Recent developments with aerial camera technology, computer
hardware, and software solutions have led to development of important new
methods of producing digital elevation models (DEM) directly from aerial
photography. Here’s why this is important to the geospatial profession and you.
The Traditional Way
Digital elevation (“surface”) models can be generated from overlapping aerial
imagery. Historically, this is accomplished using one of two methods. First,
they are manually digitized from stereo imagery using a digital photogrammetric
stereo workstation (also called softcopy) and typically to pre-defined grid
spacing (say every 150′, for example). Alternatively, they are automatically
generated at similar pre-defined grid spacing from stereo imagery on softcopy
workstation using principles of photogrammetry and then manually edited in
stereo to correct blunders. Any grid spacing could be used, but they are generally
not any denser than 50-300′ because of the time required to autocorrelate,
effort to edit this many points, and because this density of points was
sufficient for most mapping applications. These processes have been used for
years and work well, but they can be expensive because of the manual editing
Digital Cameras Change
With the advent of digital aerial cameras the autocorrelation of 3D (X, Y, and
Z) points from aerial imagery is being revisited. The success, speed, and
accuracy of the autocorrelation depends upon many factors. The clarity of the
image patch being analyzed in each overlapping image is quite important. Scans
of film imagery are grainy, may have scratches, dust, or other artifacts. Plus,
one scan of an area may have very different tonal qualities from frame to
frame. All of these factors limit the effectiveness of autocorrelation in film.
But digital camera imagery is superior to film scans in many important respects
improving the accuracy and effectiveness of autocorrelation. There are no
scratches, dust, or similar artifacts in a digital image. Also, tonal qualities
from image to image is much more consistent and predictable. Another very
important advantage, not apparent to the eye, is digital camera imagery has
greater bit depth, meaning each pixel can have any one of 4096 (12-bit) shades
of a single color compared to film scans that may have only 256 (8-bit)
possible shades. This difference, not perceived with our human vision, is like magic to an autocorrelator.
Aerial photography typically contains lots of content that appears very
uniform: pavement, forest canopies, cornfields, water bodies, shadows, etc.
When an autocorrelator tries to determine if an image patch from the middle of
a parking lot is the same as a different patch from the another image of the
same general area, it has trouble. Many pixels in one patch look very similar
to many pixels in the other patch. Therefore, the positional accuracy and
statistical “sameness” of the autocorrelated point can be quite low. One reason
for this is the number of possible color values for each pixel is limited to
256 for scanned film. A shadow is typically black. Only black. All black. The
parking lot is uniform gray all over. With only 256 possible shades of “black”
or “gray” there just isn’t enough dissimilarity between both patches.
However, with 12-bit digital imagery there are thousands of
shades of gray or black that any pixel can acquire. An autocorrelator can look
at a shadow and instead of seeing black, it now sees a hundred shades of black.
Instead of seeing a uniform gray parking lot, it sees a parking lot with thousands
of undulating areas of non-uniform patches. The autocorrelator can find truly
identical patches in overlapping images with much greater certainty and with
superior positional accuracy even in areas that look poor in contrast with
Two other significant developments have made important contributions to the new
autocorrelators. The first is the way digital aerial camera systems acquire the
imagery differently. Because film is no longer used, it may be more convenient
and less expensive to increase the overlap between images. More overlap means
there is an increasing likelihood the position of any particular point in an
image can be triangulated. The more triangulation possible for a point, the
more accurate its position in space can be determined. These new
autocorrelators are especially relevant to pushbroom cameras like the Leica
ADS80 system. It acquires imagery
in strips using 3 separate sensors so a single strip of imagery is actually 3
separate images with almost 100% forward overlap with each other. Normally,
frame-based camera systems acquire photography using 60% forward overlap. The pushbroom
sensors use 100% overlap among three distinct images on every flight strip,
meaning the number of points that occurs in 3 or more images is extremely high.
Therefore, dense DEMs can be generated. In fact, DEMs created from this new
breed of software are super-dense
(approximately 1 point for each pixel). So if the GSD is 6″, you could
theoretically have one DEM point on the surface every six inches. This is much
more dense than a typical LiDAR model and may rival LiDAR in positional
accuracy. At point densities this high, realistic 3D models of the earth and
structures are possible.
Second, these advantages would be impotent if the last major advancement had
not been achieved: fast, inexpensive, and distributed computation. Generating
DEM points every 3-12″ across the surface of the earth requires major CPU
cycles and resultant files are extremely large. With distributed computing the
rapid generation of the surface models using these new autocorrelators is now
possible. Software typically used for processing LiDAR data can be used to edit
the large models, and create bare earth and surface models just as they are
with LiDAR-generated DEMs. Another advantage over LiDAR-only is these surface
models are always accompanied with high-resolution imagery that can be
used to verify the quality of the surface models. Additionally, both the high
quality imagery and DEM are acquired in a single flight mission with a single
piece of equipment. This may be the least expensive way to produce these two
products. Many clients may not have the resources to procure LiDAR and produce
1-2′ contours, may already have stereo imagery of that area that could be used
to produce the similar contours much more affordably.
One important limitation of these new terrain services is they use passive
sensors, that is, they capture reflected, incident light. LiDAR sensors, on the
other hand, are active sensors and detect reflections generated using their own
light. The advantage to LiDAR then is some vegetation canopies can be
penetrated and the surface of the earth beneath measured. This is not possible
using only incident light. DEMs produced with these new autocorrelator
techniques, though extremely dense and accurate, may have voids in the bare
earth models beneath vegetation canopies.
These new autocorrelation methods hold much promise for an accurate, dense, and
affordable DEM solution for those interested in acquiring imagery and new
terrain data without flying both LiDAR and digital orthos. Stay tuned!
Automatic Terrain Extraction (eATE) and BAE Systems Next-Generation Automatic Terrain
Extraction (NGate) are a couple systems that are now capable of producing these
next-generation image-based DEMs. Aerial Services is evaluating this software now.
Results will be reported here when complete. In the meantime, if you are interested in learning more
about such services, please feel free to contact us.