Why Worry?

• Predictive models of vegetation-environment relationships are an important first step in mapping vegetation classes at regional scales.

• There are many modeling techniques available for building maps.

• Because different models may produce different maps, attention to model-choice is important.

Objective

• Compare Random Forest (RF) and Gradient Nearest Neighbor (GNN) modeling techniques with respect to:

1) classification accuracy

2) class area representation

3) spatial patterns

The West Cascades

Asheville

The West Cascades

Mapping Methods

– We map NatureServe's Ecological Systems – Using GNN and RF models built from

– 8,109 records from our plot database– and mapped explanatory variables, selected from 115

possible layers

– At a 30m grain

Landsat Bands, transformations, texture

Climate Means, seasonal variability

Topography Elevation, slope, aspect, solar

Disturbance Past fires, harvest, insects and disease

Location X, Y

Soil Parent Material e.g., Ultramafic rocks, sandstone, basalt, etc.

Methods: Random Forest• One Classification Tree:

|MTM100 < 21.1095

DEM < 679.825

MTM300 < 13.874

SLPPCT < 29.3858

YFIRE < 3.15519

CANOPY < 42.085930725365 3415

4323

8793 4847

5767

Methods: Random Forest

• A whole forest of classification trees!

• Each tree model is built from a random subset of explanatory variables and input data.

• When the model is applied to mapped data, each tree ‘votes’ on which Ecological System a pixel should be.

|TM100 < 22.9069

TM100 < 19.0223 FOG < 0.5

TM100 < 33.82934108 5120

5977 86398622

|ANNHDD < 2766.21

SLPPCT < 10.3216 STDTM100 < 16.1739

ANNHDD < 3469.43STDTM100 < 46.7235

9148 5675 3517 4192 4607 5832

|TM200 < 22.9549

STRATUS < 201.108

TM200 < 35.1356

4156

5559 8269

8694

|MTM300 < 28.9494

MTM300 < 13.8683 IDSURVEY < 0.5

MTM300 < 43.8013

3922 4672

6770 8947

5136

|DECMINT < 48.1696

MTM700 < 27.8564

DECMINT < -214.708DECMINT < -276.567

R5700 < 145.622

R5700 < 185.313

4280 36687896 4737

6104

8506 5086

|TM100 < 22.9069

TM100 < 19.0223 DISTNF:b

4108 5120

5833 8480

Methods: Adjusting The Random Forest Map

• The RF model may favor some classes to maximize overall accuracy. – Over-mapping some systems– And under-mapping others

• We can map the votes for the under-mapped systems, creating single-system probability maps.

• ...which can be used to expand their area in the final map.

Methods: Adjusting The Random Forest Map

Single System Map of: Mediterranean California Dry-Mesic

Mixed Conifer Forest

(2) calculate

axis scores of pixel from

mapped data layersstudyarea

(3) find nearest-

neighbor plot in

gradient space

(4) impute nearest

neighbor’s value to

pixel

Methods: GNN

gradient space geographic spaceCCA

Axis 2(e.g., Climate)

CCAAxis 1

(e.g., elevation, Y)

(1)conductgradient

analysis ofplot data

The Maps

Without Landsat TMRF

RF_ADJ

GNN

With Landsat TMRF_TM

RF_ADJ_TM

GNN_TM

Results

RF:

83.8%91.8%

RF_TM:

82.5%91.0%

RF_ADJ:

82.9%91.0%

RF_ADJ_TM:

82.5%90.4%

GNN:

82.5%89.7%

GNN_TM:

78.6%87.5%

Top #: Accuracy, Bottom #: Fuzzy Accuracy

MC

Me

s M

ixC

on

MC

DM

Mix

Co

n

NP

Dry

PS

ME

NR

M P

IPO

% o

f are

a

0

5

10

15

20

PlotRF

RF_ADJRF_TM

MC

Me

s M

ixC

on

MC

DM

Mix

Co

n

NP

Dry

PS

ME

NR

M P

IPO

% o

f are

a

0

5

10

15

20

PlotRF

RF_ADJRF_TM

RF_ADJ_TMGNN

MC

Me

s M

ixC

on

MC

DM

Mix

Co

n

NP

Dry

PS

ME

NR

M P

IPO

% o

f are

a

0

5

10

15

20

PlotRF

RF_ADJRF_TM

RF_ADJ_TMGNN

GNN_TM

MC

Me

s M

ixC

on

MC

DM

Mix

Co

n

NP

Dry

PS

ME

NR

M P

IPO

% o

f are

a

0

5

10

15

20

Plot

MC

Me

s M

ixC

on

MC

DM

Mix

Co

n

NP

Dry

PS

ME

NR

M P

IPO

% o

f are

a

0

5

10

15

20

PlotRF

MC

Me

s M

ixC

on

MC

DM

Mix

Co

n

NP

Dry

PS

ME

NR

M P

IPO

% o

f are

a

0

5

10

15

20

PlotRF

RF_ADJ

MC

Me

s M

ixC

on

MC

DM

Mix

Co

n

NP

Dry

PS

ME

NR

M P

IPO

% o

f are

a

0

5

10

15

20

PlotRF

RF_ADJRF_TM

RF_ADJ_TM

Class Area Representation

RF

RF

_AD

J

RF

_TM

RF

_AD

J_T

M

GN

N

GN

N_T

M

0

20

40

60

80

% o

f la

ndsc

ape

Largest Patch Indexp < 0.001

RF

RF

_AD

J

RF

_TM

RF

_AD

J_T

M

GN

N

GN

N_T

M

0.0

0.2

0.4

0.6

0.8

edge

s /

cells

Edge Densityp < 0.001

RF_ADJ:Accuracy OK

Area good

RF:Most Accurate

Area lousy

Coarse-grained

RF_TM:Accuracy OK

Area lousy

RF_ADJ_TM:Accuracy OK

Area good

GNN:Accuracy OK

Area good

GNN_TM:Least accurate

Area good

Fine-grained

? ? ? XX X

Conclusions

• No single map is perfect.

• Each has its strengths.

• ...and weaknesses.

• The maps vary most with respect to class areas, and pattern.

• Unfortunately, we lack reference data for pattern.

• And yet, we still need to choose ‘the best’ technique for the GAP vegetation maps.

Discussion

• If you were choosing which methods to use to build a GAP map, which one seems best to you?

Why?• Acknowledgements:

– USGS GAP analysis program– LEMMA research group at Oregon State

University

Landscape Ecology Modeling Mapping & Analysis