By Gerry Mitchell, P.Geo, President, PhotoSat
In an effort to find a faster way to produce elevation surveys from satellite photos, PhotoSat geophysicist Michael Ehling and I accidentally discovered a novel way to greatly improve the accuracy and resolution of satellite topographic survey results.
It was 2007, during the peak of the natural resource boom and PhotoSat could not keep up with the demand from Vancouver mining companies who needed accurate satellite survey data for their projects in remote parts of the globe. Without accurate ground surface surveys the mining engineers couldn’t produce reports of ore body volumes. Without the engineering reports the companies couldn’t report their mining discoveries to a booming stock market waiting expectantly for their news.
Interactive photogrammetric processes
Michael and I had been watching how photogrammeters produced elevation surveys from stereo satellite photos since 2004, when stereo IKONOS satellite photos first became available. PhotoSat was buying stereo IKONOS satellite photos from Space Imaging, now part of DigitalGlobe.
We were reformatting the photos so that the photogrammeters could produce elevation surveys using computer systems that had been designed for processing stereo photos taken from airplanes. They were using highly interactive processes and were taking an average of 150 hours to produce satellite surveys for 100 square kilometer projects.
Michael and I could see that the processors spent most of their time interactively measuring the matches between identical features on pairs of satellite photos. The photos had been taken with the satellite looking at the same area on the ground from different directions. By identifying identical ground features on each of the photos, and precisely measuring their locations, the elevations of the features can be computed.
When Michael and I asked if the photo feature matching could be done automatically we were told that the automatic process usually didn’t work, but when it did, editing the results took more time than doing the matching interactively, so no one used it. As geophysicists we were intrigued by what looked like an interesting technical challenge.
Oil and gas seismic processing tool box
In the 1980’s and 90’s when I was working as a Geophysicist in oil and gas exploration I processed a lot of seismic data. Oil and gas seismic survey data is used to image geological formations thousands of meters below ground in the search for oil and gas. Seismic data processing has always been one of the most complex and computer intensive data processing fields, with expenditures of billions of dollars annually.
Over the past 50 years seismic processors have developed an immense array of data processing tools, including many automatic image matching tools, and I thought that we could probably apply these to the satellite photos.
Michael tested seismic processing image matching tools on stereo IKONOS satellite photos for several months in 2007. He had to format the digital satellite photos so that they would look like seismic data to the seismic processing systems, run tests, and then reformat the results to look like photos again.
We were in search of a faster way to produce the survey results that the photogrammeters were spending hundreds of hours to produce. We were testing with a pair of IKONOS satellite photos that had already been processed by the photogrammeters so that we could compare our results with theirs.
Initial PhotoSat processing test results were amazing
After three months of testing we had our first real success. We were astounded by the results. We could see many fine topographic details on our test data that were simply not visible at all in the photogrammetric processing.
We continued to refine the process over the next few months until we had produced satellite survey results that were over three times as accurate as the photogrammetric processing and had much more topographic detail. The initial process took over 100 hours of computer processing time to process 100 square kilometers, so we had not really found a faster way to produce the results, but completely unexpectedly, we had found a way to produce better results.
New PhotoSat Workstation built from scratch
Now, nine years after our initial accidental discovery we still have a team of researchers and software engineers improving our satellite processing system.
Several years ago they replaced the seismic processing system with a computer system built from scratch to efficiently apply the seismic algorithms and processes to satellite photos. This system, the PhotoSat Workstation, was designed to harness the processing power and speed of Graphics Processing Units (GPUs). The GPUs process numerical data a thousand times faster than CPUs. Older software that is retrofitted to use GPUs typically shows speed improvements of two to five times.
It took several years and several million dollars of software development, but since our initial discovery in 2007 we have successfully created an automatic process that produces satellite surveys much faster than the photogrammeters, with much higher accuracy and better topographic detail.