Landsat classification © Team. SSIP 20052 © Team Delia Mitrea – Technical University of...

13
Landsat classification © Team

Transcript of Landsat classification © Team. SSIP 20052 © Team Delia Mitrea – Technical University of...

Landsat classification

©Team

SSIP 2005 2

©Team

Delia Mitrea – Technical University of Cluj-Napoca, Romania

Sándor Szolyka – Budapest Tech, Hungary

Imre Hajagos – University of Szeged, Hungary

Szabolcs Berecz - Budapest Tech, Hungary

Gergely Grósz – University of Veszprém Georgikon Faculty of Agricultural, Hungary

SSIP 2005 3

The Problem• Input: Landsat images of terrain, plus sample

images of fields, sea, forests or etc.

• Aim: Segmentation of scene based on texture and colour.

• Output: Label scene.

SSIP 2005 4

The Solution

Solution 1. - Histogram matching I.Step 1. Decompose the image into small

cells. Step 2. Compute the histogram in the RGB

levels (All grid has three (red, green, blue) histograms.).

Step 3. Classification based on the correlation of histograms.

Step 4. Segment the image.

SSIP 2005 5

The Solution

SSIP 2005 6

The Solution

Solution 2. - Histogram matching II.Convert the histograms to a greyscale.

(Y=0,299 R+0,587 G+0,114 B)

SSIP 2005 7

The Solution

Solution 3. – Markov Random Fields•Statistics based classifier algorithm.

•Uses spatial information.

•Driven by energy minimization.

SSIP 2005 8

The SolutionSolution 4. - Texture-based recognition

Features used:

•Average edge frequency (density)

•Average edge contrast

•GLCM (Gray Level Cooccurrence Matrix) homogeneity

•GLCM (Gray Level Cooccurrence Matrix) entropy

•GLCM (Gray Level Cooccurrence Matrix) variance

•GLCM (Gray Level Cooccurrence Matrix) energy

SSIP 2005 9

The Solution

Solution 4. - Texture-based recognitionStep 1. Learning

• Select a known region int the image (forest mountains or water)

• Compute GLCM features and edge-based features

• Store the feature vector in the training set for the corresponding class

SSIP 2005 10

The SolutionSolution 4. - Texture-based

recognitionStep 2. Recognition

•Select an unknown area in the image in order to classify it: forest mountains or water

•Compute the GLCM features and the edge-based features

•Compare the feature vectors with the data int he training set: euclidean distance

•Use the k-nn method and decide the class

SSIP 2005 11

The Solution

Solution 4. - Texture-based recognition

SSIP 2005 12

References• M. Berthod, Z. Kato, S. Yu, J. Zerubia: Bayesian

imageclassification using Markov random fields. Image and Vision Computing,14(1996): 285-295, 1996.

• Z. Kato: Multi-scale Markovian Modelisation in Computer Vision withApplications to SPOT Image Segmentation. PhD thesis, INRIA SophiaAntipolis, France, 1994.

• Z. Kato, J. Zerubia and M. Berthod: Satellite image classification using amodified Metropolis dynamics Proc. IEEE International Conf. on Acoust., Speechand Sig. Proc., vol. 3, pp. 573-576, San Francisco, CA, March 23-26,1992.

The End

Thank you for your attention!