Está en la página 1de 115

Unit 1: Introduction

Level:
Beginner
Time:

This unit should not take you more than 1 hour


Resources:

A licence of Definiens Developer (Version 7.0 was used to develop


these units).

An image you can load into Definiens for this exercise. For this unit
the Landsat 7 image orthol7_20423xs100999.img (Row: 204 Path: 23
Date 10/09/1999) over North Wales and available from the Landmap
service is recommended.
By the end of this unit you should:
Be aware of the purpose of the Definiens Developer software and the
types of projects the software can be used for.
Know how to create a new project within Definiens and the options
available for creating a project.
Know the main elements of the Definiens Developer user interface.
Know how to change the viewing properties.

1.1. Background
This section will provide you with a brief outline of the purpose of the
Definiens software and an overview of the collection of software tools
Definiens have to offer. For a more detail on Definiens products please visit
the Definiens website http://www.definiens.com or contact them directly.
1.1.1 Purpose
The purpose of the Definiens Developer is to facilitate the development of
object oriented rule based classification procedures. Therefore, rather than
simply independently classifying the individual pixels within the scene the
image is split (segmented) into regions representing objects within the scene.
Working with objects rather than pixels have numerous benefits over
traditional pixel based analysis, for example, the spatial relationship between
objects can be represented or the shape of an object analysed. Definiens
Developer provides an easy to use (although relatively step learning curve)
interface to represent the classification rules and visual scripting interface
(processes) to control the segmentation and classification process.
1.1.2 The Suite of Definiens Tools
Definiens Developer is one of a number of tools which Definiens produce
which together form the Definiens Enterprise Image IntelligenceTM suite
(Figure 1).

Figure 1.1. Definiens Enterprise Image Intelligence Suite, Client and Server software.
(Source: Definiens.com)

The software can be divided into three categories End-user, Developer and
Server-side, where the use of each depends on your role within the image
processing chain.
The end-user products include Definiens Architect, Definiens Analyst and
Definiens Viewer. Definiens Viewer is the simplest of the three and intended
for a user to simply view results that have been previously processed.
Definiens Analyst allows a use to import and execute fully automatic
classification processes (previously developed) and view the results. Finally,
Definiens Architect provides a framework within which a user can run fully and
semi automatic image analysis programs, written using Definiens processes,
and provides a mechanism for manual correction of results.
Definiens Developer encapsulates the functionality of the Viewer, Analyst and
Architect products but with the additional functionality to develop the ruleware
(classification algorithms) required to process image data.
Finally, the Definiens eCognitionTM Server provides functionality to process
your images through a dataset type infrastructure where the ruleware
(developed using Definiens Developer) can be executed simultaneously
across a series of servers. Additionally, this infrastructure supports automatic
tiling and stitching of images and results to allow very large images to be
efficiently processed.
1.1.3. Help while using Definiens
Definiens provide customer (and user to user) support through their online
forums (http://forum.definiens.com/index.php) where you can post your
problems or read previous answers. Additionally there is a section where you
can download sample rulesets. Definiens also make a number of documents
available online including presentations, case studies, white papers and
scientific papers (http://www.definiens.com/resource-center_61_24_0.html).

Additionally, your installation of Definiens Developer contains sample data


and a user guide alongside a technical reference guide which contains a
wealth of information/

1.2. Technical Specification


For a full technical specification of the software please refer to the Definiens
User Guide and Reference Guide. These documents are usual install
alongside the software C:\Program Files\Definiens Developer 7.0\UserGuides.

1.3. Understanding the User Interface


1.3.1. Starting Definiens Developer
Definiens Developer is normally available from the Windows start menu:
Start > All Programs > Definiens Developer 7.0
1.3.2 The User Interface
Once started, you will be presented with an interface similar to the one shown
in Figure 1.2.

Figure 1.2. The User interface for Definiens Developer 7.0.

Where the interface is different, from Figure 1.2, you will need to toggle the
view buttons
. The view buttons result in the following interfaces,
Figure 1.3, but for this series of units you will only require the Developer
interface (4) as shown in Figure 1.2.

1) The workspace interface

2) Analysis Interface

3) Results Interface
4) Developer Interface
Figure 1.3. The interface views available with Definiens Developer.

1.3.2.1

Components of the User Interface

Figure 1.4. The user interface with the components labelled.

Data Viewer: The image and classification data viewer. The viewer allows
you to view the imagery you are classifying, including manipulating the band
order and image stretching.
Process Tree: The window within which you develop your ruleset script.
Class Hierarchy: The window displaying the classes you develop.
Image Object Information: This window displays selected feature values for
a selected object.
Feature View: This window displays a list of all the available features within
Definiens Developer and allows the current image objects to be coloured
(green high values and blue low values) given their value for a select feature.
1.3.2.2

Toolbar Icons

Table 1.1 provides a glossary of the icon available on the various toolbars
within Definiens Developer.
Icon

Description
File Toolbar
Create New Project
Open Existing Project

Save Project
New Workspace
Open Workspace
Save Workspace
Predefined Import
View Settings Toolbar
Workspace view
Analysis View
Results View
Developer View
View Image data
View Classification
View Samples (for Nearest Neighbour Classification)
Feature View
Toggle Object Means and Pixel Data
Toggle Object Outlines
Toggle Polygons
Toggle Skeletons
Toggle Image View and Project Pixel View.
Single Layer (Grey Scale)
Mix Three Layers RGB
Show Previous Layer
Show Next Layer
Select Layers to be Displayed and Image Stretch
View Navigation Toolbar
Delete Level
Select Level For Display
Down a Level
Up a Level
Tools Toolbar
Object Information
Object Table
Undo process
Redo process
Class hierarchy

Process tree
Feature View
Managed customised features
Toggle manual editing toolbar
Zoom Toolbar
Manual Editing Toolbar
Single Selection
Polygon Selection
Line Selection
Rectangle Selection
Cut Object
Merge Object Selection
Merge Selected Objects
Clear Merge Object Selection
Filter Classes for Multiple Image Object Selection
Classify Image Objects
Samples Toolbar
Select Samples
Drag and Click Brush to Assign Image Object Samples
Sample Editor
Sample Selection Information
Toggle Sample Navigation
Table 1.1. A glossary of toolbar icons.

1.4. Creating Your First Definiens Developer Project


The first step when using Definiens Developer is to load your data into the
software. During these units this will always be done through the creation
of an individual project. If you later require to process very large datasets or a
large number of individual images a workspace provides a convenient and
efficient way to store and manage these data. Please refer to Definiens
Developer user guide for information about workspaces.
To create your new project either select File > New Project or select the new
project icon (
). You will be presented with the following dialog box (Figure
1.5a) through which you will enter your datasets and parameters to create
your
project.
In
this
example
you
will
load
the
image
orthol7_20423xs100999.img (Figure 1.5b), by selecting Insert next to the
Image Layer List. Once you have loaded your data you need to define the

layer aliases, as shown in Figure 1.5b, where bands 1-6 correspond with
aliases BLUE, GREEN, RED, NIR SWIR1 and SWIR2, respectively. To bring
up the layer properties dialog double click on each image band in turn, or
select the band and select on the edit button.

a) Empty Create Project Dialog

b) Create Project Dialog with Layer


Properties
Figure 1.5. Shows the create project dialog.

You can also add thematic information in the list below the image layers list
which could be used during your classification and segmentation, for example
a polygon shapefile of building. The next step is to give your project a name,
in this case call it Example 1 and check the projection information for your
image has been correctly read. If this information is incorrect then you need to
check the Pixel size (unit) on the right-hand side. The Pixel size (unit)
should be set to auto and the unit to meters and the Use geocoding option
ticked on. You also have the option of re-sampling you imagery to a resolution
of your choice using the Resolution (m/pxl) dialog box and to select a subset
using the Subset Selection button, which present a dialog similar to the one
shown in Figure 1.6. To select a subset you can either draw a red box on the
image in the dialog or provide the pixel limits of your subset. Before finalising
the project and selecting OK we will subset the image (as shown in Figure
1.6), where minimum X is set to 4600, the maximum X is set to 5200, the
minimum Y is set to 4400 and maximum Y is set to 5000.

Figure 1.6. Create a subset of the datasets for the project.

Now click OK to create the project and you will move back to the Definiens
Developer interface, Figure 1.7.

Figure 1.7. The Definiens Developer interface once the project has been loaded.

1.5. Using the Display Options within Definiens Developer


1.5.1. Using the Zoom Toolbar.
Once the project has been loaded you can pan and zoom around the data, in
the display region, using the zoom toolbar, shown below in Figure 1.8.

Figure 1.8. Zoom Functions toolbar

If the zoom functions toolbar is not displayed you can turn it on using the
View>Toolbars menu.
1.5.2. Selecting bands for Display
To select the layer(s) to be displayed you need to use the Edit Image Layer
Mixing dialog, Figure 1.9, available via the following icon

Figure 1.9. The Edit Image Layer Mixing dialog.

Using the Layer Mixing drop down menu you can select the number of layers
to be mixed in the display and then by selecting the individual layers you may
turn then on and off (or increase the weight), Figure 1.20.

a) One layer grey level


b) 3 layer RBG model
Figure 1.20. Selecting the layers for display.

c) Six layer mixing

Also, you can adjust the equalisation (or stretch) of the data layers being
displayed using the Equalizing drop down menu. The available options are
Linear (1.00%), Standard Deviation (3.00) Gamma Correction (0.50),
Histogram and Manual.
1.5.3. Multiple Views
Definiens Developer also allows you to split your display, therefore allowing
you to have multiple views of the same data. This functionality is available
from the Window menu (Figure 5.21.).

Figure 5.21. The Window Menu.

Here the current display can be split horizontally and/or vertically and once
split can be linked to provide views which automatically move together.
Once you have split your screen by selecting the window you wish to change
the same tools as outlined above can be used to manipulate the display
properties in each of the different views.

1.6. Conclusion
In summary, you should now be able to open Definiens Developer, create a
project and manipulate the display to view the data as you wish. The following units
will take you through the segmentation and classification of the
imagery you have loaded into your project and some more advanced features
of the Definiens software.

1.7. Exercises

1) Experiment with the layer properties, such that you can view each image
band individually and then a number of 3 and 6 band mixings. Observe how
the different land cover types visually change as you change the band
mixings.
2) Using the layer combination of your choice (R: NIR G: SWIR1 B: RED, is
recommended) experiment wit the image equalisations available. Again,
observe how the various land cover types change to these changes.
3) Produce a four way split of the display (i.e., a vertical and horizontal split)
and set each region to different viewing properties. Finally, link all four
together (side by side).

Unit 2: Image Segmentation


Level:
Beginner
Time:

This unit should not take you more than 1.5 hours
Resources:

A licence of Definiens Developer (Version 7.0 was used to develop


these units).

An image you can load into Definiens for this exercise. For this unit
the multispectral Landsat 7 image orthol7_20423xs100999.img and its
corresponding panchromatic scene o20423_pan.tif (Row: 204 Path:
23 Date 10/09/1999) over North Wales and available from the
Landmap Service is recommended.
Processes:

RulesetTemplate.dcp

Chessboard_Segmentation.dcp

Quadtree_Segmentation.dcp

Multiresolution_Segmentation.dcp

SpectralDifference_Segmentation.dcp

ContrastSplit_Segmentation.dcp

ContrastFilter_Segmentation.dcp
By the end of this unit you should:
Be able to apply each of the segmentation techniques available with
Definiens Developer to an image.
Be aware of the difference between the various segmentation
algorithms and the types of objects (size and shape) they each
produce.

2.1. Introduction
Segmentation is always the first step of any process within Definiens
Developer as it generates the image objects on which the classification
process will be performed. The important part is for the segmentation process
to identify objects which a representative of the features you wish to classify
and are distinct in terms of the features available within Definiens (e.g.,
spectral values, shape, texture).

2.2. Setup a Project


As with all work within Definiens Developer the first step is to create a project
containing all the datasets required for the study. Although, it is important to
keep an eye on the size of the images you are creating a project with, as
Definiens Developer can become very slow with very large datasets, due to

the number of objects generated during the segmentation process. Therefore,


a subset of the inputted images will be created once again. Your project
should have the same parameters as that shown in Figures 2.1.a and 2.1.b.

a) The project parameters


b) The subset parameters
Figure 2.1. The parameters for setting up the Definiens Project.

Please note the order in which the image files have been loaded, i.e., the
panchromatic band first, as this will decide on the image resolution for the
project. In this case the 25 m multispectral landsat 7 data will be resampled to
the 15 m of the panchromatic data.
Once you have matched your project window to those shown in Figure 2.1.
select OK and create your project.

2.3. Setup your data display


For these exercises it is recommend that you split the display horizontally
(Window menu), where one of the windows contains the panchromatic data
(Figure 2.2a) and the other the multispectral data using the band combination
NIR, SWIR1 and RED as the red, green and blue components (Figure 2.2b),
respectfully (Figure 2.3).

a) The layer mixing properties for the


b) The layer mixing properties for the
panchromatic display
multispectral display.
Figure 2.2. The layer mixing properties for the split windows.

Figure 2.3. The Definiens Developer interface with the project and display parameters defined

2.4. Setup your Process Tree


The process tree (
) will contain the script that you produce to control the
processes (algorithms) which run and the order in which they are executed. It
is important to keep the script that you produce during your segmentation and
classification procedures as organised as possible, this will allow you to
understand what you have done when you come back to it. With this in mind
Figure 2.4 contains the template to which you should aim to adhere to.

Figure 2.4. Template Process Tree.

To insert a process right-click within the process tree window and the
following menu will appear, Figure 2.5. Select Append New and the Edit
Process dialog will appear, Figure 2.6.

Figure 2.5. Process tree context menu.

Figure 2.6. Edit process dialog

The Edit Process dialog is made up of 6 elements. Each of which will


become clear as you move through the notes.
Name:
The name of the process. This can either be manual entered or
automatically provided by the software. A good convention is to only manually
edit the name where nothing else within the process has been changed,
otherwise use the automatic. Finally, the note icon ( ) allows a comment to
be written about the process.
Algorithm:
The algorithm to execute. This drop down menu allows you to select
the algorithm you wish to execute, there is an extensive list of algorithms a
number of which will be used during these units.
Image Object Domain:
The image object domain defines the object(s) on which the algorithm
will be executed. The drop down box and Parameter box allow the level to be
selected. The following button (all objects) allows a class(es) to be defined
while the final button (no condition) allows a rule to be used, for example,
area > 20 m2.
Loops & Cycles:
It is possible to allow a process to form a loop, often in the form of a
while loop and the tick box allows this to be selected.
Algorithm Description:
A simple description of the algorithm you are using.
Algorithm Parameters:
These are the parameters which are associated with algorithm which
has been selected.
To recreated the template shown in Figure 2.4, edit the name of the process
to be Process Template and select the comments button and enter This is a
template ruleset which most process trees will adhere to., the rest of the
process should be left unchanged. Select OK, you have now created your first
process which will simply execute any process which is create beneath it.
To create the next process Segmentation right-click of the process you have
just created and select Insert Child, this will create a new process under your
previous process. Edit the name of this new process to be Segmentation and
select OK. Select the Segmentation you have just created right-click and
select Append New, edit this name of this process to be Classification. Now
repeat this to add the processes Merge and Export.
To move processes you can drag and drop them while holding down the left
mouse button. To place a process under another process drag and drop
holding down the right mouse button.

Finally, you can save and load your process independently of your project
(although, your process is saved within the project), this is done by rightclicking within the process tree window and selecting Save Rule Set.
Alongside the contains of your process tree this will also save any classes or
customised features you have created, which are associated with you
process.

2.5. Multi-Resolution Segmentation


The first and most general segmentation technique available within Definiens
Developer is the Multi-Resolution segmentation. To insert this algorithm within
your process tree right-click on your Segmentation process in the template
you previous entered and select Insert Child. Within the following dialog box
select the algorithm as multiresolution segmentation. If this algorithm is not
available scroll to the bottom of the list and select more and move the
algorithms you wish to have in the list to the right-hand column. You should
now be presented with the dialog shown in Figure 2.7. Table 2.1 describes the
parameters available for this segmentation algorithm.

Figure 2.7. The Multi-Resolution segmentation process.

Table 2.1. An overview of parameters for segmentation


Parameter
Description
Level Name
Name of the level in the hierarchy created by the
segmentation.
Image Layer
Weights

Increases the weighting of the layer when calculating


the heterogeneity measure used to decide whether
pixels/objects are merged. Zero ignores the layer.

Thematic Layer
usage

If any thematic layers are available, allows thematic


layers to be turned on and off individually for use within
the segmentation.

Scale Parameter

Controls the amount of spectral variation within objects


and therefore their resultant size. Has no unit.

Shape - Colour

A weighting between the objects shape and its spectral


colour whereby if 0, only the colour is considered
whereas if > 0, the objects shape along with the colour
are considered and therefore less fractal boundaries
are produced. The higher the value, the more that
shape is considered.

Compactness

A weighting for representing the compactness of the


objects formed during the segmentation.

The multiresolution segmentation creates objects using an iterative algorithm,


whereby objects (starting with individual pixels) are grouped until a threshold
representing the upper object variance is reached. The variance threshold
(scale parameter) is weighted with shape parameters (with separation of
shape and compactness parameters) to minimize the fractal borders of the
objects. By increasing the variance threshold larger objects will be created
although their exact size and dimensions is dependant on the underlying data.
To run the segmentation process, leave the parameters at their default values
and click execute but it is recommended that you give your level a suitable
name. A common level name convention is to number them, starting with
Level 1. Once you are happy with the parameters and have executed the
process you should have successful completed your first segmentation.
Once the segmentation has been executed, select the Show or Hide Outlines
icon (
) and the outlines of the objects (segments) created will be displayed
over the image. Making sure you have the cursor in the cursor mode rather
than the zoom mode, select the objects (with either the outlines turned on or
off) in turn. Using the Image Object Information window (
), you will see
the values for features associated with selected object (e.g., band values)
displayed.
To select the features displayed in the Image Object Information window,
right click within the Image Object Information window and select Select
Features to Display. By double clicking on the feature, you can move them
from one side of the displayed menu to the other. The right hand side contains
those features which will be displayed when selecting an object.
2.
2.5.1 Simple Exercise
The aim of this exercise is to get used to the multi-resolution segmentation
and experiment with the parameters outlined above. The process is the same
as the one you implemented above but the segmentation parameters can be
altered almost indefinitely, but your task is to alter the parameters in such a
way as to achieve the best segmentation over the scene. At this point in time,
you may wish to ignore part of the image (e.g., the field areas) and

concentrate on the upland areas or forests and then switch to concentrate on


another area. You should find that each of these areas require slightly
different segmentation parameters, which we will deal with in later units.
Spend approximately 30 minutes experimenting with different segmentation
input parameters and observe the differences in the images objects created.
Initially try the following:

High Scale factor


Low Scale factor
High Shape weighting
Low Shape weighting
High Compactness weighting
Low Compactness weighting
Use just the panchromatic layer (set all other layer to weight zero).

To remove your segmentation and try new parameters, you need to delete the
level before re-executing your segmentation process, this is done using the
Delete Level icon (

).

An example is shown in Figure 2.8. However, there is no definitive answer as


to whether one segmentation is better than another and the final selection
depends upon whether you are satisfied that the objects you are interested in
classifying are adequately delineated.

Figure 2.8. Segmentation of the Landsat imagery.

2.6. Quad-tree Segmentation

A quad-tree segmentation creates regular square objects, where the size is


defined by the objects variation. Unlike the multi-resolution segmentation, the
objects are created by dividing larger objects until the resultant objects are all
within the upper boundary of allowed variation. As with the multi-resolution
segmentation, the variation at which a final object is created is defined using a
scale parameter.
2.6.1. Simple Exercise
To create a quad-tree segmentation, follow the same procedure used to
create the multi-resolution segmentation but select the quad-tree
segmentation algorithm instead. A scale factor of 20 is recommended for the
segmentation, although others might be more appropriate. Try several scale
factors for representing a) the fields, b) the uplands, c) forests and d)
compromise for all surfaces. Write these down below.
Scale Factor Fields: ___________
Scale Factor Uplands: ____________
Scale Factor Forests: ______________
Scale Factor Compromise for all surfaces _____________

2.7. Chessboard Segmentation


A chessboard segmentation is the simplest segmentation available as it just
splits the image into square objects with a size predefined by the user. The
segmentation does not consider the underlying data and therefore when large
objects are created, the features within the data you are trying to classify will
not be delineated. This segmentation tends to be used in more advanced
processes where segmentation is undertaken in a number of steps combined
with a classification. An example is the tree crown delineation algorithm
created within Definiens Developer for delineation of crowns within mixed
forests in Australia (Bunting and Lucas 2006).
2.7.1. Simple Exercise
To perform a chessboard segmentation, setup the process in the same way
as the multi-resolution and quad-tree segmentations, but select the
chessboard segmentation.
To being with, use a value of 1 for the
segmentation (This will generate objects of 1 x 1 pixel) and then progressively
increase this value. Notice, in each case, how the boundaries and spectral
information of the underlying data are ignored.

2.8. Contrast Split Segmentation


The aim of this algorithm is to split bright and dark objects using a threshold
that maximises the contrast been the resulting bright objects (consisting of

pixel values above the thresholds) and dark objects (consisting of pixel values
below the threshold). The algorithm aims to optimize this separation by
considering different pixel values, within the range provided by the user
parameters, with values selected based on the inputted step size and
stepping parameter. Table 2.2 provides a list of the parameters for the
algorithm.
Parameter
Chessboard
Tile Size

Description
If no level is already present then a chessboard
segmentation is undertaken to generate a set of large
objects which are iterated through during the
segmentation process.

Minimum
Threshold

The minimum grey-level value that will be


considered for splitting

Maximum
Threshold

The maximum grey-level value that will be considered for


splitting.

Step Size

The sizes of the steps the algorithm will use to move from
the minimum threshold to the maximum threshold.
Large values will make the algorithm quicker to calculate
but smaller values will tend to produce better results.

Stepping Type

Either:
Add Calculate each step by adding the value in the
scan step field.
Multiply Calculate each step by multiplying by the value
in the scan step field.

Image Layer

The image layer where on which the algorithm will be


applied.

Class for Bright


Objects

The class the brighter objects (above the threshold) will


be given.

Class for Dark


Objects

The class the darker objects (below the threshold) will be


given.

Contrast Mode

The method the algorithm uses to calculate contrast


between bright and dark objects. The algorithm uses the
mean of possible bright border pixels and the mean of
possible dark border pixels to calculate either the edge
ratio or edge difference which can be used to define the
contrast between two objects. Alternatively, the object
mean (of all pixels) within the bright and dark pixels can
be used.

Execute
Splitting

If yes objects will be split while if no only the threshold


value will be calculated.

Best Threshold

A variable (see units 8) to store the threshold that


maximises the contrast.

Best Contrast

A variable to store the best contrast calculated.

Minimum Rel.
Area Dark

The minimum (relative) area identified as dark objects for


the segmentation to be performed. Range 0-1.

Minimum Rel.
Area Bright

The minimum (relative) area identified as bright objects


for the segmentation to be performed. Range 0-1.

Minimum
Contrast

The minimum contrast threshold for the segmentation to


occur.

Minimum Object The minimum object size for the segmentation to take
Size
place.
Table 2.2. The parameters associated with the contrast split segmentation.

To execute this algorithm you will need to create two classes, one for the
bright objects and one for the dark objects. To do this within Definiens
Developer right-click within the class hierarchy window and select New Class,
you do not need to enter any parameters at this point, so just select OK.
2.8.1. Simple Exercise
As with the other segmentation algorithms try to achieve the best
segmentation of the landscape you can using this algorithm. Remember to
investigate all the parameters to observe there effect of the final
segmentation.

2.9. Spectral Difference Segmentation


This is a merging algorithm where neighbouring objects with a spectral mean
below the threshold given (maximum spectral difference) will be merged to
produce the final objects. To use this segmentation algorithm you are required
to already have a segmentation (level) in place you cannot create a new level
using this algorithm.
2.9.1. Simple Exercise
Using one of the segmentations you have previously generated extend your
process to include a spectral difference segmentation process. Remember, to
add a process right-click and select Append new process, in this case on
the previous segmentation process. Again try to achieve a segmentation that
you think is the best for the landscape within the scene.

2.10. Contrast Filter Segmentation

The contrast filter segmentation uses a combination of two pixel filters to


create a thematic raster layer, with the values no object, object in first layer
(filter response 1), object in second layer (filter response 2), object in both
layers and ignored by threshold. Finally, the raster segmentation is converted
to Definiens image objects using a chessboard segmentation. The parameters
for the algorithm are outlined in Tables 2.3, 2.4 and 2.5.
Parameter
Chessboard
Settings

Description
The chessboard segmentation parameters for producing
the final segmentation from the filter results.

Layer

The image layer on which the filters will be applied.

Scale 1-4

The scale parameters can be used to define membership


to the various classes. See the reference manual for
more information.

Gradient

A minimum gradient threshold

Lower
Threshold

Pixels with a filter response below this threshold will be


assigned to the ignored by threshold class.

Upper
Threshold

Pixels with a filter response above this threshold will be


assigned to the ignored by threshold class.

Table 2.3. The main parameters for the contrast filter segmentation.

The contrast filter segmentation allows the definition of shape parameters to


control the outputted image objects, the param
Parameter
Shape Criteria
Value

Description
Larger values reduce the inclusion of irregularly shaped
objects.

Working on
Class.

The class (see class assignment) on which the operation


will be applied.

Parameter
Enable Class
Assignment

Description
If set as no, the remaining parameters are not used.

No Objects

The class to be pixels with the value no objects will be


given.

Ignored by
Threshold

The class to be pixels with the value Ignored by


Threshold will be given.

Object in First
Layer

The class to be pixels with the value Object in First Layer


will be given.

Table 2.4. The shape parameters for the contrast filter segmentation.

Object in
Second Layer

The class to be pixels with the value Object in Second


Layer will be given.

Object in Both
layers

The class to be pixels with the value Object in Both


layers will be given.

Table 2.5. The classification parameters for the contrast filter segmentation.

2.10.1.

Simple Exercise

As with the previous simple exercises experiment with this algorithm to


achieve a segmentation on the image provided. Although, as you will see this
algorithm does not produce results on par with the other algorithms outlined in
this unit for the image subset provided.

2.11. Conclusion
Following the completion of this unit you should have knowledge of all the
segmentation processes available within Definiens Developer and
implemented each of the algorithms on the image provided.

2.12. Exercises
1) Decide on the most appropriate segmentation algorithm for segmenting this
scene. As you are doing this think of what elements you think provide a good
segmentation and how the different characteristics of the various algorithms
could be used to achieve the segmentation you require.

Unit 3: Nearest Neighbour Classification


Level:
Beginner
Time:

This unit should not take you more than 1.5 hours
Resources:

A licence of Definiens Developer (Version 7.0 was used to develop


these units).

The multispectral Landsat 7 image orthol7_20423xs100999.img and


its corresponding panchromatic scene o20423_pan.tif (Row: 204
Path: 23 Date 10/09/1999) over North Wales and available from the
Landmap service.
Processes:

NN_Classification_Process.dcp
By the end of this unit you should:
Be able to complete all the steps require in the process tree to
complete a classification (using the nearest neighbour classifier) within
Definiens Developer.
Be aware of the parameters and features to aid the nearest neighbour
classification.
Be aware of the classification, merge and export processes.

3.1. Introduction
Within in this worksheet you will create a nearest neighbour classification of a
segmented Landsat 7 image of the area around Llyn Brenig, Denbigh Moors,
North Wales (Figure 3.1). This area contains extensive tracts of upland heath
and bog as well as coniferous forest plantations and grasslands at various
levels of improvement.

Figure 3.1. Ordnance Survey Map of the study area


(http://www.ordnancesurvey.co.uk/getamap).

The Nearest Neighbour (NN) classifier is a supervised classification approach


whereby for each of the classes required, training samples are located and
used to classify all remaining (unknown) objects in the image. The NN
classifier has been used successfully for many classification problems (e.g.,
tree species; Leckie et al., 2005).

3.2. Create a project


As with all work within Definiens Developer the first step is to create a project
containing all the datasets required for the study. A subset of the inputted
images will be created once again. Your project should have the same
parameters as that shown in Figures 3.2.a and 3.2.b.

a) The project parameters


b) The subset parameters
Figure 3.2. The parameters for setting up the Definiens Project.

Please note the order in which the image files have been loaded, i.e., the
panchromatic band first, as this will decide on the image resolution for the
project. In this case the 25 m multispectral Landsat 7 data will be resampled
to the 15 m of the panchromatic data.
Once you have matched your project window to those shown in Figure 3.2.
select OK and create your project.

3.3. Setup your Class Hierarchy

For classification, the first task is to create the classes you require and (in this
case) to insert the Nearest Neighbour Feature into each class. To create a
class, you require the Class Hierarchy window (shown in Figure 3.3) to be
open. If the window is not already visible, then click on the

icon.

Figure 3.3: Class Hierarchy Window before inserting any classes.

To insert a class, right click in the Class Hierarchy window and select Insert
Class (Figure 3.4).

Figure 3.4. Inserting a class into the class hierarchy.

This provides you with an empty Class Description (Figure 3.5).

Figure 3.5. Empty Class Description.

The next step is to edit your class description by first giving your class a
name. For example, give the class the name Water and assign a blue

colour. When you have done this, insert and name new classes of Forest,
Other vegetation and Not Vegetation. You should then have four classes
inserted and named:

Water
Forest
Other Vegetation
Not Vegetation

After giving each class a name, select an appropriate colour for each. This
can be anything you wish, although the final classification will be easier to
understand and interpret if you chose a logical colour (e.g., Green for Forest).
Next, the features (e.g., mean object spectral response) to be used for
classification (in this case, the standard nearest neighbour algorithm) need to
be inserted into the class. To do this, right-click on the and (min) and select
Insert new Expression (Figure 3.6).

Figure 3.6. Inserting a new expression into the class.

This will present the window (Figure 3.7), where you need to select Standard
Nearest Neighbour and click Insert.

Figure 3.7. Selecting the expression to be used for the classification.

Your resulting class description should be similar to that shown in Figure 3.8
for the forest class.

Figure 3.8. The resulting class description to be used for the classification.

The same procedure now needs to be repeated for the remaining three
classes so that you end up with a classification hierarchy similar to that shown
in Figure 3.9.

Figure 3.9 The final class hierarchy.

To select the features used for the nearest neighbour classification use the
Edit Standard NN feature Space function, Figure 3.10a, where initially you
should just use the mean spectral values of the objects, Figure 3.10b.

a) The menu for editing the NN feature


b) The dialog for selecting the features within the
space
NN feature space
Figure 3.10. Editing the features used within the nearest neighbour classification.

3.4. Setting up the Process Tree


3.4.1 Segmentation
As with all classifications in Definiens Developer, the first task is to perform a
segmentation. In this case the use of a multi-resolution segmentation is
recommended, although others could be investigated.
As with previous units, it is recommend you create an outline within your
process tree mirroring that outlined in Figure 3.11. Remember, a process is
created by right-clicking in the process tree window and select Append
Process or Insert Child Process.

Figure 3.11: Process outline and Segmentation process.

To create the process that performs the segmentation, right-click on the


Segmentation process you have already created, select Insert Child Process,
and then the algorithm Multiresolution segmentation.
Choose the
parameters shown in Figure 3.12 and once you have entered these
parameters, click on Execute to perform the segmentation.

Figure 3.12. Parameters used for the segmentation of the Landsat image.

Note, that the layer weighting for the panchromatic band (PAN) has been
increased to 2. This is in take advantage of the extra spatial resolution of the
panchromatic band, 15 m rather than 25 m of the multispectral.
3.4.2. Classification
To run the classification, you need to add a classification process to your
process tree. This is achieved by right-clicking on the process you named
Classification and selecting Insert Child Process. Edit the new process such
that it is similar in appearance to that shown in Figure 3.13. To select multiple
classes, use the Shift and Control keys as you would in Windows Explorer.

Figure 3.13. The process parameters to be used for the classification.

After inputting the parameters into the process, click on the OK button at the
bottom, you need to select samples before performing your classification.
Your process tree should now be similar to that shown in Figure 3.14.

Figure 3.14: The process tree after the inclusion of the classification process.

3.4.3 Merge Result


The next step is to setup the processes which will merge your classification so
that all neighbouring objects of the same class will form single objects. It is
important to merge your classification to identify complete objects. For
example, once merged you can query the lake to find its complete area. To
merge the result you will need to enter a merge process for each class (Figure
3.16; Insert Child), the merge parameters for the Forest class are shown in
Figure 3.15. The class for merging is defined using the Image Object Domain
where the class of interest is defined, if you were to select multiple classes all
the select classes would be merged, removing the boundaries and
classification of these objects.

Figure 3.15. The process parameters to merge the Forest class.

To save time, once you have created you first merge process you can copyand-paste (ctrl-c, ctrl-v or right-click on the process) this process to duplicate it
and then edit the class you wish to merge.

Figure 3.16. The process tree including the merge processes.

3.4.4. Export Result


Finally, we usually wish to export the classification result from Definiens into a
GIS for further processing or the production of a map. Therefore, our final
process will be to export the classification to an ESRI shapefile, Figure 3.17.

Figure 3.17. Process parameters to export the classification as a shapefile.

To select the classes to export you again edit the Image Object Domain,
remember these parameters define the image objects the process will be
applied to. The name of the outputted shapefile has been defined as
Classification while the features to be exported are the area (of the image
object) and the class name. Area is found under Object Features > Shape >
Generic while class name is found under Class-Related features > Relations
to Classification > Class name. For the class name feature you will need to
create it, right-click on the Create new Class name and select Create, leave
the parameters as their default values and just select OK. The shapefile will
output to the directory within which your project is saved, if you have not yet
saved your project then the shapefile will be outputted to the directory
containing the input imagery.
You final process tree should then be the same as the one shown below in
Figure 3.18.

Figure 3.18. Final process tree.

3.5. Select Samples to Train Classifier

The next stage is to select the samples for each of the four classes, you need
to have executed the segmentation process before undertaking these steps If
you are unsure of the distribution of ground cover types, please refer to the
shapefile LlynBrenig_BasicLandcover.shp. To create a sample, you need to
first activate the tool for sample selection (Select Samples) as shown in Figure
3.19.

Figure 3.19 Activating sample selection.

Once you have activated sample selection, highlight the class you wish create
a sample for in the class hierarchy window. Either double click on the objects
you wish to select as samples or hold down the Shift key and use a single
click. To unselect a sample, repeat the process of selection for each chosen
object.
To aid the selection of your samples, Definiens Developer offers two windows
(both available from the menu in Figure 3.19) of information based on the
selected samples. Firstly the Sample Editor window (Figure 3.20) and
secondly the Sample Selection Information window (Figure 3.21).

Figure 3.20. Sample Editor Window

The Sample Editor provides a visual comparison of two classes using a range
of selected features. In Figure 3.20 the Forest and Water classes are
compared using the object means from each spectral band of the Landsat
data. When an object is selected, a red arrow is displayed to illustrate where
the object mean fits in relation to the mean of the other samples. To change
the displayed features, right-click within the main window and select Features
to Display or, if you only want the features being used within the NN
calculation, select Display Standard Nearest Neighbour Features.

Figure 3.21: Sample Selection Information Window

The Sample Selection Information window displays information on the NN


membership boundaries and the distances to the other classes of the selected
object from the selected samples. To select the classes to be displayed, rightclick within the window and select Select classes to Display. Classes
displayed in red have, for the selected object, an overlap in the distance
measure and therefore the samples may need to be re-analysed and altered.
To set the threshold at which a class is highlighted in red, right-click in the

Sample Selection Information window and select Modify critical sample


membership overlap. The threshold ranges from 0 - 1, and 0 highlights any
overlap.
Once you have selected your samples, you should have an image which is
similar in appearance to that shown in Figure 3.22. Bear in mind that the
selection of samples does not have a correct answer. Just select the samples
you consider to be most representative of the classes you wish to separate
and which also give the best separation in the Sample Editor and Sample
Selection windows.

Figure 3.22: Samples selected for the nearest neighbour classification.

3.6. Run your Process.


Now execute the classification process you previously created (right-click and
select execute on the classification process). You should now have a nicely
classified image similar to that in Figure 3.18. If you are unhappy with the
classification, repeat the procedure but select more or alternative samples
before reclassifying the image. To re-run the classification, open the process
and simply click on execute or select the process and press F5 or right-click
on the process and select execute.

Figure 3.18. Final classification of the Landsat data.

3.6.1. Merge the Image Objects


Once you are happy with your classification, execute the merge image objects
processes you have previous created, your results should appear similar to
those shown in Figure 3.19b.

a) Before merging
b) After Merging
Figure 3.19. Before and after merging classes.

The purpose of merging is to create a final classification which is a closer


representation to the objects in the scene. For example, we will now be able
to calculate the area of the whole lake. But, be aware that merging the image
objects removes your samples, as the segmentation will have changed.
3.6.2. Export the Results
Finally, run the process to export the results. This will result in an ESRI
shapefile and allow the creation of a map such as the one shown in Figure
3.20.

Figure 3.20. A map produced using ESRI ArcMap from the result of the Definiens
classification

3.7. Feature Space Optimization Tool


To refine the classification further Definiens Developer offers an automated
feature, the feature space optimization tool, to automatically identify the
features which best separate the classes for which samples have been
selected. To use this feature delete your classification (delete level) and re-run
the segmentation process. You will also need to re-select your samples as
these will be deleted each time you change the segmentation (i.e., merge or
delete level).

Once you have selected your samples open the feature space optimization
tool, Figures 3.21 and 3.22.

Figure 3.21. The NN classification menu.

Figure 3.22. The Feature space optimization tool.

To use this tool, select the features you wish to compare Initially try the
mean, standard deviation and the pixel ratio but later try other combinations.
Then select calculate, once the calculation has finished then select advanced
to see which features offered the best separation, and Apply to the Std NN to
use within the classification.
You can now run your classification step.

3.8. Conclusions
Following the completion of this unit you should now understand basis the
process of classification within Definiens Developer where future examples
will simply build more complex classification and segmentation routines.

3.9. Exercises
1) Experiment with different segmentation parameters, both within the multiresolution segmentation and the other segmentation algorithms. Be aware
that you will have to select new samples each time you delete the level.
2) Experiment with different sets of features within the standard NN feature
space. (Classification > Nearest Neighbor > Edit Standard NN feature
space)
3) Experiment with different sets of features and maximum dimension levels
within the feature optimisation tool. (Classification > Nearest Neighbor >
Feature Space Optimisation)

References
Leckie, D.G., Gougeon, F.A., Tinis, S., Nelson, T., Burnett, C.N., & Paradine,
D. (2005). Automated tree recognition in old growth conifer stands with high
resolution digital imagery. Remote Sensing of Environment, 94, 311-326.

Unit 3: Nearest Neighbour Classification


Level:
Beginner
Time:

This unit should not take you more than 1.5 hours
Resources:

A licence of Definiens Developer (Version 7.0 was used to develop


these units).

The multispectral Landsat 7 image orthol7_20423xs100999.img and


its corresponding panchromatic scene o20423_pan.tif (Row: 204
Path: 23 Date 10/09/1999) over North Wales and available from the
Landmap service.
Processes:

NN_Classification_Process.dcp
By the end of this unit you should:
Be able to complete all the steps require in the process tree to
complete a classification (using the nearest neighbour classifier) within
Definiens Developer.
Be aware of the parameters and features to aid the nearest neighbour
classification.
Be aware of the classification, merge and export processes.

3.1. Introduction
Within in this worksheet you will create a nearest neighbour classification of a
segmented Landsat 7 image of the area around Llyn Brenig, Denbigh Moors,
North Wales (Figure 3.1). This area contains extensive tracts of upland heath
and bog as well as coniferous forest plantations and grasslands at various
levels of improvement.

Figure 3.1. Ordnance Survey Map of the study area


(http://www.ordnancesurvey.co.uk/getamap).

The Nearest Neighbour (NN) classifier is a supervised classification approach


whereby for each of the classes required, training samples are located and
used to classify all remaining (unknown) objects in the image. The NN
classifier has been used successfully for many classification problems (e.g.,
tree species; Leckie et al., 2005).

3.2. Create a project


As with all work within Definiens Developer the first step is to create a project
containing all the datasets required for the study. A subset of the inputted
images will be created once again. Your project should have the same
parameters as that shown in Figures 3.2.a and 3.2.b.

a) The project parameters


b) The subset parameters
Figure 3.2. The parameters for setting up the Definiens Project.

Please note the order in which the image files have been loaded, i.e., the
panchromatic band first, as this will decide on the image resolution for the
project. In this case the 25 m multispectral Landsat 7 data will be resampled
to the 15 m of the panchromatic data.
Once you have matched your project window to those shown in Figure 3.2.
select OK and create your project.

3.3. Setup your Class Hierarchy

For classification, the first task is to create the classes you require and (in this
case) to insert the Nearest Neighbour Feature into each class. To create a
class, you require the Class Hierarchy window (shown in Figure 3.3) to be
open. If the window is not already visible, then click on the

icon.

Figure 3.3: Class Hierarchy Window before inserting any classes.

To insert a class, right click in the Class Hierarchy window and select Insert
Class (Figure 3.4).

Figure 3.4. Inserting a class into the class hierarchy.

This provides you with an empty Class Description (Figure 3.5).

Figure 3.5. Empty Class Description.

The next step is to edit your class description by first giving your class a
name. For example, give the class the name Water and assign a blue

colour. When you have done this, insert and name new classes of Forest,
Other vegetation and Not Vegetation. You should then have four classes
inserted and named:

Water
Forest
Other Vegetation
Not Vegetation

After giving each class a name, select an appropriate colour for each. This
can be anything you wish, although the final classification will be easier to
understand and interpret if you chose a logical colour (e.g., Green for Forest).
Next, the features (e.g., mean object spectral response) to be used for
classification (in this case, the standard nearest neighbour algorithm) need to
be inserted into the class. To do this, right-click on the and (min) and select
Insert new Expression (Figure 3.6).

Figure 3.6. Inserting a new expression into the class.

This will present the window (Figure 3.7), where you need to select Standard
Nearest Neighbour and click Insert.

Figure 3.7. Selecting the expression to be used for the classification.

Your resulting class description should be similar to that shown in Figure 3.8
for the forest class.

Figure 3.8. The resulting class description to be used for the classification.

The same procedure now needs to be repeated for the remaining three
classes so that you end up with a classification hierarchy similar to that shown
in Figure 3.9.

Figure 3.9 The final class hierarchy.

To select the features used for the nearest neighbour classification use the
Edit Standard NN feature Space function, Figure 3.10a, where initially you
should just use the mean spectral values of the objects, Figure 3.10b.

a) The menu for editing the NN feature


b) The dialog for selecting the features within the
space
NN feature space
Figure 3.10. Editing the features used within the nearest neighbour classification.

3.4. Setting up the Process Tree


3.4.1 Segmentation
As with all classifications in Definiens Developer, the first task is to perform a
segmentation. In this case the use of a multi-resolution segmentation is
recommended, although others could be investigated.
As with previous units, it is recommend you create an outline within your
process tree mirroring that outlined in Figure 3.11. Remember, a process is
created by right-clicking in the process tree window and select Append
Process or Insert Child Process.

Figure 3.11: Process outline and Segmentation process.

To create the process that performs the segmentation, right-click on the


Segmentation process you have already created, select Insert Child Process,
and then the algorithm Multiresolution segmentation.
Choose the
parameters shown in Figure 3.12 and once you have entered these
parameters, click on Execute to perform the segmentation.

Figure 3.12. Parameters used for the segmentation of the Landsat image.

Note, that the layer weighting for the panchromatic band (PAN) has been
increased to 2. This is in take advantage of the extra spatial resolution of the
panchromatic band, 15 m rather than 25 m of the multispectral.
3.4.2. Classification
To run the classification, you need to add a classification process to your
process tree. This is achieved by right-clicking on the process you named
Classification and selecting Insert Child Process. Edit the new process such
that it is similar in appearance to that shown in Figure 3.13. To select multiple
classes, use the Shift and Control keys as you would in Windows Explorer.

Figure 3.13. The process parameters to be used for the classification.

After inputting the parameters into the process, click on the OK button at the
bottom, you need to select samples before performing your classification.
Your process tree should now be similar to that shown in Figure 3.14.

Figure 3.14: The process tree after the inclusion of the classification process.

3.4.3 Merge Result


The next step is to setup the processes which will merge your classification so
that all neighbouring objects of the same class will form single objects. It is
important to merge your classification to identify complete objects. For
example, once merged you can query the lake to find its complete area. To
merge the result you will need to enter a merge process for each class (Figure
3.16; Insert Child), the merge parameters for the Forest class are shown in
Figure 3.15. The class for merging is defined using the Image Object Domain
where the class of interest is defined, if you were to select multiple classes all
the select classes would be merged, removing the boundaries and
classification of these objects.

Figure 3.15. The process parameters to merge the Forest class.

To save time, once you have created you first merge process you can copyand-paste (ctrl-c, ctrl-v or right-click on the process) this process to duplicate it
and then edit the class you wish to merge.

Figure 3.16. The process tree including the merge processes.

3.4.4. Export Result


Finally, we usually wish to export the classification result from Definiens into a
GIS for further processing or the production of a map. Therefore, our final
process will be to export the classification to an ESRI shapefile, Figure 3.17.

Figure 3.17. Process parameters to export the classification as a shapefile.

To select the classes to export you again edit the Image Object Domain,
remember these parameters define the image objects the process will be
applied to. The name of the outputted shapefile has been defined as
Classification while the features to be exported are the area (of the image
object) and the class name. Area is found under Object Features > Shape >
Generic while class name is found under Class-Related features > Relations
to Classification > Class name. For the class name feature you will need to
create it, right-click on the Create new Class name and select Create, leave
the parameters as their default values and just select OK. The shapefile will
output to the directory within which your project is saved, if you have not yet
saved your project then the shapefile will be outputted to the directory
containing the input imagery.
You final process tree should then be the same as the one shown below in
Figure 3.18.

Figure 3.18. Final process tree.

3.5. Select Samples to Train Classifier

The next stage is to select the samples for each of the four classes, you need
to have executed the segmentation process before undertaking these steps If
you are unsure of the distribution of ground cover types, please refer to the
shapefile LlynBrenig_BasicLandcover.shp. To create a sample, you need to
first activate the tool for sample selection (Select Samples) as shown in Figure
3.19.

Figure 3.19 Activating sample selection.

Once you have activated sample selection, highlight the class you wish create
a sample for in the class hierarchy window. Either double click on the objects
you wish to select as samples or hold down the Shift key and use a single
click. To unselect a sample, repeat the process of selection for each chosen
object.
To aid the selection of your samples, Definiens Developer offers two windows
(both available from the menu in Figure 3.19) of information based on the
selected samples. Firstly the Sample Editor window (Figure 3.20) and
secondly the Sample Selection Information window (Figure 3.21).

Figure 3.20. Sample Editor Window

The Sample Editor provides a visual comparison of two classes using a range
of selected features. In Figure 3.20 the Forest and Water classes are
compared using the object means from each spectral band of the Landsat
data. When an object is selected, a red arrow is displayed to illustrate where
the object mean fits in relation to the mean of the other samples. To change
the displayed features, right-click within the main window and select Features
to Display or, if you only want the features being used within the NN
calculation, select Display Standard Nearest Neighbour Features.

Figure 3.21: Sample Selection Information Window

The Sample Selection Information window displays information on the NN


membership boundaries and the distances to the other classes of the selected
object from the selected samples. To select the classes to be displayed, rightclick within the window and select Select classes to Display. Classes
displayed in red have, for the selected object, an overlap in the distance
measure and therefore the samples may need to be re-analysed and altered.
To set the threshold at which a class is highlighted in red, right-click in the

Sample Selection Information window and select Modify critical sample


membership overlap. The threshold ranges from 0 - 1, and 0 highlights any
overlap.
Once you have selected your samples, you should have an image which is
similar in appearance to that shown in Figure 3.22. Bear in mind that the
selection of samples does not have a correct answer. Just select the samples
you consider to be most representative of the classes you wish to separate
and which also give the best separation in the Sample Editor and Sample
Selection windows.

Figure 3.22: Samples selected for the nearest neighbour classification.

3.6. Run your Process.


Now execute the classification process you previously created (right-click and
select execute on the classification process). You should now have a nicely
classified image similar to that in Figure 3.18. If you are unhappy with the
classification, repeat the procedure but select more or alternative samples
before reclassifying the image. To re-run the classification, open the process
and simply click on execute or select the process and press F5 or right-click
on the process and select execute.

Figure 3.18. Final classification of the Landsat data.

3.6.1. Merge the Image Objects


Once you are happy with your classification, execute the merge image objects
processes you have previous created, your results should appear similar to
those shown in Figure 3.19b.

a) Before merging
b) After Merging
Figure 3.19. Before and after merging classes.

The purpose of merging is to create a final classification which is a closer


representation to the objects in the scene. For example, we will now be able
to calculate the area of the whole lake. But, be aware that merging the image
objects removes your samples, as the segmentation will have changed.
3.6.2. Export the Results
Finally, run the process to export the results. This will result in an ESRI
shapefile and allow the creation of a map such as the one shown in Figure
3.20.

Figure 3.20. A map produced using ESRI ArcMap from the result of the Definiens
classification

3.7. Feature Space Optimization Tool


To refine the classification further Definiens Developer offers an automated
feature, the feature space optimization tool, to automatically identify the
features which best separate the classes for which samples have been
selected. To use this feature delete your classification (delete level) and re-run
the segmentation process. You will also need to re-select your samples as
these will be deleted each time you change the segmentation (i.e., merge or
delete level).

Once you have selected your samples open the feature space optimization
tool, Figures 3.21 and 3.22.

Figure 3.21. The NN classification menu.

Figure 3.22. The Feature space optimization tool.

To use this tool, select the features you wish to compare Initially try the
mean, standard deviation and the pixel ratio but later try other combinations.
Then select calculate, once the calculation has finished then select advanced
to see which features offered the best separation, and Apply to the Std NN to
use within the classification.
You can now run your classification step.

3.8. Conclusions
Following the completion of this unit you should now understand basis the
process of classification within Definiens Developer where future examples
will simply build more complex classification and segmentation routines.

3.9. Exercises
1) Experiment with different segmentation parameters, both within the multiresolution segmentation and the other segmentation algorithms. Be aware
that you will have to select new samples each time you delete the level.
2) Experiment with different sets of features within the standard NN feature
space. (Classification > Nearest Neighbor > Edit Standard NN feature
space)
3) Experiment with different sets of features and maximum dimension levels
within the feature optimisation tool. (Classification > Nearest Neighbor >
Feature Space Optimisation)

References
Leckie, D.G., Gougeon, F.A., Tinis, S., Nelson, T., Burnett, C.N., & Paradine,
D. (2005). Automated tree recognition in old growth conifer stands with high
resolution digital imagery. Remote Sensing of Environment, 94, 311-326.

Unit 4: Rule Based Classification


Level:
Beginner
Time:

This unit should not take you more than 1 hour


Resources:

A licence of Definiens Developer (Version 7.0 was used to develop


these units).

The multispectral Landsat 7 image orthol7_20423xs100999.img and


its corresponding panchromatic scene o20423_pan.tif (Row: 204
Path: 23 Date 10/09/1999) over North Wales and available from the
Landmap Service.
Processes:

Rulebased_Classification_Process.dcp
By the end of this unit you should:
Know how to create a rule-base classification within Definiens
Developer.
Know about the difference between absolute and fuzzy thresholds.
Know how to create a customized feature within Definiens Developer to
represent a band ratio or relationship (e.g., NDVI).

4.1. Introduction
Following on from the previous unit, you will now implement a more
detailed rule-based classification by using thresholds manually defined within
the class hierarchy rather than a nearest neighbour classification.
This unit uses the same Landsat 7 subset of Llyn Brenig in the Denbigh
Moors, North Wales, although it now aims to identify more classes to increase
the detail of the habitat classification. The aim of the unit is to provide you
with experience in entering thresholds for a rule-based classification and
creating the corresponding processes. You are not expected to identify any
thresholds as these will be given and the next unit will cover techniques
used commonly to identify these.

4.2. How to Define a Threshold in Definiens Developer


Definiens Developer deals with two types of threshold,
1) Absolute Thresholds
2) Fuzzy Thresholds
where each is defined in a similar way within a class description, of a class
within the class hierarchy. The absolute threshold (e.g., > 20 or < 30 or = 15)

is the simplest. To define one of these thresholds within the Class


Description, right-click on and(min) (as when including the NN classifier) and
navigate through the tree structure to find the feature with which you wish to
create a threshold.
Right-click on this feature and select Insert
Threshold(Figure 4.1) to create the threshold.

Figure 4.1: Inserting a threshold for a chosen feature.

When selected, you will be presented with the window shown in Figure 4.2.
Here, you set the threshold and the operator (e.g., <, , =, > or ).

Figure 4.2: Edit Threshold Condition Window.

Within the class description you can add as many of these thresholds as you
require. You can also include and and or statements, as shown in Figure
4.3. By default, all the features you introduce are considered within an and
statement and therefore all thresholds have to be met for the object to be
classified. On the other hand, if the statement is an or statement, only one of
the thresholds needs to be met for the object to be classified. By combining
these statements (as shown in Figure 4.3), more complex class descriptions
can be developed.

Figure 4.3: A class description using both and and or statements.

To edit the and(min) to or(max), right-click on the and(min) (Figure 4.4) and
select Edit Expression.

Figure 4.4: Editing the and(min) expression.

Within the resulting window (Figure 4.5), select or(max) and click OK. To add
and(min) operators beneath the or(max) (as in Figure 3), right-click on
or(min) as before and select Insert new Expression. From the list of features
(see Figure 1), you will find the same operators (at the bottom) shown in
Figure 5. By selecting and(min) and then adding other features/thresholds
under this operator, you can create structures similar to those in Figure 4.3.

Figure 4.5: Select Operator Expression for the class description.

Fuzzy thresholds differ from absolute thresholds as they allow a degree of


uncertainty to be included. So, objects classified using fuzzy thresholds might
be associated with the following values:
Forest = 0.8
Water = 0.7
Urban = 0.2

In this example, the object is assigned to the class forest but the fuzzyness of
other classes (water and urban) will also be allocated within Definiens
Developer to give a fuller picture of the contents of the object. Since the
introduction of the processes into the functionality of Definiens Developer,
careful consideration needs to be given to the use of the fuzzy logic
thresholds. Therefore, for most of these units, only absolute thresholds are
included.
To create a fuzzy (membership function) threshold, follow the same process
as outlined above but rather than selecting Insert Threshold in Figure 4.1
select Insert Membership Function. You will then be presented with a new
window (Figure 4.6) where you can select a membership function (not all of
these are fuzzy) and the corresponding thresholds.

Figure 4.6. Creating a Fuzzy Membership Function.

4.3. Create a project


As with all work within Definiens Developer the first step is to create a project
containing all the datasets required for the study. A subset of the inputted
images will be created once again. Your project should have the same
parameters as that shown in Figures 4.7.a and 4.7.b.

a) The project parameters


b) The subset parameters
Figure 4.7. The parameters for setting up the Definiens Project.

Once you have matched your project window to those shown in Figure 4.7.
select OK and create your project.

4.4. Setup a Customised Feature


The first step to undertake is to create a customised feature within Definiens
Developer to calculate the Normalised Vegetation Difference Index (NDVI),
Equation 4.1.

NDVI =

NIR RED
NIR + RED

Equation 4.1. The normalised vegetation difference index.

To setup the customised feature enter the feature view (


; object features >
Customised) and Select Create new Arithmetic Feature this will produce a
dialog in which you enter your customised feature, Figure 4.8a. Enter the
NDVI into this customised feature, Figure 4.8b, and select OK.

a) Empty Edit Customised Feature


b) Customised Feature for NDVI
Figure 4.8. Edit Customised Feature.

4.5. Classification Hierarchy


The next step is to create the class hierarchy shown in Figure 4.9.

Figure 4.9. Class Hierarchy

Tables 4.1 4.6 give the thresholds for each class. Note that when an upper
and a lower boundary are required, a membership function (see explanation
of fuzzy logic) can be used (see Figure 4.10).
Acid Semi Improved Grassland
Leave empty No rules
Table 4.1. Rules for the class Acid Semi Improved Grassland.

Bog/Heath
Mean GREEN
>
Mean GREEN
<

30
42

Table 4.2. Rules of the class Bog/Heath.

Forest
Mean NIR
<

100

Mean SWIR1
Mean NDVI
Mean NDVI

<
>
<

40
0.3
0.6

Table 4.3. Rules for the class Forest.

Improved Grassland
Mean NIR
>
100
Mean NDVI
>= 0.5
Table 4.4. Rules for the class Improved Grassland.

Not Vegetation
Mean NDVI
<= 0.275
Table 4.5. Rules for the class Not vegetation.

Water
Mean NDVI
<=

0.05

Table 4.6. Rules for the class Water.

Figure 4.10. Setting a membership function with an upper and lower bound.

4.6. Create Process Outline


4.6.1. Segmentation
As with all classifications in Definiens Developer, the first task is to perform a
segmentation. In this case the use of a multi-resolution segmentation is
recommended, although others could be investigated, and the segmentation
parameters will be the same as those used in the previous unit.

As with previous units, it is recommend you create an outline within your


process tree mirroring that outlined in Figure 4.11. Remember, a process is
created by right-clicking in the process tree window and select Append
Process or Insert Child Process.

Figure 4.11: Process outline and Segmentation process.

To create the process that performs the segmentation, right-click on the


Segmentation process you have already created, select Insert Child Process,
and then the algorithm Multiresolution segmentation.
Choose the
parameters shown in Figure 4.12 and once you have entered these
parameters, click on Execute to perform the segmentation.

Figure 4.12. Parameters used for the segmentation of the Landsat image.

4.6.2. Classification
The classification process is similar to the previous unit but here each
class will be classified with a separate classification process and the
classification will only be performed on the objects which are remaining to be
classified. Therefore, you need to update your process tree to appear like the
one in Figure 4.13, please make sure you have the same order as shown as
the order is important for the classification to work correctly.

Figure 4.13. The process tree including the classification processes.

Figure 4.14. shows the parameters for the classes Water and Not Vegetation.
Make sure that you match these parameters, paying attention to the Image
Object Domain for the Not Vegetation classification process which restricts the
classification to only those object which are currently unclassified.

a) The classification for Water.


b) The classification for Not vegetation
Figure 4.14. Example classification process parameters.

By classifying the scene in this way the aim is to initial remove those elements
when can be easily identified and classified, in this case water, and remove it
from the scene before classifying the next class.
4.6.3. Merge and Export Image Objects
The merging and exportation operation is, again, the same as the one used in unit
3 but with the inclusion of the extra classes. Therefore, your final
process tree should be like the one shown in Figure 4.15.

Figure 4.15. The final process tree.

4.7. Run Process


Once you have run the process, either step by step or executing the parent
process which will execute the complete algorithm you will have a result like
the one shown in Figure 4.16a and a shapefile, which can be imported into a
GIS, Figure 4.16b.

a) The classification shown in Definiens


b) The classification shown in ESRI ArcMap
Figure 4.16. The classification result.

4.8. Conclusion
Following the completion of this unit you should now be aware of the ability
to define an object oriented rule based classification within Definiens
Developer. Using a rule based classification can allow you to encode your
expert know, for example Lucas et al (2007) developed an object oriented rule
based classification for upland habitats within Wales using Definiens
Developer to encode the expert knowledge of ecologists. One of the problems
with rule based classification is to define the rules used within the

classification the next unit will go through the techniques available with
Definiens to aid the development of these rules.

4.9. Exercises
1) Experiment with different segmentation algorithms and parameters, you
should not have to edit the thresholds you have already entered to reclassify
the resulting segments but you may notice varying levels of accuracy between
different segmentations.
2) The classification which has been produced during this unit is
superficially OK but when viewed in more detail contents numerous errors. Try
to improve the quality of the classification through the refinement of the
existing rules.
3) In addition to the rule used within the classification there maybe other
features available within Definiens Developer which could aid the
classification. Review the feature available and try to include extra features (or
remove currently used features) from the classification to try and improve the
result. Please refer to the reference guide for details of other features.

References
Lucas R.M. Rowlands A., Brown, A., Keyworth, S and Bunting, P. (2007).
Rule-based classification of multi-temporal satellite imagery for habitat and
agricultural land cover mapping. International Society for Photogrammetry and
Remote Sensing, 62(3), 165-185.

Unit 5: Threshold Identification


Level:
Beginner
Time:

This unit should not take you more than 2 hours


Resources:
A licence of Definiens Developer (Version 7.0 was used to develop
these units).
A copy of the image Identification_of_Thresholds.tif generated for
this exercise.
Processes:
ClassifyImage_example.dcp
ClassifyImage.dcp
By the end of this unit you should:
Be aware of the different tools and methods available within Definiens
Developer to help you identify suitable thresholds to undertake a
classification.

5.1 Introduction

Through this unit, you will go over a number of techniques to aid the
identification of thresholds. To illustrate the techniques more easily and
simply, an artificial image (Figure 5.1) has been created and will be used
throughout this unit. Afterwards, you can try these techniques on actual
data acquired by remote sensing instruments.

Figure 5.1. Artificial image created to illustrate the different techniques of threshold
identification.

5.2. Getting started within Definiens Developer


5.2.1 Create the project.
As before, the first step is to create your project using the image
Identification_of_Thresholds.tif as input, where Bands 1, 2 and 3 should be
named Red, Green and Blue respectively. When viewing the image use no
stretch to see the same image as shown in Figure 5.1.
5.2.2. Create process and perform segmentation
As with previous projects, the first step is to create your process outline
(Figure 5.2) and perform a multi-resolution segmentation. The parameters for
the segmentation are shown in Figure 5.2, through the automatic name
provided.

Figure 5.2. Process outline and segmentation parameters.

5.2.3. Create and include a class hierarchy in the process tree.


Initially, when creating the class hierarchy, create empty classes (without
associated features) as shown in Figure 5.3. Through the processes outlined
below, you will identify and create the required thresholds within the classes.

Figure 5.3. The class hierarchy to be duplicated within your project.

You also need to create new processes to perform the classification once you
have created the rules within your class hierarchy. You can do this in two
ways:
1) Create an individual classification process for each class as in the previous unit
, or
2) Create a single process and edit while developing the rules and finally
select all classes and classify them in one process once the rules have been
developed (Figure 5.4).

5.2.4. Create the merging processes.


These are the same as the processes created in the previous unit and
need to be created for each of the classes within the hierarchy, as shown in
Figure 5.4.

Figure 5.4. The final process.

5.3

How to identify a threshold:

The next consideration is how to develop the rules required for classification
of the image. This seems difficult to start with but will become a lot easier with
perseverance. To help identify thresholds, a series of functions/options are
available which are:

The Feature View


The Sample Editor
The Feature Space Optimization
The Object Information Window
The Sample Selection Information
The order of classification defined in the process.

But, also the extent to which you know your imagery in terms of.
The range of values.
What you are seeing. (e.g., What is vegetation type X likely to be
doing at the time of image capture?)
The nature of the objects you are trying to extract (e.g., in the form of
a model such as a hill and valley model for tree crown delineation).
Interpreting the colour you can see within the image. For example, if
the object is yellow in the image, which bands need to be used for
classification?
But above all it comes down to experience!! So, take your time going through
the following exercises and consider how the features and options outlined
above help. Experiment with each of these and decide which ones you are
most comfortable with and use these. Note, that you quite often produce a

different result using these different methods but there is no right answer and
the most important consideration is that your classification works and is
appropriate to your application.
5.3.1 The Feature View
The feature view window (Figure 5.5) can be used to colour the objects (using
a colour bar) within the scene based on a single feature. The upper (green)
and lower (blue) bounds can be edited manually. Moving these upper and
lower bounds until only the area of interest is in the coloured area allows the
upper and lower bounds to be identified. These values can then be inserted
as a rule into the appropriate class.

Figure 5.5. Feature view window.

To test this form of threshold identification, go to the Feature View window


within your Definiens project. If the Feature View window is not already
available, click on the Feature View icon (
) which will reveal the required
window. Once you have identified the Feature View window, navigate to the
Brightness feature using:
Object features > Layer Values > Mean > Brightness
Upon identification of the Brightness feature, right-click and select Update
Range and then select the check-box at the lower left corner of the feature
view window (if not already ticked). Figure 5.6 shows the result of this
operation and you should now try to identify a single threshold (or pair of) to
separate as many of the colours in the image as possible.

Figure 5.6. Feature view selection and the image colouring

Write down the brightness thresholds in the table below. Note, that not all
classes may be identified using the brightness feature.
Object
Black object
Blue object
Green object
Orange object
Red object
Yellow object
White
background

Thresholds

The Brightness Feature:


The brightness feature in Definiens Developer is defined as
the sum of the mean values ( C ) of each (selected) layer for
the object divided by the number of layers ( n L ).

b=

1 nL
Ci
n L i =1

To edit the layer used to contribute to the brightness feature,


navigate through the following menus
Classification > Advanced Setting > Select Image Layers for
Brightness
Now, try identifying further thresholds for the objects which could not be
separated using the brightness feature by using other features. Once you
have identified thresholds for each of the classes, add these thresholds into
the class descriptions and run the classification process.

5.3.2 Sample Editor


One of the disadvantages with the Feature View method of threshold
identification is that it is only possible to examine features one at a time. By
using the Sample Editor and selecting samples (in the same way as you did to
train the NN classifier), multiple features can be compared.
Before using the Sample Editor, delete your existing classification (using
Classification > Class Hierarchy > Delete Classification). If you have fused
the classes previously, delete the level and resegment. Now, turn on the
Sample Selection and select samples for each of the classes (in the same
way as with the NN classification preparation). Once you have created your
samples, open the Sample Editor window and select the features you wish to
compare by right-clicking within the window and select Select Features to
Display . Then, move the features you wish to display in the feature editor
to the right-hand side of the window and click OK. Now, using the drop down
boxes at the top of the window (Figure 5.7), select the two classes you wish to
compare. Note, that the features you wish to display need to match those
which are specified to be used for the NN classifier.

Figure 5.7. Sample Editor Window.

From Figure 5.7, you can see that the black and yellow classes have a good
separation using the features brightness, mean red and mean green but a
reduced separation in the mean blue feature. Therefore, you can start to get
a feel for where suitable thresholds may exist. By continuing the process
through comparing the black class to all others, you should be able to identify
feature rules or combinations of these that separate the classes of interest.

Again, identify features with their thresholds to separate the given classes and
list below. These may differ from those you might have listed using the
Feature View. After you have defined these, add the new thresholds to the
hierarchy and classify your image.
Object
Black object
Blue object
Green object
Orange object
Red object
Yellow object
White
background

Thresholds

5.3.3. Feature Space Optimization


Within Definiens Developer, another useful function for identify features that
provide best separation of classes is the Feature Space Optimization tool.
The function is available through the icon
or by navigating through
Classification > Nearest Neighbour > Feature Space Optimization. To use the
function, you again need to create samples representing each of your classes
and then run the Feature Space Optimization afterwards. The Feature Space
Optimization window will be similar to that in Figure 5.8.

Figure 5.8. Feature Space Optimization Window.

The first step is to select the classes you wish to consider. In Figure 5.8, all of
the available classes have been selected but you can select a subset of
classes if you want to focus on these. Second, select the features you wish
to consider for the separation, and select the level (if appropriate) you wish to
work on (Levels are discussed in the next unit so for the moment, you dont
need to worry about this). Finally, you need to select the number of

dimensions you wish to consider, which equates to the maximum number of


features you want to be use together to identify the best separation of your
classes.
Once you have entered those parameters, click on Calculate and you will
notice that numbers appear in the lower left box. These indicate a) the best
separation distance and b) the number of dimensions (features) used to arrive
at that separation. To see which features were used to identify a separation,
click on Advanced and a dialog similar to that shown in Figure 5.9 will
appear.

Figure 5.9. Advanced results window for the Feature Space Optimization.

In Figure 5.9, the most significant information is within the textbox where, as
you saw in the previous window (Figure 5.8) and in the displayed graph, 5
dimensions produce the best separation of the classes. By scrolling down to
the Dimension 5 information you can discover the features which produced
the separation. You could now, by using the Apply to Std NN button, add
these classes to the standard nearest neighbour and use the nearest
neighbour classifier but here we are identifying thresholds so we will not do
this.
Now you have identified the features which give the best separation for all
classes, experiment to identify those features which are most suited for the
separation of individual or groups of classes. After which, use that knowledge
and the two techniques above to refine the thresholds required for the
classification.

List these in the table below.


Object
Black object
Blue object
Green object
Orange object
Red object
Yellow object
White background

Thresholds

5.3.4. Object Information Window


Another interface Definiens Developer offers for exploring the data is the
Object Information window, which is accessed using the
icon. When an
object is selected, this window will display the values for the selected features.
To select the features you wish to have displayed, right-click within the
window and select Features to Display. To help the identification of
thresholds (once you have an initial classification), you can go through the
objects you judge to be in error and find the reason for the errors. Through
this approach, you can adjust your thresholds accordingly and subsequently
refine your classification.
Within one of your previous classifications, examine objects to see whether
the classification is correct. If not, then use the information extracted from
objects to refine your classification. Write down the new rules in the table
below.
Object
Black object
Blue object
Green object
Orange object
Red object
Yellow object
White
background

Thresholds

5.3.5. Order of classification defined in the Process Tree


The order of classification does not really aid the identification of thresholds
but it does provide another layer of logic that you can include within your
classification hierarchy and associated processes. By using the classification
process and limiting the objects being considered for classification, the rules
within your hierarchy can be made simpler. For example, in the previous unit
, the class description for Acid Semi Improved Grassland was left
empty but by classifying objects to the other classes first and restricting the
classification of Acid Semi Improved Grassland to the currently unclassified

objects, this class can be identified. This class might otherwise be very
difficult and complex to identify because of the variation in the data values
associated with the broad range of vegetation types that is likely to exist within
this class.
5.3.6. Knowing your imagery
One of the most important aspects of classification is to know what you are
viewing and equally what you are not viewing within the imagery. For
example, in the Landsat 7 imagery for North Wales you have used for the
previous units, the date of the imagery is important as the vegetation
behaves differently at different times of the year and will therefore need a
different set of rules at different times. Equally, with temporal data from
different seasons these variations can be exploited for identifying and
classifying the land cover.
Also, in knowing your imagery and the objects you wish to classify you may be
able to think of them in the form of a model. For example, when trying to
identify tree crowns, it is useful to visualise the image as conforming to a hill
and valley model, where the crowns form the hills. This can be used to
identify seeds at the crown tops (brightest parts of the image on the hill tops)
which can be expanded to identify the crown edges (in the valleys).
5.3.7. Interpreting the colour within the image
The image you are seeing on the screen is displayed (for the most part) as a
Red, Green and Blue (RGB) composite and therefore, if the object looks red
on the screen you know it must have a large contribution from the channel you
are displaying as red. From this observation, you can use the channel in red
in the classification. Figure 5.10 shows the RGB colour space and by
considering the colours you observe in the image and in this figure, you can
start to establish which channels are contributing to the appearance of the
image as displayed in a particular colour combination. Note that when using
this approach, consider also the stretch you are applying to the image to
enhance the display as this can change the colour you see and the contrast
between features.

Figure 5.10. RGB Colour model.

5.3.8. Experience
Finally, and perhaps the most important thing to recognise, is that it takes
experience to become good at identifying thresholds and developing the
processes and methods which fit around those thresholds and which form
your classification. The more imagery you gain experience with, the better
you will become at classifying and youll be able to apply your knowledge from
one set of imagery to the next.
Another aspect of classification that should be considered is the ability of your
developed rule bases and processes to be applied to imagery other than
those to on which they have been developed. Ideally, this should be done
such that no or minimal adjustments are applied.

5.4. Conclusions
Following the completion of this unit you should now be aware of the tools
and concept through which you can identify the thresholds you will require to
classify a scene using a rule-base.

5.5

Exercises

1) Experiment with different segmentation algorithms and parameters. You


should not have to edit the thresholds you have already entered to reclassify
the resulting segments.

Unit 6: Working with Levels


Level:
Intermediate
Time:

This unit should not take you more than 1.5 hours
Resources:

A licence of Definiens Developer (Version 7.0 was used to develop


these units).

A copy of the image crowns_forest_image.tif generated for this


exercise.
Processes:

LevelsExampleProcess.dcp
By the end of this unit you should:
Be aware of the concept of levels within Definiens Developer and how
they can be used to increase the concepts represented through the
classification.
Know how to use the enclosed by class process.

6.1. Introduction
Within this unit, you will learn how to use Levels within Definiens Developer
and some of the features which allow interaction between levels. These
features increase the knowledge available within the system as different
scales of information are used.
To illustrate the use of Definiens Developer levels you will use an artificial
image that has been created for this unit (Figure 6.1). Within this image,
the green objects represent trees (herein referred to as Level 1) and a second
level (herein referred to as Level 2) will be created to represent the forest
extent. To identify the forest extent, the use of more complex processes will
be required to fill in the gaps between the crowns to create the forest mask.

Figure 6.1. Image to be used for classification.

6.1.1 What is a Definiens Level?


In its simplest form, a Level in Definiens describes a set of objects
represented at a particular scale. Levels form a hierarchy where, as shown in
Figure 6.2, the lowest level contains the smallest scale (objects) and then
higher levels progressively increase in scale (object size). Through this
approach, hierarchical relationships can be described such as a tree is within
a forest or a house is within a town.

Figure 6.2. Hierarchical Structure of Definiens Developer Levels.

6.2. Getting started within Definiens Developer


6.2.1. Create the project
As before, the first step is to create your project using the image
crowns_forest_image.tif as input, where bands 1, 2 and 3 should be named
Red, Green and Blue respectively. Make sure the unit it set to pixels and
geocoding is turned off. When viewing the image use no stretch to see the
same image as shown in Figure 6.1.

6.2.2. Create process and perform segmentation


As with previous projects, the first step is to create your process outline
(Figure 6.3) and perform a multi-resolution segmentation to create Level 1.
The parameters for the segmentation are given in Figure 6.3.

Figure 6.3. Process outline and multiresolution segmentation parameters.

6.2.3. Class hierarchy and setting of thresholds.


Create a class hierarchy which contains classes representing background,
crowns and forest (Figure 6.4).
Note that parameters (Figure 6.5) are
assigned only to the crown class.

Figure 6.4. Class Hierarchy.

Figure 6.5. Class Description of the Crown Class.

6.2.4. Classification and Merging at Level 1


Implement and execute the components shown in Figure 6.6 into your
process tree under the Classification > Level 1 and Merge and Tidy
processes.

Figure 6.6. Process for classification and merging at level 1.

The new part of this process is the chessboard segmentation of the


background following fusion for two classes. This is done to separate the
background into small objects which can then be used within the upper layer
to fill in the gaps between crowns and form a region forest containing the
crowns.
6.2.5. Creation of Level 2
The next process is to copy Level 1 to create Level 2. To create the process,
insert a new process under the Creation Level 2 process and input the
parameters shown in Figure 6.7.

Figure 6.7. Parameters from copying Level 1 to create Level 2.

6.2.6. Classification of Level 2


Reproduce the processes shown in Figure 6.8 into the process hierarchy
under the Classification > Level 2 process but notice the first process has a
restriction that object needs a border to an object within the class crowns. To
define this restriction click on the no condition button in the Edit Process
dialog and edit the resultant dialog to match that shown in Figure 6.9.

Figure 6.8. The classification processes for Level 2.

Figure 6.9. The classification process to identify objects within a border to a crown.

The feature Rel. border to Crowns is used rather than Border to Crowns as
it is normalised and independent of the object size and border length and
therefore creates a more stable threshold.
6.2.7.

Merge and Tidy the result

The final part of the classification is to fuse and tidy the result. To do this, you
need to reproduce the processes shown in Figure 6.10.

Figure 6.10. Processes to tidy and merge to give the final result.

You should already be familiar with the fusion process but take note of which
processes require execution on Level 1 and Level 2. To switch the Level,
remember to use the Parameter button next to the drop down box.
The new process here fills in the gaps within the areas of forest so when
executing, it is worth stepping through the processing and executing one step
at a time to observe the workings of each of the processes. The parameters
required for the process which fills the gaps are given in Figure 6.11.

Figure 6.11. Parameters for the process which fills any gaps within the forested areas.

6.3. Results
Once the process has been executed you should a result at each level as
shown in Figure 6.12.

a) The result at Level 1


b) The result at Level 2
Figure 6.12. The results of the classification process.

6.4. Conclusions
Following the completion of this unit you should be aware of the concept of
levels within Definiens Developer and how to implement them and represent
the relationships between the levels. You have also come into contact with
another process, in this case the fill enclosed by class process.

6.5. Exercises
1) Experiment with different segmentation strategies when creating a new
level. Figure 6.12 demonstrates a multi-resolution segmentation process
which will create a new Level above the existing one.

Figure 6.12. A segmentation process which creates the segmented layer as a new level
above the existing one.

2) Examine and experiment with the other features which allow interaction
between objects within and between levels (e.g., Relations to sub objects and
Relations to super objects). Note that super objects are those on the level
above while sub objects are those on the level below.
3) Explain below why the class background on Level 1 cannot be fully fused to
create one large object.
______________________________________________________________
______________________________________________________________
______________________________________________________________
______________________________________________________________
______________________________________________________________
______________________________________________________________
______________________________________________________________
______________________________________________________________
______________________________________________________________
______________________________________________________________
______________________________________________________________
________

Unit 7: Putting it all together


Level:
Intermediate
Time:

This unit should not take you more than 4 hours


Resources:

A licence of Definiens Developer (Version 7.0 was used to develop


these units).

The multispectral Landsat 7 image orthol7_20423xs100999.img and


its corresponding panchromatic scene o20423_pan.tif (Row: 204
Path: 23 Date 10/09/1999) over North Wales and available from the
Landmap Service.
Processes:

LandcoverClassificationExample_provided.dcp

LandcoverClassificationExample.dcp
By the end of this unit you should:
Be able to put together a real world land use classification.
Be able to demonstrate that you can calculate thresholds for
classification for real world images.
Be aware of the process to grow a class from an identified core.

7.1. Introduction
The aim of this unit is to allow you to pull together the skills you have
developed within Definiens Developer to produce a single more complex
example. The process outline will be provided, with a segmentation and initial
classification of elements such as water to illustrate some more advanced
features but you will be require to identify and enter the thresholds for the
classification of the scene.

7.2. Setup the Project


To setup the project you are required to load the following layers, Figure 7.1a,
create a subset, Figure 7.1b.

a) The project layers and alias


b) The project subset.
Figure 7.1. The project parameters.

Please note the inclusion of the two thematic layers. The first defines the area
of the image where cloud is present and the second defines the upland and
lowland areas of the scene and will be used for segmentation.

7.3. Setup the Process Tree


7.3.1. The segmentation
The first step is to set out your process tree with the default structure, Figure
7.2.

Figure 7.2. The default Process Tree.

The initial segmentation is performed with the parameters shown in Figure


7.3, please notice the inclusion of the thematic layer to define the cloud
covered area. Please note the order of the layers in the Figure.

Figure 7.3. Segmentation parameters.

Once executed, you should observe that the segmentation process has
identified the areas defined within the shapefile defining the cloud cover area.
The following step is to classify these as such and ignore them for the
remainder of the classification process. The classification is performed with
reference to the thematic layer (Figure 7.4) and results in the following
process tree, Figure 7.5.

b) The features which references the


thematic layers
Figure 7.4. The classification of cloud.

a) The classification description

Figure 7.5. The process tree, including the cloud classification.

Once the cloud has been removed from the scene, the follow segmentation
process, Figure 7.6a, will be added to the process tree, Figure 7.6b. Please
note the use of the second shapefile to separate the lowland and upland
regions of the scene. Also, note that the segmentation is being performed at
Level 1 and the Level Usage parameter is set to Use current.

a) The segmentation parameters

b) The process tree, including the second


segmentation step.
Figure 7.6. The second segmentation process to define the upland and lowland regions.

Once segmented the classes of upland need to be defined using the thematic
layer, Figure 7.7.

a) The classification description

b) The classification process.

Figure 7.7. The classification process and description to define the regions of upland and
lowland.

The final part of the segmentation is to segment within the upland and lowland
regions to produce the segments for classification. The process and
parameters are shown in Figure 7.8.

a) The process tree


b) The segmentation parameters
Figure 7.8. The process and segmentation parameters upland and lowland regions

These steps result in a segmentation which varies depending on the land


cover. So, in the lowlands where land units are defined by fields and hedges
the segmentation needs to be coarser to pull out these broad similar regions.
While, in the upland regions the land units are much smaller the segmentation
is much finer allowing these regions to be classified.
7.3.2. The classification process.
The classification process will be defined under the Classification of
Landcover process, Figure 7.2. The first part of this process tree has been
provided, below, but you are required to input the remaining processes and
rules for the classes shown in Figure 7.9.

Figure 7.9. The classes to be identified.

Within the Groups tab of the class hierarchy classes can placed in a hierarchy
allowing the relationships between the different classes to be defined. For
example, all the forest classes have been placed under the Forest class.
Therefore, Definiens Developer is aware that the Broadleaf Forest, Coniferous
Forest and Young Coniferous Forest classes are all types of forest. Once
classified if you collapse the Forest group all these forest regions will be
coloured as forest. But, be aware that if you merge the Forest class all the
sub-classes will be merged forming only a single class and removing
information from your classification.
If you are unsure of the classes to be identified please refer to the shapefile
Landcover_classification.shp.
The first step within this classification is to identify the Not Vegetation regions
within both the upland and lowland regions. The NDVI has been calculated
using a customised feature, you will need to create (see earlier unit), and a
threshold of NDVI < 0.25 has been identified (enter within the class
description of the class Not Vegetation) to separate the Not Vegetation
regions, Figure 7.10.

Figure 7.10. The process tree for the classifying the Not Vegetation regions.

Within the Not Vegetation regions the area of Water have been identify,
using the rules SWIR2 < 15 AND NDVI < 0.1, but when you run these rules
you will notice that not all the areas of Water have been identified. This is
because there are still some small regions of cloud over the lake. To correctly
classify these regions we will grow the Water class using the Water Grow
class. The Grow Water class contains the rules shown in Figure 7.11, where
the new rule Rel. border to Water > 0 defines that to be a member of the
class Grow Water the object needs to have a border to a Water object.

Figure 7.11. The class description of the Grow Water class.

By defining the process tree as shown in Figure 7.12 the Grow Water class is
iteratively classified 10 times (10x: for all, Figure 7.13) where the identified
Grow Water objects are assigned to the Water class in between each
iteration.

Figure 7.12. The process tree to classify the regions of water within the scene.

Figure 7.13. Using a process to loop a group of process 10 times.

Finally, the Water regions are merged and the remainder of the classification
will concentrate on the vegetation within the scene, where you are require to
develop you our process and rules.
7.3.3. Tidy and exportation.
As with the previous processes the final steps are to merge and tidy
classification classes and export to results for use within a GIS. Therefore,
based on the knowledge gained within the previous units developed these
parts of the process.

7.4. Results.
Once you have completed the classification and executed the tidy and
exportation processes you should have results similar to those shown in
Figure 7.14.

a) The classification with Definiens


b) The classification with a GIS.
Figure 7.14. The results of the classification.

Finally, it is recommended that you check you classification process with the
model process developed (LandcoverClassificationExample.dcp). To do this
open another instance of Definiens Developer and setup the same project
structure, then right-click within the process tree window and select Load
Rule Set, Figure 7.15. You may now execute this process and should gain
the same result as the one shown in Figure 7.14. Observe how each object
can have membership to multiple classes (use the membership to feature
and the object information window) as fuzzy membership functions have been
developed for each of the added classes.

Figure 7.15. Importing a saved rule set.

7.5. Conclusions
Following the classification of this scene you should now be confident in
performing you own classifications including several structured classes and
many of the classification features available within Definiens. It is recommend
that you look through the reference guide within your installation of Definiens
to observe the large number of features available for you during classification.

7.6. Exercises
1) The classification ruleset provided is a fuzzy classification; therefore each
object has a membership to all the classes. Look up the classification stability
feature and observe the objects which are on the boarder between two
classes.

Unit 8: Calculating Image Thresholds


Level:
Advanced
Time:

This unit should not take you more than 3 hours


Resources:

A licence of Definiens Developer (Version 7.0 was used to develop


these units).

The multispectral Landsat 7 image orthol7_20424xs240799.img (Row:


204 Path: 24 Date 24/07/1999) over South Wales and available from
the Landmap Service.
Processes:

CalculatingThresholdsExample.dcp
By the end of this unit you should:
Be able to develop Definiens processes which calculate features from
the image.
Be able to implement iterative processes to loop through a series of
objects and/or grow an object from its core.
Be able to use variables during the classification.

8.1. Introduction
A limitation of the methods which have so far been presented is that the
thresholds for classification have be identified manually. This is time
consuming and thresholds can vary between images and even across a single
image. With appropriate image pre-processing atmospheric correction and
topographic correction many of these differences can be corrected for but not
for all. Therefore, this unit will demonstrate the how Definiens Developer
can calculate thresholds from the imagery and use it for classification, in this
case for cloud and shadow detection.

8.2. Setup the Project.


To set up the project you need to define the same layer alias and subset as
that shown in Figure 8.1.

a) The layers and alias


b) The project subset
Figure 8.1. Setting up the project.

8.3. The Underlying Concept of the Classification


The methodology of the process you are to develop is an iterative process
that spatially breaks down the problem allowing the thresholds used within the
process to be calculated separately across the image. Therefore, the first step
is to split the image into a large chessboard where each segment (or tile) will
be classified individually (1 to 42), Figure 8.2.

Figure 8.2. The iterative process of processing the image.

Once a segment of the chessboard has been selected an upper and lower
quartile within the segment will be calculated and used as the thresholds for
identifying seeds for the clouds and the shadows within the scene. A fine
segmentation is then performed on the segment and once these seeds (or

cores) have been identified the remainder of the scene will be used to
recalculate these threshold values. The seeds will then be grown to the limit of
these new thresholds. Finally, the cloud and shadow objects will be merged
and the process will move on to the next large segment until all parts of the
image are processed.

8.4. Setup the Process Tree


8.3.1. Setting up the looping process
The first step is to setup up the large chessboard segmentation and iterating
through the segments, but the standard process tree is still used, Figure 8.3.

Figure 8.3. The standard process tree.

The segmentation process within this structure is a large chessboard


segmentation, Figure 8.4.

Figure 8.4. The large chessboard segmentation process

Figure 8.5, shows the process tree which you should have up to this point.

Figure 8.5. The process tree.

You then need to define the iterative process which will allow each segment to
be selected in turn, Figure 8.6. A while loop needs to be set up to loop while
the number of unclassified objects is greater than 0. Within this loop the first
process selects a single object. The Find domain extreme process will be
used, Figure 8.7., where the object with the maximum value in its Y location is
selected. Where multiple objects have the same value one will be selected at
random, as the option Accept Equal Extrema has been set to no, otherwise
all objects with the same value would be selected together. The select object
will be given the class _active. The preceding underscore is used to define a
class which is used for processing and not a classification class.

Figure 8.6. The iterative process.

Figure 8.7. The find domain extreme process and parameters.

Once an object has the class _active it can then be processed individually
using the Image Object Domain filter within a process. Finally, once the
processing has finished all the remaining _active objects need to be removed
to allow the loop to terminate.

8.3.2. Implementing the processing stage.


The next stage is set up the template within the loop to allow the processing of
the individual segments, Figure 8.8.

Figure 8.8. The standard template within the loop.

Within the segmentation process of the template a quad-tree segmentation


with a scale parameter of 40, Figure 8.9.

Figure 8.9. The quad-tree segmentation process.

The process you will use to calculate the thresholds is the compute statistical
value process which allows the number, sum, minimum, maximum, mean,
standard deviation, median and quantile to be calculated. The value from this
calculation is outputted into a variable, Definiens Developer supports the
concept of variables within the process tree. Definiens offer 5 variable types,
Scene, Object, Class, Feature and Level, where the scope of the variable is
defined by the type. For example, an object variable will be created for each
individual object, while a Level variable will be defined for an individual level
(i.e., a different value can be stored for each level) and a scene variable is
defined for the whole project (i.e., only one value for the whole project). For
this project you will only use scene variables and calculate the quantile from
the compute statistical value process. To calculate the quantile you first need
to define the quantile you are interested in, for example the 90 % quantile, to

increase the flexibility of the process we will define a variable to store this
value.
The classification processes, Figure 8.10, starts by defining the set of
variables to be used during the classification. The first LowerQuantile
contains the quantile threshold used to calculate the threshold for the shadow
seeds. While the second UpperQuantile is used to define the quantile for the
cloud
seeds.
Finally,
the
LowerQuantileBrightness
and
UpperQuantileBrightness need to be defined to store the brightness
thresholds to be used for classification. To setup a variable the update
variable process needs to be used, Figure 8.11. The initial values for the
variables is shown in Table 8.1.

Figure 8.10. The process tree to set up the variables.

Figure 8.11. The process parameters for setting up a variable.

Variable
LowerQuantile
UpperQuantile
LowerQuantileBrightness
UpperQuantileBrightness

Initial Value
5
88
0
0

Table 8.1. The initial values of the required variables.

The next stage of the process, Figure 8.12, is to compute the threshold values
into the LowerQuantileBrightness and UpperQuantileBrightness variables.
Note, the Image Object Domain specifies the _active class, therefore the
values are only computed over objects which have the class _active and will
therefore vary across the scene, as each chessboard segment is selected in
turn.

a) The process tree

b) The process parameters

c) The process parameters for an if statement


Figure 8.12. Computing the threshold values.
Following the calculation of the thresholds (UpperQuantileBrightness and
LowerQuantileBrightness) two if statements are provided to identify
particularly bright tiles and to lower the quantile threshold to take into account
this situation, Figure 8.12c. Otherwise, the threshold used for classification of
the cloud cores will be too high, resulting in an under estimation of the cloud
cover.
Once these thresholds have been calculated the classification of the classes
Cloud and Shadow can be completed to produce their respective seeds.
Figure 8.13 shows the class descriptions of the two classes. Note, the

variables UpperQuantileBrightness and LowerQuantileBrightness are used in


place of the threshold. Finally, a classification process is added to the process
tree, followed by merging processes for the two classes, Figure 8.14.

a) Cloud Class Description


b) Shadow Class Description
Figure 8.13. The class descriptions for Cloud and Shadow classes.

Figure 8.14. The process tree.

The next stage is to grow the cloud and shadow seeds to identify the full
extent of the clouds and their shadows, this will require another loop. But first,
the thresholds UpperQuantileBrightness and LowerQuantileBrightness need
to recalculate to provide the thresholds to terminate the loop used to grow the
seeds. The same process as before is used to calculate the new upper and
lower quantiles of the _active class. The new value calculated will differ from
the one previous calculated because the identified cloud and shadows seeds
no longer have the class _active and therefore not included in this
calculation. To define the loop create a new process, below calculate
threshold, and tick on the Loop while something changes option, Figure 8.15.
The elements within the loop will now be inserted as child processes.

a) Process Tree

b) Process parameters

Figure 8.15. The process tree and process parameters to set up the loop.

Once the loop has been defined, the classes Cloud Grow and Shadow
Grow need to be created, where the class description will be the same as
Cloud and Shadow but for the inclusion of the relative border features to
restrict the classification objects to those bordering the Shadow and Cloud
features, Figure 8.16.

a) Cloud Grow Description


b) Shadow Grow Description
Figure 8.16. The class descriptions for the Cloud Grow and Shadow Grow classes.

Following the classification of Cloud Grow and Shadow Grow the two class
need to be assigned to the Cloud and Shadow classes, before being merged
and any remaining _active class objects assigned to _processed, Figure 8.17.
These three steps, classification, assign and tidy will happen for each iteration
of the loop, where the loop will continue until all the objects fits the rules have
been identified.

Figure 8.17. The process tree.

The next step is to tidy the classification, which consists of three steps. The
first is to assign all the _processed objects to be unclassified and then merge
them. The next is to fill any holes in the cloud or shadow object with an area
less than 20000 m2. To do this the fill enclosed by class process was used,
Figure 8.18, where all the unclassified objects with an area less than 20000m2
enclosed by cloud are assigned (i.e., use class description = no) to the class
cloud. This is repeated for the shadow class. Finally, the Shadow and Cloud
classes are merged and exported (remember to export the class names),
Figure 8.19.

Figure 8.18. The fill enclosed by class process parameters.

Figure 8.19. The process tree to tidy and export the classification.

8.5. The Results


Once you have run the process you will have a map of the clouds and their
shadows, Figure 8.20a, and a shapefile containing these objects, Figure
8.20b.

a) Within Definiens Developer


Figure 8.20. The results

b) From ArcGIS

Finally, to increase your understanding of the process you can make use of
the Update View option (right-click on a process) which is available on every
process and will update the view in the data window after a process has
executed. This will allow you to watch the progress of your classification.
Initially, select Update View on the processes which selects an active object,
the merging of the cloud shadow seeds and the merging of the cloud shadow
during the grow. Now execute the process and you can watch your
classification being performed. Beware of over using this feature as updating
the view is a slow process and can significantly increase the processing time.
For example, switching on the three updates as suggested will double the
processing time for this algorithm.

8.6. Conclusions
From this worksheet you should be aware some of the more advanced
process and functions available within Definiens, including growing a class,
using variables and calculating thresholds.

8.7. Exercises
1) Although the method superficially works well there are numerous small
errors when area of vegetation have been included in the cloud mask;
Develop rules to remove this mis-classification.
2) Currently if a segment (from the chessboard segmentation) does not
contain any cloud or shadow objects will be identified as such regardless. Add
in extra if statements to try to remove or reduce this problem.

3) Select a new subset and try the classification on the new subset to check
for the algorithm robustness.

También podría gustarte