Simulated effects of co-registration errors: implications related to … · 2018-08-02 · Simulated effects of co-registration errors: implications related to estimating forest characteristics

Simulated effects of co-registration errors: implications related to estimating forest

characteristics using remotely sensed data

INGY Coeur d’Alene March 13, 2017 John Hogland, David Affleck, & Nathaniel Anderson

Previous Work

Previous Work

Error

Spatial Outputs

Preprocessing

Modeling

Issues

• Inconsistent image acquisition dates

• Plot layout• Size• Sampling intensity• Small trees

• Co-registration errors• GPS• Imagery

NAIP Shift GPS (8m)

NAIP Shift GPS & Image (8m, 7m)

Landsat Shift GPS & Image (8m, 60m)

Co-registration Errors

• What are the impacts

• How to address co-registration errors

• Better horizontal accuracy

• Larger plot area

How To Simulate

• Field perspective• Time intensive

• Relationship between spectral and plot measurements

• Limited in geographic extent

• Image perspective• Assume perfect relationship between plot and mean

spectral values

• Introduce random spatial shifts

Simulations Overview

• Random Locations

• Circle vs square

• Introduce shift

• Extract spectral values

• Regress values

• Growing overlap

Info Explained = 𝑂𝑟𝑖𝑔𝑖𝑛𝑎𝑙 𝑎𝑟𝑒𝑎−𝑟𝑒𝑔𝑖𝑠𝑡𝑟𝑎𝑡𝑖𝑜𝑛 𝑒𝑟𝑟𝑜𝑟 𝑎𝑟𝑒𝑎

𝑂𝑟𝑖𝑔𝑖𝑛𝑎𝑙 𝐴𝑟𝑒𝑎

56%69%77%

Extracting & Aggregating 56 55 52 29 10

45 59 57 33 24

46 60 57 44 25

48 63 40 33 26

30 50 41 24 20

65 56 55 52 29

63 45 59 57 33

62 46 60 57 44

66 48 63 40 33

59 30 50 41 24

Worst Case Scenario

Random Noise (0-255)

% overlap = Info Explained = R2

Images

Landsat 8 City Ag

NAIP

Forest Forest & Ag WaterRandom

Simulations• 7 images

• 5000 locations

• 2 random shifts • GPS (7m)• Image (NAIP: 6 cells, Landsat: 2 cells)

• Extract spectral values

• Regress against one another

• Record intercept, slope, and R2

• Repeated (0-50 cells by 5)

𝑦 = 𝛼 + 𝛽𝑥

Record

Results

GPS = 7mImage = 6m (6 cells)

GPS = 7mImage = 6m (6 cells)

GPS = 7mImage = 6m (6 cells)

GPS = 7mImage = 60m (2 cells)

GPS = 7mImage = 60m (2 cells)

GPS = 7mImage = 60m (2 cells)

Conclusion

• We can minimize effect of co-registration

• Requires sampling a bigger area

• Practical limitations in the field

• Co-registration errors introduce bias

Next Steps

• What proportion of plot extent needs to be sampled?

• What is the best layout within the plot extent?

• How to use the estimate of bias to refine predictions?

Contact Information

John Hogland, Biological Scientist

Rocky Mountain Research Station

800 East Beckwith

Missoula, MT 59801

Phone: (406) 329-2138

email: jshogland@fs.fed.us