data access through spatial search.

A special tool is available within SOCET SET 3D feature extraction to perform polygon topology for Arc/Info polygon coverage type of GIS data extraction. This tool is designed to extract polygonal features with shared edges efficiently. It performs "operator assisted" topology checking on-the-fly in 2.5D space. It is "user assisted" because users must follow certain rules to delineate features. The rules cover six cases: (1)isolated simple polygon; (2)splitting polygon; (3)attach to one existing polygon; (4)attach to multiple existing polygon; (5)separated multiple edges; (6)inside polygon; and (7)outside polygon. As long as users are performing delineation that is one of the above seven cases, the polygon topology tool will perform topology checking and insure topological consistency in the 3D topological feature database. Operators do not need to do anything extra other than following the rules. This tool makes this type of GIS data extraction from imagery very efficient.

SOCET SET 3D feature extraction provides a set of fence tools to manipulate and query features, including (1)polygon clipping, which can click features using a polygon; (2)trim/extend, which solves dangling nodes problems; (3)linear feature intersection and stacked nodes generation, which are mainly used for linear features such as streets and highways; (4)database manipulation such as copying, deleting, moving, grouping, updating attributes on a list of features selected either by attribute query or by fence polygon selection.

SOCET SET 3D feature extraction provides special tools for texture patch, which is mainly used for 3D visualization. Other special tools include generic features and model placement. This tool, similar to Arc/Info 8.0 division, can extract large number of generic feature such as street lamps efficiently. 3D models are used for volumetric feature extraction such as building. Dimensional attributes can be computed on-the-fly. There are more special tools such as Volume Create etc. All tools are designed and optimized for efficient 3D GIS feature extraction from imagery.

SOCET SET is tightly integrated with Esri’s products such as Arc/Info, ArcView, and SDE.

SOCET SET 3D feature extraction is designed to work with different database management systems such as SDE or SOCET SET’s 3D topological feature database system. Feature databases become more valuable and useful when they are integrated with non-spatial data. SDE provides an interface to integrate feature databases with non-spatial data, and puts feature databases directly under database management system control. The 3D feature extraction with the SDE link takes full advantages of SDE for real time applications. With this system, users can extract 3D features directly from controlled imagery. This system allows multiple operators to extract 3D features to the same feature database, update the feature database while it is being used for real time applications, and has no limits on the feature database size. 3D features are tightly integrated with non-GIS data to maximize their value and application.

The best file format for transferring data between systems is Shapefile. SOCET SET supports 3D shapefile export and 2D and 3D shapefile imports. A shapefile stores non-topological geometry data and attribute data of a feature in a data set. SOCET SET imports 2D and 3D shapefiles and writes them into a SOCET SET 3D feature database. For 2D shapefiles, SOCET SET will automatically compute the Z coordinates of all vertices from a terrain database. Shapefile export and import not only converts geometric data, but also attribute data.

Terrain data can be shared between SOCET SET and Esri products with the ArcGrid format, which is an ASCII file storing grid terrain data.

The authors wish to acknowledge their sincere appreciation to Dr. Stewart Walker for his numerous comments and editing.

Aerial photographs taken from airplanes or satellites - such as the IKONOS satellite - can be used to create, update, and revise GIS databases. Photogrammetric software, such as SOCET SET, is required to extract vector and terrain data from the imagery in a format suitable for GIS. When 2 or more images of the same region are available, stereo capabilities of the photogrammetric tools can generate three-dimensional data sets, where every vertex has a unique Z elevation value. Data extracted from imagery is exceptionally accurate compared to alternatives, such as map digitizing or digitizing from orthophotos. Modern photogrammetric packages automate many processes, such as terrain generation and image registration, enabling high-throughput production lines. SOCET SET directly edits, without the need of translation, SDE vector databases.

Marr, D. (1982). Vision. W. H. Freeman and Company, San Francisco, California, 1982.

Ohta, Y. and T. Kanade. (1985). Stereo by intra- and inter-scanline search using dynamic programming. IEEE Transactions on Pattern Analysis and Machine Intelligence. Vol. PAMI-7, No. 2, pp. 139-154.

Okutomi, M. and T. Kanade. (1992). A locally adaptive window for signal matching. International Journal of Computer Vision. Vol. 7, No. 2, pp. 143-162.

Olander, N. (1998). Modeling Sensors in Software. Proceedings of ISPRS Commission II Symposium.

Stevenson., P.J. (1995). The problem of data conversion. Geo Info Systems, February, pp. 28-32.

Zhang, B. and S. Miller. (1997). Adaptive Automatic Terrain Extraction. Proceedings of SPIE, Volume 3072, Integrating Photogrammetric Techniques with Scene Analysis and Machine Vision (edited by D. M. McKeown, J. C. McGlone and O. Jamet). pp. 27-36.

Zhang, B., N.F. Olander, F.C. Paderes, S. Miller, and A.S. Walker. (1998). Automatic TIN Generation from Imagery. Proceedings of ASPRS-RTI annual conference.