Tree crown segmentation in three dimensions using density models derived from airborne laser scanning

Figures & data

Table 1. Summary of the field data used for validation

Plot	Stem density (numberha^−1)	DBH Average (mm)	DBH Min (mm)	DBH Max (mm)	Proportion Pine (%)	Proportion Spruce (%)	Proportion Deciduous (%)
1	613	297	49	469	0	97	3
2	420	258	50	397	2	84	14
3	643	298	47	514	0	93	7
4	462	258	86	405	0	97	3
5	348	287	175	435	14	86	0
6	523	282	43	498	52	48	0
7	643	172	42	465	31	62	7
8	485	281	55	421	0	82	18
9	352	314	42	507	84	3	14
10	436	301	72	509	0	94	6

Figure 1. Overview of the segmentation algorithms using templates (i.e. tree crown density models).

Table 2. Algorithm summary

Input	Unit	Action	Output
Files with ALS data of manually delineated trees crowns	Do for ALS data from each tree crown in the training dataset	For the associated crown density raster: loop the raster and aggregate laser return values to raster cells	Templates
ALS data	Do for each raster cell within the laser-scanned area	Set raster cell value to the height value of the laser return contained in the raster cell with the maximum distance from ground	DSM
DSM	Do for each raster cell of the laser-scanned area	Set raster cell value to one if the maximum height (DSM) is above a height threshold (2 m) followed by morphological closing	Canopy area raster
DSM; canopy area raster	Do for each raster cell within the laser-scanned area	Set raster cell value to the DSM. If the value of the DSM is zero but the canopy area raster is one, interpolate a height using the smallest window needed to find a neighboring raster cell with canopy DSM values above zero	Canopy Height Model (CHM)
CHM; ALS data; templates (one for each tree class)	Do for each raster cell within the laser-scanned area	Derive a tree crown model using surrounding data and calculate B value, one for each of the templates, and save the maximum B value in the raster cell	Model Fit Surface (MFS)
MFS	Do for each raster cell within the laser-scanned area	Calculate a new value of the model fit surface using a Gaussian kernel, which produces a smoothed model fit surface	Smoothed MFS
Smoothed MFS	Do for each raster cell within the laser-scanned area	Watershed segmentation: starting at each raster cell, create a path by traveling to a neighboring raster cell in the steepest direction	Initial segmentation
CHM; ALS data; templates; Initial segmentation	Do for each raster cell within the laser-scanned area	Derive a tree crown model using surrounding data only from locations within the same segment, and calculate B value, one for each of the templates, and save the maximum B value in the raster cell	CMF Surface
CMF Surface	Do for each raster cell of the laser-scanned area	Derive a model fit surface using a Gaussian kernel	Smoothed CMF surface
Smoothed CMF surface	Do for each raster cell within the laser-scanned area	Watershed segmentation: starting at this raster cell, create a path by traveling to a neighboring raster cell in the steepest direction	Final 2D segmentation
Final 2D segmentation; ALS data	Do for each crown segment within the laser-scanned area	Find the tree center position, tree height, crown area, and other variables derived from laser data within a crown segment	File with tree attributes
ALS data; Final 2D segment	Do for each crown segment within the laser-scanned area	Derive 3D points with model fit coefficient	MF strings
MF strings	Do for each 3D point of the model fit strings within each 2D segment	Find cluster centres using the mean shift algorithm	3D points with cluster identification

Figure 2. The templates created in the training phase using low-altitude ALS data.

Figure 3. (a) Model fit (MF) strings with colors set by the model fit coefficient (B value) and gray points showing the original point cloud, (b) the original point cloud with colors set by the identification numbers of 3D segments. the original point cloud was selected using a polygon obtained from 2D segmentation using high-altitude ALS data.

Figure 4. DSM with 0.05 m resolution as the background image, and tree crown segments from the 2D segmentation algorithm using ALS data from low altitude (150 m above ground level) as the input. Field measured tree positions shown as yellow dots. the boundary of the field plot (80 m diameter) is shown with a green circle.

Figure 5. Clustering of 3D point cloud produced from high-altitude ALS data (1450 m above ground level).

Figure 6. Histograms showing tree detection results from 2D segmentation using high-altitude ALS data (white bars for all field trees and gray bars for field trees that were linked to trees detected in the ALS data).

Figure 7. Histograms showing tree detection results from 2D segmentation using low-altitude ALS data (white bars for all field trees and gray bars for field trees that were linked to tree detected in the ALS data).

Figure 8. Histograms showing 3D segmentation tree detection results using high-altitude ALS data (white bars for all field trees and gray bars for field trees that were linked to tree detected in the ALS data).

Figure 9. Histograms showing 3D segmentation tree detection results using low-altitude ALS data (white bars for all field trees and gray bars for field trees that were linked to tree detected in the ALS data).

Table 3. Results from tree detection using 2D segmentation and high-altitude ALS data: omission errors affect (in percentage) on the number of trees (Number), on the basal area (Basal area), and on the stem volume (Volume); commission error (in percentage) of all trees and of only tall trees (ALS tree height >10 m)

	Omission error (%)			Commission error (%)
Plot	Number	Basal area	Volume	All trees	Tall trees
1	9	5	4	9	8
2	12	10	9	7	7
3	17	7	6	10	10
4	2	1	1	6	5
5	2	2	1	13	8
6	7	4	3	11	9
7	49	11	8	17	12
8	12	6	5	12	11
9	15	3	3	34	17
10	7	6	6	13	13

Table 4. Results from tree detection using 2D segmentation and low-altitude ALS data: omission errors affect (in percentage) on the number of trees (Number), on the basal area (Basal area), and on the stem volume (Volume); commission error (in percentage) of all trees and of only tall trees (ALS tree height >10 m)

	Omission error (%)			Commission error (%)
Plot	Number	Basal area	Volume	All trees	Tall trees
1	4	1	1	5	4
2	4	3	3	4	3
3	14	5	4	6	6
4	3	2	2	7	4
5	2	2	2	11	3
6	5	3	3	7	5
7	38	8	5	15	3
8	9	4	3	5	4
9	10	2	1	46	7
10	5	6	5	10	10

Table 5. Results from tree detection using 3D segmentation and high-altitude ALS data: omission errors affect (in percentage) on the number of trees (Number), on the basal area (Basal area), and on the stem volume (Volume); commission error (in percentage) of all trees and of only tall trees (ALS tree height >10 m)

	Omission error (%)			Commission error (%)
Plot	Number	Basal area	Volume	All trees	Tall trees
1	7	3	3	17	7
2	11	10	9	21	8
3	15	5	4	23	10
4	3	2	2	15	6
5	1	1	1	35	8
6	5	4	4	26	10
7	24	4	3	41	17
8	10	6	6	28	11
9	3	1	1	69	23
10	5	5	4	31	14

Table 6. Results from tree detection using 3D segmentation and low-altitude ALS data: omission errors affect (in percentage) on the number of trees (Number), on the basal area (Basal area), and on the stem volume (Volume); commission error (in percentage) of all trees and of only tall trees (ALS tree height >10 m)

	Omission error (%)			Commission error (%)
Plot	Number	Basal area	Volume	All trees	Tall trees
1	1	0	0	61	5
2	0	0	0	67	3
3	1	1	0	66	7
4	1	1	1	62	4
5	1	0	0	67	3
6	0	0	0	73	6
7	1	0	0	81	25
8	1	1	0	72	6
9	0	0	0	92	45
10	0	0	0	77	14

Table 7. Mean values of plot results from tree detection using 2D segmentation and low-altitude ALS data: omission errors affect (in percentage) on the number of trees (number), on the basal area (basal area), and on the stem volume (volume); commission error (in percentage) of all trees and of only tall trees (ALS tree height >10 m)

		Omission error (%)			Commission error (%)
Dataset	Algorithm	Number	Basal area	Volume	All trees	Tall trees
High altitude	2D segmentation	13	6	5	13	10
Low altitude	2D segmentation	9	4	3	12	5
High altitude	3D segmentation	8	4	4	31	11
Low altitude	3D segmentation	1	0	0	72	12

Figure 10. Stem volume for all trees and for detected trees using high-altitude 2D segmentation in stem diameter intervals on all field plots.

Figure 11. Stem volume for all trees and for detected trees using low-altitude 2D segmentation in stem diameter intervals on all field plots.

Figure 12. Stem volume for all trees and for detected trees using high-altitude 3D segmentation in stem diameter intervals on all field plots.

Figure 13. Stem volume for all trees and for detected trees using low-altitude 3D segmentation in stem diameter intervals on all field plots.

Figure 14. Tree heights from high-altitude ALS data (1450 m above ground level) and tree heights from low-altitude ALS data (150 m above ground level) for automatically connected trees on 10 sample plots (2200 linked trees), coefficient of determination (R²), standard deviation of differences (STD) and mean value of differences (Bias).

Figure 15. Tree heights from individual trees obtained from high-altitude ALS data (1450 m above ground level) and field-measured trees (214 linked trees), coefficient of determination (R²), standard deviation of differences (STD) and mean value of differences (Bias).

Figure 16. Tree heights for individual trees obtained from a helicopter at low altitude (150 m above ground level) and field-measured trees (221 linked trees), coefficient of determination (R²), standard deviation of differences (STD) and mean value of differences (Bias).