Transcript Satellite image classification SSIP 08, Vienna

Satellite image
classification
SSIP 08, Vienna
Tamas Blaskovics
Michael Glatz
Korfiatis Panagiotis
José Ramos
University of Szeged
Vienna University of Technology
University of Patras,
Porto University
Task
Satellite Image Classification:
Input: Landsat images of terrain, plus sample
images of fields/ sea, forest etc
Aim: segmentation of scene based on texture
(and color)
Additional goal: intenfication of key features
such as cave openings etc
Output: labeled scene
Satellite image classification
Input Image
k-means Unsupervised
Segmentation
or
MRF Semi- Supervised
Segmentation
or
MRF Unsupervised
Segmentation
Output Image
Area classification ( User Interaction)
or
Area classification ( Automated)
Dataset
20 images aquired with the IKONOS Satellite.
)
(http://www.satimagingcorp.com/satellite-sensors/ikonos.html
Speed on Orbit
7.5 kilometers per second
Speed Over the Ground
6.8 kilometers per second
Revolutions Around the
Earth
14.7, every 24 hours
Altitude
681 kilometers
Resolution at Nadir
0.82 meters panchromatic; 3.2 meters multispectral
Resolution 26° Off-Nadir
1.0 meter panchromatic; 4.0 meters multispectral
Image Swath
11.3 kilometers at nadir; 13.8 kilometers at 26° off-nadir
Equator Crossing Time
Nominally 10:30 AM solar time
Revisit Time
Approximately 3 days at 40° latitude
Dynamic Range
11-bits per pixel
Image Bands
Panchromatic, blue, green, red, near IR
Method 1/2
Step 1: Image Segmentation
The RGB image was converted to L*u*v color space
Two unsupervised methods were used:
MRF segmentation ( Kato et al. )
EM step
ICM
K-means
Parameters:
User defined: # of regions, β, temperature.
Method 2/2
Step 2: Class Characterization
User defined
User chooses the desired region for classification
The first order statistics (mean, variance, skewness, kurtosis, range) are
calculated for a ROI around the selected image
Automated
Skeletonization technique was applied for each segmented region
A sliding ROI (21 x 21) was used to extract first order statistics
K-nearest neighbor classifier was used (NN)
Segmented area is also calculated
Features evaluated
Segmentation Stage:
•Intensity value channel U
•Intensity value channel V
Classification Stage:
•Mean value
•Standard deviation
•Kurtosis
•Skewness
•Range
GUI
Result – image segmentation 1/2
Result – image segmentation 2/2
Comments
 Visual
evaluation seems to present good
results
 No serious evaluation was conducted
 Segmentation process is slow
 Dataset is too small to construct robust
learning process
Future developments
 Segmentation
process
Evaluation of more complex techniques
 Classification
process
Bigger training database
Other texture features
Different classifiers
 Evaluation
Use of ground truth and shape differentiation
metrics
THE END
Thank you!