Introduction

to
IMAGE ANALYSIS
1

WORKSHOP

AT.769015
DR.TARAVUDH TIPDECHO
W212/W202 (10:00-12:00)
19 August 2009

Dr.Taravudh Tipdecho, SET

 What is it?
2

 Image analysis is the extraction of meaningful infor

mation from images; mainly from digital images by m
eans of digital image processing techniques.
 In general, there are two types :
 Pixel-based image analysis
 Object-based image analysis.

Dr.Taravudh Tipdecho, SET

 Pixel-based image analysis
3

 In classical image classification approach, the unit is a

single pixel (called pixel based approach). This pixel-b
ased approach utilizes spectral information of the pix
els to classify the image. Normally the different physi
cal properties of earth objects have different spectral i
nformation and can be classified by this pixel based ap
proach.
 The ability of this analysis is limited when objects hav
e similar spectral information.
 This is based on “binary theory”, YES/NO (0/1)

Dr.Taravudh Tipdecho, SET

 So,
4

Different land cover types with distinct spectral pattern,

the pixels cluster of these land cover types will be formed
separately.
Dr.Taravudh Tipdecho, SET
 In addition,
5

In the case of land cover types with similar spectral pattern,

the pixels cluster will overlap in the feature space. It is call
“mixed pixels”. These area cannot be classified correctly in
pixel based image analysis. (it may be, if + spatial information)
Dr.Taravudh Tipdecho, SET
 Object-based image analysis
6

 It combines spectral and spatial information to classif

y the image. This spatial information includes texture
information of the object itself and also the context inf
ormation that describe the relationship between the o
bjects being classified and the remainder of the image.
 Object based image analysis is based on “Fussy Theor
y”, 0-1, 0 means exactly NO, while 1 is exactly YES.

Dr.Taravudh Tipdecho, SET

 Feature space
7

A graph that show values of the feature vector is called feature space.
Dr.Taravudh Tipdecho, SET
 Image Classification
8

 The basic assumption of image classification is a specific pa

rt of the feature space corresponds to a specific class. Once t
he classes have been defined in the feature space, each ima
ge pixel can be compared to these classes and assigned to th
e corresponding class.
 If classes do not have distinct clusters in the feature space, i
mage classification can only give results to a certain level of
reliability.
 The principle of image classification is that a pixel is assign
ed to a class based on its feature vector.

Dr.Taravudh Tipdecho, SET

 In RS, there are 2 basic approaches
9

 A pixel-based approach is based on conventional statistical

techniques such as parallelepiped, maximum likelihood, an
d minimum distance procedures. This approach performs
well but the ability for resolving inter-class confusion is limi
ted. One pixel can only be classified into one class.
 Object-based approach, not only the spectral information
will be used as classification information, but the texture an
d context information in the image will be combined into cl
assification as well.

Dr.Taravudh Tipdecho, SET

 Pixel-based image analysis:
10

 It consists of two kinds of method

 Unsupervised, is used when no external information ab
out landcover or area interest. It does not utilize training
data as the basis for classification. This examine unknow
n pixels in an image and aggregate them into the number
of classes based on natural grouping. There are numerous
algorithms such as K-Means
 Supervised, is used when the analyst supervises the pix
el categorization process by specifying. There are three ba
sic stages: training stage, classification stage, and accurac
y assessment.

Dr.Taravudh Tipdecho, SET

 Training stage
11

 To yield acceptable classification results, training data must

be most representative and complete. This means training s
amples for each information class should be typical and cov
er all the spectral variability for that information class.

Dr.Taravudh Tipdecho, SET

 Classifier:
12

Minimum distance to mean classifier

 Mean for each information

class is first determined

 Unknown pixel is calculated

based on distance to mean

and assign them to the
closest class.
 Some pixel are further than

distance threshold defined,

they are named as
“unknown” or “unclassified”

Dr.Taravudh Tipdecho, SET

 Classifier:
13

Parallelepiped
 This classifier is sensitive to
variance by considering
the range of spectral values
in each category training set.
 Overlap area is a problem.

Dr.Taravudh Tipdecho, SET

 Classifier:
14

Maximum likelihood
 It is based on a normalized estimate
of a probability density function
of each class.

The ellipses around the mean of

the training of the training samples
are equiprobability contours. Pixels
are classified by calculating the
probability and with the highest
probability pixels are classified to that
information class.
Dr.Taravudh Tipdecho, SET
 Accuracy Assessment
15

 In this context, accuracy means the label of agreement betw

een label assigned by classifier and the class allocation on t
he ground collected by user as test data.

Dr.Taravudh Tipdecho, SET

 Object-based image analysis
16

 In object-based analysis is based on segments or objects, no

t only single pixel.
 The classifier in object is based on fuzzy logic using membe
rship to express an object’s assignment to a class (0.0 -> 1.0
)
 The degree of membership depends on the degree to which
the object fulfill the class-describing condition.
 The first step is always to form the processing unit by image
segmentation.

Dr.Taravudh Tipdecho, SET

 Image Segmentation
17

 Objects formed by grouping pixels according to a certain cri

terion of heterogeneity and homogeneity by image segment
ation.
 Image segmentation split into two main domains:
 Knowledge driven (top down) – user already know what he
want to extract from the image but he does not know how to pe
rform extraction. By formulating a model, the system tried to fi
nd the best methods of image processing to extract them.
 Data driven (bottom up) – a kind of data abstraction or data
compression ,as with clustering method at first the generated s
egments are only image object primitives. It is up to the user to
determine what kind of real world objects.

Dr.Taravudh Tipdecho, SET

Object-based image analysis requires ?
18

 It requires complete segmentation of an image to obtain image ob

ject primitives as basic processing units. Therefore the following
part should be considered:
 Global Thresholding, control of meaningful threshold is very impo
rtant.
 Region Growing Algorithm, cluster pixel starting from a limit of s
ingle seed point. This algorithm basically depends on a set of seed poi
nts.
 Texture Segmentation Algorithm, characteristic features are ext
racted from the textured input image and ranged from spatial freque
ncy.
 Knowledge-based approach, try to incorporate knowledge derive
d from training areas or other sources into the segmentation process.

Dr.Taravudh Tipdecho, SET

 Fuzzy Theory
19

 Fuzzy theory allows a greater flexibility because compared

with binary theory which can only have two extreme values
of yes or no, and one pixel only belong to one information cl
ass.
 Fuzzy set theory allows one pixel to hold several non-zero
membership grades for different information classes.

Dr.Taravudh Tipdecho, SET

 Fuzzy Classification
20

 A fuzzy system used for classification generally comprised o

f three main steps,

2.Fuzzification 1. input

3.Fuzzy rule base

4.Defuzzification 5.output

Dr.Taravudh Tipdecho, SET

 Fuzzification
21

 Fuzzification, describes the transition from crisp system t

o a fuzzy system. It assigns membership value between 0 an
d 1 to each feature. The membership value is defined by so c
alled membership functions: monotonic, triangular, trapez
oidal and bell-shaped. Each membership function generate
s different membership grades.

Dr.Taravudh Tipdecho, SET

 Fuzzy Rule Base
22

 Fuzzy rules are “if-then” rules. Operators can be used to c

ombine the fuzzy rules. The basic operator are “AND” and “
OR”.
 “AND” represents the minimum, meaning that the minimu
m value of all rules defined the return value.
 “OR” represents the maximum value, meaning that the max
imum value of all rules defines the return value.
 A fuzzy rule base delivers a fuzzy classification, which consi
sts of discrete return values for each of the classes.

Dr.Taravudh Tipdecho, SET

 Defuzzification
23

 Defuzzification is a reverse process of fuzzification. It del

ivers a crisp classification. To produce land cover classificat
ion results like maps, the fuzzy results have to be translated
back to a crisp value, which means that an object is either a
ssigned to a class or not.

Dr.Taravudh Tipdecho, SET

 Fuzzy classification for object based
24

 In object based image analysis the classifier uses a degree of

membership to express an object’s assignment to a class. Th
e membership value usually lies between 1.0 and 0.0, where
1.o expresses full membership and 0.0 absolutely non-mem
bership.
 The degree of membership depends on the degree to which
the objects fulfill the class-describing conditions.
 This classification lead to express uncertainty about the clas
s’s description.

Dr.Taravudh Tipdecho, SET

 Procedures, you need to know
25

Remote sensing data

Data preprocessing/ georeferencing / etc.
Pixel based image analysis (supervised / unsupervise
d)
Object based image analysis (segmentation)
Classifier selection / Modeling
Accuracy assessment
Map composition
result

Dr.Taravudh Tipdecho, SET