Digital Orthophotography - Principles, Project design Issues,Utility, Accuracy, Economics John Michael, Senior Project Manager Intermap Technologies 2 Gurdwara Road, Suite 200 Nepean, Ontario, Canada K2E 1A2 Marek Krolikowski, Digital Photogrammetric Applications | Abstract Of The Paper & The Profile of The Speaker | Speaker Index | Paper Title Index | Standards Digital orthorectified imagery derived from scanned aerial photography has assumed prominence in the marketplace as the preferred "land base" for many AM/FM/GIS applications. This renown has precipitated numerous different sets of specifications but no "Standards" for the preparation of uniform image products. Most agencies contracting the production of orthorectified image sets rely on existing accuracy specifications originally designed for line maps. Others have argued that by virtue of the characteristics of digital imagery, such specifications do not apply. It is our contention that these accuracy specifications are appropriate but require supplementary qualification and discussion of accuracy testing procedures. The United States National Map Accuracy Standards (Office of Management and Budget 1941) and the ASPRS Accuracy Standards for Large-Scale Maps (ASPRS 1990) are the two most influential statements of map accuracy in use today. The ASPRS Standards are the more rigorous of the two and contain complete definitions and testing procedures. The National Map Accuracy Standards stipulate that 90% of all well defined features not be displaced in excess of 1/30 inch for scales larger than 1:20,000 and 1/50th of an inch for smaller scales. (A well defined feature is one that can be plotted on the map within 1/100 inch.) The ASPRS Standards establish limiting root mean square errors in X and Y for populations of check points. Given that these two Standards represent the only reasonable approaches to quantifying map accuracy, digital orthophotography must conform to either. Both of these Standards have been formulated to apply to mapping at discrete scales. The argument that digital image maps are independent of scale hinges on the fact that scale is simply a measure of representation (usually in hardcopy). Digital image maps may be viewed or output at any scale, therefore scale is immaterial to the data. However, imagery should be described as a given resolution prepared in accordance with mapping practices of a given scale. For example: orthophoto imagery comprised of 50 centimeter pixels derived from 1:30,000 scale photo and compiled to 1:10,000 scale mapping standards. The pixel size must be such that it will support measurements necessary to verify the desired accuracy. The existing Standards both apply to the cumulative result of all error associated with the processes of map production (photography, survey control, aerial triangulation and map compilation). Mapping specifications usually dictate parameters and techniques for the production of interim products. These specifications are consistent to the extent that when followed using accepted professional practice, they will result in a map product which meets or exceeds the overall Standard. To this point in time, certain of the component processes in digital orthophoto production have not been defined, and the results quantified and verified, to the point of general acceptance. That is to say that the processes have not been subjected to a rigorous validation procedure. All of the same parameters which apply to photogrammetrically-compiled line maps apply to digital orthophoto. However, there are additional parameters unique to digital orthophoto. These considerations include the scanner, scanning resolution, resampling algorithms, and the digital elevation model (DEM). The Digital Elevation Model The DEM itself poses a problem in terms of specification since it must be designed to adequately characterize a surface. Such design cannot be quantified without a measure of surface roughness. A quantification of surface roughness is only possible following analysis of a DEM. Therefore, DEMs are specified based on empirical evidence of their adequacy for certain applications. The DEM is preprocessed before being input to the rectification procedure. In most cases a rectangular grid of elevation points is generated or a TIN is created based on a reduced number of points. This approach greatly simplifies and expedites the estimation of elevation for each pixel which is used in the rectification process. The interpolation procedure does however, degrade and smooth the DEM. The vertical accuracy of each point has a direct impact on the horizontal accuracy of the rectification. The impact of the Z value accuracy varies with the distance from the principal point of the exposure being rectified. As one moves away from nadir (the principal point of the photo), the effect in X and Y of a height error increases. DEMs are sometimes specified based on their appropriateness for generating contours of a given interval. The contour interval, in turn, implies certain accuracy e.g. "ninety percent of all contours and elevations of points interpolated from contours shall be accurate within one-half of the contour interval". The relationship between DEM characteristics and contour interval is dependent upon the terrain. Once again, we are faced with attempting to quantify the characteristics of a DEM without a measure of topographic roughness. The DEM characteristics include: Compilation strategy Grid Profiles Progressive sampling Random Density Interval between profiles Sampling interval along profiles Breakline Capture Type Hydrographic Transportation Physical Spot Heights Attributes Hard Soft Measurement Static measurement Dynamic measurement Infill by interpolation Stereocorrelation Density Editing strategy Source Photography Scale Camera focal length The number of permutations of these individual characteristics is virtually unlimited. The characteristics of the DEM should change with the terrain. Therefore within a single map sheet or tile, the density of points and their relative disposition should be variable. It is quite possible to produce a digital orthophoto which meets accuracy criteria in general, and passes the stipulated tests, but still contains localized anomalies. The most common occurrence of such problems is in areas adjacent to abrupt changes in elevation such as cliffs, retaining walls, bridges, embankments etc. Elevated highways, flyovers and bridges frequently cause problems in large scale orthophoto products. The problems manifest themselves as small "smears" or distortions in the rectified images. A severe example of such a distortion is shown in Figure 1 below. Figure 1. Localized Distortion There is considerable diversity of opinion regarding the DEM requirements for the purposes of orthorectification of aerial photography. There is agreement that the DEM need not be as dense or as accurate as that required to support 3D modeling and contour generation. There is not agreement as to a formula for relating the two types of DEMs. There is very little published material describing rectification results based on differing DEMs. The dearth of information arises from the high cost associated with the experiments and particularly, the high cost associated with field verification. We undertook to test the accuracy of digital orthophotos which were prepared using identical inputs with the exception of the DEMs. The test site was a 4 km by 4 km area in central Massachusetts. The terrain varied considerably in the area as did the ground cover. The range of relief was approximately 150m. The aerial photography was 1:30,000 scale scanned at 12.5 (m. The rectifications resulted in imagery with a ground sample distance of 50 cm. 190 check points were established and pugged on the diapositive. The points were uniformly distributed over the area. We calculated RMSE values for the XY vector between the coordinates measured on an analytical plotter and those derived from the rectified image. The RMSE values were calculated separately for the points falling in the flatter areas and for the rougher areas within the image. Figures 2 depict the test image and a portion the level 1 DEM. Figure 2. Test Image   Our hypothesis held that there is a direct relationship between the density of the DEM and the accuracy of the resultant image. We tested DEMs compiled on nominal scan line spacings of 45m, 90m, 135m and 300m. We also compiled breaklines commensurate with each of these grid densities. The accuracy of the check points in the resultant images ranged from 1m to 3m RMS. While our results support the hypothesis, the relatively small difference between 45m and 300m scan line spacing in the DEM was surprising. We expected a much larger range of error. The result of the test led us to question the validity of the experimental design. We concluded that the sample size and the spatial distribution of check points was such that it would not bias the results. Similarly, the means used to establish the check points was unrelated to terrain. We concluded that the breaklines were largely responsible for the results because they provided a strong characterization of the surface. The orthorectification process required the DEM to be in the form of a regular grid. The density of the points along the breaklines was high relative to the density of points along scan lines. When the DEM was interpolated to a regular grid, the high density of breakline points in conjunction with the interpolation algorithm yielded a good terrain model. Therefore, the adequacy of a terrain model for rectification is not simply a case of density of "mass points" or spacing between "scan lines". It is very much a case of quality breaklines. The nominal spacing of mass points can be quite coarse so long as the breaklines provide good characterization of the surface. Large Scale Distortions Large scale rectifications in urban areas result in interesting problems which are not addressed solely by diligent collection of breaklines. In these circumstances the resolution of the imagery is usually very high. The pixel size is very small relative to the magnitude of change in Z which occurs at retaining walls, bridges, flyovers, tunnel entrances etc. This relationship can precipitate severe image smearing. If the rectification process is dependent solely on a regular grid of elevations the resultant image will contain bizarre distortions. These distortions can be minimized by the collection of breaklines and gridding the DEM to a very high density. The distortions will not however, be eliminated. There are several strategies employed to mitigate the effects of this problem. The first is the collection of three dimensional polygons which model the elevated structure, for example a bridge deck or bridge ramp. These polygons are orthorectified and the feature appears in the correct position in the output image. This process necessarily involves overwriting pixels in the original image and creating "voids" where no image data exists in the output image. These voids can sometimes be filled with imagery from adjacent images. From a practical perspective, this strategy is difficult and costly to implement. The second strategy involves the collection of breaklines such that the elevated structure is rectified to its correct position and then the interactive removal of the distortions which arise around the structure. An example of the implementation of this strategy is shown in Figure 4 below: Figure 4-1 Example of distortions adjacent to bridge deck.   Figure 4-2 Distortions eliminated by image editing.   Scanning and Resampling Considerations The scanning resolution influences the precision and accuracy of measurements which can be made on the resultant image. The scanning process is analogous to a resampling process in that it generalizes the intensity of light transmitted through a diapositive over a given area. The area is dependent upon the aperture setting on the scanner or the focus of light on the CCDs within the array. The area is not necessarily identical in size and shape to the output pixel. Where the CCD photo-sites are separated in space aliasing results. The generalization of intensity over the area of the pixel produces some fascinating effects. Small features which are in high contrast to the background will be emphasized. An example of this phenomenon is shown in Figure 5 below. Figure 5. Lines on tennis court visible on 50 cm pixels.   There is error inherent in the scanning process and the subsequent image formation. Scanners rely on high precision mechanical components or a reseau to "calibrate" the digital image. The aerial camera calibration information is used in an affine transformation to remove systematic error introduced by the scanner. For the purposes of creating digital orthophotography, the scanner does not have to produce images of geometric accuracy commensurate with the final product accuracy. It has been shown that the resolution of standard aerial film is approximately 27 lp/mm. One resolution element is thus approximately 37 (m/lp. For true maintenance of photographic resolution content a scanning resolution of approximately 15 (m is necessary (Light 1993). Scanning at a resolution higher than 15 (m will not yield additional information nor will it enhance interpretability of the image. Scanning resolutions of 20 to 30 (m are common practice since they achieve good results while providing economy in terms of file size. The rectification process is comprised in part of a digital resampling procedure. This resampling relies on an assessment of the intensity values of a group of pixels (from one to sixteen) to assign a value to a new pixel. The procedure creates a blank array of pixels representing a fixed size on the ground and systematically assigns intensity values to the array. The resampling process analyzes the surrounding (input) pixels regardless of their size. Therefore it is important that the pixel dimensions of the input image be equal to or smaller than those in the output image. In practice, similar pixel sizes in the input and output images will produce the best results. Bilinear or cubic convolution resampling algorithms are normally employed in orthorectification software packages. The cubic convolution algorithm is acknowledged to produce superior results and it is sometimes stipulated in orthophoto specifications. In practice, given input and output pixels of similar dimension, it is difficult to discern any difference between the two outputs. Digital orthophoto imagery derived from aerial photography does not support automated classification the same way multispectral satellite data does. Illumination effects are such that meaningful classification cannot be achieved, particularly using black and white photography. Therefore, the importance of the resampling is limited to preserving spatial resolution and contrast. Orthophoto for Map Revision Orthophoto is an ideal tool for assessing the completeness and correctness of vector data. Overlaying vectors on imagery immediately draws one's attention to areas of change, errors of omission and geometric inconsistencies. In practice however, the use of digital orthophoto for monoscopic line map revision is not as simple an exercise as one would think. There are three issues which arise: mismatches between existing vectors and the image; displacement of vertical structures; and the lack of stereo viewing as an aid in photo interpretation. Each of these problems causes the operator to question the validity of data, prompts analysis and necessitates decisions. When the existing vectors do not fit the image data exactly, one must decide whether to fix the existing vectors or adjust the new information compiled from the image. This decision can result in significant spatial inconsistencies over the area. It is very difficult for an operator to leave inconsistencies unfixed. Displacement of vertical structures can result in the operator missing information which is obscured by 'building lean". In some cases the operator must interpret the footprint of the building from the shape of the roofline, which may or may not be valid. The lack of a stereo view can also make interpretation more difficult. In some cases it can be difficult to differentiate a flat roof from an adjacent area of asphalt for example. When one is faced with inconsistencies between existing vectors and a new image the first question is always the provenance of the data sets and the specifications governing the inputs and processes used to create both. If map revision was the objective of the orthophoto it should have been prepared in accordance with specifications resulting in an accuracy exceeding that of the original vector maps. Each of these problems is minimized through operator experience. Nevertheless, the efficiency of the revision process is not as high as one would expect. Conclusions The same techniques and practices applied to line mapping are appropriate to image mapping. The specifications simply must be elaborated to describe component processes and products. The key issues which must be addressed within the project specifications are: Photo scale, orientation and coverage (forward overlap and sidelap), Film type, Control density and accuracy, Aerial triangulation adjustment algorithm, and accuracy, Scanner and scanning resolution, DEM character, density, accuracy, and format, Output pixel dimensions, Output image accuracy, Output image accuracy verification, Output image radiometric characteristics, Image mosaicking strategy, Output file format and media, Project documentation and reporting. Each of these parameters implies technique and must be fully consistent with all other, related parameters. The tendency in designing an orthophoto project is to err on the side of caution by using conservative estimates of the parameters describing the component parts of the project. The most common manifestation of this tendency is in respect of the DEM. The elevation information typically exceeds the requirement of the rectification process. Assuming identical inputs and techniques, orthophoto accuracies are equivalent or superior to those achieved in line mapping. Verification of this accuracy represents a problem owing to the resolution limitations inherent in a digital image. As one magnifies the digital image to the point where individual pixels are evident, feature context can be lost. It can be difficult to identify the individual pixel which represents a "well defined" feature. Further, distance measurements may only be made to a precision of ( 2 pixels. Therefore, where specific accuracies must be verified, a suitable pixel size must be specified. In spite of the limitations and uncertainties associated with digital orthophoto it is destined to change the way we think of maps and the way we undertake geographic analysis. Digital imagery is now an integral part of GIS and one that is growing in importance. Bibliography John Michael is a mapping specialist with twenty years of hands-on experience in all phases of photogrammetric mapping programs from design to data analysis. He holds a B.Sc. in geography with a specialization in surveys and mapping and an M.B.A. with a concentration in international business. John has lived and worked in Canada, Nigeria, Egypt, Tanzania, Nepal and Jordan. Each of these assignments involved the implementation of a major mapping program. He has also worked on short-term assignments in a number of other countries as a consultant to the World Bank, national and state governments, international aid agencies and private sector corporations. His consulting assignments have included automation design for production systems, environmental GIS programs, land and real estate registration systems, and the development of mapping standards and specifications for GIS program implementation. Marek Krolikowski is a photogrammetric mapping specialist with over ten years of experience in photogrammetric digital mapping and GIS. He holds an M.Sc.Eng. degree in photogrammetry, as well as a diploma in photogrammetric operations. Marek has undertaken much development work for processing of vector and raster data, and integration and fine-tuning of photogrammetric data capture systems. He has managed and provided training on several projects in Mexico, Poland and Russia. References The Manual of Photogrammetry, American Society of Photogrammetry, Fourth Edition, 1980. Engineering and Design Photogrammetric Mapping, Department of the Army, U.S. Army Corps of Engineers, Engineer Manual 1110-1-1000, 1993. Standards for Digital Orthophotos, National Mapping Program Technical Instructions, U.S. Department of the Interior, U.S. Geological Survey, 1993. Specifications for Aerial Survey Photography, Interdepartmental Committee on Air Surveys, Energy, Mines and Resources Canada, 1982. National Aerial Photography Program, U.S. Department of the Interior, U.S. Geological Survey, 1993. Report to the Specifications and Standards Committee, American Society for Photogrammetry 1982. The National Aerial Photography Program as a Geographic Information System Resource, Donald L. Light, U.S. Geological Survey, Photogrammetric Engineering and Remote Sensing Vol. 59, July 1993. Film Cameras or Digital Sensors? The Challenge Ahead for Aerial Imaging, Donald L. Light, U.S. Geological Survey, ACSM/ASPRS Annual Meeting, Reno, Nevada. 1994. Digital Orthophoto Production Issues, Robert C. Shanks, Hammon, Jensen, Wallen & Associates Inc. ASPRS/MAPPS Mapping and Remote Sensing Tools for the 21st Century Conference, Washington, D.C. 1994. Creating Digital Orthophotos Requires Careful Consideration of Project Design Elements, John Michael, INTERA Information Technologies Corp., Earth Observation Magazine, February 1994. Computer Processing of Remotely-Sensed Images An Introduction, Paul M. Mather, John Wiley & Sons 1987. Geographic Information Systems, A Management Perspective, S. Aronoff, WDL Publishers 1989. Figures 1, 4-1 and 4-2 courtesy of The Boston Water and Sewer Commission, Boston, Massachusetts. Figures 2, 3 and 5 courtesy of The Commonwealth of Massachusetts, Executive Office of Environmental Affairs, Massachusetts. | Abstract Of The Paper & The Profile of The Speaker | Speaker Index | Paper Title Index | CGIS HOME PAGE CONTENTS