Object-oriented Land Cover Classification in an Urbanizing Watershed

1
Object-oriented Land Cover Classification in an
Urbanizing Watershed

• Erik Nordman, Lindi Quackenbush, and Lee
  Herrington
• SUNY College of Environmental Science and Forestry

2
Objectives

• Create a land cover classification
• Suitable for ArcHydro pollution model
• Up-to-date
• High spatial resolution
• Emphasis on impervious surface

Introduction Objectives Study
Area Methods Results Discussion Conclusions
3
Study Area
Introduction Objective Study
Area Methods Results Discussion Conclusions
4
Methods Imagery

• Satellite QuickBird (DigitalGlobe)
• 2.44 m multispectral resolution
• 4 bands (3 visible NIR)
• Created NDVI layer
• Collected over 2 dates
• Half on each date
• May and June 2005

Introduction Methods Imagery Software
Classification Results Discussion Conclusions
5
Detail of Imagery
Upper Lake
Introduction Methods Imagery Software
Classification Results Discussion Conclusions
Carmans River
6
eCognition Object-oriented classification

• Uses spectral, textural and thematic information
• Segmentation into homogeneous polygons (objects)
• Can vary the size (homogeneity) of polygons at
  different levels

Introduction Methods Imagery Software
Classification Results Discussion Conclusions
7
Impervious Cover

• Critical to analyzing runoff and pollution
• Challenges
• High spatial resolution
• Individual roads, houses
• Tree canopy covers roads

Introduction Methods Imagery Software
Classification Results Discussion Conclusions
8
Impervious Cover

• Solution
• Use road vector layer
• ALIS data set
• For public safety
• NYS GIS Clearinghouse
• 10 meter buffer

Introduction Methods Imagery Software
Classification Results Discussion Conclusions
9
Level 2 Segmentation
Introduction Methods Imagery Software
Classification Results Discussion Conclusions
10
Level 1 Segmentation
Introduction Methods Imagery Software
Classification Results Discussion Conclusions
11
Detail ALIS roads in Level 2 segmentation
Introduction Methods Imagery Software
Classification Results Discussion Conclusions
12
Classification

• Classes based on TR-55
• Impervious
• Includes roads, driveways, roofs
• Tree, Grass
• Wetlands
• 3 classes woody, emergent, tidal
• Also from thematic layers
• Bare, Water

Introduction Methods Imagery Software
Classification Results Discussion Conclusions
13
Classification

• Attributes used in classification included
• Color and Brightness
• Area
• Shape Index and Compactness
• GLMC heterogeneity
• Proximity to objects in other classes

Introduction Methods Imagery Software
Classification Results Discussion Conclusions
14
Introduction Methods Results Discussion Conclusion
s
15
Introduction Methods Results Discussion Conclusion
s
16
Accuracy Assessment

• Reference data
• Digital orthophotos
• Acquired April, 2004
• Leaf-off
• Stratified random sample, 727 points

Introduction Methods Results Discussion Conclusion
s
17
Accuracy Assessment

• Overall 73.9
• Users accuracy of key classes
• Impervious 73.4
• Tree 74.5
• Grass 66.7

Introduction Methods Results Discussion Conclusion
s
18
Discussion

• Accuracy comparable to other studies
• ALIS road layer successfully used to aid
  classification

Introduction Methods Results Discussion Conclusion
s
19
Discussion

• Seasonality
• Imagery leaf-on
• Orthophotos leaf-off
• Affected agreement between classification and
  reference data
• Scrub vegetation
• Confusion among bare, grass and tree classes

Introduction Methods Results Discussion Conclusion
s
20
Discussion

• Accuracy Assessment
• Response unit 1 pixel in classified image
• Response unit should be object, not pixel

Introduction Methods Results Discussion Conclusion
s
21
Conclusions

• QuickBird and eCognition produced a highly
  detailed classification
• Adequate for pollution and economic models
• Thematic layers proved useful

Introduction Methods Results Discussion Conclusion
s
22
Acknowledgements

• IAGT
• Provided satellite imagery
• NYS Department of State Division of Coastal
  Resources
• Provided financial support