mapping current vegetation in the pacific coast states with gradient nearest neighbor imputation...
TRANSCRIPT
Mapping Current Vegetation in the Pacific Coast States with
Gradient Nearest Neighbor Imputation
Landscape Ecology, Modeling, Mapping, and Analysis (LEMMA) team (www.fsl.orst.edu/lemma):
Janet Ohmann1, Matt Gregory2, Ken Pierce1, Tim Holt2, Heather May2, Emilie Grossmann2
Collaborators:Jeremy Fried3, Jimmy Kagan4, Ken Brewer5, Miles Hemstrom6, Melinda Moeur7, Tom DeMeo7, Gary Lettman8, Mike Wimberly9
1USDA FS, PNW, Ecosystem Processes; 2Oregon State University, Forest Science Department; 3USDA FS, PNW, Forest Inventory and Analysis; 4Oregon State University,
Institute of Natural Resources; 5USDA FS, Remote Sensing and Applications Center; 6USDA FS, PNW, Focused Science Delivery; 7USDA FS, Region 6; 8Oregon Department of Forestry;
9South Dakota State University
Outline...
• Motivations for GNN
• Key attributes of GNN
– Canonical correspondence analysis
• Overview of GNN applications (or focus on GAP Ecological Systems mapping???)
• Areas of research interest
Primary motivation (GNN niche): supply missing data for analysis and modeling of forest ecosystems at the regional
level(not estimation of sub-population totals)
-Gradient Nearest
Neighbor Method
Satelliteimagery
GISdata
Landscape vegetation map
Fuelmodels,wildlifemodels,
etc.
Fuel maps
Fieldplots
Predicted future
landscapes
Stand and landscape
simulators (FVS-FFE,VDDT, TELSA, etc.)
Fire behavior models
(FARSITE, FLAMMAP)
Fire effectsmodels(FOFEM,
CONSUME)
Habitat maps
Etc.
Criteria for maps of current vegetation in the Pacific Northwest• Spatially complete (spatial pattern, small geographic areas)
• Consistent across large, multi-ownership regions
• Rich in detail on species composition and forest structure
– Reasonable covariance structure among attributes
• Suitable for input to stand and landscape simulation models
• Flexibility to meet a variety of analytical needs
– Too expensive to develop multiple vegetation maps at the regional scale
– Advantages in using consistent information for multiple analyses, by multiple agencies
COLA
CLAMS,GNNfire
GNNFire
GNNFire
GNN projects
GNN mapping in the Pacific Coast States• Analysis of forest policy effects
via landscape modeling and scenario analysis (CLAMS, COLA, IMAP)
• Regional risk assessments (WWETAC, RSAC 250-m study)
• Assessing fire hazard, planning fuel management, modeling fire behavior and effects (GNNFire)
• Regional inventory and monitoring (FIA, NWFP Effectiveness Monitoring)
• National Forest planning, BLM cumulative effects analysis
• Conservation planning (GAP)
• 3-D visualization using computer gaming technology (JFSP)
GAP
Overview of Gradient Nearest Neighbor Imputation
(GNN)
Components of GNN Imputation (Current Incarnation)• Statistical model:
– direct gradient analysis (canonical correspondence analysis)
– explanatory variables: satellite imagery, other GIS layers
– response variables = abundance (basal area, importance value, cover, etc.) of ‘species’ on plots (FIA, CVS, etc.)
– other vegetation variables retained with plot-map link
• Similarity measure:
– Euclidean distance in n-dimensional gradient space, where n = first n axes (usually 8) from CCA
– Axes weighted by their explanatory power (eigenvalues)
• Imputation method:
– Single nearest neighbor (k=1), to maintain covariance structure
– Summary statistics of multiple neighbors
Why Canonical Correspondence Analysis???(ter Braak 1986)
• Multivariate predictors and multivariate response
– Results in a weight for each of many spatial variables, based on its relationship with the multiple response variables
• Grounded in ecological theory, widely used in practice
– Heuristic mathematical approximation to a Gaussian response curve of species along environment gradients, but robust to violations (Palmer 1993)
– Used primarily for exploratory and descriptive studies, not prediction or hypothesis testing
• Robust to sparse data matrices (species on plots), common across regions with long gradients and high species turnover (Palmer 1993)
• Robust to multicollinearity among explanatory variables (Palmer 1993)
• Alternative models can be specified, depending on objectives
Canonical Correspondence Analysis
(CCA) Algorithms• LC = linear
combination site scores (used in GNN)
• WA scores = weighted averaging site scores
• End result: ordering of plots (LC scores) along n orthogonal axes such that most similar plots are nearest one another
Assign species scores as weighted average of LC scores
Arbitrarily assign LC scores
Assign WA scores as weighted average of species scores
Create LC scores as predicted values from
multiple regression
Start
Stop
Any change in scores?Yes
No
(adapted from Palmer 1993)
• Guru #1: yes
• Guru #2: no
• Guru #3: both
• Guru #4: dumb question
• Guru #5: why do you want to know?
Ask the ordination guru...
“Is CCA non-parametric?”
Gradient Nearest Neighbor MethodPlot data
ClimateGeologyTopographyOwnership
Remotesensing
PredictionSpatial data
Plot locations
Direct gradient analysis
Plot assigned to each pixel
Statistical model
Imputation
PixelPSME
(m2/ha)CanCov (%)
Snags >50 cm
(trees/ha)
Old-growth index
Etc...
1 11 3 7.4 0.27 ...
2 79 97 2.1 0.82 ...
Regional Plot Data
Source n (OR) n (WA)
FIA (nonfederal) 385 445
BLM (BLM) 99 --
CVS (Natl. Forest) 279 1,596
Ecology (Natl. Park) -- 52
Total 763 2,093
CoastalOregon
EasternWashington
1
2 3 4
5 6 7* 8 9
10 11 12
13 Plotlayout
(~1 ha)
Vegetation data:
• Live trees
• Snags
• Down wood
• Understory vegetation
Landsat TMBands, transformations,
texture
Climate Means, seasonal variability
Topography
Elevation, slope, aspect, solar
Disturbance
Past fires, harvest, I&D
Location X, Y
Ownership Public, private
EasternWashingto
n
CoastalOregon
Explanatory Variables
(2) calculate
axis scores of pixel from mapped data
layers
(3) find nearest-
neighbor plot in
gradient space
Axis 2(climate)
gradient space geographic space
Axis 1(Landsat)
(1)conductgradient
analysis ofplot data
field plots study
area(4)
impute nearest
neighbor’s grounddata to
mapped pixel
The imputation component of GNN
GNN model specification
SpeciesSpecies
+ structure
Structure
Image segments
(polygons), watersheds
(imagery not used)Median-
filtered√ √
Unfiltered √ √
Coarsegrain
Finegrain
Model response variables
Spatial grainof Landsatvariables
Emphasison speciescompositi
on
Emphasis on forest structure
‘Tuning’ of GNN models
Accuracy assessment (‘obsessive transparency’)• Local-scale accuracy (at plot locations) via cross-validation:
– Confusion matrices
– Kappa statistics
– Correlation statistics
• Regional-scale accuracy:
– distribution of forest conditions in map vs. plot sample
– range of variation in map vs. plot sample
• Spatial depictions:
– Variation among k nearest neighbors
– Distance to nearest neighbor(s) (sampling sufficiency)
• General findings re. GNN map accuracy:
– Excellent for regional patterns and amounts, imperfect for local sites (mid-scales???)
– Appropriate for regional planning and policy analysis
Results:
What factors are associated with regional gradients in forest composition
and structure?
What factors are associated with vegetation gradients?
(variation explained in GNN models, % of total inertia)
Subset of explanatory variables
Species composition(‘species’ model)
Stand structure(‘species-size’ model)
Coastal
Oregon
CentralOrego
n
Calif.Sierra
Coastal
Oregon
CentralOrego
n
Calif.Sierra
Topography* 2.5 11.1 8.2 3.0 6.7 7.8
Climate 8.0 19.3 13.2 8.6 11.8 12.2
Location 5.0 9.9 5.1 4.9 5.8 4.5
Disturbance:
Landsat -- -- -- 12.8 12.1 9.5
Ownership
-- -- -- 5.5 3.3 ?
Full model 10.0 22.7 16.5 23.9 19.5 18.2
* Includes elevation.
Dominant Gradients in Forest Structure (coastal Oregon)
Explanatory Variables:yellow=climate, pink = topography, white =
location, blue = Landsat TM, red = disturbance
Dominant Gradients in Forest Structure (California)
Explanatory Variables:yellow=climate, pink = topography, white =
location, blue = Landsat TM, red = disturbance
ANNPRE
ANNSWANNVP
AUGMAXTCONTPRE
DECMINT
SMRTP X
Y
TC1
TC2 TC3
TC4
TC5
ADTC1
ADTC3ADTC4
DEM
PRR
SLP CHG FIRE
Axis 1
Axis 2
QUCH2
Tree Species Positions Along Dominant Gradients (California)
red = hardwoodsblue = high-elevation
Axis 1
Axis 2
low precip.,
southeast
high precip.,
northwest
low elevation, warm, high
moisture stress
high elevation, cold, low
moisture stress
ABCO
ABMA
ACMA3
AECA
ALRH2ARMECADE27
CONU4
FRLA
JUGLA
JUOC
JUOS
LIDE3
PIAL
PIAT
PIBAPICO
PICO3
PIJE PILA
PIMO
PIMO3
PIPO
PISA2
PLRA
POBAT
POFR2
POTR5
PSME
QUDO
QUGA4
QUKE
QULO
QUWI2
SALIX
SEGI2
TABR2
TOCA
TSME
UMCA
Structure
Axis 1
Coastal Oregon: Dominant Gradients in Vegetation and Environment
low elevation Young forests, open canopies,hardwoods, private lands
Old forests, closed canopies, public lands
Species
Axis 1
Species
Axis 2interior climate
high elevationmaritime climate
Dominant Gradients in California (Scores on CCA Axes)
Axis 1,species and structure
Axis 2, species
high elevation,cold, low moisture stress
low elevation, warm, high moisture stress
high precip.,
northwest
low precip., southeast
Axis 2, structure
high greenness, high wetness, high precip.
low greenness, low wetness, low precip.
Species Gradients(Linked to Environment)
CCA axis 1(climate)
CCA axis 2(elevation)
Maritime
Interior(Valley)
Forest Types
Picea sitchensis
Tsuga heterophylla
Quercus woodlands
Abies amabilis/procera
Dry T. heterophylla/
mixed evergreen
High
Low
Paci
fic
Oce
an
(Ohmann and Gregory, in press)
Forest Structural Conditions
Young
Old
• Coast: linked to disturbance history and ownership
• Cascades: confounding of environment, disturbance, ownership
GNN Kappa Statistics for Species Presence/Absence
Tree speciesCoastal OR
Central OR
Casc.
East. WA
CA Sierr
a
OR Blue Mtns.
Abies amabilis -- 0.60 0.67 -- --
Abies concolor/grandis 0.39 0.54 0.57 0.51 0.45
Abies procera 0.34 0.50 -- -- --
Juniperus occidentalis -- 0.72 -- 0.40 0.43
Pinus contorta 0.60 0.77 -- 0.60 0.44
Pinus ponderosa -- 0.74 0.53 0.56 0.52
Pseudotsuga menzieseii 0.22 0.81 0.43 0.59 0.45
Thuja plicata 0.11 0.57 0.53 -- --
Tsuga heterophylla 0.27 0.68 0.49 -- --
Tsuga menzieseii -- 0.69 0.58 0.53 0.43
Acer macrophyllum 0.25 0.56 0.24 0.27 --
Alnus rubra 0.32 0.31 -- -- --
Example applications of GNN maps
-Gradient Nearest
Neighbor Method
Satelliteimagery
GISdata
Landscape vegetation map
Fuelmodels,wildlifemodels,
etc.
Fuel maps
Fieldplots
Predicted future
landscapes
Stand and landscape
simulators (FVS-FFE,VDDT, TELSA, etc.)
Fire behavior models
(FARSITE, FLAMMAP)
Fire effectsmodels(FOFEM,
CONSUME)
Habitat maps
Etc.
Coastal Landscape Analysis and Modeling Study (CLAMS): a simulation approach
Current policy
Alt A
Alt B
Alt C
Natural Processes
Landowner Behavior
t =1BiophysicalResponse
t =n
Landscape/ Watershed Condition
t =1t =n
Socio-economicResponse t =1
Coast Range Ecosystem
t =n
Conceptual model
Map of current vegetation needed for:
• Stand and landscape simulation models
• Response models for wildlife, aquatic, timber
• Requires:
– tree list for each pixel
– species and structure
– reasonable covariance structure
– fine-scale pattern
100-Year Change in Vegetation Classes
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1.4
1.6
1.8
2.0
2.2
1996 2046 2096
Mil
lio
n a
cre
s
Open Forest Remnants
Broadleaf Small Mixed/ Conifer
Medium Mixed/ Conifer Large Mixed/ Conifer
Very Large Mixed/ Conifer
1996(GNN)
2096 projected(base policy)
(Spies et al., in press)
Northern Spotted Owl Habitat Capability Index
• Nesting capability(patch level)
– Trees/ha >100 cm dbh
– Diameter Diversity Index
• Foraging capability(patch/landscape level)
– Canopy height
– Diameter Diversity Index
– Habitat availability within 2.2 km
1996(GNN)
2096 projected(base policy) (McComb et al. 2002)
* Wimberly and Ohmann (2004)
Coastal Oregon:
Change in Large Conifer Forest, 1936-1996 *
1936 1996 1996 1996
% change
Shrub/Tree Regeneration0 – 25 yearsOpen Mid-
heightShrub
Grass/Forb1 – 15 years
Closed Herbland
Interior Ponderosa Pine150 years or more
Late-Seral, Single-Layer Forest
Interior ponderosa pine150 years or more
Late-Seral, Multi-layer Forest
Interior Ponderosa Pine40 – 85 years
Stem Exclusion Forest
Interior Ponderosa Pine
20 – 60 years Forest Stand
Initiation
Interior Ponderosa Pine
75 – 175 yearsForest Understory
Reinitiation
Crown Fire
MixedSeverityFire
Insects or Disease
GroundFire
GroundFire
Ground Fire Insects or
Disease
Growth and Development
Growth and Development
Growth and Development
Growth and Development
Growth and Development
Growth and Development
VDDT state-and-transition model for warm, dry ponderosa pine forest
100-year change in ‘Giant Tree’ forest, Central Oregon
Cascades• Pixel data aggregated to strata defined by 5th-field
HUC, owner class, potential vegetation type, cover type, structure class
• Aggregated data input to VDDT state-and-transition models for scenario analysis
• Approach adopted by IMAP across Pacific Northwest
(Hemstrom et al.)
Fuel Models in Yosemite (GNN species-size model)
FLAMMAP InputsFLAMMAP Inputs
Canopy bulk densityCanopy bulk density
Fuel Fuel modelmodel
Moderate Moderate Fuel Moisture,Fuel Moisture,10 mph Wind10 mph Wind
Very Low Fuel Very Low Fuel MoistureMoisture25 mph Wind25 mph Wind
FLAMMAP OutputsFLAMMAP Outputs
Mapping Ecological Systems (ESs) for Gap Analysis Program
• Relatively new national classification of existing vegetation, based on floristics (Comer et al. 2003)
• Use GNN ‘species model’ to map plant communities
• Classify plots into ESs
– Guidance from qualitative descriptions and LANDFIRE ‘sequence table’ rules based on species relative abundance
– Species often are site indicators (mesic vs. dry) and/or understory species, inconsistently measured on inventory plots
– Sensitive to minor shifts in relative abundance, often caused by disturbance and/or localized microsites (within-plot variability)
• ‘Mask’ GNN (forest) model with nonforest maps from other sources
Forested Ecological Systems of the Blue Mountains Ecoregion, Eastern
Oregon
Kappa statistics for 18 forest Ecological Systems
• % correct: 52%, % ‘fuzzy’’ correct: 79%
• Many ESs are rare (only 4 are >5% of forest area)
• Many ESs are similar (mixed-conifer species)
• Confusion with nonforest ESs not shown
Ecological SystemKapp
aFuzzykappa
NRM W. larch 0.03 0.81
RM aspen 0.00 0.60
CP w. juniper 0.60 0.88
EC mesic montane mixed-conifer 0.67 0.86
NP mountain hemlock 0.00 0.00
NRM dry-mesic montane mixed-conifer 0.33 0.68
NRM subalpine 0.26 0.66
NRM mesic montane mixed conifer 0.21 0.86
RM lodgepole pine 0.05 0.61
NRM ponderosa pine 0.42 0.58
RM subalpine dry-mesic spruce-fir 0.38 0.59
RM subalpine mesic spruce-fir 0.14 0.43
RM subalpine-montane limber-bristlecone pine 0.62 1.00
MRM montane Douglas-fir 0.33 0.88
RM poor site lodgepole pine 0.08 0.68
EC oak-ponderosa pine 0.83 0.85
IMB mountain mahogany 0.00 0.00
CP = Columbia PlateauEC = East CascadesIMB = Inter-mountain basinsNP = North PacificNRM = northern Rocky MountainRM = Rocky Mountain
Area of Ecological Systems from plot-based estimates and GNN prediction
(Blue Mtns.)
0
5
10
15
20
25
30
35
Nor ther n Rocky
Mountain
Wester n Lar ch
Savanna
Rocky Mountain
Aspen For est and
Woodland
Columbia P lateau
Wester n J uniper
Woodland and
Savanna
East Cascades
Mesic Montane
Mixed-Conif er
For est and
Woodland
Nor th P acifi c
Mountain
Hemlock For est
Nor ther n Rocky
Mountain Dr y-
Mesic Montane
Mixed Conif er
For est
Nor ther n Rocky
Mountain
Subalpine
Woodland and
P ar kland
Nor ther n Rocky
Mountain Mesic
Montane Mixed
Conif er For est
Rocky Mountain
Lodgepole P ine
For est
Nor ther n Rocky
Mountain
P onder osa P ine
Woodland and
Savanna
Rocky Mountain
Subalpine Dr y-
Mesic Spr uce-Fir
For est and
Woodland
Rocky Mountain
Subalpine Mesic
Spr uce-Fir For est
and Woodland
Rocky Mountain
Subalpine-
Montane Limber -
Br istlecone P ine
Woodland
Middle Rocky
Mountain
Montane Douglas-
fi r For est and
Woodland
Rocky Mountain
P oor Site
Lodgepole P ine
For est
East Cascades
Oak-P onder osa
P ine For est and
Woodland
Inter -Mountain
Basins Mountain
Mahogany
Woodland and
Shr ubland
Columbia Basin
Foothi l l Ripar ian
Woodland and
Shr ubland
Notaly tr ees
Perc
ent
of
Fore
ste
d A
rea
OBSERVED FROM PLOTS PREDICTED FROM GNN
GNN-predicted occurrence of Juniperus occidentalis in the Central Oregon Cascades
Species model (tree species)(n=1415, kappa=0.72)
Ongoing areas of research for gradient modeling...
• More and better data (e.g., LIDAR, nonforest vegetation)
• Map evaluation tools (spatial error depictions, assessing spatial pattern)
• Comparison of alternative statistical models (imputation, other)
• Isses of spatial scale and pattern
– Fine-scale heterogeneity and pattern: What is real and what is useful? Optimal spatial resolution? Pixels vs. polygons?
– Multi-scale and hierarchical modeling approaches, e.g. partial bivariate scaling (Thompson and McGarigal 2002), hierarchical variance partitioning (Cushman and McGarigal 2003, 2004)
• Ecological research
– Ecological characterization
– Linkages to stand and landscape models for ecological analysis
• Technology transfer:
– Database and software systems to support imputation mapping
– Tools for serving and interacting with maps (web server, visualization using computing gaming software)