|
|
|
I'd always wanted to work for NASA. I was interested in science from my earliest days. The science books by Isaac Asimov fired my imagination more than his fiction did, and I read as many of his books as I could find. It was Lester del Ray's classic "Rockets Through Space" that first fired my imagination about space travel. I was even a member of the Planetary Society and the L5 Society. One year I audited an astronomy course at Columbia University, and I discovered NASA had a branch office associated with the college, he Goddard Institute for Space Studies, where Robert Jastrow was the director. I applied for a job and received an offer as a scientific programmer, doing data analysis for the Landsat 2 satellite. The satellite flew in a medium altitude polar orbit (920 km or 570 miles), maintaining a sun synchronous attitude with the surface of the Earth, flying overhead at roughly 9:30 am local time. The means the illumination angle in latitude was always the same. The satellite has a telescopic camera that scanned the Earth as it flew overhead, sweeping the view in latitude as the craft traveled in longitude. Notice how the picture provided above is slanted. This is mostly due to the orbital inclination of 81 degrees, but effects from the rotation of the Earth underneath the satellite during the flight time of the picture had to be accounted for as well. The telescope's beam was split into four parts which went through individual narrow band pass filters, providing four spectral data points for the same location simultaneously. Two of these values determined the infrared intensity (both near infrared), one was in the visible red region and the forth was in the visible green region. The spatial resolution of each pixel was about 80 m (250 ft) when contrast was normal, which is fine for agricultural and geologic use but is inadequate to be of use for military intelligence (high contrast linear features, like a paved road through a corn field, could usually be seen down to 10 m or 33 ft). The collected data was transmitted to Earth in real time if within range of a ground station, or recorded in the satellite for later transmission to a US receiving station. Receiving stations were also built in several other countries around the world, such as Brazil, Canada, Iran, Italy and Zaire,. Once received, the data was stored on magnetic tape reels, and these tapes were then sent to us, where the data was input to Fortran programs for analysis. The analysis was used to identify the material seen by the satellite, and different materials could be assigned arbitrary colors in the final image. This is how maps like the one shown above were made. The technical details of the categorization algorithms were based on the relative intensities between the spectral values for each pixel intensity in a six dimensional phase space. First, we calibrated the data given the known response of the optical sensors. Then we normalized the readings of all the intensities so that differences in brightness across the image would not skew references to past history. Brightness was defined as the square root of the sum of the squares of each color. The dimmest pixel's value was subtracted from all the other pixel's, producing a zero point. The brightest pixel's magnitude then divided all the other data points, producing a maximum magnitude of 1. Thus modified, the six ratios between the four readings for each point were calculated. To be explicit about that, let's call the four spectral reading for each point w, x, y, and z, going from longest wavelength to shortest. The first dimension in phase space is the ratio w/x, the second is w/y, and the third w/z. The fourth and fifth are x/y and x/z respectively. Finally, the sixth dimension is y/z. Pixels which give rise to phase values that cluster near other phase values are grouped together and identified as probably the same substance. Let's explain that by a physical example. Vegetation is typically very reflective in the green region and dim in the visible red, but bright in both infrared bands. Dry soil is brighter in the visible red then in the green, and only moderate in the infrared. Wet soil is somewhat dimmer in the visible bands, and somewhat more dimmer in the infrared. Bodies of water thicker than a few centimeters are dark in all bands. In this way the various types of substances seen can be identified, and to an extent even the conditions that affect them. For example, dehydrated corn is dimmer in both infrared bands then healthy corn, but insect damaged corn may be dimmer in only one infrared band. Unfortunately, given the limitations of the earlier versions of these satellites, by the time insect damage could be detected from the satellite and the data processed, the crops were already very badly damaged. Libraries existed of known values in phase space for most of the possible
results. This was done by the collection of "ground truth" in correlation
with what the satellite saw. During the early stages of the project's
development (before my time there), after an image was processed that yielded
unknowns in phase space, data would be collected locally. This was done in
various forms, from high flying U2 aircraft carrying high resolution cameras and
equipment similar to what the satellite used, to low flying planes with movie
cameras to graduate students in jeeps with notebooks and pencils. The
libraries of phase space values were then updated accordingly. A few
years later, to commemorate my involvement with this project, I built a 1/48th scale
model of a U2, specially painted in NASA colors and marked as Earth Resource
Aircraft No. 4. A brass plaque labels the diorama simply as "The Truth."
Just as an aside, here's a photo of my model:
However, we used other information to improve the accuracy of our analysis, and continually research new methods. The results of the phase space calculations were combined with masks of the location data. That is, pixels located close together both in phase space and geographically were assigned a higher confidence of being the same substance as pixels nearby in phase space but far apart in real space. In effect, we could use this data to redefine what was meant by "nearby" in phase space dynamically. For example, a field of wheat one one side of the image and a field of straw on the other might be sufficiently close in phase space to be identified as the same substance, but if forced to be identified by the geographic constraint might be recognized for what they individually were. Or, a cluster of a dozen pixels might be identified as sagebrush but a cluster of several hundred pixels might be identified as prairie. If the dozen pixels were between the prairie pixels and an area confidently identified as concrete, it could be identified not as sagebrush but as a mixed border of prairie and concrete, possibly at an airport. Other areas involved looking at regions of phase space by the center of gravity method, seeing how the distribution of pixels changed as a function of directional differentials. Or were identifications better when the distribution was assumed to be flat? How does the shape of the clusters in phase space affect classifications (for example, urban rooftops show much greater variation in the infrared than they do in the visible)? Does making several passes through the data using restricted dimensional information yield the same results as processing all the dimensions in one pass? Does processing four quadrants of the image separately give the same results as processing them all together? What information can be gleaned by the differences? Such questions made for ongoing experiments. The following links allow interested persons can learn more about the art and science of remote sensing and the current Landsat program in general. |
