Use of Aerial Videography in Habitat Survey and Computers as Observers Leonard Pearlstine University...

Use of Use of AerialAerial Videography in Habitat Survey Videography in Habitat Survey

and Computers as Observersand Computers as Observers Leonard Pearlstine University of Florida

Land Cover Classification

Landsat TMLandsat TM Digital CameraDigital Camera

Landsat TM Schinus Signature

0.00

0.20

0.40

0.60

0.80

1.00

1.20

1 21 41 61 81 101

121

141

161

181

201

221

241

Reflectance Value

Re

lati

ve

Fre

qu

en

cy

TM Band 4

TM Band 2

TM Band 3

Digital Camera Schinus Signature

0.00

0.20

0.40

0.60

0.80

1.00

1.20

0

21

42

63

84

10

5

12

6

14

7

16

8

18

9

21

0

23

1

25

2

Reflectance Value

Re

lati

ve

Fre

qu

en

cy

Layer 1

Layer 2

Layer 3

Texture

The spatial (statistical) distribution of gray tones.The spatial (statistical) distribution of gray tones.-Haralick et al. 1973

Desirable Texture Characteristics

• Angularly independent

•Invariant under gray level transformations

•Simple algorithms

Brazilian Pepper

Measures of Edge

DensityDensity MagnitudeMagnitude

Rate of ChangeRate of Change

““Visual discrimination of pattern is based primarily on clusters or Visual discrimination of pattern is based primarily on clusters or lines formed by proximate points of uniform brightness” -Julesz 1962lines formed by proximate points of uniform brightness” -Julesz 1962

Edge Signatures

Edge Signatures

Multivariate Discrimination

Logistic Regression selected for

• Heteroscedastic Variances• Dichotomous Classification

Reference

Classified

Producer’s Accuracy

34% 71% 61%

User’s Accuracy 98% 97% 80%

Logistic Regression

Commission Error 16% 21% 19%

Reference

Classified

Logistic Regression – No Schinus Images

Omnidirectional Variogram

Compute Homogeneity Index Image

Pasture

Trees canopy

Grass

Individual Trees

0

1000

2000

3000

4000

5000

6000

7000

1 9 17 25 33 41 49 57 65 73 81

grass

Indiv. Trees

pasture

Trees

Edge Textures Application Interface

Birds Detection and Counting

Video Still Image showing Birds Colony of approximately 150 birds

Template Matching

•Identify Bird Template(s)

•Area Based Matching (e.g. Correlation Matching)

9x9 bird Template

Area Based Matching(correlation Matching)

•Compute The correlation Coefficient between Template and Reference Image as:

R(x,y) = ΣΣ(T’(x’,y’)*I’(x’+x,y’+y))

Where:

T~(x,y) = T(x,y) – T & I~ (x+x’,y+y’) = I(x+x’, y+ y’) – I

T and I are the mean under the Template and reference windows respectively.

Correlation Image

•Bright values indicates Template and Reference images match and Birds Existence

Correlation Image

Reference Image

Threshold and Identify Birds

Different Threshold can be used.•High Threshold Missing Birds (Increase Omission errors).

•Low Threshold Add noise and other features as Birds (Increase Commission errors).

Threshold = 140Birds Count = 153Actual Birds = 150

Missing Birds

No Birds

Progressive Scan Video Image

Progressive Scan Video Image with Bird Pattern Matching

Birds Count Application Interface

Conclusions

•Characterizations of edge can effectively discriminate vegetation classes.

• Multivariate discrimination using logistic regression substantially improved accuracies.

• The logit model successfully identified Schinus terebinthifolius and excluded most other vegetation types.

Conclusions

• Additional work needs to be done to separate Sabal palmetto signatures from Schinus.

• “Big white birds” can be effectively discriminated in even low quality videography.

• Larger sample sizes over a greater geographic extent and with additional species will be needed before these procedures can be considered operational.

Conclusions

• The modeling approach develop in this dissertation provides an effective procedure for rapid and consistent identification of target species from aerial imagery.