**Canopy Segmentation using Airborne LiDAR**

Outline

Data Overview

Goals

Primary Segmentation Methods

Application to DHM

Application to RGB Imagery

Edge Detection and Filtering

Application to DHM

Application to RGB Imagery

Findings and Statistics

Conclusions

Future Work

Data Overview

Hemlock Forest

Ground truth data for 43 trees collected during the 2012 Share Collect

Less dense comparatively

Only about 45min away

Basic Segmentation Methods

Watershed Method

Application to WASP RGB Imagery

THANK YOU!

Data Display

Rasterization by 1 meter bins

Digital elevation model

Digital surface model

Digital height model

WASP visible imagery

Digital Surface Model

Digital Elevation Model

Goals

Segment airborne LiDAR data into individual trees

Evaluate basic segmentation methods and compare results

Provide suggestions for future work based on findings

XY Euclidean Distance

XYZ Euclidean Distance

Watershed Method

Invert height model

"Fill" inverted model

Change filter sizes based on image size

XY Euclidean Distance

Find maximum points

Calculate euclidean distance to max points using X and Y

Find minimum euclidean distances

XYZ Euclidean Distance

Find maximum points

Calculate euclidean distance to max points using X, Y, and Z

Find minimum euclidean distance

Application to DHM

XY Euclidean Distance

XYZ Euclidean Distance

Watershed

Watershed

XY Euclidean Distance

XYZ Euclidean Distance

Red Channel Only

Green Channel Only

Blue Channel Only

Watershed

XY Euclidean Distance

XYZ Euclidean Distance

Watershed

XY Euclidean Distance

XYZ Euclidean Distance

Watershed

XY Euclidean Distance

XYZ Euclidean Distance

Edge Detection Methods Applied to RGB Imagery

Matlab Sobel Filter

Envi Sobel Filter

Edge Detection on Red Channel

Edge Detection on Green Channel

Edge Detection on Blue Channel

Matlab Sobel Filter

Matlab Sobel Filter

Matlab Sobel Filter

Envi Sobel Filter

Envi Sobel Filter

Envi Sobel Filter

Edge Detection Applied to DHM

Matlab Sobel Filter

Envi Sobel Filter

1

2

3

4

5

6

7

8

9

11

13

15

17

19

21

22

23

24

Basic Segmentation Results for RGB Imagery

Watershed

XY Euclidean Distance

XYZ Euclidean Distance

779 segments total

Over segments by a factor of 18

3358 segments total

Over segments by a factor of 78

3358 segments total

Over segments by a factor of 78

Basic Segmentation Results for Red Channel

Basic Segmentation Results for Blue Channel

Basic Segmentation Results for Green Channel

767 segments total

Over segments by a factor of 18

Watershed

XY Euclidean Distance

XYZ Euclidean Distance

2096 segments total

Over segments by a factor of 49

2096 segments total

Over segments by a factor of 49

Watershed

868 segments total

Over segments by a factor of 20

XY Euclidean Distance

XYZ Euclidean Distance

4097 segments total

Over segments by a factor of 94

4097 segments total

Over segments by a factor of 94

Watershed

837 segments total

Over segments by a factor of 20

XY Euclidean Distance

XYZ Euclidean Distance

2366 segments total

Over segments by a factor of 49

2366 segments total

Over segments by a factor of 49

Edge Detection Results for the DHM

Edge Detection Results for RGB Imagery

Matlab

In general, 3% of pixels fell in borders

4602 in red channel

4639 in green channel

5161 in blue channel

Under segmentation

75% fell between a light and dark region

20% fell between two light regions

5% fell between two dark regions

Conclusions from Basic Segmentation

1. Trees with similar heights get clustered together

2. Trees with similar colors get clustered together

3. Unique heights and colors separate individually

4. RGB Imagery tends to over segment while DHM under segments

Trees of Similar Height get Clustered Together

Trees cluster by age and species

Lack of conifer trees means lack of distinct height changes

Potential Solutions

Begin with a lesser dense forest to test methods

Visually insert breaks before segmenting

Make use of point cloud to add another dimension

Trees with Similar Colors get Clustered Together

Single species' tend to have similar colors across trees

Trees of similar age tend to have similar color

Color varies with height

Potential Solutions

Attempt to zoom in to clusters of similar colors to reveal smaller changes in color

Begin with a lesser dense forest for testing

Begin with a more obviously variant forest for testing

Unique Heights and Colors Segment Individually

Segmentation methods rely on distinct changes in numbers across a field

Visually easier to see as well

RGB Tends to Over Segment while DHM Tends to Under Segment

Changes in individual tree canopies - eg. leaf clusters, branches, etc...

Rasterized form of DHM limits data

Potential Solutions

Create an averaging filter for RGB imagery before filtering

Use smaller bin sizes when rasterizing DHM

Conclusions from Edge Detection

1. Easier to detect between low ground and high trees

2. Difficult to find edges within similar clusters

Easier to Detect Between Low Ground and High Trees

More noticeable value change

Sobel is known as the "derivative filter"

Potential Solutions

Use for separating ground from trees

Use as initial phase not as sole segmentation method

Difficult to find Edges Within Similar Clusters

Not enough change within clusters

Similar trees grow closer together

Potential Solutions

Zoom in closer for similar clusters

Use leaf-off data

Use edge detection as initial clustering only

**Potential Future Work**

Collect leaf-off data for Hemlock Forest

Create DHM with smaller bin size

Attempt segmentation using spectral data

Attempt segmentation by height level

Attempt segmentation within zoomed in clusters

Attempt segmentation on blurred RGB imagery

WASP visible imagery for site

Basic Segmentation Results for DHM

58 total clusters

4 trees segmented individually

3 clusters of 7+ trees

67 total clusters

8 segmented individually

Remaining trees clusters in groups 3 or under

67 total clusters

4 segmented individually

7 trees grouped in clusters of 3 or more

Watershed

XY Euclidean Distance

XYZ Euclidean Distance

Matlab

Envi

64 pixels, or 4%, marked as edges

Under segmentation

Edge Detection Results for RGB Imagery

Envi

When segmented via euclidean distance, the Sobel results gave:

6464 segments in the red channel

7652 segments in the green channel

6464 segments in the blue channel

When segmented via Watershed , the Sobel results gave:

980 segments in the red channel

1089 segments in the green channel

1042 segments in the blue channel

Segmented using Watershed - 36 segments

10

12

14

16

18

20

25

25

26

27

27

28

29

30

31

32

33

34

35

36

37

38

39

40

all trees in bigger clusters are within 3m of each other in height

http://www.gisresources.com/confused-dem-dtm-dsm/

Ashley Miller

Committee Members: Jan van Aardt, Paul Romanczyk, Bob Kremens

**Ashley Miller**

Committee Members: Jan van Aardt, Paul Romanczyk, Bob Kremens

Committee Members: Jan van Aardt, Paul Romanczyk, Bob Kremens