Automated extraction and validation of Stone Pine (Pinus pinea L.) trees from UAV-based digital surface models

Figures & data

Figure 1. Proposed methodology for the automated extraction of individual Stone Pine trees.

Figure 2. Overview of the study area.

Figure 3. Orthomosaic of the study area, and sample images from the study area.

Figure 4. Test sites selected from the study area (The red polygon in each test site presents the coverage of the reference data collected during fieldwork; the blue triangles represent the positions of GNSS receivers deployed for HMLS measurements).

Figure 5. (a) DSM of test site #3. (b) Probability map of local maxima. (c) Probability map of local minima. (d) Enhanced probability map of local maxima. Notice how unnecessary local maxima information is suppressed in (d) compared to (b).

Figure 6. (a) DSM of test site #3; (b) Probability map of orientation symmetry combined with enhanced local maxima; (c) Output binary map.

Figure 7. (a) DSM of test site #3; (b) Automatically extracted regions; (c) Stems of individual trees.

Figure 8. Georeferenced HMLS point clouds. (a) Test site #1; (b) Test site #2; (c) Test site #3; (d) Test site #4; (e) and (f) Sample 3D view of the HMLS point cloud and a detailed view of Stone Pine trees.

Figure 9. (a) a sample HMLS point cloud; (b) Detection of ground points; (c) segmented objects; (d) the output of cylinder fitting process; (e) and (f) Cylinder fitted to a single stem using least squares.

Table 1. Parameter settings of the proposed methodology (the operators $\left[.\right]$ and $\left[.\right]$ rounds a floating number to the lowest preceding and the largest succeeding integer value, respectively).

Methodology Section	Parameter	Value
Enhancing Probability Map of Local Maxima	$h_{\mathrm{m}\mathrm{a}\mathrm{x}}^1$ – $h_{\mathrm{m}\mathrm{a}\mathrm{x}}^8$	10 cm−80 cm
	$h_{\mathrm{m}\mathrm{i}\mathrm{n}}^1$ – $h_{\mathrm{m}\mathrm{i}\mathrm{n}}^{10}$	1 m−10 m
Combining Orientation Symmetry with Enhanced Local Maxima	r^1 = $\left[50\mathrm{c}\mathit{m}/\mathrm{GSD}\right]$	5 pixels
	r^K= $\left[6\mathit{m}/\mathrm{GSD}\right]$	61 pixels
	${\alpha }_1$ – ${\alpha }_4$	2–5
	σ	5 pixels
	$c_{\mathrm{otsu}}^1$ – $c_{\mathrm{otsu}}^{10}$	2–11
Extraction of the Boundaries of Trees	λ1 – λ2	1–1
	ρ	0.1
	υ	−0.7
	τiter	200
	τdist= $\left[40\mathrm{cm}/\mathrm{GSD}\right]$	5 pixels

Figure 10. Pixel-Based (left column) and object-based (right column) performances of the proposed methodology.

Table 2. The elapsed time of each section of the proposed methodology.

Process	Elapsed Time min: sec
Enhancing Probability Map of Local Maxima	00:08 (≈2%)
Combining Orientation Symmetry with Enhanced Local Maxima	00:36 (≈8%)
Extraction of the Boundaries and Stems of Trees	06:38 (≈90%)
Entire Processing	07:22 (100%)

Figure 11. Visual outcomes of the proposed methodology (The results of the test sites #1-#4 are provided from top to bottom, respectively; the input DSM, the extracted regions for trees, and the visual output of the accuracy assessment are given from left to right, respectively).

Table 3. Pixel- and object-based performance of the proposed methodology.

Test Site	Pixel-Based Performance(%)			Object-Based Performance(%)
	Precision	Recall	F1-score	Precision	Recall	F1-score
#1	77.1	86.2	81.4	96.7	87.4	91.8
#2	85.3	94.7	89.8	99.6	98.2	98.9
#3	83.1	95.7	89.0	100	99.6	99.8
#4	82.4	95.5	88.5	100	97.5	98.7
Overall	83.1	94.2	88.3	99.3	96.3	97.7

Figure 12. Graphical representation of the error vectors (note the scale of error vectors shown).

Table 4. 2D positional accuracy of automatically extracted stem locations.

Test Site	# of Stems in Reference	# of Stems Automatically Detected	# of Stems Automatically Selected for Comparison	Min. (m)	Max. (m)	Median (m)	RMSE (m)
#1	199	129	121	0.06	5.77	0.51	1.11
#2	285	219	217	0.06	5.26	0.32	0.86
#3	283	253	248	0.08	0.99	0.30	0.37
#4	197	174	171	0.05	3.09	0.40	0.52
Overall	964	775	757	0.05	5.77	0.36	0.72

Table 5. Pixel-Based comparison of the previous approaches and the proposed methodology (the best and second-best performances across approaches are illustrated in green and orange colors, respectively).

Table 6. Object-Based comparison of the previous approaches and the proposed methodology (the best and second-best performances across approaches are illustrated in green and orange colors, respectively).

Figure 13. Comparison with the previous studies (The results of the test sites #1-#4 are provided from top to bottom, respectively; the visual outputs of the approaches in Popescu and Wynne (2004), Swetnam and Falk (2014), Dalponte et al. (2015b), and ours are given from left to right, respectively; green, red, and blue colors represent TP, FP, and FN pixels, respectively).

Table 7. Pixel-Based comparison of the previous approaches based on orientation symmetry and the proposed methodology (the best and second-best performances across approaches are illustrated in green and orange colors, respectively).

Table 8. Object-Based comparison of the previous approaches based on orientation symmetry and the proposed methodology (the best and second-best performances across approaches are illustrated in green and orange colors, respectively).

Popescu, S.C., and R.H. Wynne. 2004. “Seeing the Trees in the Forest.” Photogrammetric Engineering & Remote Sensing 70 (5): 589–604. doi:10.14358/PERS.70.5.589.

 Web of Science ®Google Scholar

Swetnam, T.L., and D.A. Falk. 2014. “Application of Metabolic Scaling Theory to Reduce Error in Local Maxima Tree Segmentation from Aerial LiDar.” Forest Ecology and Management 323 (7): 158–167. doi:10.1016/j.foreco.2014.03.016.

 Google Scholar

Dalponte, M., F. Reyes, K. Kandare, and D. Gianelle. 2015b. “Delineation of Individual Tree Crowns from ALS and Hyperspectral Data: A Comparison Among Four Methods.” European Journal of Remote Sensing 48 (1): 365–382. doi:10.5721/EuJRS20154821.

 Web of Science ®Google Scholar

Data availability statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.