Overview and Lessons from the Alberta Project

Transcription

Overview and Lessons from the Alberta Project
Enhanced Forest Inventory
A case study in the Alberta foothills
Roger Whitehead & Jim Stewart
CFS, Canadian Wood Fibre Centre
Glenn Buckmaster
West Fraser Mills, Hinton Wood Products
Mike Wulder, Joanne White,
Gordon Frazer1 & Geordie Hobart
CFS, Pacific Forestry Centre
1Current affiliation – GWF LiDAR Services, Victoria, BC
1
Outline

Site & Data Sources

What we did

Model predictions

Validation & discussion

What we‟re working on now
2
Study Area & Data Sources

Hinton FMA
– West Fraser Mills, Hinton WP
– 988,870 ha;
– 185,000 AVI polygons

LiDAR & AVI data
– Alberta ESRD /WF- HWP

Ground Calibration data
– HWP‟s established network of
Permanent Growth Sample Plots
3
LiDAR data

Alberta ESRD
provided HWP with
full FMA coverage
– small footprint (30 cm)
– 0.75 +/- points/m2
– multiple discrete return
(max 4 returns)
– collected 2004–2007
– pt cloud, CHM, DEM
4
Data Cloud  Canopy Metrics

Used USDA FS freeware
package FUSION/LDV to
– tile, grid & calculate canopy
metrics on 25m X 25m grid
13,665,234 grid-cells
– „forest type‟ assigned from
AVI polygon level
• Conifer. Deciduous, Mixed
5
Ground calibration/model training…

WF-HWP maintains >3200 PSP
 empirical yield curves

we used 735 of those plots to
train prediction models
– date of last measure & GPS quality.
– used HWP mensuration / calculations for
top ht, volumes, BA & trees/m3
– biomass from Lambert (2005) & Ung
(2008)

separate models for each forest
type using area-based approach
– conifer, deciduous, & mixedwood
6
LiDAR–based Prediction of Attributes

Used Random Forests (“R”)
to create prediction models
– Top height, Co-dominant
& Mean height
– DBHq & BA
– Total Volume &
Merchantable Volume
Total Above Ground
Biomass (tonnes/ha)
– Above Ground Biomass
– Mean piece-size (trees/m3)
Partners: WFM - Hinton Wood Prod.; Alberta SRD; CFS–PFC; UBC
7
Mapped as GIS raster layers


25m cell level
AVI Polygon level
33 m3/ha
384 m3/ha
14 m3/ha
247 m3/ha
331
m3/ha
Merch. Volume (m3/ha)
For ~1 million ha
Hinton FMA
525 m3/ha
276 m3/ha
164 m3/ha
0 m3/ha
8
9
10
11
12
So… are any predictions “correct’?
Weight-scaled volume from 272 cutblocks harvested since LiDAR acquisition
compared to predictions from LiDAR vs. Cover Type Adjusted Volume Tables
Block Size
(m3 X1000)
Source of
Prediction
Predicted Volume
– Scaled Volume
Statistically
significant?
<5
n = 138
LiDAR
CT Vol. Table
-6.7%
-23.7%
No
Yes
5 – 10
n = 76
LiDAR
CT Vol. Table
+1.8%
-17.4%
No
Yes
10 – 15
n = 25
LiDAR
CT Vol. Table
-1.2%
-22.3%
No
Yes
15 – 20
n = 15
LiDAR
CT Vol. Table
-4.4%
-23.5%
No
Yes
>20
n = 18
LiDAR
CT Vol. Table
+6.6%
-17.4%
No
No
Vol.T. underestimated scaled volume by 19.8%
LiDAR overestimated scaled volume by 0.6%
Information courtesy
Hinton Wood Products
13
Why are the Volume Tables so far off ?

Volume Table predictions
– rely on AVI polygon height
– don‟t handle within-polygon
variability well

Polygon-level LiDAR
predictions
– don‟t rely on age or SI50
– aggregate cell-level
predictions

What about the bias?
– “It‟s the operational
planner‟s fault”
14
The problems with using existing PSPs

The 735 PGS plots we
used were…
– not well-distributed across
variation in LiDAR metrics
– biased to young, even-age
conifer stands

customized sample design
should  better models
Frazer et. al, 2011
Partners: WFM - Hinton Wood Prod.; Alberta SRD; CFS–PFC; UBC
15
Structurally-guided sample design
PGS plots used
“Structurally-guided” sample
White et al, 2013
16
Required sample size will depend on…



acceptable error
confidence level required
# of “forest types” modeled
ACCEPTABLE ERROR
CONFIDENCE LEVEL
REQUIRED SAMPLE SIZE
(per “forest type”)
5%
95%
386
10%
95%
96
10%
90%
68
White et al, 2013
17
Sampling Intensity influences cost…

HWP has maintained 3202 plots since
1950s, specifically  empirical yield curves

LiDAR prediction models used only 735 of
these plots  better volume predictions

LiDAR-specific sample design should
– more cost-effectively  still better results
– train models to predict attributes wanted
• CBD? ht to live crown? understory? etc.
18
What we’re working on now…

Evaluating model improvements with
structurally-balanced sample design

“Best Practices” Guideline (31 March, 2013)
– support “standards” for LiDAR-enhanced inventory

Support acceptance of LiDAR in Forest
Management Plans & AAC determination
– Complex yield curves from LiDAR rasters (fireorigin stands)  Woodstock  TS Analyses
19
What we’re working on now…

High resolution LiDAR &
digital imagery with SGM
– HWP re-flight proposal
– UBC/CFS to explore…
• “growing” the inventory
• object-based predictions
• species & “product profiles”
Images courtesy – Steve Platt, Strategic Group, Campbell R, BC
20
Strategic  tactical  operational
LiDAR rasters

Link to FPInnovations
Value Maximization &
Decision Support
– net value @ cell &
polygon-level
– cost-benefit across full
value chain
 Discussion session #3
21
Linking Block Planning to Mill Needs
3 - 5 TPM
< 2 TPM
3 - 8 TPM
22
Linking Compartment Planning to Mill Needs
Column A
Column B
Column C
Column D
Column E
Trees Per Metre
Potential
Planning Unit
Range of 68% of the log profile
harvest area (ha)
MEAN
Lower
Higher
GALL 2
1,919
2.1
0.9
3.4
CONK 21
3,577
2.4
1.1
3.6
CONK 20
3,692
2.6
0.9
4.3
CONK 8
2,733
3.5
1.4
5.7
GALL 13
1,973
3.8
1.7
6.0
CONK 4
1,939
3.9
1.5
6.4
BURL 7
2,351
4.0
1.7
6.3
CONK 10
2,088
4.1
1.9
6.3
BURL 8
1,869
4.5
1.9
7.0
BURL 11
2,348
4.8
2.3
7.4
BURL 6
1,978
4.9
2.0
7.8
BURL 21
2,893
5.7
3.0
8.4
CONK 14
4,832
6.3
4.2
8.4
BURL 1
2,049
6.4
3.7
9.2
23