Dr. Matt Nolan

Institute of Northern Engineering
University of Alaska Fairbanks

 

Home

Curriculum Vitae

McCall Glacier

Photography

2007

2008

2009

2010

2011

2013

Older Projects

Personal

Contact

 

 

2013 Photography

 

Introduction

I believe that airborne Digital SLR Photogrammetry is an awesome tool for arctic field sciences that is likely to revolutionize our scientific methods and capabilities. This fall I was able to fly a number of airborne photogrammetry campaigns throughout the Alaskan Arctic and the Fairbanks area using a DSLR Photogrammetry system I built and this page presents an overview of what I've learned and why I think it is so cool.

Based on the testing I've done thus far by comparing time-series of my data, airborne DSLR Photogrammetry can achieve these specs:

- Spatial resolution: 5 cm to 50 cm ground sample distance (GSD, or pixel size); 10-20 cm is usually the optimal resolution in terms of flying hassle, costs, and file sizes
- Absolute x,y,z positioning with no ground control: <50 cm
- Absolute x,y,z positioning with ground control: ~2 cm (or as good as ground control)
- Random and systematic errors: <20 cm
- Orthoimage co-registration to DEM: Perfect (made with same data)
- Rapid Results: full resolution ortho of large projects (e.g., 4 GB DEM, 16GB ortho) can be processed to low-resolution DEM within 24 hours; smaller areas same day
- Standard Results: large projects (e.g., 4 GB DEM, 16GB ortho) require 3-4 days CPU time at full resolution with fast 32 CPU machine

Below are some applications which I think these maps would be useful

One-time efforts:
- Base imagery maps down to 5 cm resolution for GIS analysis related to science, infrastructure, etc
- Base DEMs down to 5 cm resolution for GIS base, hydrologic modeling, topographic analysis, etc
- Example uses: base layers for vegetation maps, infrastructure identification, locating field sites and transects, 3D terrain visualization, project outreach, glacier extent outlines, counting caribou, etc.

Time-series of maps with these characteristics can be used for:
- Measuring glacier volume change, crevasse depth and change, ice-cored moraine change, and surface velocity fields
- Detecting and measuring permafrost thaw slumps
- Measuring aufeis volume change
- Measuring sea ice topography and ice berg volume above waterline
- Measuring coastal erosion
- Monitoring mining progress and making hazard assessments
- Measuring snow pack thickness
- Measuring NDVI and other vegetation changes seasonally or annually
- Monitoring infrastructure degradation
- Monitoring construction progress

The near-realtime applications might include:
- Detecting and measuring ice jams topographically (that is, where the dams are forming, their height and potential severity, etc)
- Determining the extent of current flooding (mostly through orthoimagery analysis)
- Monitoring unstable slopes for rock and debris avalanches
- Measuring earthquake displacements, damages, and avalanche volumes
- Determining wildfire extents and damage

My system specs:

- Nikon D800 ~$3000
- Nikkor 24mm lens ~$2000 (almost any glass 20-30mm would work)
- Trimble 5700 DGPS ~$3000 (used on ebay; any L1/L2 GPS would work)
- Cirrus Design Intervalometer ~$1000
- Garmin 495 GPS (in flight nav) ~ $500 (used on ebay; any VFR GPS would work)
- IMU ~ $0 (We don't need no stinking IMUs...)
- Waypoint GrafNav (GPS PPP processing) ~$4000 (there is freeware for this too)
- TopoFlight (flight planning) ~ $800
- Agisoft Photoscan Pro (photogrammetric processing) ~ $550
- An airplane, camera port, sensor mount, power, etc ~ $20k - $2M
Total system cost (minus aircraft) ~$15k

What's new here?

Film photogrammetry has been around for since WW2 but was a painstaking process requiring enormous expertise and substantial ground control to tackle projects like these. Lidar and digital photogrammetry using purpose-built cameras have been around for 20 years or more, requiring equipment investements of $400k - $1M with digital workflows that are at best cumbersome and usually require some ground control. What's new here is that I can do these same things now with a stock DSLR camera and other common equipment costing in total less than $15,000 within a workflow that is easy, fast, accurate, and precise with no ground control, allowing me to make an awesome map within 24 hours of picking a target. Advances in camera technology, GPS technology, and processing technology over just the past few years have made this possible, my contribution is simply identifying those technologies and making them work together as a package in a way that maximizes their benefits and then demonstrating that it actually works by using it on Alaskan targets. As an example of how this translates into cost, I (or rather the NSF) paid $400k for a lidar map of my glaciers in the eastern Brooks Range in 2008. I believe a commercial contractor could now make the same map for $40k using DSLR Photogrammetry, a factor of 10x less money, while delivering a product that is 144x higher resolution (25 cm pixels vs 3m pixels) and also includes an orthorectified image that perfectly drapes over the terrain map. My direct costs, not driven by profit, of doing it myself are substantially lower than this. So from what I can tell thus far, there is now no technical or scientific advantage to using lidar in any application above tree line and thus no reason to pay the extreme costs for it, the only major disadvantage currently is finding a commercial vendor to do it. But more than just saving money, the low-cost of DSLR Photogrammetry means that it is now affordable to make time-series of such maps, and such time-series open up many new possibilities for use in analyzing and understanding changes in topography. That is, we're not just doing the same thing for less money, we are now able to do brand new things.

Why this page?

I'm excited about what this technology can do for us and so want to see it get used. At the moment, I am one of the few people in Alaska doing this sort of thing, and I'm not really intending at this point to turn this into my full time job for other people. But until the commercial geophysical community catches up and drops its prices by 10x, I'm happy to help others out by acquiring data for their projects to further improve my techniques and to further push the frontiers of the emerging field of low-cost, high-accuracy, high-precision map making with ordinary DSLR cameras. I'm also happy to help people build and use their own system, but I am fond of this old addage when describing what it takes to succeed at it: On the first day of medical school the doctor says to his class "I can teach you how to take out an appendix in a day, but it will take me four years to teach you what to do if something goes wrong". The same is true of airborne geophysics, unless you want this to become a major part of your career your time is probably better spent hooking up with someone who has made that commitment to allow you to focus on mastering the science that really interests you. Trust me, I know - though it literally only took me a day this summer to build my system once I had the parts, it took me four years of effort before that day to figure out how to troubleshoot everything that could go wrong.

Disclaimer

The materials presented here are confidential, privileged and copyrighted, describing intellectual property of the University of Alaska Fairbanks. The contents of this page, including text, calculations, figures and analysis, are not to be used for any commercial or promotional purpose without express written consent of Dr. Matt Nolan of the University of Alaska Fairbanks. Any calculations, measurements, or analysis contained herein are to be considered provisional and should not be used for any operational, commercial or scientific purpose.

 

A quick introduction to photogrammetric data products

The basic data products are a digital elevation model (DEM) and orthoimage. The DEM is basically a digital topographic map -- it's like draping an excel spreadsheet over the earth and filling each spreadsheet cell with the average elevation of the terrain that cell covers. An orthoimage is an image mosaic corrected for perspective and scale using topography, such that it appears to be taken looking straight down on each pixel and the distance between objects in the orthoimage is accurate and consistent. So, for example, even though an indvidual image from the camera shows the side of the building along with its roof, giving it the sense of leaning over, the orthoimage shows only the roofs of buildings and one can accurately measure the distance between roofs. The accuracy of this terrain correction is dependent on the accuracy and resolution of the DEM used in the correction.


Here is an orthoimage I made of Ft Knox mine near Fairbanks on October 6 2013. It is composed of about 1000 individual images from my camera, taken vertically looking through a hole in the bottom of my plane.


The same individual images make this digital elevation model (DEM, or digital surface model DSM depending on preference). Here topography is color coded, with high elevations shown in red and low in blue.


Here is the same topography shown as a shaded relief image, with fake sun causes shadows on a grayscale terrain.


Because I mapped the same area a week later, I can subtract the two maps and assess what changed between them. This image is the difference of these two topographic maps, with no change indicated by the light green-yellow color. More about difference maps later.


Because the data are digital, I can view them at any 3D perspective and zoom in, using any image drape I like. Here is a perspective view of the mine and its big pit. Move your mouse over the image to see the same topography overlain with the change detection image. Here you will see that most of the changes observed were in the pit.

 

A quick introduction to data quality

Here are a couple of examples showing resolution and accuracy of my maps to give a quick sense. Mouse-over the images to see them change.


Mouse-over the image to visually see the difference in topography over one week at Ft Knox mine near Fairbanks. Near center you will see a lot of rock has been removed by an excavator loading a dump truck (you can even see the vehicles at the edge of the pit). You will also see a small landslide has occurred just left of center. What's more important, however, is how little the surrounding area has changed. That is, the lack of difference in my measurements over a one week period demonstrates how little error my measurements have.


Here is the same scene, but quantifying the differences over one week. Green indicates no change, yellows-reds indicate a positive change, and blues indicate loss. You can see the landslide at left of center has lost rock at top (blue), which then accumulated at bottom (red). You can also now more clearly see how accurate and repeatable these data are by looking at the rocks that spilled over the side (yellow) of the excavation (blue). Mouse-over the image to pick out the yellow rocks on the lower ledges, then find them again in the previous image pair to see how they accumulated over a week. For scale, these ledges are 3m-4m wide and those rocks are about the size of your head.


Here I have draped the orthoimage over the topography of caribou! Mouse-over the image to the topography more clearly. In the mouse-over, color represents slope, and the blue-greens are polygonal-ground common in permafrost and the bright reds are the steep slopes of caribou bodies. You can even tell which are sitting down and which are grazing!

 

Assessment of accuracy

Pretty pictures like the ones above are fine, but for scientific use we want to know accurate the maps are. As of October 2013, I have only begun to rigorously analyze the accuracy of my data, as my first real acquisitions only began in September. One of the main obstacles to assessing accuracy is that there is little data to compare to that is of comparable or superior accuracy, so it is more difficult than you might imagine to determine accuracy, especially on the landscape scale in remote regions where available ground control is only a few points or none at all. In any case, here are some examples of a more quantitative assessment than just visual inspection.

I made a map of McCall Glacier in the eastern Brooks Range on September 10 2013. I compared this to lidar that I commissioned commercially several years earlier. The goal of making maps like these is to measure the change in glacier volume over time, so the trick in comparing these maps for accuracy is to avoid comparing regions with glaciers in them. That is, we look at the rock areas that are presumably not changing to assess accuracy between maps.


These are my flight lines over McCall Glacier in September 2013 draped over Goggle Earth. McCall Glacier is the large one at center that flows north (up).


Here I have subtracted a 2008 lidar map from my 2013 photogrammetry of McCall Glacier. The light blue color indicates little change and is mostly mountain. The reds indicate the most change (about 15m) and are found near the glacier termini.


Here is the same difference image draped over Google Earth topography, this time showing areas I used to assess the accuracy of my map by comparing areas where presumably no change has occurred, that is mountains (purple polygons). The standard deviation of difference in these areas was about 35 cm. The mean varied but was always within 20 cm of zero. This is a fantastic comparison, especially considering there are significant differences in the way that lidar and photogrammetry sample the earth surface to determine an elevation for each pixel.

To put these results into context, I did the same comparison between my lidar data. I paid to have these data commercially acquired in 2008 (3 times), 2009, 2010, and 2011. The highest quality data were acquired in one of the 2008 acquisitions and the 2011 acquisition. Comparing these same rock areas as in the previous figure, I found standard deviations of about 30 cm also and means of about up to 20 cm, essentially the same as my photogrammetry. Comparing this 2008 lidar to two poorer quality lidar acquisitions made in 2008 (acquired about a month apart) yielded standard deviations of 60 cm and 55 cm. So it seems to me that my photogrammetry meets or exceeds the accuracy of what was delivered by a major commercial geophysical contractor that specializes in lidar.

Though 35 cm accuracy is a fantastic comparison, I suspected my data were even better than this, largely because of inherent differences between lidar and photogrammetry. So I began making repeat maps at test sites near Fairbanks so that I could compare my maps to my maps. My favorite test site is the Ft Knox gold mine, about 20 minutes outside of town, because it has steep terrain on the same scale as McCall Glacier valley, it has broad flat areas, and there is a lot of man-made change that is easy to distinguish from the background of no change.


This is a difference map of my photogrammetric maps taken a week apart at the Ft Knox goldmine (as are the next several images). This is a flat area used to store tailings. The red line on the terrain is about 400 meters long and the difference in elevation between the two maps is shown in the plot above it. horizontal grid lines here are 5 cm apart. I do not believe there is any airborne technique that can achieve a better repeatability than this, especially considering that some of this difference is likely real due to the traffic by heavy trucks here. The rainbow colors near the bottom left are individual gravel push scars made by a bulldozer.


Here is a 2 km transect across some "lakes", where effluent from the gold-separation process is drained; the surface here is solid but perhaps floating. The grid lines are again 5 cm, with a standard deviation of difference <10 cm.


Here is an inclined road leading into the pit. The large changes at bottom are the results of blasting, where the rock has been lifted and loosened by the blast to facilitate removal by trucks. It appears that during the blast some rocks landed on the road and were subsequently bladed flat. You can see the results of this grading on the road as a green-yellow transition, and in the plot by a 50 cm displacement. Gridlines here are 10 cm. The bulk of the road has a standard deviation <10 cm, clearly revealing the 50 cm signal.


This difference image has the bottom of the pit at left and the transect goes across the same blastfield and road at right. Horizontal grid lines on the plot are now 1m. The bulk of the difference on the steps leading down to the pit is <20 cm standard deviation. The spikes seen on the transect over these steps are largely due to the resolution and difference in origin of the pixels. That is, because the pixels are 23 cm and the flat part of the steps are 3-4m, some of the pixels are averaging the break in slope and thus average the flat part of the step with slope below it. Because the ground is pixelated slightly differently between acquisitions, the pixels at the slope break are slightly offset and sampling different elevations. If the pixels were 5 cm, I suspect these spikes would disappear or be greatly reduced in magnitude because smaller pixels would resolve the slopes better. I've tested that by reducing the resolution, but have not yet tried imaging here at higher resolution. Nevertheless such spikes are not seen in any areas without such large changes in slopes so I do not believe they are system noise. There could also be a slight mis-registration between maps, as I have not done any fine co-registration to shift the two maps laterally. Such sub-pixel shifts would also likely remove a bit of these spikes.


Here are the buildings at the processing facility, highlighting the same issue as with the benches in the previous image. Ticks are again 1m. The difference in roof elevations is again ~10 cm. The spikes along the building edges are again caused by the 23 cm resolution of the topography not being able to sharply resolve the building widths. That is, in these maps, the building widths can only be integer multiples of 23 cm and because the pixels in each map have a different origin they measure the building shape slightly differently.


Here is the entire scene, about 6 km across. Ticks are 1m again. You can find no large-scale warps or ramps across this distance and the difference again is <20 cm in unchanged areas. The spikes at right are in a forested area, as dense canopies are difficult to resolve at 23 cm (same issues as the steps above) but in any case are <1 m. Some of these spikes are real, as in the week between acquisitions in October, the leaves fell off the trees. There is a slight color shift left-right within the pit, indicating either that some sub-pixel fine registration needs to be done, or it could indicate a real change due to slope instabilities.


At center here is a tailings pile created by a conveyor that dumps the rocks at the center of the cone. If there was any large lateral misfit between the maps, this cone would show different colors due to the apparent change in height caused by the misfit. That is it is uniformly red indicates that the registration is very good, better than subpixel. These maps were made with no ground control.


Another assessment of accuracy comes from the photogrammetry software. This image reports the difference between the airborne camera locations that I input into the software and the locations the software determined were the best fit to the bundle adjustments and interior orientations for one of the Ft Knox acquisitions. The standard deviation of difference is about 14 cm for this map, and this is typical of all of the ones that I have made. GPS processing of in-flight position results in a noise floor of 5-15 cm, and I believe this to be the major systematic error in the system. Given that the repeatability (a proxy for the random error) of my own results seems to be in the 15 cm range or better based on the analysis above, there is some consensus in these metrics that 15 cm is a suitable number to report for both systematic and random variation in this technique. I believe that the level of error puts these methods into the Phenomenal range, compared both to the current state of the art and to the needs of most studies of Arctic landscape.

That's captures the essence of my work thus far in determining the accuracy and precision of my methods. While I plan more rigorous analysis this winter, I suspect the final numbers will still be in the 15 cm range, or better.

 

Example Applications

Here I present some screenshots of my work this fall to give a sense of the applications that this technique can be applied to.


Here are the flight tracks for most of the missions I flew this fall. The only major one not here is the Big Ram lake mission, which is located about 20 miles west of where the green lines first intersect (Arctic Village) about 3/4s of the way to McCall Glacier.


Glaciers


This is my orthoimage draped over Google Earth terrain. I made this the night I acquired the images and sent it out to a few friends. Several responded about what beautiful weather it was, which I took as a great compliment since they did not realize that this is not an oblique image taken from the window of the plane but rather 3D visualization of the data. (The weather actually sucked over the glacier and I was happy to make it home!)


Here is a photo comparison I made at McCall Glacier, comparing a 1958 image taken by Austin Post to ones I've taken in 2003 and 2008.


Here is the 2013 comparison of taken from the perspective of the 1957 Austin Post photo location. Mouse-over the image to see the change between 2008 and 2013.


Here is my orthoimage of McCall Glacier (flowing left to right) from September 2013. There was a recent snowfall, depositing up to about 20 cm of snow. Mouse-over to see the difference image with the 2008 lidar.


Here is a difference image with 2008 lidar. This steep headwall represents one of the most challenging photogrammetric environments -- steep topography with fresh snow cover in deep shadow. Mouse-over to see the 2013 ortho drape. With this technology we now have the ability to understand the dynamics of both small hanging glaciers (which are essentially impossible to measure in the field) but also small changes to them. The reds indicates change in their small termini, as nunataks become exposed.


Here is geomorphology in action. A small rockslide occurred here in 2008, where meltwater from the Hanging Glacier above washed out a lot of rock down the side of the moraine, spilling it onto McCall Glacier. Mouse-over to see the drape. With time-series of these data, we have the opportunity to measure landslides of nearly any size, and over a 10 or 100 year period we could come up with excellent rates of such erosion.


This is an ice-cored moraine that McCall Glacier (red area to the right) left behind. As ice gets exposed to the sun, it melts and the large rocks covering it slide, exposing more ice, like a conveyor belt. There is essentially no way to measure such dynamics on the ground, as this is quite treacherous terrain. Mouse-over to see the drape.


Just one of my favorite views of McCall Glacier, taken from the perspective of our camp site, in the foreground.

 

Permafrost Degradation


One of my main goals of the past 4 years was to make these maps on the Seward Peninsula for the National Park Service. Shishmaref is on the coast at left near the bay and Cape Krusenstern is at top right. Here I have overlain my flight lines from early September over Google Earth, along with one of the orthophoto mosaics I made of the lower block. These are pretty huge blocks resulting in enormous files, but it worked!


Mouse-over to see the comparison. Can you tell which is my mosaic and which is Google's? There is a slight misregistration between the two (about 5m), but I suspect that mine (the rollover image) is the one that is more accurately placed, despite not having any ground control.


One of our goals in this research was to find and track thaw slumps. These occurs when ice rich ground beneath the tundra begins to melt, causing the tundra to detach and slide downhill. Mouse-over to see a shaded relief image. You can see shrubs at the river bottom for scale.


These detachments are only several meters high at the headwall, and you can see smaller bits of tundra rafting downhill. This one is located in the Noatak River valley. Mouse-over to see the drape.


Here is a close-up of the a headwall and polygonal ground above it. Mouse-over to see a slope image (colors represent slope, with red indicating steep slopes, like at the headwall) that highlights the detail and the fact that we are measuring polygonal ground topography. The surface width of these polygon edges are roughly 10-50 cm and probably less than that in height. My previous analysis of accuracy resulted in saying the noise floor is about 15 cm, but clearly these features are being resolved despite being below that noise level. Probably a better way to phrase my accuracy assessment is to say that the repeatability of my maps is about 15 cm (like a systematic error) but the noise level is about 5 cm (like a random error). But I really don't care that much about how to phrase it, the point is that this is awesome data that can resolve ice wedges and sorted circles! Perhaps even cooler is that if my accuracy assessment is correct, these maps may virtually eliminate the need for ground control in vegetation or other field surveys, since they are likely more accurate than the GPS most people take into the field and I can make 100,000,000 measurements in the same time it takes someone to land out here and set up a tent to acquire their 20 ground measurements.

 

Coastal Erosion


I flew a few flight lines by eye over Cape Kruzenstern after finishing the Shishmaref work. The cape is characterized by a series of shorelines that get older as they get further back. Archeologists apparently like to hang out here, because there is a similar age progression of artifacts, as ancient peoples kept moving forward as the new benches formed. So I guess I'm not measuring coastal erosion here, but coastal aggradation. In any case, mouse-over the image to see colorized terrain, and the fact that the shoreline is all one color blue, as it should be since it all at the same elevation -- sea level.


Here are some of the benches. Mouse-over to see the terrain.


Here's some erosion! Tundra bluffs are melting and falling onto the beach, forming a blob that will soon get washed away. We can not only measure coastal erosion with these techniques, but count driftwood!


The benches are pretty subtle features.


Here is coastal erosion in action, with tundra clods falling off the bluff as the ice beneath as the waves undercut them and ice within them thaws.

 

Vegetation Mapping


This is an oblique view of a long-term vegetation mapping site in the Sheenjek River near Last Lake (at left). I acquired this for a colleague on my way back from McCall Glacier. It took about 30 minutes to acquire. Mouse-over to see the shaded relief.


Any idea what this is? Mouse-over to see that it is a small pingo with spruce trees on it! In open canopies like this, we can actually measure tree-line migration and increase in biomass. This level of detail also means that we can create synthetic stream-channel networks that will route around the trees, something impossible to do with the currently available USGS DEMs at 30 m resolution which barely see forests let alone individual trees. I also think its cool that there's something going on with the vegetation surrounding the upper arc of the pingo, not sure what causes that. I thought at first maybe the spring building the pingo might be making the ground too mushy for spruce trees, but that would make more sense if it was the lower arc. Maybe the pingo used to be much bigger? But I'm not sure pingos can get smaller like that, unless its completely disappeared and reformed?


This is part of a large block on the Wind River near Big Ram Lake, about 20 miles west of Arctic Village, for colleagues at USFWS. I was very happy to see how well this turned out, because the entire valley was snow covered and beneath a thick overcast when I flew it. Such conditions create flat light and can eliminate the contrast needed for photogrammetry. My awesome camera combined with some prior expertise with photography allowed me to capture the scene with no real issues. This means that we can measure changes in snow depth, among other things, provided they are at least about 10 cm or more different. Mouse-over to see colorized terrain.


Here is a closeup of the the Wind River. Mouse-over to see the image drape. Those are trees on the left. You can see a lot of noise where the open water exists (use the mouse-over), but the almost-featureless white snow on the river banks still creates accurate topography. This tells me that we should be able to map ice-jam dynamics on larger rivers.


Here is a hippie farm run by some friends just outside of Fairbanks. Mouse-over to see a shaded relief.


Here is a closeup of one of their fields. They were able to identify the plants from this. The forest looks like a wall because this is an orthophoto, which is an image which tries to be a top view everywhere; that it looks like a wall is actually a compliment to the accuracy of the processing. Mouse-over to see the topography. In the Arctic, the combination of using an orthophoto to identify plants and high resolution terrain to determine the solar aspect and slope on which they grow, could presumably allow one to get at very site-specific plant-microclimate relationships. Though I haven't done it yet, I see no technical obstacles to doing this with NDVI by adding a second camera to do NIR imaging simultaneously. Imagine making such measurements throughout the year, including winter -- unprecedented! (as far as I know, but I'm a glaciologist...).


Here is an individual image from the camera. Here you can see the edges of the trees. The only pixel that is truly orthometric here (if that really is semantically possible) is the one in the center of the frame looking straight down (and because the plane is always a bit tilted even that pixel doesn't exist). The process of orthorectification uses the underlying topographic measurements to pretend that I flew straight over the top of every pixel. This process forces all pixels to be the same GSD (ground sample distance), which means the resolution of some pixels gets degraded so that they all can be the same without oversampling.


Here is a crop of that image at 100%. You can tell their broccoli has gone to flower. I included this image just to make the point that when draping an orthoimage over topography, some image resolution is lost by the visualization process. There is a lot more detail in the the raw photos themselves, which one can use for identification of objects, plants, caribou, etc, but you cant accurately measure distances with them.

 

Caribou (and other critters)


Any idea what those things are sticking up from the landscape? Mouse-over to see that they are caribou. I acquired this for some colleagues at NPS to see if we could use this technology to rapidly and accurate assess caribou numbers and biomass.


Here is a slope image. Because the caribou stick out, they have steep slopes. The blue and yellow swirly patterns are polygonal ground. You can convince yourself you can tell which are sitting or standing.


Here are the same caribou, visualized this time using a technique designed specifically to highlight weirdness on the terrain. This is about the extent of my fooling around with this, but I think it gives a good sense of what's possible.

 

Forest Fires


Here is a top view of a forest fire scar from 2013 at Lake Minchumina. The lake shoreline is at the bottom of the image.


This the same burn. This acquisition was part of an early system development test, I flew the lines by eyeball from high elevation so it's not as high res or accurate as most everything else here. But it's clear from this that we can accurately map the extent of the fire as well as its severity. I haven't processed the data to full resolution, but its conceivable to me that we could actually count the number of burned trees lying on the ground.

 

Infrastructure Development and Degradation


Here is a new building under construction next to mine (the green area down from center) at UAF. Mouse-over to see the drape.


You may notice that the sides of the buildings have no image texture -- this is a deliberate part of an orthometric image, as described previously. There are other 3D visualization techniques that will allow for the sides of the buildings to show up, but I haven't messed much with them yet. Mouse-over to see the colorized height. They have already backfilled the foundation. I have a time-series of construction progress here.


Here is where I park my plane, on the East Ramp of Fairbanks International Airport. With a time-series of these, I can easily tell who's been out flying. Mouse-over to see the colorized terrain. I haven't looked for it yet, but if there was any subsidence more than 5 cm anywhere on this airport, I'm sure I could find it. Same goes with roads, parking lots, etc.

 

Mining


I didn't know much about mining before I began using Ft Knox gold mine near Fairbanks as a test site, but it's fascinating. There is so much going on here that we can measure, it's like fast-forwarding geomorphology. This is the main pit, mouse-over to see colorized topography.


Here what the pit looks like from below...


Here is a blast field all loaded up and ready to explode. Mouse-over to see what it looked like after the explosion a week later.


Here's the same blast field in colorized topography. Mouse-over to see how the explosion changed the surface.


Here's that same blast field at top left. Mouse-over to see the change. Also pay attention to how little change there is in other areas, this should give you a good sense of how accurate these maps are.


Here is a difference image with the same view, showing the change over 1 week. I measured the amount of volume increase (red) and loss (blue) in the feature at bottom right, and they came out to within 0.05% (!!!) indicating that no rock was removed here, it was just pushed downhill by a bulldozer (which can be seen at work in closeups).


This enormous landslide occurred at the Bingham Canyon mine in Utah. It would be easy to measure the volume of this slide considering how well we can measure tiny ones, but it might also be possible to predict such hazards through time-series of high-resolution maps.


Here's a much smaller slide at Ft Knox. Mouse-over to see the mass movement as a difference image.


Mouse-over to see before and after.


Here we can see the cause of the slide -- a small fault running diagonally up through the center of the image and to the top of it. Mouse-over to see the rock thrust up from beneath the lower part of the fault causing the rock above to subside and the rock on the sides to rotate towards it. Notice how little change there is in the rest of the scene.


Here I believe we are detecting precursory motion to the landslide. I have stretched the color scale of this difference image of two maps made 2 and 3 weeks before the slide to highlight subtle detail. The fault is below the bench in the lower-left quadrant of the image, running diagonally downward, highlighted by small upward motion shown in yellow beneath it. Mouse-over to see a shaded relief of a single map to identify the fault line's location is coincident with this motion. If I made these maps at 10 cm instead of 23 cm, and done some subpixel co-registration, I think we should be able to find subtle motions like this even more clearly.


Here's a closeup of the parking lot. You can tell who called in sick that day...


Here is a lot where they store their tires. Mouse-over to see how these piles changes. I counted nine missing tires.


I found six of the missing tires here. Mouse-over to see the change.


I found several tire dump sites like this one, but not the missing 3 tires. You can tell they were not put here because there was no change between images. Mouse-over to see the difference image.

This mine is only 20 miles from Fairbanks, but there is no technical reason this same level of accuracy could not be achieved at any mine site, even mines much smaller than this one.

 

Conclusions

I think this is only the beginning for DSLR Photogrammetry (or DSLeRP...?). Presumably things will only get better from here as a both supply and demand increase. I'm currently using the fastest computer I could buy as of August 2013 and I'm still maxing out 32 fast processors for days at a time for a single large project, but as computer processing continues to improve these processing times will continue to decrease. Digital camera technology continues to improve at a rapid pace as well. I think we are probably already pretty close to achieving the limits of accuracy, but future advances in GPS positioning technology may reduce our current noise of about 15 cm down to perhaps 5 cm. In any case, if nothing changed I think we have an awesome tool to exploit here for years to come and I think within a few years this technology will be commonplace within the commercial geophysical community, and this can only help drive overall accuracies, workflow, and costs in the direction we want. I think our biggest gap right now is in actually handling and using these data scientifically, as even small orthoimages and DEMs at these resolutions can overwhelm the capabilities of the computers on most peoples' desks. So I think there is a lot of potential to do some really neat science with these airborne methods, and I'm excited about the future. In the meantime, I think the best thing we can do is exploit our current capabilities to help drive demand and lay down baseline measurements now so that we can begin understanding changes in the earth surface at unprecedented detail in the near future.

 

Click here to see 2014 results



 

 

 

 

(c) 2010 Matt Nolan.