y = ax 2 + bx + c Poly fit standard error 104 bands

Atlanta city-model derived from a WorldView-2 multi-sequence acquisition
Spectral classification
of WorldView-2 multi-angle sequence
N. Longbotham, C. Bleilery, C. Chaapel,
C. Padwick, W. J. Emery, and F. Pacifici
Outline
This presentation illustrates the unique aspects of the WorldView-2
satellite platform by combining multi-spectral information with multiangle observations
The previous presentation dealt with very high spatial resolution
imagery with multi-angle observations
What can we do with this kind of data set?
Four experiments have been carried out to investigate the
classification contribution of multi-angle reflectance (MAR) as well as
different feature extraction data sets (reducing the large size of the raw
data space)
2
Methodology (1/2)
13 Multispectral Images
13 Panchromatic Images
Digital Surface Model
Atmospheric Correction
Nadir Multispectral
Multi-angle Multispectral
13 Multispectral True-Ortho Images
Polynomial Multispectral
Principal Component Analysis
3
Methodology (2/2)
Nadir
Multispectral
Multi-angle
Multispectral
8 bands
104 bands
Polynomial
Multispectral
y =32bands
ax2 + bx + c
Poly fit standard error
4
PCA
10 bands
Coastal
Blue
Atmospheric Correction (1/2)
Green
Yellow
Red
Red Edge
NIR1
NIR2
5
Coastal
Blue
Atmospheric Correction (2/2)
Green
Yellow
Red
Red Edge
NIR1
NIR2
6
Information Sources
The MAR contains a partial bidirectional reflectance distribution
function (BRDF) over a single satellite track at a single sun
angle
Objects with pitched surfaces, such as trees and residential
roofs, will present a different observational cross-section at
each angle
Surfaces with varying reflectance in both time and angle can be
described by an error term that encapsulates the variation of a
pixel through the multi-angle sequence
7
Coastal
Blue
Partial BRDF - over a single satellite track
Green
Yellow
Red
Red Edge
NIR1
NIR2
8
Coastal
Blue
Pitched surfaces
Green
Yellow
Red
Red Edge
NIR1
NIR2
9
Varying reflectance in both time and angle (1/2)
Differentiates land-use of similar spectral
signature
– low vs. high volume traffic roads
Multi-angle spectral variability
– stationary vehicles
10
Varying reflectance in both time and angle (2/2)
11
Four Experiments
The most-nadir multi-spectral image is used as base-case
Exp. 9
Nadir Multispectral
Exp. 10
Exp. 11
Exp. 12
X
Multi-angle Multispectral
X
Polynomial Multispectral
X
Principal Component Analysis
X
12
Classification and Validation
Flat Roof
15 classes of interest have been selected representing
a wide variety of both natural and man-made landcovers, including different kind of roof, roads, and
vegetation
Pitched Roof
Concrete
Pavement
Parking Lot
Healthy Vegetation
Stressed Vegetation
Training: 50 samples per class
Dormant Vegetation
Validation: 90,000 of independent samples
Soil
Evergreen Trees
Deciduous Trees
Each of the classification experiments are conducted
using the Random Forest algorithm
Parked Cars
Recreational
Shadow
Water
13
Nadir
Multi-angle
Polynomial
PCA
Exp. 10
Exp. 11
Exp. 12
Exp. 12
X
X
X
X
14
Exp. 12
Exp. 9
Exp. 11
Exp. 11
Exp. 10
Exp. 10
Exp. 9
Exp. 9
Results (1/2)
Flat Roof
55.4
62.8
54.6
57.2
Pitched Roof
35.6
54.8
63.2
65.8
Concrete
84.7
94.2
89.1
91.0
Pavement
54.1
80.5
80.2
84.0
Parking Lot
76.4
82.0
90.7
89.3
Healthy Vegetation
94.1
96.7
96.0
96.2
Stressed Vegetation
94.7
94.8
91.4
90.4
Dormant Vegetation
80.8
80.8
82.0
85.4
Soil
90.2
94.8
92.5
96.7
Evergreen Trees
77.2
85.6
92.4
93.0
Deciduous Trees
85.9
94.8
95.8
91.0
Parked Cars
28.0
43.0
64.7
49.6
Recreational
88.8
96.2
96.7
97.1
Shadow
89.3
94.8
93.0
89.9
Water
95.1
95.9
83.4
83.2
Results (2/2)
Flat Roof
Pitched Roof
Concrete
Pavement
Parking Lot
Healthy Vegetation
Stressed Vegetation
Dormant Vegetation
Soil
Evergreen Trees
Deciduous Trees
Parked Cars
Recreational
Shadow
Water
15
Detail
•
Parked Cars
Flat Roof
Pitched Roof
•
Concrete
Empty Parking Spots
Pavement
•
Parking Lot
Pitched Roofs
Healthy Vegetation
•
•
Stressed Vegetation
Deciduous Trees
Dormant Vegetation
Soil
Stressed/Dormant Grass
Evergreen Trees
Deciduous Trees
•
Road
Parked Cars
Recreational
Shadow
Water
16
Feature Contribution
17
Conclusions
This study showed that there is significant improvement in
classification accuracy available from the spectral data in a multi-angle
WorldView-2 image sequence.
Four spectral classification experiments were separately presented
using a nadir multi-spectral image, the full multi-angle multispectral
data set, and two feature extraction techniques.
The multi-angle spectral information provided 14% improvement in
kappa coefficient over the base case of a single nadir multispectral
image.
Specific classes benefited from the unique aspects of the multi-angle
information:
– The classes car and highway are of particular interest
18
2011 IEEE GRSS Data Fusion Contest
Data Fusion Session:
•
•
WHEN: Tuesday, July 26, 08:20 - 10:00 AM
WHERE: Ballroom C
Data Fusion Technical Committee meeting:
• WHEN: Tuesday, July 26th, 5:30 to 6:30 PM
• WHERE: East Ballroom A