It’s was serendipitous when I found out a friend of mine would be teaching a Machine Learning class — her specialty — and her lectures would be recorded for posterity and she generously offered to let me “play from home.” How could I not seize the opportunity?
During the fourth lecture, when she posted a slide showing work on orbital classification of land features. The strip on the left is an image taken from the Earth Observing-1 (EO1) spacecraft using its Hyperion instrument. The strip in the middle is the result of a computer-derived classification of the image.
With my interest piqued, I spent the better part of a Sunday sleuthing around to find out more about the spacecraft, what it was observing, and how the results were derived. Today, I’m just interested in the gadgets. When I have a better handle on the analysis of images, I’ll post more.
“The Landsat Program is a series of Earth-observing satellite missions jointly managed by NASA and the U.S. Geological Survey.” They’re chartered with remote sensing, e.g., collecting information about Earth from space. Landsat 5 (launched 1984) and Landsat 7 (launched 1999) are still currently operational. (Landsat 6 failed to launch.)
Why I mention Landsat 7 first is its instrument, the “Enhanced Thematic Mapper Plus,” has been calibrated to better than 5%, meaning Landsat 7 serves as the benchmark for other instruments, like…
Earth Observing-1 was launched on November 21, 2000 into an orbit such that it flies in formation with Landsat 7. By following in the same orbital track, and within a minute later, both EO-1 and Landsat 7 effectively observe the same things. Landsat 7 is thus being used to validate EO-1’s technologies.
EO-1 has three major banks of instruments to test:
- ALI (Advanced Land Imager) – a spectrometer designed to eliminate some of the additional components (like the cooling system). Compared to Landsat 7’s ETM+ (Enhanced Thematic Mapper), it’s smaller, lighter, uses less power, and provides better coverage.All of these are good things.
LandSat 7 vs ALI, Source: 
- LEISA (Linear Etalon Imaging Spectral Array) – This corrects for degradation of surface images occurring from atmospheric variability from water vapor and airborne particulates.
- Hyperion – a pair of spectrometers (near-IR, 400-1000nm; shortwave IR, 900 – 2500nm) that collectively cover a selection of 220 wavelengths. The instrument was developed in parallel with the ALI. What makes this instrument interesting is it provides the potential for a richer analysis. For example, on the left, LandSat has enough granularity to distinguish among grass, hardwoods, softwoods and “I dunno.” Hyperion can improve upon this by offering more specific data that can be later analyzed down to specific major species.
LandSat compared to Hyperion, Source: 
Cool acronyms are essential!
The EO-1 spacecraft has a couple of 12MHz processors (8 MIPS), one assigned to primary spacecraft operations, the other for processing images. The processor and limited memory (not enough to boot Windows Vista) limit the data processing that can occur. In the very first image, it was noted only twelve bands can be accessed at a time.
-  Kiri Wagstaff, CS461 Winter 2008, Lecture 4.
-  Rebecca Castaño, Learning Classifiers for Science Event Detection in Remote Sensing Imagery, iSAIRAS, Sept 2005
-  Earth Observing-1 Baseline Mission History
-  Landsat 7 Program page
-  Advanced Land Imager Results – ALI, GSFC Systems Engineering Seminar, June 2002
-  Advanced Land Imager Results – Hyperion, GSFC Systems Engineering Seminar, June 2002
Excellent post! Your background research is spot-on. 🙂
Ah… classification of satellite imagery. Fun (frustration) fun (frustration) fun (frustration) 🙂
It will be interesting to see your findings.
Are you going to stick to unsupervised classification?
Will you be trying out neural nets?
Examining how the number of control (ground-truth) points you use affects the accuracy of the classification?
Using raw imagery or correcting for light leakage between pixels?
Comparing satellite-derived imagery to aerial photography of the same sites to determine how much information you can’t extract from a satellite image?
Jumping backwards through hoops while balancing plates on your head and juggling chainsaws? 🙂
Comments are closed.