Photogrammetric Point Clouds
Origins of digital point clouds:
• Basics have been around since the 1980s.
• Images had to be referenced to one another. The user had to specify either where the camera was in space or what parts of images overlapped.
– Referencing each image to one another was a fairly tedious process.
A DEM generated in 1989 from Spot-1 satellite
imagery.
• Resolution was too coarse for forest estimates.
• Computing power wasn’t
available for high resolution reconstructions.
Early work was done in the 90s on a
landscape scale with stereo satellite images
In the mid 2000s there was a revolution..
• Algorithms were developed to identify the same spot in multiple pictures.
– This meant you didn’t have to tell the computer how the two images were aligned with one another
– Now you could feed in dozens or even hundreds of
images at a time and the computer could figure out how they related to one another.
• This process is known as Structure-from-Motion
The motion refers to the “movement”
of a point from one photo to the next.
Structure from Motion (SfM)
• Allows us to take images at any angle and any distance and stitch them together with no user input.
• Traditional stereo photogrammetry required rigid flight lines, 2/3rds overlap, and no oblique photos.
– SfM actually performs better when photos are taken haphazardly, since they can capture more angles than just a top down view. SfM can be taken at any
angle, and so can penetrate the canopy better
Old stereo photos needed to be a top down perspective.
• This cut down on flight time
• Allowed for less rigid flight plans
• Reduced the need for expensive camera equipment (off the shelf cameras work well)
•
The combination of which makes SfM very cheap and particularly useful with drones.
Structure from Motion (SfM)
Autonomous drones are becoming more and more common for this kind of data
acquisition.
Many off the shelf drones allow you to program a flight path they can follow using GPS navigation.
-Fixed wing aircraft time can often exceed $150 dollars per hour.
-Drone time is practically free
Tools
• Free: 123DCatch, reconstructs models from photos taken at different angles in your browser, on a
phone, or tablet.
Tools
• Free: Insight3D, open source modeling
Tools
• Cheap ($50): Agisoft Photoscan
– Good for aerial photos
– Allows you to scale and georeference photos easily – We use it here
When taking point cloud measurements, you need to either scale or georeference
your images
• Otherwise the computer doesn’t know how big things are in the model, and all measurements are arbitrary.
Note the scale bars
With small objects, like single trees..
• Walk around the object in a full circle for about two laps taking about 30-40 pictures at different angles.
• Use a low resolution camera to cut down on processing time.
Once the point cloud is constructed the object needs to be isolated..
This can be
tricky and time consuming,
but is sure easier than
slicing the tree up into pieces
Finally, we can ‘solidify’ the object and
measure volume
One day soon..
We may be able to snap a few images in the field and determine sawlog sizes within each tree, accounting for things like stem form and bend.
This is already done at many mills, but with laser scanning
One day not so soon..
-Drones may be able to navigate through the forest.
Other uses for small object point clouds..
• Paleontology
• Archeology
• Video Games
• CSI
• Facial Reconstruction
• CGI
Aerial point clouds are derived the same way
Flying for point cloud data
• Photos that involve point cloud reconstruction generally need to be very high resolution.
– ~15cm pixels is good
• Color or multispectral photos work better than
black and white, since there’s just more information available.
- A 15cm color infrared image used to create the point cloud used in
Monday’s lab
• Photo point clouds of leaf off hardwoods are unreliable.
– The computer can see the tree and the ground beneath it in almost the same spot, and gets confused.
Flying for point cloud data
• Low resolution photos can result in height
underestimates, since the tippy tops of trees aren’t clear in the photos
– Fortunately this can be accounted for with regression modeling, provided you know there’s a problem.
Flying for point cloud data
-Many post-WW2 historical aerial photos are just on the borderline.
There’s one huge problem..
• In a closed canopy, a passive camera can’t see the ground. So we can’t know actual tree height.
How tall are these trees?
Since the photo can’t see the ground
beneath them, we don’t know
Solution 1: Use a public DEM
• The USGS puts out 10m resolution DEMs for the entire country based on mostly radar.
• You can try to use these DEMs as ground height to determine tree height
– PROS: Public 10m DEMs are available everywhere for free.
– CONS: 10m resolution isn’t really enough to measure tree height unless the ground is super flat.
-Most 10m (33ft) patches of forest will have ups and
downs of a meter or more, screwing up your height measurements.
-Shifting from one 10m cell to another can result in an instant dramatic change in tree height, which doesn’t really exist in the forest.
Solution 2: Try to find ground points
• If the terrain is flat and you have enough gaps in the canopy, you can try to find ground points in the photos.
• This can be done by filtering out vegetation by color, and by removing sharp vertical peaks.
– Pros: doesn’t require any external data.
– Cons: Doesn’t work in rough terrain or with a closed canopy.
•You can then use these points to create your own DEM
Solution 3: Use a LiDAR DEM
• LiDAR can see beneath the forest canopy, yielding a very detailed DEM.
– Pros: Really the only way to make quality measurements – Cons: If you flew LiDAR already why bother with stereo
photos?
• LiDAR is way more expensive than photos.
The DEM Problem
• Many people feel aerial photogrammetric point clouds are only useful when complimenting LiDAR.
• It’s been proposed that stereo photos would be a cheap way to ‘update’ old LiDAR point clouds.
Forestry Uses
• Obviously you can measure tree heights.
• You can correlate many other forest biometrics with point could metrics
– Point cloud metrics include things like mean point height, point height at different percentiles, canopy closure, ect..
• Regression equations can be developed to predict things like stem volume, tree density, basal area..
-Basal Area -Diameter Distributions -Stem Density (trees/acre) -Merchantable Volume -Sawlog Size - % Live Crown
-Height -Biomass/ Carbon Stock - % Hardwood/Softwood These are the biometrics Irving predicts from point clouds.
From Point Clouds to Inventories
Mean Ht
Median Ht 80th
Percentile Ht
Max Ht
1. We break the forest into plots (10x10m cells).
2. We take a bunch of
measurements of each plot’s points.
3. We plot that against field measurements on the same plots, and fit a regression line.
4. We predict forest attributes on new plots.
For example..
-Then we use the regression equation to predict stem volume on new plots, across the entire forest.
In reality we use machine learning algorithms instead of simple linear
regression..
Random forest regression
Deep neural networks
Finally, you can generate rasters from each plot’s prediction. Each plot is a raster pixel
-These rasters can be used to create stand maps, harvest plans, and forest inventories without having to go into the field.
Another problem..
• Photogrammetric point clouds can change dramatically based on flight parameters.
– Flying at different times of day can result in different
penetration of the sun into the canopy. This means that two point clouds of the same forest may end up
different!
Most predictive models therefore are only good for the specific set of photos they
were developed on
• Thanks to this problem and the DEM problem, photogrammetric point clouds have been
somewhat sidelined in favor of LiDAR.
5x5m pixels colored by tree height A SfM point cloud
An aerial photo
In conclusion:
Pretend Tom Cruise is a tree..
Who needs field crews when you’ve got evil spider drones?