U'S' Geological Survey Earth Resources Operation Systems EROS Data Center

1
U.S. Geological SurveyEarth Resources Operation
Systems (EROS) Data Center
World Data Center for Remotely Sensed Land Data
2
USGS EROS DATA CENTERLand Remote Sensing from
Space Acquisition to Applications


Earth Observation Satellites
USGS National Archive Challenge
Data Applications

• Land Cover
• Environmental Monitoring
• Emergency Response
• Fire Danger Rating
• DOI Land Management
• Natural Hazards
• Coastal Zones

• Declassified Systems
• Landsat 1-5,7
• NOAA - POES
• Shuttle Radar
• TERRA (1999)
• NASA-EOS (1999)
• High Resolution Systems

• Preserve
• Provide Access
• Process
• Reproduce
• Distribute
• Hold in Trust

Expanding to over 18 million images of the earth!
3
USGS EDC Data Holdings

• Aerial Photographs
• 1940-present
• U.S. coverage
• gt 9 million frames
• Scale 1-2 meter

4
USGS EDC Data Holdings

• Landsat Satellite Images
• 1972-present
• gt 18 million frames
• Global coverage
• 15-80 meter

5
USGS EDC Data Holdings

• AVHRR Satellite Images
• 1987-present
• Global coverage
• 1 km resolution

6
Using Landsat satellite imagery to estimate
agricultural chemical exposure in an
epidemiological study
Susan Maxwell, PhD (USGS EROS Data
Center) Interface 2002, Montreal, Canada
Collaborators Dr. Jay Nuckols, EHASL,
Colorado State University Dr. Mary Ward,
National Cancer Institute Eric Smith, EHASL,
Colorado State University Leanne Small,
EHASL, Colorado State University
7
Why use satellite imagery?

• Traditional methods of collecting chemical
  exposure data dont work well (environmental/biolo
  gical sampling, questionnaires)

Spray drift
Dust

• Agriculture Chemicals
• Fertilizers
• Pesticides

Drinking water
8
Why use satellite imagery?

• Cancers generally take several years to develop,
  therefore need to reconstruct historical exposure
  
• Our approach use Landsat imagery to create
  historical land use/crop type maps integrate
  with other data (chemical use, soils, wind, etc.)
  to estimate exposure

9
Metric Development Transport Modeling
(Ward et al. Environmental Health Perspectives,
2000)
10
Why Landsat ?

• Longest running satellite sensor (1972-current)
• Successful crop type mapping applications
  (AGRISTARS, etc.)
• Appropriate spectral bands (visible, near
  infrared, middle infrared)
• Appropriate spatial resolution (30-80 meter)
• Inexpensive (compared to higher resolution data
  sets)

11
Crop Type Classification - Sheldon, NE
12
Case Study Mapping Corn

• Chemicals used on corn (nitrogen, atrazine) have
  been associated with several cancers and birth
  defects

13
Traditional classification methods are not
appropriate

• Only want CORN
• BIG Data Sets
• Large geographical regions
• File size
• 500 Mb/image
• Multi-year

30 years
14
Traditional classification methods are not
appropriate (cont.)

• Usually need ground reference data expensive,
  difficult to get for historical data
• Time-consuming process

15
Crop characteristics

• Corn dominates

16
Crop characteristics

• Large, homogeneous fields
• Spectral characteristics differ from other major
  crops (soybeans, alfalfa, winter wheat, etc.)
• Spectrally similar to deciduous trees, riparian
  area

17
Case Study Mapping Corn

• Initial method software was developed to .
• Use existing land cover maps (NLCD) to eliminate
  non-row crop classes (spring grains, hay/pasture,
  trees, urban, wetland, etc.)
• Use existing USDA acreage estimates to target
  specific geographic region (i.e., county) to
  collect training statistics
• Use maximum likelihood algorithm to classify the
  entire image
• Use the Mahalanobis distance image in combination
  with USDA acreage estimates to identify cut-off
  for highly likely corn, likely corn and
  unlikely corn

18
Method cont.

• Use existing land cover maps (NLCD) to eliminate
  non-row crop classes (spring grains, hay/pasture,
  trees, urban, wetland, etc.)

19
Method cont.

• Use USDA acreage estimates to target specific
  geographic region (i.e., county) to collect
  training signature

20
Method cont.

• Use the Mahalanobis distance image in combination
  with USDA acreage estimates to identify cut-off
  for highly likely corn, likely corn and
  unlikely corn

Highly Likely Corn
Mahalanobis distance image
Likely Corn
21
Mahalanobis Distance Threshold

• ...

• ...

• ...

• ...

• ...

• 131

• 1066.3

• 59082.1

• 100.7

• 2

• 132

• 417.2

• 59499.3

• 3

• ...

• ...

• ...

• ...

• 1787

• 0.4

• 82893.2

• 3

22
Results

• gt80 average accuracy
• Higher errors occur when
• Spectrally similar cover types in same area
  (millet, sorghum)
• Image date is too early in growing season
• Non-parametric signature (clouds/haze,
  irrigated/non-irrigated corn)

23
Thank You
Susan Maxwell maxwell_at_usgs.gov