Land Cover Mapping for the Southwest Regional Gap Analysis Project

1
Land Cover Mapping for the Southwest Regional Gap
Analysis Project
UGIC Conference, 2003 Sherwood Hills Resort, Utah

• John Lowry
• College of Natural Resources
• RSGIS Laboratory
• Utah State University
• Collaborators Doug Ramsey, Lisa Langs,
• Gerald Manis, Jessie Kirby, Wendy Rieth, Marie
  Ducharme

2
Presentation Overview

• I. Project Description Participants
• II. Mapping Methodology
• III. Comparison with 1995 GAP Vegetation Map
• IV. Timeline Summary

3
I. Project Description Participants
What is GAP?
4
State-Based GAP Projects

• State-based vegetation classification systems
  (cover type legends)
• State-based mapping methods
• State-based mapping area

5
Salt Desert Shrub in the 4 Corners
6
Objectives for SWReGAP

• Regional vegetation classification system (land
  cover legend)
• Regionally standardized data and mapping methods
• State mapping responsibilities defined by
  eco-regional areas
• Improvements in vegetation land cover map

7
II. Mapping Methodology

• A) Nature of the Classification Problem
• Spatial Resolution Extent
• Thematic Resolution
• B) Predictor Data
• Variables Used to Predict Land Cover
• C) Training Samples
• Sample size and adequacy
• D) Properties of the Classifier
• Basics of Decision Tree Classifiers
• Tools developed

8
Classification Problem Spatial Resolution

• 85 Landsat 7 scenes
• 30 meter resolution
• Three seasons spring, summer fall

9
Classification Problem Spatial Extent

• 40 Mapping zones
• Spectrally consistent
• Eco-regionally distinct

10
Classification Problem Thematic Resolution
NatureServe Ecological Systems
NVC Formation
NVC Alliance
NVC Association
NVC Class/Subclass
1,800 units
10 units
5,000 units
700 units
300 units
MRLC 2000 Proposal
Gap Analysis Program
National Park Mapping
(Natural/Semi-natural types)
(Slide Courtesy Pat Comer, Nature Serve)
11
Ecological Systems

• Groups of plant communities and sparsely
  vegetated habitats unified by similar ecological
  processes, substrates, and/or environmental
  gradients.
• Manifest in a landscape as a spatial aggregation
  at an intermediate scale (10 ha 100,000 ha),
  persisting for 100 or more years.

(Slide Courtesy of Pat Comer, NatureServe)
12
Inter-Mountain Mixed Salt Desert Shrub System
Shadscale Shrubland Alliance
Four-wing Saltbush Alliance
13
Predictor Data Spectral Characteristics
14
Predictor Data Elevation
15
Predictor Data Aspect and Landform
16
Training Data Sources
3000 air photo interpretation sites US Forest
Service
17
Training Data Sources
3000 air photo interpretation sites US Forest
Service
5200 Sites from other organizations (USGS
Landfire BLM)
18
Training Data Sources
3000 air photo interpretation sites US Forest
Service
5200 sites from other organizations (USGS
Landfire BLM)
7800 field work RSGIS Lab in collaboration
with BLM UDWR
16000 total sample sites
19
Properties of the Classifier Decision Trees

• Data-mining software for decision-making and
  exploratory data analysis, also called CART
  (Classification and Regression Trees (Breimann
  1984))
• Identify complex relationships between multiple
  independent variables to predict a single
  categorical occurrence
• Predictor variables may be categorical or
  continuous
• Recursively splits the predictor data cloud
  to create prediction rules or a decision tree
  with branches nodes
• Several software packages SPLUS, See5/C5, CART,
  R
• Gaining popularity remote sensing land cover
  classification (Brown de Colstoun et. al, 2003
  Larwrence Wright 2001 Hansen et al. 2000
  Friedl Brodley 1997 Hansen, Bubayah Defries
  1996))

20
Mining the Predictor Layers
21
(No Transcript)
22
Example Splits on 2 variables
23
Example Tree Output for 2 Variables
24
Example Rules Output
See5 Release 1.17 Wed Apr 23 134202
2003   Options Rule-based
classifiers   Class specified by attribute
dep'   Read 7097 cases (10 attributes) from
t3.data   Rules   Rule 1 (17, lift 45.4)
band01 1 band03 gt 115 band03
lt 122 band05 lt 81 band06 lt
1419 -gt class 1 0.947   Rule 2 (9,
lift 43.6) band01 1 band02 lt
102 band03 gt 115 band03 lt 118
band04 lt 117 band06 lt 1419
-gt class 1 0.909   Rule 3 (6, lift 42.0)
band01 13 band03 lt 110
band05 lt 73 band07 4
Generated with cubistinput by EarthSat
Training samples 10260 Validation samples
2565 Minimum samples 0 Sample method
Random Output format See5   dep. h/mgz
n_5/trainingdata/mrgpts1.img(Layer_1)   Xcoord i
gnore. Ycoord ignore. band01 1,2,-30
h/mgzn_5/img_files/sum30cl.img(Layer_1) band02
continuous. h/mgzn_5/img_files/subrt.img(Layer
_1) band03 continuous. h/mgzn_5/img_files/sundv
i.img(Layer_1) band04 continuous. h/mgzn_5/img
_files/fandvi.img(Layer_1) band05 continuous. h
/mgzn_5/img_files/fabrt.img(Layer_1) band06 con
tinuous. h/mgzn_5/img_files/elev.img(Layer_1) b
and07 0,1,2,3,4,5,6,7,8,9,10. h/mgzn_5/img_file
s/landf.img(Layer_1)   dep 1,2,3,4,5,6,7,8,9,10,
11,12,13,14,15,16,17,18,19,20. h/mgzn_5/training
data/mrgpts1
25
Tools for Spatial Applying Decision Trees

• Rulemaker (SPLUS Imagine)
• Vinod Chowdary (USU, MS Computer Sci.)
• http//www.gis.usu.edu/docs/projects/swgap/rulemak
  er.html
• STATMOD (SPLUS Arcview)
• Christine Garrard (USU, MS Biology)
• http//bioweb.usu.edu/gistools/statmod/
• Imagine CART Module (See5 Imagine)
• Eros Data Center (Earth Satellite Corp)
• http//www.gis.usu.edu/7Eregap/download/C5Module/
  

26
II. Comparison with 1995 Utah GAP Vegetation Map

• Utahs Great Basin Eco-Region ( 17.5 M acres,
  300 x 120 Miles)
• Approximately 5 mosaicked Landsat 7 scenes
• 3000 sample sites (1700 USU/1300 other sources)

27
Comparison for Great Basin Eco-Region (Partial
List)
28
1995 GAP Vegetation Map
2003 GAP Veg. Map (preliminary)
29
Park Valley Example
30
1995 GAP 30 M
2003 GAP 30 M
1995 GAP Pub.1KM
31
Tooele Example
Tooele
32
Tooele
Tooele
Tooele
1995 GAP 30 M
2003 GAP 30 M
1995 GAP Pub.1KM
33
Clarkston Example
Clarkston
34
1995 GAP 30 M
2003 GAP 30 M
1995 GAP Pub.1KM
Clarkston
Clarkston
Clarkston
35
Validation, Accuracy and Appropriate Uses

• Accuracy Assessment when map is completed
• Internal Validation concurrently with mapping
  effort
• Large landscape monitoring and planningscales of
  1 100 k 1 250k

36
IV. Timeline Summary

• 2000 Work began
• 2001 Field data collection imagery acquisition
• 2002 Field data collection, imagery acquisition,
  land cover mapping began
• 2003 Field data collection land cover mapping
  continue
• 2004 Land Cover map complete. Final products
  complete

37
Comparison with 1995 Utah GAP Vegetation Map
Product

• 1995 GAP Veg. Map
• Utah state boundary
• 36 land cover types
• Re-sampled to 1 km resolution
• Distributed on CDROM

• 2003 GAP Veg. Map
• Independent of state boundaries
• Anticipated 50 cover types
• 30 meter resolution with 0.5 ha MMU
• Distributed via WWW and CDROM
• Improved Accuracy
• Nationally consistent vegetation classification
  (NatureServe NVCS)

38
Acknowledgements

• Doug RamseyProject Principle Investigator
• Lisa LangsGraduate Student
• Wendy RiethGraduate Student
• Jessie DenhamField Botanist
• Marie DucharmeField Botanist
• Gerald ManisPlant Ecologist/Mapper
• Chris GarrardProgrammer
• Rob Johnson--Cartographer
• Eric SantGraduate Student
• Chris McGintyGraduate Student
• Jarom GilbertUndergraduate Student
• Sheryl BoyackUndergraduate Student
• Wendy HurdUndergraduate Student
• Meg PoulsonField Technician
• Todd SajwajGraduate Student