Remote Sensing, Controversy and the Million Man March
After march organizers demand a more accurate crowd estimate, Boston University remote sensing experts go to work
By Farouk El-baz

The Million Man March took place on Monday, Oct. 16, 1995 along the Mall in Washington, D.C. The National Park Service announced that 400,000 people took part in the event. The Nation of Islam, organizers of the march, vehemently objected, called it a gross underestimate of what was counted on site to be a crowd 1.5 to 2 million strong, and threatened to sue the National Park Service. A controversy was brewing.
      Two days later, Michael Guillen, a colleague who teaches physics at Harvard University called to ask if satellite remote sensing techniques would apply to the estimation of the number of march participants from 35mm photographs.
      Because the answer was yes, I assembled a team of assistants and graduate students including four geologists, four geographers and two geographic information systems (GIS) experts at the Boston University Center for Remote Sensing. We proceeded to study a videotape of the march and requested the scanning of 10 color photographs obtained from the Park Police (a division of the National Park Service). Digitized data from the scanned photographs were transmitted from Washington, D.C. to our computers in Boston via the Internet.
      After displaying and analyzing the photographs, various ideas were proposed for counting the participants. After much discussion it was decided that the most efficient method was to divide the Mall area by a grid and estimate the number of people within each area based on a given crowd density.
      The area of the Mall was measured using a scaled multispectral image. The crowd density was estimated based on different degrees of packing in each square meter, ranging from six people per square meter to one person per five square meters. I estimated the highest crowd density by simply drawing one square meter on the lab floor to see how many people could easily fit inside it. This number was used for the most densely packed areas of the Million Man March such as the Capitol area, and near the half-dozen television monitors distributed throughout the Mall.
      My team worked through the night, with frequent discussions and rechecking of each step, until the result of 878,587 was reached. With an estimated error margin of 25 percent, the number could be as high as 1,098,234 or as low as 658,940. The result was announced live on Good Morning America on Thursday, Oct. 19. Instantly, the scientific research effort turned into a media event.
      I invited the National Park Service managers and Park Police technical people to visit our image analysis laboratory to see how we came up with a number that is more than twice their estimate. A meeting was convened on Oct. 24 in the presence of representatives of the Nation of Islam.
      We briefed the group on the theory and practice of crowd estimation, gave examples of our experience in estimating tree numbers in the California forest and counting sand dunes in the desert of Kuwait, and explained the remote sensing and GIS methodologies that we used to arrive at our figure.
      When it was the turn of the National Park Service to discuss how the official estimate was made, we learned that it received its figures from the Park Police, which calculated the number based on several sources, including: (1) photographic enlargements (5x7 inches) of a videotape taken from a helicopter; although 35mm photographs were obtained at the same time, their processing required time, and therefore, they were not used as a source of data; (2) the number of buses, that would have brought people from out of town, in pre-designated parking lots; and (3) the number of passengers of the Metro system, above and beyond the number of riders on a normal Monday.
      The Park Police helicopter was not allowed to fly directly above the crowds such that it would not represent a danger in the case of a malfunction. (In general, helicopters are very unstable platforms for photo acquisitions.) Therefore, both the videotape and the 35mm photographs were taken with such oblique angles that Smithsonian Institution buildings hid vast areas of the Mall, as did the trees that line up the Mall on the north and south sides. The result was a less than adequate data set.
      To make matters worse, the Park Police did not fly a photographic run between 11:30 a.m. and 3:00 p.m., although the regulation states the need for photography every hour of the day. The period between noon and 2:00 p.m. was designated "peak time" by the Million Man March organizers (highest attendance around 1:00 p.m.). In addition, an emergency had developed and a woman and her child had to be taken by the helicopter to a hospital.
      Because of these problems in the data aquisition procedures, it was clear that the Park Police photographs were not ideal for crowd estimation. At the conclusion of the meeting, the National Park Service signed the statement that "the 400,000 figure can no longer be considered final." Also, the National Park Service offered to give me the original negatives of five, 36-exposure, 35mm films to use for a recount using better data.
      The Boston University team selected twelve 35mm photographs that covered the Mall and surrounding areas. We had these digitized using an instrument capable of resolving 5,000 dots per inch. The digital data were fed to our computers and the research team was divided into groups huddled around several workstations.
      We first retrieved a 1-meter-per-pixel photomap of the Mall area from the Internet. It was necessary to use such a geometrically corrected, vertical view of the area as a base.
      To be able to overlay a grid on the photographs, it was necessary to translate the oblique views onto the vertical photomap. This was done by the time-consuming process of registering locations of fixed points, such as road intersections or corners of buildings, from the oblique view onto the vertical one. On average each photograph required the registration of 40 points from the digitized 35mm oblique view to the photomap.
      In areas where computer-assisted enlargements of the 35mm photographs allowed the identification of individuals, a count was obtained by tagging each person's head or the top of his shadow. For example, at the Washington Monument, there were 3,644 individuals.
      In most other areas, however, the crowd was so densely packed that it was impossible to distinguish individuals. In these cases, it was necessary to classify the density of people per unit area, and calculate the totals of similarly packed pixels. We again assumed that six people could stand in one square meter (the Park Police uses the more conservative estimate that each person requires 3-5 square feet). We used decreasing numbers in less packed areas down to one person per 10 square meters in sparsely populated areas.
      In addition to the use of original film negatives to obtain a better number, the main focus of the second count was to reduce the uncertainty factor as much as possible. The two factors controlling the margin of error were the exact boundaries of occupied areas of the Mall and the variations in crowd density. For the second count, use of the negatives and the photomap improved both the resolution and the geometry of the images.
      The next step was to accurately estimate the density of the crowd in various parts of the Mall. This was accomplished using two methods. The first was based on defining the occupied area and then estimating the different densities within each pixel, by assigning a different color to each density. The total number of pixels for any particular region was then calculated and multiplied by the specific density for that region to obtain the number of participants. The second method consisted of utilizing the GRASS (GIS) software, to create a 50 meter cell-grid and apply it to the registered images in order to estimate the density within each cell and multiplying that by the cell size to get the crowd number. Both methods worked well resulting in comparable numbers.
      The result of the second estimate based on the analysis of original negatives was 837,214, as announced on Friday, Oct. 27, 1995. The margin of error of 20 percent was better than in the first attempt; it could not have been further improved because of inherent problems in the way the original photographic data were acquired. The estimate was widely reported by the media and appears to have gained wide acceptance. A reference to the Million Man March on page 18 of LIFE's January 1996 issue states the number as "more than 800,000."
      If one thing became clear from this exercise, it was the realization that for an accurate and efficient crowd estimate, the photographs have to be taken at hourly intervals using two systems onboard a fixed-wing aircraft looking straight down: (1) a digital camera with ground resolution equal to the size of an average person from the flying altitude; data from such an imaging system would be fed to a computer to tally the number of pixels with the specific spectral reflectance for a quick but reliable estimate; and (2) a conventional aerial photographic camera with 9x9 inch film to obtain high definition, stereo photographs for post-event analysis and archival purposes. Only then could the National Park Service undertake the very difficult task of estimating crowds at such events.

About the Author:
Farouk El-Baz helped establish the field of remote sensing by working on the selection of landing sites for the Apollo lunar missions. As director of the Boston University Center for Remote Sensing, he has devoted much attention to the use of remote sensing data to the location of ground water resources in the desert. He may be reached at 617-353-5081.

Back