Transcript FOREST INVENTORY BASED ON INDIVIDUAL TREE CROWNS

FOREST INVENTORY BASED ON
INDIVIDUAL TREE CROWNS
Jim Flewelling
Western Mensurationist Meeting
June 18-20, 2006
OUTLINE






Perspective
Crown Segmentation
Tree predictions
Sample Frame
Estimation
Summary
Perspective







Aerial Surveys date from 1920’s and 30’s
Images for stand boundaries and attribution.
Individual crown locations and delineation since
late 1980’s.
Attribution – training process.
Research process – match trees and Images.
Lidar – huge improvement.
Limited use of sampling theory at tree level.
Crown Segmentation,
Delineation & Attribution



Identify individual crowns.
Locate center points.
Delineate crown boundaries.



(non-overlapping)
Attribute species.
Attribute height.
Individual Tree Crown (ITC) Delineation
Deep
shade
threshold
Valley
following
Rulebased
system
1995
Courtesy of Canadian Forest Service
Delineated Individual Tree Crowns
Courtesy of Canadian Forest Service
At ~30
cm/pixel,
81% of the
ITCs are the
same as
interpreted
crowns
Delineated and Classified ITCs
Courtesy of Canadian Forest Service
Predictions


Much attribution without specific data.
Goal:



DBH’s, total heights, correct species, counts.
Per-acre statistics: BA, volume, biomass.
Empirical predictions:


per-acre level is common.
tree level (matched data) as resolutions and
technology improve.
Tree Predictions - Data


Ground-measured tree crowns.
Rough plot alignment


Crown Images and Actual Trees aligned.



correlated distributions.
Research: 100% mapped, special locators.
Fixed area plots for inventory.
Sample plan.
Matched Trees & Crowns
The tree points
are then
matched up with
the tree polygons
to create
regressions used
for the inventory
calculation.
©ImageTree Corp 2006
Matched Trees and Crowns
Matched Trees and Crowns

Errors in Segmentation






One delineated crown = 2 neighboring trees.
One real tree wrongly divided into 2 crowns.
Trees entirely missed.
Ground vegetation seen as a tree.
Understory trees don’t contribute.
Technical improvements, but no absolute
solution.
Sample Frame - Ground or Map?
Individual
stand on
LiDAR image
after tree
polygon
creation. A
polygon now
surrounds
every visible
tree crown.
©ImageTree Corp 2006
Sample Frame - Ground





Traditional forest sampling.
Plots are installed on the ground.
Stand boundaries recognized in field.
Hope the stand area is correct.
Awkward to use crown information.
Sample Frame - Crown Map



Data-rich environment.
Fixed-area plots.
New or different challenges:




sample locations
tree & crown matching
stand boundaries
edge bias
CROWN BASED SAMPLE FRAME
REQUIREMENT




Trees linked to segmented crowns.
Linkage must be independent of
sampling.
BUT
Linkages need not be physically correct.
Suppressed trees need not be linked if
sampled another way.
TREE MATCHING SCHEMES

Subjective


Crown Captures ALL in tessellated area.


potential for significant bias
Expand crown area.
Trees compete to be captured.


Consider DBH, height, species …
Ground plot size > crown plot size.
Sample Locations
(Crown Map as Sample Frame)

Select fixed-area plot centers

Ground


- usual compromises plus boundary issues.
Crown map
rigorous random selection process.
 Difficult to find on the ground.


Both:

Unequivocal tree & crown matching?
Crown-based sampling scheme





Crown delineation, all stands, all crowns.
Select Sample Stands (in strata)
Randomly locate 2 plot centers on map.
GPS to those locations.
Install stem-mapped ground plots.
Challenges - plot location.


Map error + GPS error of several meters.
Process to find ground plot center on
crown map.



(x, y) plus angular shift.
Force ground plot to include selected pt?
Accept the random deviation?


Ground plot center outside of stand?
Altered probability density.
Challenges - Edge Effects

Edge bias correction SIMPLE




“Tree concentric method.”
Computer finds area of “tree-center plot”
within stand boundary.
More efficient than field-based methods.
Plot location - random error.


Minor alteration in probability density.
Computer can correct.
Estimation

Research focus is deterministic.


Attempt to remove uncertainty.
Alternative is stochastic modeling.



Each crown has multiple outcomes: trees and
species.
DBH, heights vary with outcome.
Stand prediction = sum of expectations.
Estimation (continued)

Approximate Unbiasedness (strata).


DBH distributions NOT unbiased.



Model-assisted survey estimators (regr.)
“Regression towards the mean”
Can correct for unbiased width.
Could use data from sampled stands to
improve those stands.
Summary


Attractive technology.
Best for which forest types





Irregular spatial tree distributions.
Some multi-species situations.
Detailed predictions without sampling all
stands.
Useful spatial information.
Sampling theory has been under-utilized.
Acknowledgements




Many slides were provided by Francois Gougeon
and are courtesy of Natural Resources Canada,
Canadian Forest Service.
Other slides were provided by ImageTree
Corporation.
Mike Wulder, Canadian Forest Service.
Adam Rousselle, Vesa Leppanen, Olavi Kelle,
Bob Pliszka (Falcon Informatics).
Resources






2005 Silviscan http://cears.fw.vt.edu/silviscan/
2004 ISPRS Laser-Scanner for Forest and http://www.isprs.org/commission8/workshop_las
er_forest/
ImageTree Corp. www.imagetreecorp.com
Pacific Forestry Center
http://www.pfc.forestry.ca/index_e.html
Precision Forestry Coop (U.W.)
http://www.cfr.washington.edu/research.smc/
THANK YOU
questions?