1. LiDAR Analysis of Mixed-Species, Multi-Cohort Stands
Dr. John A. Kershaw, Jr.
Western Mensurationists
June 20, 2016

2. Topics for Western Mensurationists
Influence of Data Sources and Model Approaches on Predictions from LiDAR
Influence of Cell Size on model predictions and forest-level estimates
Stand Typing using “shape-based” metrics
Individual Tree Segmentation
Copula-Based Diameter Prediction Model

3. Collaborators Graduate Students Faculty Research Scientists
Charles McPhee, UNB
Hunter Roberts, UNB
Rei Hache, UMaine
Elias Ayrey, UMaine
Faculty
Dr. Aaron Weiskittel, Umaine
Dr. Mark Ducey, UNH
Prof. Jasen Golding, UNB
Research Scientists
Dr. Mike Lavigne, Can For Service
Elizabeth McGarrigle, NS DNR

4. Boots Off-the-Ground?

5. Noonan Research Forest
1500 ha
Softwood, Intol. Hardwood, Mixedwood Stands
Many degraded stand structures
100 m sample grid
Angle count sample at each grid intersection (2M BAF)
Every 10 years
1480 ACS points
“Lam” Plots
110 fixed-area PSPs
11.28 m radius
50 X 50 mapped plots
Mixed HW/SW multistrata stand
Pure BS single strata
Pure EH single strata

6. Penobscot Experimental Forest
1619 ha
Softwood and mixedwood stands
50 years of silvicultural study
117 permanent sample plots
Nested design
15.95 m radius OS plot
7.98 m radius Sapling plot

7. First Principles Own your Work
(con)FUSION doesn’t give me much flexibility or confidence in the metrics extracted
Wrote my own R code
Canopy extractor
Ground extractor
Flexible metric generator
As did my grad students for their work
I’m a systems kind of guy
If you can’t explain you model have you really developed a model or merely summarize data?

8. First Principles

9. LiDAR – A New Paradigm in Forest Mensuration
Data-Rich Source of Information
Lots of Interest/Effort on Various Aspects of LiDAR and Forest Mensuration
Dominated by Imputation-Based Data Summary Approaches
Many Paradigms/Myths being Generated
randomForest Imputation
Fixed-Area Plots
Coregistration Requirements

10. but,….We Need a Few Curmudgeons to ask some Critical, Hard Questions

11. Influence of Data Sources and Model Approach on Predictions from LiDAR (Hache, Kershaw, Weiskittel)
Two Data Sources
Noonan Research Forest, UNB
Penobscot Experimental Forest, Umaine
Same Forest Types
Different underlying ground data
LiDAR acquisition differed in pulse density
Questions?
LiDAR data extraction
Fitting of randomForest vs NLME
Predictions of randomForest vs NLME
Fixed-area vs Angle Count Samples
What factors influence model performance?

12. LiDAR data NRF PEF Leaf-off condition Riegl LMS Q680i laser scanner
724 m altitude
28-degree scan angle
50% overlap
6 pulses/m2
PEF
Partial leaf-off condition
Optech Gemini 246 sensor
1982 m altitude
20-degree scan angle
25% overlap
1 pulse/m2

13. Models randomForest LiDAR extraction radii NLME Stand Type (HW, SW,MW)
q45LiDAR
q50Canopy
LiDAR extraction radii
10, 15, 20, 25, 30 m about sample point
NLME
Random effects on b0 and b1
Stand Type as random factor

14. Goodness-of-fit with Local Data
Data Source
Fit
NLME
Random Forest
Model
Test
Statistic 10 15 20 25 30
NRF R2
0.68
0.67
0.64
0.63
0.61
0.85
0.84
0.83
0.82
RMSE
33.4
34.4
35.5
36.4 37
22.8
23.5
24.3
24.9
25.5
Mean Bias
-0.3
-0.2
-0.1
-0.11
-0.07
-0.05
-0.01
Abs Error
25.9
26.8
27.8
28.5
29.1
17.7
18.2
19.1
19.6
PEF
0.43
0.53
0.54
0.8
45.4 41
40.6
40.8
41.1
26.7
24.7
23.8
0.33
-0.5
-0.6
0.16
0.08
0.15
0.34
32.2
31.8
31.7
18.7
16.6
17.3
16.7
17.1

15. Equivalence of Local Prediction
Equivalence Test
LiDAR Extraction Radius 10 15 20 25 30
Field
NLME{N/N+P/P}
H{0}
Reject
Mean Diff.
-0.26
-0.28
-0.25
-0.21
-0.18
Std Dev
34.5
34.95
35.92
36.74
37.37
Region of Sim
15%
10%
RF{N/N+P/P}
-0.10
-0.05
-0.04
0.01
-0.07
23.09
23.57
24.36
24.83
23.58
NLME(N/N+P/P)
0.16
0.22
0.21
0.11
14.96
14.84
14.87
15.18
15.41
25%
20%

16. Goodness-of-fit with Nonlocal Data
Data Source
Fit
NLME
Random Forest
Model
Test
Statistic 10 15 20 25 30
NRF
PEF R2
0.4
0.48
0.47
0.46
0.35
0.37
0.39
RMSE
46.3
43.2
43.1
43.6
44.1
48.3
47.5
46.7
46.6
Mean Bias
3.56
5.36
7.12
6.93
7.84
14.4
14.51
15.03
14.52
14.36
Abs Error
36.2
32.7
32.4
32.9 36
35.4
35.2
0.64
0.58
0.59
0.62
0.57
0.53
0.51
0.49
35.8
38.7
38.1
38.5
36.6
38.9
40.7
41.8
42.5
5.42
11.7
8.2
7.42
-5.39
-11.9
-14.1
-15.4
-15.5
27.8
29.7
29.4
29.8
28.9
31.4
33.5 34

17. Equivalence of Prediction Nonlocal Data
Equivalence Test
LiDAR Extraction Radius 10 15 20 25 30
Field
NLME{N/P+P/N}
H{0}
Reject
Mean Diff.
5.27
11.21
8.12
7.38
6.45
Std Dev
36.34
37.41
37.66
38.23
38.72
Region of Sim
25%
40%
30%
RF{N/P+P/N}
-3.87
-9.86
-11.89
-13.01
-13.24
38.43
39.44
40.07
40.78
20%
35%
45%
NLME{N/N+P/P}
Not rej.
5.54
11.06
8.37
7.59
6.62
10.91
10.08
9.66
50%
≤50%
-3.60
-9.59
-11.63
-12.89
-13.07
13.34
17.99
17.92
18.15
RF{N/N+P/P}
-5.38
-11.27
-8.16
-7.38
-6.51
17.94
19.31
18.24
18.21
18.14
-9.14
-21.07
-20.00
-20.48
-19.69
18.04
21.42
20.78
20.22
19.45
-3.77
-9.81
-11.85
-13.11
-13.17
19.84
20.51
20.59
21.24

18. Fixed-Area versus ACS

19. So what factors influenced model performance?

20. Effect of Cell Size on Area-Based Estimates
Total Volume as the response variable
NLME and RF models
3 variables: maxCanHT, q45Height, and Height of Max LiDAR Density
Fitted to the 100 m grid ACS plots on Noonan
Applied to LiDAR cells 2, 4, 5, 10, 20, 25, 50, and 100 m squares
Summed to stand levels
Total and 95% CI based on var between stands

21. 100 m Grid Map

22. 2 m grid Map

23. Forest-Level Totals

24. Broad Forest Typing with LiDAR
Hunter Roberts interested in detecting hemlock groves
Small patches, mostly ha in size
Detect and map these across Noonan
Used LiDAR height and intensity distributions
Weibull shape and scale parameters
Skewness
Broad Types:
EH, Sp-Fir, Mix HD, HD/SW, SW, Water, Other (Meadows, CCs, young PCT)

25. T14 Map

26. A Canopy View

27. Individual Tree Segmentation
From the canopy view it is clear that individual trees can be “seen”
Several methods to extract or “segment” individual trees have been proposed
Two approaches
Crown profile probabilities
Layer Staking

28. Crown Profile Probabilities

29. Crown Profile Probabilities
Develop a library of crown profile parameters for species of interest
Start with highest LiDAR point – assume this marks a tree
Look at next highest point
Fit a crown profile through the two points and compare shape parameter with library
Estimate the probability that shape belongs in library
Decide based on minimum probability if the two points belong to the same tree of two different trees
Assign point to a tree
Calculate crown radius and assign all points below to that tree
Repeat until all points assigned to trees

30. Noonan 50X50 plot 1

31. Noonan 50X50 plot 2

32. Results 60% of trees in mixed-species, multi-cohort stand
78% of trees in dense, spruce-dominated single cohort stand
Algorithm slow
8 hours on my mini-supercomputer
Can’t detect subcanopy trees
Could apply a density filter to identify “hotspots” of LiDAR returns to possibly segment these trees
Requires another pass through all points (16 hours for 50 X 50 m plot)

33. So what? What’s this all good for?
Tree segmentation gives us tree height distributions
Ground-based data gives us a copula of H-D
Inverting the copula, we can take the LiDAR-derived height distribution and sample to obtain a predicted diameter distribution

34. LiDAR-derived H-D Distribution

35. Noonan DBH Distribution

36. Summary of Pros/Cons PROS CONS Coregistration not required
Area-based dbh distributions quite good
Mean tree volume equivalent
Intensive computational resources required
About 60% of tree extracted
Total volume underestimated by about 15%
Not species-specific

37. Questions?