Tree Segmentation

Below, I lay out the process I worked through for my Master’s of Science in Biology to capture photogrammetric point clouds of 45, 000 trees, to segmented

out the crowns of each tree, classify each tree by species, and then map, model, and measure each tree to predict it’s biomass. Once done, I used the biomass

predictions to model the effects of forest stand structure and topographic factors on biomass distribution across an island forest landscape.

Field Location

I completed this field work in the Cypress Hills Interprovincial Park along the southern border of Alberta and Saskatachewan. Shown below is a True Colour Composite image from data captured by the Landsat 8 Operational Land Imager and overlaid on a rayshaded digital elevation model constructed from radar data from the Shuttle Radar Topography Mission (SRTM).

Field Work

To capture the imagery, I flew a DJI Phantom 4 drone. I laid out the flight plan in a serpentine manner to maximize the overlap between photographs both side-to-side between passes and front to back

Raw Drone Photographs

There are three dominant tree species in the Cypress Hills, each with a distinct growth strategy and preferred environmental niche.

Trembling Aspen (Populus tremuloides)

Lodgepole Pine (Pinus Contorta)

White Spruce (Picea glauca)

Orthomosaic

After capturing around 500 photographs per site, I georeferenced and orthorectified the images to correct

lens distortions and prepare an orthomosaic of the entire forest stand

3D Point Clouds

Using a Structure-from-Motion algorithm, I reconstructed from the photographs a 3D point cloud of the trees I flew over

From the point clouds, I created 3 models to normalize the surface.

First, a Digital Surface Model (DSM) of the top of the canopy

Second, a Digital Terrain Model (DTM) from the visible ground points, and then an interpolated (filled-in) surface of the areas unseen beneath the tree canopies. That is why some part of the model seem jagged and flat - they are the mathematical best guess of terrain I was unable to see directly.

Third, I subtracted the DTM (ground elevation) from the DSM (canopy elevation) to create a Canopy Height Model (CHM) of tree height above the ground. Kind of like pulling a wrinkled tablecloth off a table (removing the ground) and leaving the silverware on the flat table surface (leaving behind only the trees).

Digital Surface Model (DSM)

Canopy Height Model (CHM)

Below is a Bird’s Eye View of a segment of each model: left to right is the DSM, DTM, and CHM

Below here is a alpha transition from the original orthomosaic in true colour, to the rendered CHM

Groundtruthing

To verify that the height observed from the drone in the CHM are accurate, I went out to the trees themselves, manually marked them in the orthomosaic, and measured their heights.

Individual Tree Segmentation

Using the field data to feed a machine learning algorithm, each tree segment was grown in the point cloud form a local maxima (the treetop) to create a spatial polygon.

Using the colour imagery in the orthomosaic, I extracted spectral signatures of each tree species and applied an, aptly named, Random Forest classifier to assign a species to each of the 45,000 tree that I did not manually classify myself

Here’s the full process below with another alpha transition, this time from orthomosaic in true colour to segmented and classified polygons by species:

And here is a scatterplot comparing the field height I measured to the height as measured by the drone. They are somewhat accurate overall with an R2 value of 0.56 for a total of 1129 trees, which is reasonable enough from which to predict biomass relationships at a broad scale. Species-wise, Aspen was the least accurate ( R2 = 0.40, RMSE = 2.66m, n = 349), with Spruce ( R2 = 0.62, RMSE = 2.71, n = 351) and Pine (R2 = 0.43, RMSE = 1.99, n = 429).