OBIA - Tutorial-OBIA Tutorial V1 - Slides Only

OBIA Tutorial
Introduction to Object-based Image Analysis
Stefan Lang, Florian Albrecht, Thomas Blaschke

Centre for Geoinformatics (Z_GIS)
Paris-Lodron University Salzburg
Centre for Geoinformatics, Salzburg University
OBIA Tutorial
Slide 2 / 96
Disclaimer
This tutorial was built upon teaching material for courses on advanced remote
sensing, delivered by Dr. S. Lang and Prof. T. Blaschke between 2002 and
2006. Explanations are partly quotes from literature.
Its purpose it to give an introduction to the emerging field of object-based image
analysis and to provide a comprehensive overview of methods involved and
the respective background.
The tutorial is available as slide set only or slide set with additional text notes.
Both are provided in PDF (Acrobat Reader required for display).
Dissemination of this Version 1.0 of the OBIA Tutorial is free. Please obey fairuse conditions as being commonly understood.
We gratefully acknowledge the financial support from Definiens AG
(www.definiens.com), which was granted for the compilation of this tutorial.
Contact
Dr. Stefan Lang
Centre for Geoinformatics
Schillerstr. 30, Bauteil XV, 3. Stock
5020 Salzburg
+43.662.8044.5262
stefan.lang@sbg.ac.at
Please cite the tutorial as follows

Lang S, F Albrecht & T Blaschke (2006)
OBIA-Tutorial Introduction to Objectbased Image Analysis, V 1.0 Salzburg.
Centre for Geoinformatics | 2006 Lang/Albrecht/Blaschke |
OBIA Tutorial
Slide 3 / 96
Table of Contents
Chapter 1
perception
Chapter 2
theory
Chapter 3
Chapter 4
Chapter 5
Chapter 6
Image interpretation and

Basic concepts of hierarchy
Knowledge representation
Image segmentation
Object-based classification
Accuracy assessment
OBIA Tutorial
Introduction to object-based image analysis
Chapter 1
Image interpretation and perception
OBIA Tutorial
Slide 5 / 96
Outline
Visual perception
Image context
Role of experience
Pixel- vs. object-scope
Using objects
Visual delineation vs. machine-based
segmentation
OBIA Tutorial
Slide 6 / 96
Visual perception
OBIA Tutorial
Slide 7 / 96
Visual perception (2)
OBIA Tutorial
Slide 8 / 96
Visual perception (3)
OBIA Tutorial
Slide 9 / 96
Image context
contrast b/w and shape:

black line in white area
mainly shape (pattern

suppressed): chair with
zebra pattern
contrast b/w and shape

(elongated but acute):
certain stripes pattern
zebra
OBIA Tutorial
Slide 10 / 96
Role of Experience
Human vision is well adapted for complex image
interpretation tasks
Experience built up since early childhood
But human vision is challenged when dealing with remote
sensing imagery:
Applying an overhead view
Dealing with spectral characteristics beyond the visual spectrum
Working with unfamiliar scales and resolutions
-> Experience is an important prerequisite for skillful and

successful interpretation
Images of the Fortress in Salzburg, Austria
Quickbird; bands 3,2,1
Quickbird; bands 4,3,2
Aster; green, red, infrared
Color Photo
OBIA Tutorial
Slide 11 / 96
Pixel- vs. object-scape
river
spectral properties
specific form/shape
municipal park
spectral properties
specific spatial
context
From Definiens, 2004
OBIA Tutorial
Slide 12 / 96
Pixel- vs. object-scape (2)
picture element
integrated signal
treated individually,
no matter where
located
Spectral
reflectance B
Pixel
P2
(99/74)
P2
P1
P1
(183/60)
Spectral
reflectance
OBIA Tutorial
Slide 13 / 96
Pixel-based classification process
Problems
Classified image
Feature space
Source: Jensen, 2005
Raw image
Spectral values belong to more than one information class

No spatial relationships used in classification
Pixel artificial spatial unit
Artifacts (salt-and-pepper effect)
OBIA Tutorial
Slide 14 / 96

Limitations of pixel-based analysis
considering
Colour (spectral reflectance in
n Bands)
Texture (certain environment,
e.g. 3*3 pixels)
but not
Form & shape
Neighbourhood
Context
Levels
OBIA Tutorial
Slide 15 / 96
Using objects
Relation between
target objects and
spatial resolution
Increasing
importance of VHR
EO data
High level of detail
provides extended
set of target classes
Addressing these
target classes in a
Landsat-imagery
would fail
OBIA Tutorial
Slide 16 / 96
Using objects (2)

Manual
delineation
(Bog)
Pixel-based
classification
Object-based
classification
OBIA Tutorial
Slide 17 / 96
Using objects (3)
Meaningful objects
Improved reliability of statistics
Several measurements (pixels) per object

Clear boundaries
CT
DN
Augmented, uncorrelated feature

space
Texture within objects, shape,
neighbours, hierarchy
OBIA Tutorial
Slide 18 / 96
Using objects (4)
Integration of
Remote Sensing and
GIS
GIS users are used to
polygon environment
Aggregation of
information (highly
textured images like VHR
data or radar)
Modelling of scale-specific
ecological processes
through multi-scale
representation
OBIA Tutorial
Slide 19 / 96
Visual delineation vs. machine-based segmentation

Problems occurring with visual delineation visually
(may be solved via segmentation):
selection of appropriate levels of generalization
Individual delineations
placement of boundaries when there
is a graded transition
Problems that challenge machinebased segmentation:

Delineating conceptual boundaries
(e.g. the outline of an orchard,
see below)
Several possibilities for the
delineation of Forest
OBIA Tutorial
Slide 20 / 96
Visual delineation vs. machine-based segmentation (2)

Image Objects =
Landscape Units?
Usually landscape analysis
or -planning is based on
landscape units
manual interpretation
Leads image segmentation to
similar results?
CIR
interpretation
Image
segmentation
Image objects not per se

real world objects
not object recognition
Any raster layer can be used
for segmentation (e.g. DEM)
object merge can improve

appropriateness of objects
OBIA Tutorial
Slide 21 / 96
Visual delineation vs. machine-based segmentation (3)

Some more aspects
Results not necessarily more

correct, but more intuitive, more
convincing, more practical
object generation suitable for
textured or low-contrast image data
VHR-satellite imagery
Airborne optical scanner data
Airborne laser scanning (ALS) data
Synthetic aperture radar (SAR) data
Figure: M. Preiner
Semi-automated image
interpretation
Supports image understanding by
solving complex semantic problems
Monitoring of known structures
e.g. existing land use classification
can be used as pre-defined
boundaries for segmentation
Figure: E. Schpfer
OBIA Tutorial
Chapter 2
Some basic concepts of hierarchy theory
OBIA Tutorial
Slide 23 / 96
Outline
Relevance of multi-scale representation

Scales and scaled representation
Decomposition and holarchy
Multi-scale approach with remote
sensing
Hierarchical patch dynamics paradigm
OBIA Tutorial
Slide 24 / 96
Relevance of multi-scale representation

General remarks
Sensor resolution
nowadays enters the
scale of human activity
Field of potential
applications is
increasing
More and more detailed
object levels can be
represented
Complexity of analysis
tasks increases as well
Meteosat
Landsat 7
IKONOS
...
OBIA Tutorial
Slide 25 / 96
Scales and scaled representation

Scale
Refers to the size of objects
that appear in the landscape
or in a representation of it
(e.g. a map or a satellite
image)
Different Views Different Scales

Different Objects?
Different objects have

different scales
Every object has its inherent
scale
It only appears in a certain
range of scale
Depending on the
elevation of our
viewpoint we see
certain objects
Figure: M. Conradi
OBIA Tutorial
Slide 26 / 96
Scales and scaled representation (2)
Scaling Ladder
Every portion of land
contains objects of many
different scales resulting
in a series of scales
Boundaries within
the scale spectrum
Thresholds between ranges
of scale are never crisp
Level 1
(Super Level)
: Nodes
: Span
Level 0
(Focal Level)
: Branches
Level -1
(Sub Level)
OBIA Tutorial
Slide 27 / 96
Hierarchical patch dynamics paradigm
HPDP combining hierarchy theory and

patch dynamics
A hierarchical scaling strategy dealing with spatial
heterogeneity
Holon = patch = ecological unit at a

particular scale
Interaction of components
Loose vertical and horizontal coupling in structure
and function
OBIA Tutorial
Slide 28 / 96
Hierarchical patch dynamics paradigm (2)
Wu (1999)
Ecological systems as spatially nested patch
hierarchies
Dynamics of an ecological system derived from
the dynamics of interacting patches
Pattern and process are related and change with
scale
Non-equilibrium and stochastic processes do not
necessarily work against stability
OBIA Tutorial
Slide 29 / 96
Relevant range of the

scale spectrum for
landscape analysis
Grain
minimum area at which an

organism responds
comparable to resolution (spatial,
spectral, temporal) in an image
Extent
coarsest scale of spatial

heterogeneity
extent of the whole scene (total
area, bandwidths, covered temporal
duration)
OBIA Tutorial
Slide 30 / 96
Representation in various
levels simultaneously
Each knows its context
and hierarchical
neighbourhood
Object hierarchy
Taken from Lang et al., 2004
OBIA Tutorial
Slide 31 / 96
Decomposability and Holarchy
Landscape as a system
Consisting of interacting
subsystems
Separation
Hierarchical
organization
Decomposition
Separating a system into its
components according to
their scale and ordering
them hierarchically
OBIA Tutorial
Slide 32 / 96
Decomposition and Holarchy (2)
Subsystems are
rather independent
from each other
Horizontal and vertical
coupling
But still have

integrative
tendencies
Part-being of
constituting
elements
Level +1
(Higher
Level)
Level 0
(Focal Level)
Level -1
(Lower
Level)
Embeddedness
Integrative
tendencies
Self-assertive
tendencies
Coupling of
holons
OBIA Tutorial
Slide 33 / 96
Multi-scale approach with remote sensing

Definition of fundamental objects in remote sensing
images
Integrated objects vs. aggregate objects
Interaction of objects within and across scale domains
What scales should be chosen for the different objects?
At what scale should hierarchies be established?
OBIA Tutorial
Chapter 3
Knowledge representation
OBIA Tutorial
Slide 35 / 96
Outline
What is knowledge?
Cognition networks
Image understanding
Production systems vs. adaptive learning
OBIA Tutorial
Slide 36 / 96
What is knowledge?
Knowledge plays key role in image interpretation

part of remote sensing
Implicit human knowledge is supplemented with explicit
knowledge by training
Artificial intelligence distinguishes knowledge into:

Procedural knowledge (specific computational functions)
Represented by a set of rules
Structural knowledge (how concepts of a domain are

interrelated)
For image understanding in remote sensing: Are there links
established between image objects and real world
geographical features?
Rich semantic content
Represented by a semantic network
OBIA Tutorial
Slide 37 / 96
What is knowledge? (2)
Structural knowledge can be organized in knowledge

organizing systems
Realised as graphic notations such as semantic nets or
frames
Semantic knowledge representation (using inheritance
concept, e.g. is part of, is more specific than, is an
instance of) as formal framework for image analysis
Semantic net
Example:
To be created
Spe
c tr
Golf
Course
Water
Control over existing
valu al
e
low IR
connections, once
is more
is part of
specific than
established
Size
Transparency and
Pond
small
operability
compare to Pinz, 1994; p. 94
OBIA Tutorial
Slide 38 / 96
Cognition network
Cognition network
Structural knowledge
Procedural knowledge
Number and parameterisation

of segmentation layers
How to group pixels from an

image into segments
(segmentation)
System of target classes

Class definitions, stored in a
semantic net
How to assign segments from

the image to their corresponding
classes (classification)
Purpose of the cognition network
Documentation of every step and setting in the process of image understanding
Usefulness of the approach
Transparent, a suitable means for communication

Reproducible
To a high degree comprehensible
Technically transferable to other scenes
OBIA Tutorial
Slide 39 / 96
Image understanding
1.
Definition
Image understanding (IU) is a

1. Definition
2. Extentofofthe
IUimage
Image understanding (IU) is a process leading to the description
2.
3.
process leading
description
content (Reconstruction
of to
anthe
imaged
scene)
Extent of IU of the image content

Reaching form signals
reconstruction
an imaged
scene)representation
Reaching (=
form
signals (imageofdata)
to a symbolic
of the scene
(image
data) to a
content
Symbolic representation
Circumstances for IU
of the
sceneby:
content
Outcome depends on the domain of interest of the interpreter,
defined
6. Involved disciplines
Image understanding
Underlying research question

Specific field of application
Pre-existing knowledge and experience of the interpreter
processing
Image
3. Conditions for IU
4.
Output
description
Pattern recognition
Description of real-world objects and their relationships in the specific scene
Output
description
Artificial Intelligence
Resulting in thoroughly4.
described
features
(not mere listing and labelling of features)
Outcome depends on the
5.
Knowledge input
domain of interest of the
Process is driven by Description of real-world
interpreter, defined by:
5. Knowledge
input
Utilisation
of processing knowledge
objects and their
Underlying research
transformation of structural knowledge
relationships in the
question
6.
Involved disciplines
Process is driven
by
specific scene
Image processing
Specific field of
utilisation ofPattern
procedural
recognition Resulting in thoroughly
application
knowledge Artificial Intelligence described features (not
Pre-existing knowledge
transformation of
mere listing and
and experience of the
structural knowledge
labelling of features)
interpreter
OBIA Tutorial
Slide 40 / 96
Image understanding (2)
Utilization and transformation

of knowledge
Planning of the image description outcome

Scene
description
Which target objects,
scales
and classes
as conceptual
reality
(dependent on domain
of interest)
Multi-scale segmentation
Modelling (categorizing the image objects)

Modelling (categorizing
Scene descriptionthe
asimage
conceptual
objects)reality
Utilization and transformation of
knowledge
Planning of the image
description outcome and segmentation
which target objects, scales and classes
(dependent on the domain of interest)
multi-scale segmentation
OBIA Tutorial
Slide 41 / 96

Knowledge
Scene description as conceptual reality

Planning of the image description outcome and segmentation
1st match
Knowledge
transformation
Segmentation problem
Complex image
content
1st match
Domain of
interest
Input:
scale
Class description
ORM rules
Labels
body plan
Target
objects
classes
Real world scene

OBIA Tutorial
Slide 42 / 96

Knowledge
Output:
Spatial distribution
of known features

2nd match
Class system
2nd match
Spectral,
structural,
semantic
Knowledge
transformation
Object
hypotheses
Embeddedness
1st match
Input:
Class description
ORM rules
Labels
body plan
Object modelling
1st match

1st match
Complex image content
Domain of interest
Target
scale
objects
classes
Real world scene
OBIA Tutorial
Slide 43 / 96

Knowledge
Output:

Semantic system
3rd match
of known features
conceptual reality
Spatial
distribution of
categorized
objects
2nd match
Transparency
Transferability
Objectivity
Scene description
Class system
2nd match
Knowledge
transformation

2nd match
Class system
Spectral,
structural,
semantic
Object hypotheses
Embeddedness
Object modelling

1st match
Domain of interest
Target
scale
objects
classes
Real world scene
OBIA Tutorial
Slide 44 / 96

Knowledge
Output:

Semantic system
conceptual reality
3rd match
of known features
Spatial distribution of
categorized objects
Transparency
Transferability
Objectivity
Scene description
2nd match

2nd match
Knowledge
transformation
Class system
Spectral,
structural,
semantic
Object hypotheses
Embeddedness
Object modelling

1st match
1st match
Input:
Class description
ORM rules
Labels
body plan
Domain of interest
Target
scale
objects
classes
Real world scene
OBIA Tutorial
Chapter 4
Image segmentation
OBIA Tutorial
Slide 46 / 96
Outline
Short history
Image segmentation in remote sensing
Categories of segmentation algorithms
Histogram-based/pixel-based segmentation
Region-based segmentation
Edge-based segmentation
Scale space analysis - Image Objects and Blobs

Multiresolution Segmentation algorithm
Region merging and scale
Local mutual best fitting approach
Colour and shape homogeneity
OBIA Tutorial
Slide 47 / 96
Short history
1980ies: first developments of image

segmentation
Major driver: industrial vision
Rarely made use of in remote sensing until

the late 1990ies
Reason: little progress in segmentation of multiband images, algorithms not made for EO data
Since then a high number of segmentation

algorithms has been developed
Availability in commercial products made use of
segmentation of EO data more common
OBIA Tutorial
Slide 48 / 96
Image segmentation in remote sensing
Division of an image into regions so that
the whole scene is covered by regions (spatially

continuous, exhaustive segmentation)
the regions do not overlap
the regions are homogeneous within themselves
the homogeneity criteria of neighbouring regions differ
Region (token):
aggregate of pixels grouped together (directly or indirectly)
Homogeneity as overarching principle
relatively homogeneous regions reflect better the Neardecomposability of natural systems

High heterogeneity creates boundary to neighbouring
patches, low remaining heterogeneity within patches
Homogeneity criterion: grey value, colour, texture, form,
altitude, etc.
OBIA Tutorial
Slide 49 / 96
Categories of segmentation algorithms
Pixel-based or
histogram-based
Thresholding techniques
Segmentation in the feature
space
Finding homogenous
objects
Region-based
Region growing, merging
and splitting
Edge-based
Laplace filter, Sobel-operator,
representativeness,
Non-image-related/
non content expecting
Tiling image with a honeycomb
or chessboard structure
Detecting edges
between objects [and
background (matrix)]
Regions defined without
information from the
image
OBIA Tutorial
Slide 50 / 96
Objects as Tokens
Objects vs. background

(matrix) in a grey value image
tokens (token set)
Constellation token
background
brightness
Region (token): aggregate of pixel

grouped according to homogeneity
criterion (directly or indirectly)
profile
objects
profile
Image events (tokens)
OBIA Tutorial
Slide 51 / 96
Histogram-based
Histogram Thresholding:
simplest way to accomplish
exhaustive regional segmentation
Swiss cheese segmentation
for punched parts
One- or bimodal distribution of
grey values, threshold has to be
determined
Quickbird: band 1, thresholded
Quickbird: 6 classes Isodata

classification
OBIA Tutorial
Slide 52 / 96
Region-based segmentation
Region growing
Seed cells are distributed over image
Bottom up (randomly)
Top-down (content expected)
Neighbours (4- or 8-neighbourhood)

are included into region, if
they do not belong to another region yet
the homogeneity criterion H applies
Campbell, p. 346
Two neighbouring regions are unified,

if H applies
OBIA Tutorial
Slide 53 / 96
Region-based segmentation (2)
Split and Merge

Combination of coarse segmentation and
merge
Example: Quadtree
Initially: image as one object
division into 4 parts, if
H does not apply
Resulting quadtree structure
Merge of homogenous quadtree areas
OBIA Tutorial
Slide 54 / 96
Region-based segmentation (3)

Quadtree structure in a Quickbird Image of Salzburg, generated with eCognition
OBIA Tutorial
Slide 55 / 96
Edge-based segmentation
Build the
derivative
edge
brightness
brightness
Region-based segmentation makes sense when

large, compact and coherent objects occur
(blobs)
edge-based segmentation for elongated
structures
Edge: boundary between homogenous areas
line
profile
profile
OBIA Tutorial
Slide 56 / 96
Edge-based segmentation (2)
Workflow
1. Edge detection
Filtering smoothing to
decrease noise in the image
Enhancement revealing
local changes in intensities
Detection select edge
pixels, e.g. by thresholding
Closing of gaps / deleting
artefacts
Combining, extending of
lines
2. Linking the edge pixels to

form the region boundaries
OBIA Tutorial
Slide 57 / 96
Enhancement filters (examples)

Sobel operator
1 2 1
0 0 0
-1 -2 -1
Laplace filter
0 1 0
1 -4 1
0 1 0
vertical
-1 -2 -1
0 0 0
1 2 1
0
horizontal
-1 0 1
-2 0 2
-1 0 1
Compass edge
1 4 1
4 -20 4
1 4 1
-1 0 1
-2 0 2
-1 0 1
90
-2 -1 0
-1 0 1
0 1 2
45
0 1 2
-1 0 1
-2 -1 0
135
The resulting filter

is the combination
of all eight directions.
OBIA Tutorial
Slide 58 / 96
Segmentation by
representativeness
measures
Calculate a
representativeness of
each pixel for its
neighbours
The minima represent
unsteadiness and
indicate edges
Vectorised minima
delimit areas with
similar spectral
properties
OBIA Tutorial
Slide 59 / 96
Watershed segmentation
Watershed segmentation
OBIA Tutorial
Slide 60 / 96
Scale space analysis - Image Objects and Blobs

Blobs
Compact image object
usually with a clear centroid
Temporal aspect
image events
Scale space blobs

Significant within scale domains
defined by
spatial extension, brightness, scale (4D)
Detection
Repetitive Gaussian filtering
Local maxima are determined
Shapes are drawn around
centroids (circle, polygon)
Hyper blob with

image events
Annihilation
Split
Merge
Creation
from Hay stefan.lang@sbg.ac.at
et al., 2003
OBIA Tutorial
Slide 61 / 96
Multiresolution Segmentation algorithm
Design goals
Convince the human eye

Multi-resolution (strictly
hierarchic)
Similar resolution
Reproducibility
Universality
Performance (i.e. speed )
Overview
Region merging technique

Decision heuristics
Homogeneity criterion
Colour homogeneity
Shape homogeneity
compactness and smoothness

OBIA Tutorial
Slide 62 / 96
Region merging and scale
Bottom up region merging technique
Starting with each pixel being a region

A pair of regions is merged into one region,
each merge having a merging cost (degree of
fitting)
Objects are merged into bigger objects as long
as the cost is below a least degree of fitting
(scale parameter)
= the merge fulfils the homogeneity criterion
Starting points for merging distributed with
maximum distance
Pair wise clustering process considering
smallest growth of heterogeneity
Level 3
Level 2
Level 1
Establishing segmentation levels on

several scales using different scale
parameters
(e.g. 2nd level based on 1st level: larger scale

parameter results in larger objects consisting of
the objects of the 1st level)
Pixel level
OBIA Tutorial
Slide 63 / 96
Decision heuristics
Finding an adjacent object B
for an arbitrary object A for
merging them
1. Fitting: when the
homogeneity criterion is
fulfilled
2. Best fitting: when the
homogeneity criterion is
fulfilled, and the merge
between B and A produces
the best degree of fitting
compared to the merge
between A and any other
adjacent object of A
B
B
A
B
fitting
not fitting
B
B
B
best fitting
B
fitting
not fitting
OBIA Tutorial
Slide 64 / 96
Decision heuristics (2)

1.
2.
3.
4.
Fitting
Best fitting
Local mutually best fitting:

find the best fitting object B for the
object A, then find the best fitting object
C for the object B. Confirm that object C
is the object A, otherwise take B for A
and C for B and repeat the procedure.
=> find best fitting pair of objects in the

local vicinity of A following the gradient
of homogeneity
B
CC
C
C
C
B
A
B
C
A
not fitting
A
C
fitting
B
CB
C
best fitting
mutually
best fitting
Global mutually best fitting:

merge the pair of objects for which the
homogeneity criterion is fulfilled best in
the whole image
Distributed treatment order
Use starting points with maximum distance

to all other points treated before
(treatment order defined over pixels or
segments)
A
B
global
mutually
best fitting
OBIA Tutorial
Slide 65 / 96
Homogeneity criterion
Definition of the degree of fitting

Colour and shape homogeneity are weighted against each
other
Compactness and smoothness make up the shape
homogeneity and are weighted against each other
Compactness
Colour homogeneity
Shape homogeneity
Smoothness
Two objects are

similar which are near
to each other in a
certain feature space
Compactness: ideal compact form of

objects (objects dont become lengthy)
Smoothness: boundaries of the edges
dont become fringed
OBIA Tutorial
Slide 66 / 96
Homogeneity criterion (2)

Relation between boundary length l of the object and the
square root of the number n of the pixels of the object
(square root of n equals the side of a square with n pixels)
Compactness
hcompact =
l
n
hcompact =
8
l
b
5.33
4.57
Relation between boundary length l of the object and the

perimeter of the bounding box of the object (bounding
box: shortest possible boundary length)
Smoothness
hsmooth =
6.4
hcompact =
5.84
5.84
6.57
1.125
7.67
8.76
9.49
10.95
hsmooth =
1.3125
1.5
1.625
1.875
OBIA Tutorial
Chapter 5
Object-based classification
OBIA Tutorial
Slide 68 / 96
Outline
Introduction
Available features for classification
Sample- vs. rule-based classification
Fuzzy classification
Class hierarchy
Class-related features
Strengths of object-based classification
OBIA Tutorial
Slide 69 / 96
Introduction
Classifying
Assigning objects to a certain class
according to the classs description
Feature space segmented into
distinct regions which leads to a
many-to-one relationship between
the objects and the classes
Definition of the class descriptions
Available object features
Sample-based or rule-based
classification strategy
OBIA Tutorial
Slide 70 / 96
Available features for classification

layer values
mean
std-dev
geometrical properties
size, shape,
textural properties
layer value texture (e.g. mean of sub objects: std-dev)

shape texture (e.g. directions of sub objects)
hierarchical properties
number of higher levels

number of super or sub objects
object features
colour
shape
texture
hierarchy
relations to classes of
neighbour objects
sub objects (relative area of )
super objects
membership to
class related features

to neighbour objects
to sub objects
to super objects
OBIA Tutorial
Slide 71 / 96
Sample- vs. rule-based classification
Sample-based classification
Define class membership by similarity to
selected samples
Nearest neighbour classifier

Object will be assigned to the
class who's samples are closest
to it in the feature space
Feature B
Sample has to be representative for its class

Use features clearly distinguishing the sampled class
from other classes
Useful approach, if knowledge

about the scenes content is limited
Feature A
OBIA Tutorial
Slide 72 / 96
Sample- vs. rule-based classification (2)
Rule-based classification
Define a class by a rule on one feature or by rules on
several features
Fuzzy or crisp rule definition
Hierarchical relations of classes
Rules can address different kinds of features
Object features
Advantages compared to sample-based

classification
Incorporation of expert knowledge in the classification
Formulation of complex class descriptions
Transparency (especially compared to neural networks)
Transferability
OBIA Tutorial
Slide 73 / 96
Fuzzy classification
Fuzzy vs. crisp classification

Uncertainties about class descriptions
Approximate real-world knowledge in its complexity
Uncertain human knowledge about the world, imprecise human
thinking
Vague (linguistic) class descriptions
Other:
Uncertainty in sensor measurements
Class mixtures due to limited (spatial and spectral) resolution
Possibilities of fuzzy classification

Express each objects membership to more than one class
Probability of an object to belong to other classes
No sharply defined boundaries between classes as in crisp

classification
OBIA Tutorial
Slide 74 / 96
Fuzzy classification (2)

Transition from a crisp to a fuzzy system
(To decide if a feature value belongs to a fuzzy set)

The fuzzy set (A):
Is a certain subset of values of the whole range of an object feature X

(e.g. NIR-band)
represents an object feature class (e.g. forest) within one object
feature
Replace boolean logic (false and true) of the membership value

by a continuous range of [0, , 1]
Define membership function (x)
Example:
Assigning to every object feature
value x a membership value
If > 0, then x belongs to the fuzzy
set A
Relation between object feature and
classification
Choice and parameterisation of the
membership function influence the
quality of the classification
Introducing expert knowledge
1
(x)
0.5
50
85
175
230 255
X
fuzzy set A
X = feature range
x = feature value
A(X) = fuzzy set
= membership value
(x) = membership function
OBIA Tutorial
Slide 75 / 96

Fuzzy rule-base
Fuzzy rule if then for assigning an object

to a class
If feature value x (of the object) is member of
the fuzzy set (e.g. associated with the class
forest), the image object is a member of the
land-cover forest
Class forest defined by

Feature X AND Feature Y
Feature X
1
0.7
x = 215
Combination of fuzzy sets to create advanced

fuzzy rules
Operator AND Minimum operation
Operator OR Maximum operation
Operator NOT inversion of a fuzzy value:
returns 1 fuzzy value
Fuzzy rule-base (combination of the fuzzy

rules of all classes) delivers a fuzzy
classification
Every object has a tuple of return values

assigned to it with the degrees of membership
to each class/degrees of class assignment
Since these values are possibilities to belong to
a class, they dont have to add up to 1 (unlike
probabilities)
x = 0.7
50
215
230
Feature Y
1
0.9
(y)
y = 90
0
90
130
y = 0.9
forest = x AND y = Min(x, y)

= Min(0.7, 0.9) = 0.7
forest = 0.7
pasture = 0.4
water = 0.05
OBIA Tutorial
Slide 76 / 96

Comparison of membership degrees
Reliability of class assignment
The higher the degree of the most possible
class, the more reliable is the assignment
Stability of classification
Stable classification for differences between
highest membership value and other values
Equal membership degrees

high values reliability for both classes:
classes cannot be distinguished with the
provided classification
Low values unreliable classification (use
threshold of a least required membership
degree to ensure quality of classification)
Defuzzification
the smaller, the more reliable

is the classification
Stability
0.5
Maximum membership degree of fuzzy

classification used as crisp classification
forest
pasture
water
OBIA Tutorial
Slide 77 / 96
Class hierarchy
Classes are embedded in a

heritage system
i.e. child
classes inherit all spectral props
from their parents
Classes can be grouped
semantically
Classification can be shown on
different semantic levels
Bright vegetation
Meadow
Football
field
Feature-based
inheritance
Urban area
Built up
Football
field
Semantic
inheritance
OBIA Tutorial
Slide 78 / 96
Classifying an object based on the

classification of other objects
class related features

to neighbour objects
to sub objects
to super objects
Class of other objects works as a

context feature
E.g. a green area (class contains parks, meadows,) is

classified as urban park because it is embedded in urban
area
Newly classified object may be a context feature

itself for other objects
Iterative Process
Possibly indeterministic or even unstable
Mutual and circular dependencies (should be
avoided if possible)
OBIA Tutorial
Slide 79 / 96
Strengths of object-based classification

Additional information can be utilized (shape, texture,
relationship to neighbours)
Overcoming texture through segmentation
Especially useful with VHSR-data and radar data
Objects: increased signal/noise ratio
Decreased number of units to be classified
Adaptable image resolution
Click and classify since based on objects
Classification: data base query
OBIA Tutorial
Chapter 6
Accuracy assessment
OBIA Tutorial
Slide 81 / 96
Outline
Definitions
Non-site specific accuracy assessment
Site-specific accuracy assessment
Error matrix
Limitations of the error matrix
Specifics of object-based accuracy
assessment
OBIA Tutorial
Slide 82 / 96
Definitions
Accuracy
Degree of correctness of a map or classification
(degree to which the derived image classification
agrees with reality or conforms to the truth)
(Foody, 2002)
Error
Discrepancy between the situation depicted on
the thematic map and reality (Foody, 2002)
Significance of accuracy for

Usefulness of thematic maps for land
management
The validity of maps for scientific investigations
OBIA Tutorial
Slide 83 / 96
Definitions (2)
Accuracy assessment
Meaningfully quantify the accuracy of digital land
cover classifications; A classification is not
complete until its accuracy is assessed
(Lillesand, Kiefer; 2000)
Comparison of
1.Pixels or polygons in a remote sensing-derived
classification (the map to be evaluated)
2.Ground reference test information (reference map)
(Jensen, 2005)
OBIA Tutorial
Slide 84 / 96
Non-site-specific assessment
Comparing area percentages of

categories in classified and
reference map (inventory error can
be assessed)
Disadvantage:
Two maps that are similar

according to non-site-specific
accuracy assessment
Inaccurate method (e.g. classes have

similar proportions but may appear in
different locations of the mapped area)
W
A = 20%
Classified Image Reference Map
Difference
42,00%
40,00%
2,00%
13,00%
17,00%
-4,00%
25,00%
22,00%
3,00%
18,00%
19,00%
-1,00%
2,00%
2,00%
0,00%
100,00%
100,00%
F = 64%
W = 16%
A = 20%
F = 64%
W = 16%
from Campbell, 2002
Example of non-site-specific accuracy assessment
Forest
Meadow
Sealed
Water
Bare Rock
OBIA Tutorial
Slide 85 / 96
Site-specific accuracy assessment
Agreement between categories in classified and

reference map at specific locations
Based on site-by-site comparison (using pixels, clusters of pixels

or polygons)
Every site can only be occupied by one class (for a clear
statement of error or correctly classified)
Calculation of error matrices

Reference
Image
Classified
Image
from Campbell, 2002
OBIA Tutorial
Slide 86 / 96
Error Matrix
Error matrix (schematic representation)
IMAGE TO BE EVALUATED
URBAN
CROP
RANGE
WATER
FOREST BARREN
TOTAL
ROW MARGINALS
URBAN
REFERENCE IMAGE
Identification of overall
errors and
misclassifications (by
category)
n n array; where n is the
number of categories
Every cell summarizes the
number of sample units
assigned to a particular
category in the
classification relative to
the actual category
CROP
RANGE
WATER
FOREST
BARREN
TOTAL
COLUMN MARGINALS
DIAGONAL ENTRIES GIVE CORRECTLY CLASSIFIED
PIXELS (RANGE CLASSIFIED AS RANGE)
SUM OF DIAGONAL ENTRIES GIVE TOTAL NUMBER
OF CORRECTLY CLASSIFIED PIXELS
Diagonal cells (upper left to

lower right): correctly
classified units
Off-diagonal cells: error in the
remote sensing classification
relative to the ground
reference information
from Campbell, 2002
OBIA Tutorial
Slide 87 / 96
Error Matrix (2)

Percentage correct (overall
accuracy)
Example of an error matrix

IMAGE TO BE EVALUATED
Sum of diagonal entries divided

by total observations
Regarding an error from two

different viewpoints
Error of omission: correct class
hasnt been recognised by the
classification process (exclusion
from category)
Error of commission: by mistake
the unit has been assigned to the
wrong class (error of inclusion)
Producers accuracy
Consumers accuracy
CROP
RANGE
WATER
150
21
17
730
93
14
RANGE
33
121
320
WATER
18
FOREST
23
BARREN
TOTAL
PA%
30
225
66.7
115
21
973
75.0
23
54
43
594
53.9
11
83
126
65.9
81
12
350
13
483
72.5
39
15
11
115
191
60.2
TOTAL
248
979
451
134
555
225
1748
CA%
60.5
74.6
71.0
61.9
63.1
51.1
URBAN
REFERENCE IMAGE
Errors of omission
Errors of commission
URBAN
CROP
FOREST BARREN
Note: Percentage correct = sum of diagonal entries/total observations =

1748/2592 = 67.4%; CA, consumers accuracy; PA, producers accuracy
Error of omission
Error of commission
Producers accuracy
Consumer
Consumers accuracy
from Campbell, 2002; modified
Accuracies of individual categories

Producers accuracy (a measure of omission error): indicates probability of a
reference unit being correctly classified
Consumers accuracy (users accuracy, a measure of commission error):
probability of a classified unit on the map actually representing that category on
the ground
OBIA Tutorial
Slide 88 / 96
Error Matrix (3)
Kappa coefficient ()
Need for a more objective accuracy assessment
Compensation of the effect of chance agreement
Example:
Random assignment of pixels into 3 classes
Results in 33% correctly classified pixels (= overall
accuracy)
4 classes 25% correct
5 classes 20% correct
Kappa coefficient
Measure of difference between observed agreement
(between a classification and reference information) and
agreement by chance
OBIA Tutorial
Slide 89 / 96
Error Matrix (4)

Kappa coefficient (Khat)
= (observed expected)/(1 expected)

Observed agreement expected agreement
=
1 expected agreement
When the observed agreement approaches 1 and the chance agreement
approaches
0, agreement
then approaches
1
If:
Observed
1
Then: 1
Expected agreement 0
= 1 perfect agreement between classification and reference data
= 0 agreement is not better than a random classification
Explanation of the formula:
Observed agreement = overall accuracy

Expected agreement = sum of the products of the consumers accuracy (CA) and
the producers accuracy (PA) of each class
Quelle: Foody, 2002; S. 188
OBIA Tutorial
Slide 90 / 96
Limitations of the error matrix

Error matrix not a standardized measure
Many different indices, none of them ideal for every single problem
Used samples
Sampling design (stratified, random) and sample size can be limiting

factors
Type of error
thematic error vs. error of location
Accuracy of the reference data accuracy
Ground truth data is a classification itself and may contain error

Often remotely sensed data is used as a surrogate for ground data
No assessment of the spatial distribution of error

Severity of an error is equalized
Minor errors between relatively similar classes vs. major errors between
very different classes
Attempts to represent continua by a set of discrete classes
Problem with mixed pixels containing more than one class
Classes are not mutually exclusive within this pixel (which is a problem for
site-specific accuracy assessment)
OBIA Tutorial
Slide 91 / 96
Specifics of object-based accuracy assessment

Need for a different handling of object-based accuracy assessment
Thematic assessment of object-based classification
Site-specific approach (making use of random points within objects)

Selecting objects prior to training process to be used as reference
information
Selecting small test areas where every single object is assessed in terms
of its label
Geometry assessment of image objects
Visually checking classified images against manual delineation

Quantitative assessment with GIS overlay techniques
Difficulties in object-based accuracy assessment
No 100% geometrical fit between reference objects and objects to be

evaluated (due to different ways of segmentation and delineation)
When using a fuzzified rule base thematic classes are not mutually
exclusive
Accuracy is a matter of geometrical and semantic agreement
OBIA Tutorial
Slide 92 / 96
Specifics of object-based accuracy assessment (2)
Visually checking classified images against manual

delineation
Classification result of Test Cell #6
Test Cell 6 - classified
16,9%
26,5%
3,3%
structured
shaded_veg
meadow
street
53,3%
Resulting thematic content
Subset of areal
photo with a
randomly selected
15x15m cell
15,8%
27,0%
3,5%
structured
shaded_veg
meadow
53,7%
street
Design: D. Hlbling
Test Cell 6 - digitised
Visual delineation result of Test Cell #6
OBIA Tutorial
Slide 93 / 96
Specifics of object-based accuracy assessment (3)

Quantitative assessment of
object geometry accuracy
Defining a tolerance (spatial overlay

strictness) to allow minor differences
in the object outlines
Testing of objects
Classification object has the same

extent as the reference polygon
(stable object)
Reference object doesnt match and is
split into several objects of the
classification
Complex geometry with objects not
having the same boundaries
Characterisation of object
relationship (object fate, similar to
object-based change analysis)
Good objects
Expanding objects
Invading objects
Tolerance buffer:
Selected digitised
polygon with
0.5m buffer
Case 1:
Object stays
the same
Case 2:
Object splits into
n sub-objects
Case 3:
Complex geometry
with expanding and
intruding objects
Approach needs further development

Good obj
Expanding
obj
Invading obj
OBIA Tutorial
Slide 94 / 96
Outlook
A next version of this tutorial is under preparation.

We will include the following topics:
Scale-specific vs. scale-adaptive segmentation
Class modelling
segmentation and classification
as a cyclic process
Basics of CNL (cognition network language)
More on object fate
object-based change detection
and object-based accuracy assessment
Application showcase (forestry, urban studies,
habitats and wetlands, etc.)
OBIA Tutorial
Slide 95 / 96
References
AJ
Note that with a few exceptions only
literature is listed from which text was taken
Figures in brackets ([]) indicate the respective
chapters
OBIA Tutorial
Slide 96 / 96
References
LZ
Note that with a few exceptions only
literature is listed from which text was taken
Figures in brackets ([]) indicate the respective
chapters

OBIA - Tutorial-OBIA Tutorial V1 - Slides Only

Uploaded by

Copyright:

Available Formats

OBIA - Tutorial-OBIA Tutorial V1 - Slides Only

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

OBIA - Tutorial-OBIA Tutorial V1 - Slides Only

Uploaded by

Copyright:

Available Formats

OBIA Tutorial

Introduction to Object-based Image Analysis

Stefan Lang, Florian Albrecht, Thomas Blaschke

Please cite the tutorial as follows

Image interpretation and

Centre for Geoinformatics | 2006 Lang/Albrecht/Blaschke |

Introduction to object-based image analysis

Centre for Geoinformatics, Salzburg University

Centre for Geoinformatics | 2006 Lang/Albrecht/Blaschke |

Centre for Geoinformatics | 2006 Lang/Albrecht/Blaschke |

Visual perception (2)

Centre for Geoinformatics | 2006 Lang/Albrecht/Blaschke |

Visual perception (3)

Centre for Geoinformatics | 2006 Lang/Albrecht/Blaschke |

contrast b/w and shape:

mainly shape (pattern

contrast b/w and shape

Centre for Geoinformatics | 2006 Lang/Albrecht/Blaschke |

-> Experience is an important prerequisite for skillful and

Quickbird; bands 3,2,1

Quickbird; bands 4,3,2

Aster; green, red, infrared

Centre for Geoinformatics | 2006 Lang/Albrecht/Blaschke |

Pixel- vs. object-scape

Centre for Geoinformatics | 2006 Lang/Albrecht/Blaschke |

Pixel- vs. object-scape (2)

Centre for Geoinformatics | 2006 Lang/Albrecht/Blaschke |

Pixel- vs. object-scape (3)

Pixel-based classification process

Source: Jensen, 2005

Spectral values belong to more than one information class

Centre for Geoinformatics | 2006 Lang/Albrecht/Blaschke |

Pixel- vs. object-scape (4)

Using objects (2)

Centre for Geoinformatics | 2006 Lang/Albrecht/Blaschke |

Using objects (3)

Several measurements (pixels) per object

Augmented, uncorrelated feature

Centre for Geoinformatics | 2006 Lang/Albrecht/Blaschke |

Using objects (4)

Centre for Geoinformatics | 2006 Lang/Albrecht/Blaschke |

Visual delineation vs. machine-based segmentation

Problems that challenge machinebased segmentation:

Centre for Geoinformatics | 2006 Lang/Albrecht/Blaschke |

Visual delineation vs. machine-based segmentation (2)

Image objects not per se

object merge can improve

Centre for Geoinformatics | 2006 Lang/Albrecht/Blaschke |

Visual delineation vs. machine-based segmentation (3)

Results not necessarily more

Introduction to object-based image analysis

Centre for Geoinformatics, Salzburg University

Relevance of multi-scale representation

Centre for Geoinformatics | 2006 Lang/Albrecht/Blaschke |

Relevance of multi-scale representation

Centre for Geoinformatics | 2006 Lang/Albrecht/Blaschke |

Scales and scaled representation

Different Views Different Scales

Different objects have