Comparing nonlinear and nonparametric modeling techniques

Download Report

Transcript Comparing nonlinear and nonparametric modeling techniques

MODELLING TECHNIQUES
FOR MAPPING IN
FOREST INVENTORIES
Gretchen Moisen, Tracey Frescino
US Forest Service, FIA
Whining from the
applications side of
the fence
Data, data
everywhere...
...and not a thought
to think.
Outline
1.
2.
3.
4.
5.
Need for new info
Data
Models
Maps and applications
Now what
Need for new information:
Traditional reports
• Inventory status and
trends in forested
ecosystems nationwide
1928 McSweeney-McNary Act
1978 Renewable Resources Act
1998 Farm Bill
• Regional estimates of
forest area, tree volume,
growth and mortality
Research to develop new products………
In addition to estimates of
population totals…….
• Make maps! Show how
forest resources are
distributed throughout the
landscape
• Use those maps: wildlife,
fire, harvest….
• Automate data retrieval,
visualization, and
analysis tools
• Build web-based
delivery systems
• Just do it
Need for new information:
Development of an interdisciplinary system
QUESTIONS
FIELD DATA
• Dialogue with users, define
problems
• Build data base, prepare data
DIGITAL DATA
•
MODELS
EVALUATION
DELIVERY
Build and test models
• Test products in real
applications
• Get it out and get feedback
Outline
1.
2.
3.
4.
5.
Need for new info
Data
Models
Maps and applications
Now what?
Data
Six Ecoregions
•
•
•
•
Regional diversity
Forested ecoregions
Within state bounds
Sample across all
owners
Data:
Plot-level Response Variables
Catagorical:
• Forest/nonforest class
• Select forest type
Continuous:
• Basal area
• Biomass
• Crown cover
• Growth
• QMD
• Stand age
• TPA
• Volume
Data:
Sample plots
UT1
F: 821
NF: 533
UT2
F: 829
NF: 491
Data:
Sample plots MT1 (F: 1277 NF: 294)
MT2 (F: 1612 NF: 2108)
Data:
Sample plots
AZ1
F: 712
NF: 135
Process:
Many RS-based Predictor Variables
• Raw imagery: TM, MODIS,
AVHRR
• NLCD
30 m resolution
19 classes, 8 broad groups
• DEMs: elevation, aspect, slope,
hillshade, topographic class
• Spatial coordinates
• Other: Soils, TEUs, Precip
Outline
1.
2.
3.
4.
5.
Need for new info
Data
Models
Maps and applications
Now what?
Models:
Establishing relationships with predictors
• Extract data from
each layer at
each FIA location
• Build a model for
each FIA variable
Example: Tree cover ~ f(Cover-type, Elev, Aspect, Slope)
Models:
Predicting over large areas
Through the final model, use
………cover type
……….elevaton
……….aspect
……….slope
to predict
……….crown cover
over unsampled areas
Models
Response
discrete x
continuous x
interactions
Forest type
NLCD
Elevation
X,Y
Basal area
Soils
Aspect
Elev, Asp, Slope
Biomass
Slope
NLCD(others)
Crown cover
Hillshade
Growth
X
QMD
Y
Age
TPA
Volume
Models:
Simple Benchmarks
• Discrete variables
Yhat=NLCD class
• Continuous variables
Yhat=mean(Y) w/i
NLCD classes
• SIMPLE..is it enough?
Numerous model building tools…..
GAM
CART
p


f (x)  g  a0   f i (x i ) 
i 1


f (x)  am , forx  Rm
1
MARS
f ( x)  a 0 
f
K m 3
ijk
 f (x )   f
K m 1
i
i
K m 2
(x i , x j , x k )  ...
ij
(x i , x j ) 
ANN
 l

K


f l (x)   l  w2 kl k   w1 jk x j   j    k .
 k 1

 j 1



Model Test Using Simulated Data
CART
LM
X1,…, X10 ~ Unif(0,1)
Y = 2sin(π*X1*X2) +
.4(X3-.5)2 +
.2(X4) + .1(X5)
GAM
MARS
ANN
Residual Plots: BIOTOT in UT2
NLCD
GAM
CART
MARS
ANN
Overview of Analyses
Responses Continuous: BIOTOT,
CRCOV, QMDALL,
STAGE
Discrete: F/NF, F1/F2
Predictors
NLCD, AVHRR,
topography, UTMs
Technique
NLCD, GAM, CART,
MARS, ANN
Evaluation
Criteria
Continuous: RMSE, PWI,
RHO, Runtime
Discrete: PCC, Kappa,
Runtime
Evaluation criteria
Modeling
Continuous: RMSE, PWI,
RHO, Runtime
Discrete: PCC, Kappa,
Runtime
System
Data preparation
requirements?
Nest modelling and
prediction within a GIS?
User
Do the maps help solve
real problems?
Can users drive?
Models fuel estimation and
EDA as well?
Outline
1.
2.
3.
4.
5.
Need for new info
Data
Models
Maps and applications
Now what?
Building maps:
F/NF, BA, CRCOV, VOL, STAGE, QMD
Fishlake Applications
Build and test large-scale
models predicting…
- Presense of cavity
nesting birds
- Elk calving sites
…using FIA-generated
maps of habitat predictor
variables
Tom Edwards, Randy Schultz
Applications:
Web Delivery
• JPEG preview
• PDF map
• Build a map
(Generate a map
based on
user-defined
criteria)
Warning: These maps are prototypes under development.
They are NOT final products
Tracey Frescino,
Frank Spirek
http://www.fs.fed.us/rm/ogden/index.html ► Techniques Research
Applications:
Interactive Display Environment
• Interactive tool for
visualize,
summarize, and
query resource
information
Tracey Frescino
Outline
1.
2.
3.
4.
5.
Need for new info
Data
Models
Maps and applications
Now what?
Future Work:
Refining Interdisciplinary System
QUESTIONS
• Continue dialogue
FIELD DATA
• Refined retrieval system
DIGITAL DATA
• New predictor variables
MODELS
EVALUATION
DELIVERY
• Streamlined modeling box
• NFS test applications
• Refined web-based delivery
Future Work:
New Applications
• Prediction for new
applications: assessment
of resources lost to wildfire
or I&D, extension to other
wildlife species
• Improved precision on
population estimates
• Improved analyses