Efficacy of DEM creation using Photogrammetry when encountering tree canopy
Purpose
This brief is to cover at a high level the problem posed by tree canopies when attempting to generate a DEM using photogrammetry. Photogrammetry is based on images and therefore can only derive data from what is ‘seen’ by the camera. Consequently, if all the camera sees is the top of the tree canopy, then any derived point cloud data will reflect this.
If we ‘cut’ down the trees and remove the canopy artifacts present in our cloud data, then there are visible gaps in the resulting point cloud of the surface. These gaps can be small; the result from a solitary tree, where sufficient image overlap will provide enough side angles to see the base, or if the canopy is narrow (hedge) and again the imagery overlap can see each side.
What do we do when there is larger coverage by surface area? And what if any is the difference in results between conifer and deciduous canopy? What should be the expectation provided to the consumer of the data as to the accuracy / quality of the DEM when encountering various canopies?
This high level overview of the topic is provided solely as introductory guidance for our pilots.
Method
A suitable target area of about 12 acres consisting of about 20% forest was chosen for our test. This target surface area has a surface elevation delta of about 20 meters. Time of day was about 1300 hours (GMT -5), which provided high solar irradiance of our surface target.
An eBee RTK was used and the imaging payload was Canon S110. All data is in the WGS 84 datum for entire project analysis as it will be satisfactory for our purpose. The mission altitude was flown to achieve ground resolution of 5cm / pixel. Mission overlap was set to 70% lat / 80% long and the target area flown with perpendicular legs. When the mission was started we experienced winds aloft of ~6 knots.
(Note: However, within a few minutes of flight the winds increased to 12-14 knots for the balance of the flight. The wind was quartering, so the resulting crab effect created by the crosswind effectively reduced by an unknown percentage the lat/long overlap.)
Initial processing was conducted using Pix4D. Settings that deviate from the standard 3D Map template are: Image Scale was set to 1 Original Size, point density set to High, no default classification of data points (as this will be handled separately).
Orthomosaic
Pictured below (Image 1) is the ortho derived from the mission. This image is provided to orient our readers to the target area. In the image there is visible a mixture of tree variety; conifer, deciduous (both with leaves and without).
The areas of interest are marked.
Area A: Canopy with leaves intact
Area B: Canopy mostly clear of leaves.
Area C: Solitary tree (Approximately 30m in height).
Area D: Row of trees, clear of leaves.
Image 1
Point Cloud
In the point cloud below (Image 2), we can see the gaps in our surface model below the canopy tops. Again the main areas of interest that were listed above are highlighted below. Notice the areas with missing point cloud data. Some of these areas are unresolvable by Pix4D at the GSP of 5 cm. This could be due to relative height of tree to aircraft or could be due to poor visibility of the surface by the camera positions, all leading to poor matches. Some of these areas could tightened up by changing overlap and better meteorological conditions. But the purpose here is to study the areas below canopy surfaces, so our resulting point cloud will be adequate for this purpose.
Image 2 – Original Point Cloud from Pix4D
Post processing to remove canopy
In the next image we have removed the data points representing the canopy tops and cleaned up bushes and small vegetation along the edges of the forested areas. Not removing the bushes / vegetation from the forest edges would have the effect of skewing the DEM by showing false high spots.
Image 3 – Refined Point Cloud data with RGB color
And the point cloud viewed by elevation.
Image 4 – Point Cloud colored to provide initial elevation model view.
Holes in the Point Cloud
As you can see in image 3 above, once the tree canopy in area A is removed, there are zero ground points from which to estimate the surface below. Therefore, the area underneath this canopy could be an elevated mound, or it could be a pit 500ft deep. The bottom line: it cannot be determined from the data.
In area B we have resolved some ground elements that ground truthing reveal are accurate. This area contains defoliated deciduous forest. Sufficient amount of resolution of the ground below was visible through the tree canopy so that we are able to derive some surface estimation. It is still not a complete surface or as perfect as the surrounding fields, but at least we have some ground points when creating a DEM.
In area C has sufficient imagery from many sides to remove the tree and still have an accurate point cloud of the surface. This elevation is accurate and trust worthy. The resulting hole in the data is from a low shrub surrounding part of the base of the tree.
In area D we can see that the removal of the canopy points from the point cloud leaves an accurate ground surface. Again the holes in are not due to the inability to see the ground but ground truthing reveal are a result of shrubs growing at the base of this hedge of trees. Using the shadows visible in the image one can see the original location of the trees.
DEM
From the refined point cloud represented in Image 3, an DEM was created. This DEM was created by:
decimating the point cloud at a 5-1 rate (From 28 million points to ~6.5),
and smoothing the field surface by using a sufficiently sized grid sampling rate,
removing any spikes over 0.1 meter,
cover (estimate) the missing areas the estimated curvature is used.
Image 5 – DEM generated from refined Point Cloud
Analysis
Viewing Image 5 above we can validate that the final estimated elevation model surface within section A is incorrect. The absence of reference ground points in this location forced the computer to just draw a surface from one side of the point cloud to the other. Physical verification of the elevation in Area A showed that this is a creek bottom sloping from the left side to match the grade at the right hand side. This area would require the manual collection of ground points to be added to the point cloud before a DEM is processed.
Area B is much more accurate and indeed does have the rapid elevation change as indicated on the left side. The general area here is flat as is indicated in the profile view image on the right.
Area C shows that the small area around the solitary tree filled correctly.
Area D also filled correctly, however the point cloud data representing the bushes at the bottom of these trees had to be manually removed to provide a better profile for the DEM conversion. Without this step the DEM would indicate a higher elevation at all points of the tree bases. This would lead to induced error in the generated DEM.
Takeaway / Conclusion
Pilots will need to observe their target area / surface and make field accommodations in flight planning to best handle missions in areas with trees. Some elements can be addressed by flight planning; however, photogrammetry cannot handle every situation.
Some areas and surfaces may require LIDAR as the best solution to meet the customer’s needs. Always set customer expectations so that they have an accurate and thorough understanding of what the capabilities are and any limitations that may exist.
