Jerry W. Wicks, Ph.D.
President , Senecio Software Inc.
&
Professor Emeritus of Sociology, Bowling Green State University

David A. Swanson, Ph.D.
Senior Social Scientist, Science Applications International Corporation
&
Dean, Helsinki School of Economics and Business Administration
International Center - Mikkeli

Robert K. Vincent, Ph.D.
Professor, Department of Geology, Bowling Green State University
&
CEO, GeoSpectra Corporation

José Luiz Pereira De Almeida
Vice-President, Software Engineering, Senecio Software Inc.

Abstract

The notion of using remotely sensed data in population work has existed for well over a half-century. In the past two decades, a growing number of research efforts have systematically explored the feasibility of using aerial and satellite imagery to generate population estimates. To-date, these efforts have met with limited success. The recent convergence of three technological developments has created a unique opportunity for advancement in the area. This paper discusses these developments, while outlining plans for the design and testing of procedures capable of producing current population estimates from remotely sensed data.

____________________________
Paper presented at the Canadian Population Society meetings, June 9-11, 1999, Bishop's University and University of Sherbrook, Lennoxville, Quebec.

A demographer's fantasy of flying over an area obtaining immediate and accurate population estimates is tantalizingly played out time and again in movies and on television by Hollywood writers and producers. Box office hits such as the 1992 movie Patriot Games, and the 1998 hit, Enemy of the State, suggest that certain U.S. government agencies controlling private satellite systems containing high resolution optics, cameras and computerized tracking systems are capable of incredible feats such as reading automobile license plates, locating and recognizing individuals, and tracking fast moving objects in the air or on the ground. These technological stunts are always portrayed as being carried out in real time and, of course, usually from a secret site in Washington , D.C.

While they are not counting people, the technology hinted at in these movies clearly suggests its feasibility. Illustrations more germane to the work of demographers has been played out over the decades in a variety of television shows, in particular Star Trek. How often have we heard either the second-in-command or the ship's computer inform the captain that, "onboard sensors show 1 billion inhabitants residing in fifteen cities located across the planet's surface." Or, "except for one dwelling containing two life forms, there are no other habitable structures on the planet." As fanciful as these images are, they work their way into our thinking, becoming a part of our popular culture and dreams. Eventually they work their way into scientific disciplines. At some point we begin asking ourselves, naively perhaps, "Why can't we obtain population estimates using remotely sensed data?"

The notion of using remotely sensed data to identify housing structures for population work has existed for well over a half-century in the professional literature. (Carls, 1947) The U.S. Census Bureau, for example, explored the feasibility of using aerial photography in an effort to locate human habitats in remote rural areas in order to reduce the decennial census undercount. Urban sociologist and geographers have also long recognized the spacial nature of cities. The notion that remote sensing can measure spatial attributes of the urban landscape which can then be digitized and analyzed using GIS systems is commonly accepted. (Cowen and Jensen, 1998) And over the last two decades a number of research efforts have explored the feasibility of using remotely sensed data to obtain population estimates. (Iisaka and Hegedus, 1982; Lo, 1995; Lo and Chan, 1980)

While these efforts have shown promise, they have all lacked the necessary precision to be useful in the commercial sector. There are three reasons these efforts have met with only limited success: 1) most have lacked high resolution imagery, 2) critical pixel-level information (i.e., height of the z-axis) has been unavailable, and 3) sophisticated pattern recognition software has only recently come on the market. All three of these limits have now been lifted as a result of soon-to-be-available high resolution stereoscopic satellite imagery, the availability of automatic, every-pixel, elevation extraction software, and sophisticated image processing, pattern recognition software capable of combining multi-spectral data with elevation information in the grouping of digitally distinct pixels into the classification of spatial attributes of the urban landscape.

Satellite Remote Sensing from 1972-2000 and the Coming of High Resolution Stereo Imagery

The era of civilian remote sensing from satellites began with the launch of LANDSAT I (then called ERTS-1 for Earth Resources Technology Satellite) in July, 1972, which bore the MSS (MultiSpectral Scanner) sensor into orbit with 4 spectral bands (half in the visible and half in the reflective infrared wavelength region) and a spatial resolution of 80-meters. For the first time from space, geologists were able to map a limited number of compositional differences in surface soils and rocks, such as ferric oxides as a class of minerals (Vincent, 1972). Ten years later, in 1982, the LANDSAT TM (Thematic Mapper) sensor was orbited with 7 spectral bands (6 with 30-meter resolution in the visible/reflective infrared wavelength regions and 1 with 120-meter resolution in the thermal infrared wavelength region), producing significantly better capabilities in compositional mapping. In the mid-80's, the French SPOT satellite was launched with 4 spectral bands and the first digital stereo capability (imaging the same area on the ground from two different observation angles), at spatial resolutions of 10-meters and 20-meters. The significance of stereo data is that height or elevation information can be extracted from digitized stereo images. In the 1990's there have been additional satellites launched by Japan, Europe, Russia, and India, some with spatial resolutions ranging from 3-meters to 5-meters.

This year and next (i.e., 1999-2000), there will be four advanced technology civilian satellites orbited that will significantly broaden the number of applications for satellite remote sensing. Two of them (one from Space Technology Development Corp. with 20-meter spatial resolution and the other from Orbital Sciences with 8-meter resolution) will be hyperspectral sensors with an excess of 200 spectral bands each. These sensors will greatly aid any problem, such as geological mapping, that depends on unique chemical compositional differences among mappable units. Although multispectral information is useful for discrimination of buildings from trees or other chlorophyll-rich objects, two spectral bands alone (red and reflective infrared) are sufficient for that task. Though helpful, hyperspectral sensors do not appear to be as important as other types of sensors for mapping buildings.

The other two satellites, to be orbited this year, will offer stereo viewing with 1-meter spatial resolution, but with modest numbers of spectral bands (5 or less). These sensors, one called IKONOS by Space Imaging Inc. (the IKONOS first launch attempt on April 27, 1999 failed) and the other called OrbView 3 by Orbital Sciences, Inc. (Montesano, 1997), will greatly aid any problem for which height discrimination at high spatial resolution is necessary.

The remote sensing task of greatest importance to the estimation of population density is the mapping of the three-dimensional boundaries of buildings. Volume, which can be obtained by multiplying a building's height times its area coverage, is one of the most important characterizations of a building. A parking lot, for example, may have the same asphalt covering and the same area as a building, but height easily discriminates between building and parking lot. The greatest difference between a warehouse and an office building is usually its height-to-area ratio. The square footage of living space in a family home is proportional to its volume (height times area covered). Some have argued (Jensen, 1995; Warner et al., 1996) that a spatial resolution of 0.3-0.5meters is best for the acquisition of stereoscopic image data from which building perimeter, area, volume, and height can be extracted. On the one hand, this argument is correct if either visualization of the buildings or survey-type positional accuracy is required. On the other hand, it may not be necessary since a 3 to 5 band multispectral sensor with 1-meter resolution is likely sufficient to discriminate buildings from everything else and to estimate building volume well enough to classify buildings as homes, warehouses, office buildings, malls, etc. In short, 1-meter stereo images from a satellite sensor with 3 to 5 spectral bands may provide more than adequate information from which to estimate population.

Why is it important to have a satellite sensor, rather than an airborne sensor, to perform a remote-sensing task that has worldwide applications? First, aerial photography is not allowed to be exported from some countries, yet satellite images fall under the U.S. Open Skies Policy and can be collected anywhere. Second, a 1-meter satellite sensor usually covers a wider field of view than an aerial photo, meaning that satellite data are usually less expensive than aerial photos per square mile, even when the spatial resolution is the same in both cases. For instance, 1:40,000 scale NAPP photography collected by the USGS can provide 1-meter data from a scanner that digitized the photos with a 25 micrometer spot size for a scanning cost of approximately $200, plus a collection cost of approximately $1,500 dollars by an airplane capable of flying at 20,000 ft altitudes. Each photo covers about 8 km x 8 km. A 1-meter-resolution satellite image from IKONOS or OrbView3 will cover a larger area per image (at least 10 km x 10 km), cost less per square km, and be available on a more regular basis than aerial photography.

Automatic Extraction of Elevation Data from Digitized Stereo Imagery

Why has it not been feasible in the past to map buildings from suitable digital imagery? The image processing software for automatically extracting elevation data from digitized stereo images has only been commercially available in the last three or four years, and all but one or two of those software packages are incapable of extracting an elevation for every pixel in the overlap region of a stereo image pair, which leads to a requirement for imagery at finer resolution than the actual elevation data resolution. One software package that is capable of automatic, every-pixel, elevation extraction, which is a practical requirement for this project (so that the elevation data has the same posting interval as the resolution of the source imagery), is ATOM™ (Automatic Topographic Mapper) from GeoSpectra Corporation of Bowling Green, Ohio (Vincent and Roberts, 1987; Vincent and Pleitner, 1987). It is the ATOM™ software package, which performs automatic correlation of the right and left images of a digitized stereo image pair that will be employed for this proposed project. The input to ATOM™ is a pair of digitized stereo images. The outputs from ATOM™ are an every-pixel-posted Digital Elevation Model (DEM) and a perfectly co-registered Digital Orthophoto (DOP) for the overlap region of the stereo pair (Vincent, 1997). Other than the operator's selection of fiducial, match points, and control points, ATOM™ is automatic.

Finally, with the commercial availability of ER Mapper, a sophisticated image processing, mapping and pattern recognition package, all three of the components missing in past efforts to generate population estimates from remotely sensed data have been brought together. The combination of high resolution imagery from space and sophisticated image processing software may make it possible for the first time to seriously test the practicality of generating accurate population estimates from space.

Benefits of Remotely Sensed Population Estimates over Traditional Methods

There are two major benefits of the satellite-based system over existing methods of population estimation. Traditional estimation methods exhibit high levels of accuracy, but they also have persistent shortcomings that can only be resolved by directly linking methods with substantive socio-economic and demographic dynamics (McKibben and Swanson, 1997). A major benefit of Satellite-based, spatially explicit data is in providing just such a linkage (Geoghegan et al., 1998), yielding not only improved accuracy, but also a basis from which population estimates can be understood and explained in terms of their underlying substantive dynamics, an important feature for end-users. The second major benefit is to be found in the high frequency with which estimates can be updated. The EarlyBird satellite, for instance, circles the earth every 94 minutes, revisiting locations every 36 to 120 hours, depending on latitude. No existing population estimate system offers a data collection process that allows for such a high update frequency.

Test Areas

The initial test will be conducted against a sample of on-the-ground data in areas familiar and accessible to the investigators: 1) sampled blocks in the rapidly growing northwest area of Las Vegas, Nevada; 2) One Block Group in Nye County, Nevada, which represents the Amargosa Valley, a rural area of southern Nevada about 100 miles from Las Vegas in which approximately 250 housing units existed in 1990 and 320 in 1997; 3) A Block Group in Wood County, Ohio, a slow growing area containing a 1990 population of 788 residing in 289 housing units consisting of rural farms, trailers and single family units in a small village setting on the outskirts of Bowling Green; and 4) sampled blocks in the center of Bowling Green, Ohio with a 1990 population of 5809 residing in 2,618 housing units. The research staff has done extensive demographic field work in the Nevada test areas, primarily to develop population estimates using the Housing Unit Method (Roe, Swanson, and Carlson, 1992; Swanson, Carlson, and Williams, 1990; Swanson, Carlson, Roe, and Williams, 1995; TRW, 1998, 1997; U.S. Department of Energy, 1997a, 1997b). All ground-based data will be collected through "windshield" surveys, using well-established protocols (Swanson, 1981; TRW, 1998).

While the areas represented in the sample we propose to use are "convenient," they also provide a good range of test conditions. The Amargosa Valley area is cloud free much of the year and virtually treeless. Its population density is extremely low, housing is widely spaced, there is no "unusual" housing or structures (e.g., boats) and only one small apartment complex. The area is one in which it is easy to identify, classify and count housing units via a "windshield" survey, which will yield highly accurate data against which the satellite-based estimates will be evaluated. This area represents the "lowest hurdle" for the system we propose to build. If remote sensing cannot develop accurate estimates of housing units for this sample, we believe that the technology is not yet sufficient to do it anywhere. As a test criterion, we expect the satellite-based housing unit estimates to be within 5 percentage points of what is counted on the ground for this area.
The block group we propose to test in Las Vegas is one that has experienced explosive growth in recent years. It includes commercial and retail structures as well as single unit and multiple unit dwellings. Again, this area is cloud free, virtually treeless and easily accessed by a car, which means that it is one in which we can obtain an accurate count of housing units via a windshield survey. This area presents more of a challenge than the Amargosa Valley because of the higher housing density, mixture of structure types, and the presence of commercial and retail structures. As a test criterion, we expect the satellite-based housing unit estimates to be within 10 percentage points of what is counted on the ground for this area.

The Ohio Tracts will provide a series of test conditions considerably different than those found in Nevada. Cloud and tree coverage will be an issue for the Ohio sites. In the rural farm settings, various out-buildings and trailers will add a new dimension to the housing identification problem. In the center of Bowling Green, there is a mix of single family housing, apartments, light industry and commercial retail buildings. Due to heavy tree cover, clouds and a variety of out-building clutter, we are projecting satellite-based estimates to be within 10 percent of ground counts in the rural area, and 15 percent in the city.

Initially we will use no collateral ground-based information in attempting to estimate housing stock. This is fundamentally different than many satellite-based estimates for small areas, which do, in fact, use collateral ground-based information. For example, one of the early discussions of using satellite-based information to make small area estimates is found in Cordenas, Craig, and Blanchard (1978), who propose using Landsat imagery in combination with multi-county level sample survey data to produce county-level crop acreage estimates.

Remote Image Processing

Because the IKONOS and OrbView3 stereo sensors with 1-meter spatial resolution are slated for orbit in 1999, it is our intention to employ digitized NAPP aerial photography for this project because it is available now and can be used to simulate the spatial resolution of the above two satellite sensors.

Two stereo pairs of NAPP photos at 1:40,000 scale (one pair each for the Ohio and Nevada test areas) will be scanned with a 25 micrometer spot size, yielding digital images of 1-meter pixel sizes. Though the NAPP photos will be scanned in color, only one of the spectral bands (the visible red) will be used as input to ATOM™. A co-registered Digital Elevation Model (DEM) and a co-registered Digital Orthophoto (DOP) (which will overlay each other perfectly because they come from the same source image pair) will be extracted for each of the two stereo pairs. The resulting DEM and DOP for each stereo pair will be imported into the ER Mapper image processing software package. The ER Mapper software will then be used to merge the DEM and DOP, along with the two other spectral bands from the same source images, into one data set per stereo pair (one pair for each of the two areas investigated). The elevation data from the DEM will be treated as a fourth spectral band (along with the visible green, the visible red, and the reflective infrared bands of the "right" NAPP false-color photograph) in a stereo pair data set, which will cover the overlap region of the stereo pair.

Automatic classification of building classes will be attempted by several methods, including supervised and unsupervised classification schemes that are available in ER Mapper. For the supervised classification method, we will try both parallel piped and maximum likelihood algorithms in ER Mapper. This will require that we identify training sets in both cities for each of the building classifications that we wish to recognize. We will also try a version of the K-means unsupervised clustering method for unsupervised classification, which will require no training sets beforehand. We will then assess the accuracy of both types of classification, using a withheld group (not including the training sets for supervised classification) of buildings identified in fieldwork. We will also try to recognize pixels of various heights and place polygons around pixels of similar height (using the DEM from the ATOM™ software), then classify the sizes and shapes of these polygons with the GEHK subroutines in ER Mapper Version 6.0. Polygons too small to be houses or other buildings will be eliminated in this manner, and buildings will be classified according to volume and height/area ratio considerations.

Conclusion

Science fiction fantasies about satellite imagery may be coming true, not to the full extent that Hollywood portrays it (e.g., Patriot Games, Enemy of the State), but at least to the level where we can construct tests of their capabilities. The tests planned in the coming year for the desert Southwest and Midwestern farmland will go a long way toward telling us the extent to which this technology has matured and its potential for use in future demographic applications.

LITERATURE CITED

Carls, N. 1947. How to Read Aerial Photographs for Census Work. Washington, D.C.: U.S. Government Printing Office.

Cordenas, M., M. Craig, and M. Blanchard. 1978. "Small-Area Estimators: County Crop Acreage Estimates Using LANDSAT Data." Pp. 32-35 in Small-Area Statistics Papers, Series GE-41, No. 5. U. S. Bureau of the Census: Washington, D.C.

Cowan, D. and J. Jensen. 1998 "Extraction and Modeling of Urban Attributes Using Remote Sensing Technology." pp. 164 -188 in D. Liverman, E. Moran, R. Rindfuss, and P. Stern (Eds.) People and Pixels: Remote Sensing and Social Science. National Academy Press. Washington, D.C.

Geoghegan, J., L. Pritchard, Y. Ogneva-Himmelberger, R. Chowdhury; S. Sanderson, and B. Turner. 1998. "Socializing the Pixel and Pixelizing the Social in Land-Use and Land-Cover Change." Pp. 51-69 in D. Liverman, E. Moran, R. Rindfuss, and P. Stern (Eds.) People and Pixels: Linking Remote Sensing and Social Science. Washington, D.C.: National Academy Press.

Iisaka, J., and E. Hegedua. 1982. "Population Estimates from Landsat Imagery," Remote Sensing of Environment 12:259-272.

Jensen, J.R., 1995, Issues Involving the Creation of Digital Elevation Models and Terrain Corrected Orthoimagery Using Soft-Copy Photogrammetry, Geocarto International: A Multidisciplinary Journal of Remote Sensing, Vol. 10, No. 1, pp. 1-17.

Lo, C.P. 1995. "Automated population dwelling unit estimation from high-resolution satellite images: A GIS approach. International Journal of Remote Sensing 16(1):17-34.

Lo, C.P. and H.F. Chan. 1980 "Rural Population Estimation from Aerial Photographs," Photogrammetric Engineering and Remote Sensing 46:337-345.

McKibben, J. and D. Swanson. 1997. "Linking Substance and Practice: A Case Study of the Relationship Between Socio-economic Structure and Population Estimation." Journal of Economic and Social Measurement 23 (2): 135-147.

Montesano, A.P. 1997, Roadmap to the Future, Earth Observation, Vol. 6, No. 2, pp. 16-19.
Roe, L., J. Carlson, and D. Swanson. 1992. "A Variation of the Housing Unit Method for Estimating the Population of Small, Rural Areas: A Case Study of the Local Expert Procedure." Survey Methodology 18(1):155-163.

Swanson, D. 1981 The Housing Unit Method: A Manual for Municipal Personnel Responsible For Making Annual Population Estimates. Alaska Department of Labor: Juneau, Alaska.

Swanson, D., L. Roe, J. Carlson, and C. Williams. 1993. "Estimating the Population of Rural Communities By Age and Gender: A Case Study of the Local Expert Procedure." Proceedings of the 1993 Public Health Conference on Records and Statistics. National Center for Health Statistics. Bethesda, MD.

Swanson, D., J. Carlson, L. Roe, and C. Williams. 1995. "Estimating the Population of Rural Communities By Age and Gender: A Case Study of the Local Expert Procedure." Small Town (May-June):14-21

Thomsen, I. and A. M. Kleive Holmy. 1998. "Combining Data From Surveys and Administrative Record System: The Norwegian Experience." International Statistical Review 66(2); 201-221.

TRW, Inc. 1998. Yucca Mountain Site Characterization Project: Summary of Socioeconomic Data Analyses Conducted in Support of the Radiological Monitoring Program: April 1997 to April 1998. TRW Environmental Safety Systems, Inc., Las Vegas, Nevada, June , 1998.

TRW, Inc. 1997 Yucca Mountain Site Characterization Project: Summary of Socioeconomic Data Analyses Conducted in Support of the Radiological Monitoring Program: April 1996 to April 1997. TRW Environmental Safety Systems, Inc., Las Vegas, Nevada, June , 1997.

U.S. Department of Energy. 1997a The 1997 'Biosphere' Food Consumption Survey: Summary Findings and Technical Documentation. U.S. Department of Energy, Office of Civilian Radioactive Waste Management (DE-AC01-91RW00134). (November, 1997) U.S. Department of Energy, Yucca Mountain Project, Las Vegas, Nevada.

U.S. Department of Energy. 1997b Technical Documentation and Summary Findings of the 1997 Survey in Lincoln County, Nevada. U.S. Department of Energy, Office of Civilian Radioactive Waste Management (DE-AC01-91RW00134). (November, 1997) U.S. Department of Energy, Yucca Mountain Project, Las Vegas, Nevada.

Vincent, R.K., 1972, An ERTS Multispectral Scanner Experiment for Mapping Iron Compounds, Proceedings of the Eighth International Symposium on Remote Sensing of Environment, Ann Arbor, Michigan, pp. 1239-1247.

Vincent, R.K., M.A. True, and D.V. Roberts, 1987, Automatic Extraction of High-Resolution Elevation Data Sets from Digitized Aerial Photos and Their Importance for Energy Mapping, NCGA's Mapping and Geographic Information Systems 1987 Proceedings, San Diego, California Meeting, pp. 203-210.

Vincent, R.K., M.A. True, and P.K. Pleitner, 1987, Automatic Extraction of High Resolution Elevation Data from SPOT Stereo Images, Proceedings of the SPOT 1 Image Utilization, Assessment, and Results International Conference, Centre National d'Etudes Spatiales (CNES), Paris, France.

Vincent, R.K., 1997, Fundamentals of Geological and Environmental Remote Sensing, Prentice Hall, Publishers, Upper Saddle River, New Jersey, Chapter 6.