Scribd
Upload
22 views

Digital Image Classification

The document discusses digital image classification, including defining clusters in feature space, selecting a classification algorithm like maximum likelihood, running the classification, and validating the results with an error matrix. Spectral characteristics are used to classify land cover types from satellite imagery.

Uploaded by

lomeca
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
22 views

Digital Image Classification

The document discusses digital image classification, including defining clusters in feature space, selecting a classification algorithm like maximum likelihood, running the classification, and validating the results with an error matrix. Spectral characteristics are used to classify land cover types from satellite imagery.

Uploaded by

lomeca
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 24

Digital image classification

ITC Textbook: Principles of RS


Chapter 12: Digital image classification

Digital Image classification

The operator instructs the computer to perform an


“interpretation” according to certain conditions.

Digital image classification is based on the different


spectral characteristics of the Earth’s surface:
● Plants - crops
● Soil
● Rocks
● Human elements
o (roads, houses etc.)
● Water – suspended materials
INTERNATIONAL INSTITUTE FOR GEO-INFORMATION SCIENCE AND EARTH OBSERVATION 3
Image & feature space …1

Image space : Spatial arrangement in the image of the


measurements of the EM spectrum

INTERNATIONAL INSTITUTE FOR GEO-INFORMATION SCIENCE AND EARTH OBSERVATION 4


Image & feature space …2

Feature space : Graph that shows the DN values of the


feature vectors
(DN values per band
of a certain pixel)

3 bands
2 bands

3-dimensional
2-dimensional feature space
feature space

INTERNATIONAL INSTITUTE FOR GEO-INFORMATION SCIENCE AND EARTH OBSERVATION 5


Image & feature space …3

Scatterplot of two bands of a digital image.


Intensity at a point related to the
number of pixels at that point.

INTERNATIONAL INSTITUTE FOR GEO-INFORMATION SCIENCE AND EARTH OBSERVATION 6


Image & feature space …4

Euclidian Distance in feature space calculated by


Pythagoras – values in DN
(Euclidian) Distance between
(10,10) and (40,30) is the
40,30
square root of
(40 - 10)2 + (30 – 10)2
10,10

INTERNATIONAL INSTITUTE FOR GEO-INFORMATION SCIENCE AND EARTH OBSERVATION 7


Image & feature space …5

Clustering of classes in a feature space – result of


training process
Example: six classes

Band Y

Band X
INTERNATIONAL INSTITUTE FOR GEO-INFORMATION SCIENCE AND EARTH OBSERVATION 8
Image classification process
1. Selection and preparation of the image data
2. Definition of clusters in the feature space
3. Selection of the classification algorithm
4. Running of the actual classification
5. Validation of the result

INTERNATIONAL INSTITUTE FOR GEO-INFORMATION SCIENCE AND EARTH OBSERVATION 9


Image classification process

1. Selection and preparation of the image data


Depending on:
• Cover types to be classified
• Selection of most appropriate
sensor (spatial & spectral res.)
• Best date(s) of acquisition
(season !)
• Band combination / correlation

Landsat TM - R G B : 4 5 2
Karoi, Zimbabwe
INTERNATIONAL INSTITUTE FOR GEO-INFORMATION SCIENCE AND EARTH OBSERVATION 10
Image classification process

2. Definition of clusters in the feature space


Two approaches:

¾ Supervised classification:
Operator defines clusters during training process.
Knowledge of the area needed.
Collection of sample sets during Fieldwork !!

¾ Unsupervised classification
Clustering being done by computer based on for instance: -
- Number of user defines clusters, etc.
- ‘minimum distance to cluster centroid’ decision rule
INTERNATIONAL INSTITUTE FOR GEO-INFORMATION SCIENCE AND EARTH OBSERVATION 11
Image classification process

2. Definition of clusters in the feature space


¾ Unsupervised classification
Clustering being done by computer, based on for instance: -
Number of user defined clusters, threshold, cluster distance etc.
- ‘minimum distance to cluster centroid’ decision rule
Iteration 1 Iteration 2 Iteration 3 Iteration 4

INTERNATIONAL INSTITUTE FOR GEO-INFORMATION SCIENCE AND EARTH OBSERVATION 12


Image classification process

3. Selection of the classification algorithm …1


Types of algorithms:
¾ Box classifier (very simple)
¾ Minimum Distance to Mean
¾ Maximum Likelihood

Landsat TM - R G B : 4 5 2
Karoi, Zimbabwe
INTERNATIONAL INSTITUTE FOR GEO-INFORMATION SCIENCE AND EARTH OBSERVATION 13
Image classification process

3. Selection of the classification algorithm …2


¾ Box classifier (or parallelepiped classification)
Only upper and lower class limits defined.

Disadvantage:
A lot of overlap!

“Boxes”

INTERNATIONAL INSTITUTE FOR GEO-INFORMATION SCIENCE AND EARTH OBSERVATION 14


Image classification process

3. Selection of the classification algorithm …3


¾ Minimum Distance to Mean classifier ( M D M )
Euclidian distances from pixels to
cluster centres calculated

Disadvantage: Pixels at large


distance from cluster centre wrongly
classified.

Solution: definition of threshold


distance (circle)
INTERNATIONAL INSTITUTE FOR GEO-INFORMATION SCIENCE AND EARTH OBSERVATION 15
Image classification process

3. Selection of the classification algorithm …4


¾ Maximum Likelihood classifier ( M L )
Considers not only the centre, but
also shape, size and orientation of
the clusters.
Calculation of statistical distance
based on the mean values and
covariance matrix of the clusters.
Statistical distance is probability
value (equiprobability contours).
INTERNATIONAL INSTITUTE FOR GEO-INFORMATION SCIENCE AND EARTH OBSERVATION 16
Image classification process

3. Selection of the classification algorithm …5


¾ Maximum Likelihood classifier (ML)
Statistical distance based on equiprobability contours.
Probability
Sand
Corn
Urban Hay
Forest

Equiprobability
contours
DN
Ba d 4
n d Ban
5 DN
Probability density functions
(Lillesand and Kiefer, 1987)

INTERNATIONAL INSTITUTE FOR GEO-INFORMATION SCIENCE AND EARTH OBSERVATION 17


Image classification process

3. Selection of the classification algorithm …6


¾ Maximum Likelihood classifier ( M L )
Threshold boundaries:

Without boundaries

With boundaries

INTERNATIONAL INSTITUTE FOR GEO-INFORMATION SCIENCE AND EARTH OBSERVATION 18


Image classification process

4. Running of actual classification

TM R G B : 4 5 2 Box classification Minimum Distance to Maximum Likelihood


Karoi, Zimbabwe Factor = 10 Mean Threshold = 100 Threshold = 100

INTERNATIONAL INSTITUTE FOR GEO-INFORMATION SCIENCE AND EARTH OBSERVATION 19


Image classification process

5. Validation of the result …1


¾ Image classification should be checked and quantified
afterwards
¾ Comparison of classification result with enough real
world (field) samples:
o Random sampling
o Stratified random sampling

INTERNATIONAL INSTITUTE FOR GEO-INFORMATION SCIENCE AND EARTH OBSERVATION 20


Image classification process

5. Validation of the result …2


¾ Creation of error matrix (confusion matrix)
o Overall accuracy: Proportion Correctly Classified (PCC)
o Error of Commission: Incorrectly classified samples
o Error of Omission: Sample points omitted in interpretation

Reference classes

Classes in
classification result

INTERNATIONAL INSTITUTE FOR GEO-INFORMATION SCIENCE AND EARTH OBSERVATION 21


Image classification process

5. Validation of the result …3


¾ Creation of error matrix (confusion matrix)

Class A : 53 samples in “real world” but 61 cases in classification


in 35 classes agreement between classification and “real world”
Error or Omission: 53 – 35 = 18 / 53 * 100 = 34 %
Producer accuracy: 35 / 53 * 100 = 66%
INTERNATIONAL INSTITUTE FOR GEO-INFORMATION SCIENCE AND EARTH OBSERVATION 22
Image classification process

5. Validation of the result …4


¾ Creation of error matrix (confusion matrix)

Class A : Error of commission : 61 – 35 = 26 / 61 * 100 = 43 %


User Accuracy : 35 / 61 * 100 = 57 %

INTERNATIONAL INSTITUTE FOR GEO-INFORMATION SCIENCE AND EARTH OBSERVATION 23


Image classification process

5. Validation of the result …2


¾ Creation of error matrix (confusion matrix)

Class A : 53 samples in “real world” but 61 cases in classification


in 35 classes agreement between classification and “real world”
Error or Omission: 53 – 35 = 18 / 53 * 100 = 34 %
Producer accuracy: 35 / 53 * 100 = 66%
INTERNATIONAL INSTITUTE FOR GEO-INFORMATION SCIENCE AND EARTH OBSERVATION 24
Image classification process

Some problems in image classification


1. Spectral classes basically land cover classes (except in
arid areas)
2. Differences in time (season !!) between image acquisition
and field data collection Terrain Image
3. Shadows
4. Mixed pixels (mixels)

INTERNATIONAL INSTITUTE FOR GEO-INFORMATION SCIENCE AND EARTH OBSERVATION 25

You might also like