NIAGARA COLLEGE GIS GM

Digital Image
Processing
Introduction to Supervised Classification Assignment
#1

Marc Michael Mancino

For: Janet Finlay

January 30th, 2015

10 Jessica Drive St. Catharines Ontario L2M 6V7

(905) 935-8366 [email protected]
December 12th, 2014
GISC9216D1
Janet Finlay
Niagara College
135 Taylor Road
Niagara-on-the-Lake
Dear Janet Finlay:

RE: GISC9216 Digital Image Processing Deliverable 9216-D1

The purpose of this deliverable is to utilize the unsupervised and supervised

classification methods of the ERDAS Imagine software to manipulate and
analyze digital imagery. The report includes all the answers to all the
questions noted in the terms of reference. The process for conducting a
supervised classification has been proven useful in solidifying my
knowledge of ERDAS Imagine and digital image interpretation and further
enhances my ability to perform like a GIS expert.
Should you require further information, please do not hesitate to contact me
at your convenience. I look forward to your comments and suggestions on
this deliverable.
Sincerely,
Marc Mancino
MM/
Enclosures:1.) Formally written document
January 30th, 2015

Page |

Table of Contents
1.0 Introduction....................................................................................................................4
2.0 Background.....................................................................................................................4
2.1 - Unsupervised Classification.....................................................................................4
2.2 - Supervised Classification..........................................................................................4
3.0 Methodology...................................................................................................................4
3.1 Performing the Unsupervised Classification.........................................................4
3.2 Performing the Supervised Classification.............................................................6
4.0 - Discussion......................................................................................................................10
4.1 Advantages and Disadvantages.............................................................................10
4.2 Discussing the Classifiers.......................................................................................11
4.3 Differences.................................................................................................................11
5.0 Conclusions...................................................................................................................14
Bibliography...........................................................................................................................15
Appendix.................................................................................................................................16

List of Figures
Figure 1 - Resultant Image of Merged TIFF files..............................................................5
Figure 2 - Subset Raster of Area of Interest (AOI)...........................................................5
Figure 3 ERDAS Imagine menu displaying the process for conducting an
Unsupervised Classification (right) and its resultant image (left)...............................6
Figure 4 - Attribute Table for the Unsupervised Classification......................................6
Figure 5 - Signature Editor for Supervised Classification..............................................7
Figure 6 - Pre-Supervised Signature Alarm Layer with visible AOI (Training Sites).7
Figure 7 - ERDAS Imagine menu displaying the process for conducting a
Supervised Classification.......................................................................................................8
Figure 8 Result from the Maximum Likelihood Classification....................................8
Figure 9 Result from the Mahalanobis Classification...................................................9
Figure 10 Result from the Minimum Distance Classification.....................................9
Figure 11 - Attribute Table for the Supervised Classification......................................10
Figure 12 - Histogram showing the pixel value (x-axis) versus the frequency of that
pixel (y-axis) for the Supervised Classification signature Deep Water...................10
Figure 13 - Supervised Classification (Minimum Distance) of the AOI Using ERDAS
Imagine 2014.........................................................................................................................16
Figure 14 - Unsupervised Classification of the AOI Using ERDAS Imagine 2014....17
Figure 15 - Subset Raster Imagery for the AOI...............................................................18

List of Tables
Table 1 - Resultant Images for both Classification Types.............................................12

January 30th, 2015

Page | 1

1.0 Introduction
What is Digital Image Processing? It is the process of manipulating and analyzing
digital imagery using computer algorithms (Tutorialspoint, 2014). The scope of this
deliverable revolves around using a versatile application called ERDAS Imagine to
do so. ERDAS Imagine is a user-friendly application capable of analyzing and
manipulating digital imagery. Preliminary steps are taken to properly prepare the
digital imagery for the classification process. Of its many abilities, the
classification methods are going to process raster images for analysis. There are a
two different types of classification methods that will be used in ERDAS to
complete the tasks required in this deliverable; unsupervised and supervised.
Unsupervised is a more automated method using a preset algorithm whereas
supervised is more user-dependent. The idea is to take the digital imagery and
manipulate and analyze it to distinguish between various features in an area of
interest.

2.0 Background
2.1 - Unsupervised Classification
By definition, this classification forms clusters of pixels with similar spectral
characteristics by applying the Iterative Self-Organizing Data Analysis Technique
(ISODATA) algorithm to the selected image. In laymen terms, it uses an algorithm
preset in the program to try and determine which surface cover types are the
same, group them and classify them accordingly.

2.2 - Supervised Classification

By definition, the analyst/user defines in the imagery, homogeneous representative
samples of the different surface cover types (training sites). In laymen terms,
instead of using an algorithm to determine which surface cover types are the same
and should be classified as one type, be it water, forest or the like, the user is now
responsible for telling the program which features are the same.

3.0 Methodology
Before any data was used, all the required files were transferred to an external
hard drive. The file pathways were used to find what was required and then saved
so they could be worked on from a source other than the X: drive. This prevented
future complications and the potential corruption of data. Never work off the X:
drive!

3.1 Performing the Unsupervised Classification

Before the unsupervised classification can commence, there are a few preliminary
steps that are required. The first workshop focused on merging the six given TIFF

January 30th, 2015

Page | 2
images (1, 2, 3, 4, 5 and 7) to form a stacked image for use in the classification
process as indicated in Figure 1.

Figure 1 - Resultant Image of Merged TIFF files

In order to get a closer display of features on a large area such as this, we were
instructed to select a 512 pixel by 512 pixel area with at least one source of water,
agriculture, forest and urban area to focus on. The red circle indicator in Figure 1
is that area of interest in this case. Figure 2 displays the clipped 512 by 512 subset
raster in true colour.

January 30th, 2015

Page | 3

Figure 2 - Subset Raster of Area of Interest (AOI)

Before the unsupervised classification proceeds, the image features need to be
classified, assigned signatures and recoded. The attribute table names were
changed from their default names to names that better suit the features
represented on the image. The image was then recoded to combine similar
features that were separately classified into one feature. Using ERDAS Imagine,
we were able to process the subset raster through an unsupervised classification.
The subset raster was input with the settings recommended in the workshop and
produced a resultant image as demonstrated in Figure 3.

January 30th, 2015

Page | 4

Figure 3 ERDAS Imagine menu displaying the process for conducting an

Unsupervised Classification (right) and its resultant image (left)

Figure 4 - Attribute Table for the Unsupervised Classification

3.2 Performing the Supervised Classification

The first step in conducting a supervised classification involved the creation of
training sites. The polygon and polyline drawing tools were used to represent the
training sites. Polygons and polylines were created on the original true colour
subset raster image (Figure 2) to assist in assigning signatures to each of the
features on the image. Multiple training sites for the same feature were grouped
together using the signature editor. The signature editor, as indicated in Figure 5,
was used to group and classify each of these signatures into categories that would
distinguish different areas of the imagery from one another. The signature editor
was also able to display statistics and values of separability between the classes to
help the user determine the most optimal selection of classes and training sites. To
ensure that the signature editor would produce a representative classification, an
image alarm was created to display the potential distribution of colour for all of the
signatures, as demonstrated in Figure 6.

January 30th, 2015

Page | 5

Figure 5 - Signature Editor for Supervised Classification

Figure 6 - Pre-Supervised Signature Alarm Layer with visible AOI (Training

Sites)
The supervised classification, as shown in Figure 7, was performed, using the
signature editor, for each of the classification types: Maximum Likelihood,
Mahalanobis, Minimum Distance, Spectral Angle Mapper and Spectral
Correlation Mapper. The results for the three applicable classification types are
displayed in Figure 8, Figure 9 and Figure 10. The result from the Minimum
Distance Classification was chosen to represent the Supervised Classification.

January 30th, 2015

Page | 6

Figure 7 - ERDAS Imagine menu displaying the process for conducting a

Supervised Classification

Figure 8 Result from the Maximum Likelihood Classification

January 30th, 2015

Page | 7

Figure 9 Result from the Mahalanobis Classification

Figure 10 Result from the Minimum Distance Classification

January 30th, 2015

Page | 8

Figure 11 - Attribute Table for the Supervised Classification

Figure 12 - Histogram showing the pixel value (x-axis) versus the frequency
of that pixel (y-axis) for the Supervised Classification signature Deep
Water

4.0 - Discussion
4.1 Advantages and Disadvantages
The processes for classifying the features in a subset of digital imagery are, as
previously mentioned, the unsupervised and supervised classifications. With any
methodology, there are always advantages and disadvantages for each one, be it
with these classifications or simply outside the boundaries of this deliverable. The
unsupervised classification tends to generalize the identification of features,
resulting in clusters of pixels that only roughly match some of the actual classes or
that group completely unrelated pixels together. Although the general quality of an
unsupervised classification is poor, it does have its benefits. It is good for
identifying similar features in regions where no prior knowledge is known. If you
viewed the digital imagery of a new planet, an unsupervised classification would
probably more accurately classify the area for similar features using an algorithm
than human judgment. It gives an idea of how the results of a supervised
classification would look before having to actually perform one. There is almost no
dependence on the user to perform an unsupervised classification, therefore

January 30th, 2015

Page | 9
minimizing human error. The classes are often more spectrally uniform, but
majority of the time yields an inaccurate representation of features on an image.
The supervised classification is heavily user-oriented but can yield a much more
accurate interpretation of features on an image. The procedural step requiring the
signature editor allows the resultant attribute table to produce information classes
that are grouped, discarding the need for recoding. The user is in full control of
determining training sites for the image. Both the signature editor and training
sites can be saved as layers for later use. An image alarm can be created as a layer
to evaluate which pixels for a specific training area will belong to that class.
Depending on the type of supervised classifier that is used, the spectral
distribution of the resultant image can change drastically between each type.
Assigning training sites could result in unclassified pixels, which are then assigned
values depending on the type of supervised classifier.

4.2 Discussing the Classifiers

There are three types of supervised classifiers that apply to the extents of this
deliverable: maximum likelihood (Figure 8), mahalanobis (Figure 9) and minimum
distance (Figure 10). The maximum likelihood classifier evaluates the variance and
correlation of spectral response patterns when classifying an unknown pixel and
assumes a normal distribution, but tends to overclassify signatures with large
values in the covariance matrix. It is considered the most accurate classifier
because it considers the most variables (Janet Finlay, 2015). It uses many variables
during processing but uses too few classes for a normal distribution, therefore over
representing agriculture. The forest feature (dark green) was under represented
and some of it was merged with agriculture (orange) when it clearly wasnt
characteristic of being agriculture (blocky farmland). The roads in the urban
feature (yellow) are undistinguishable and look like a big, blotchy mess. Finally,
there is a fireworks display of shallow water (cyan) pixels in the middle of the deep
water (dark blue) feature, which isnt representative at all.
The mahalanobis classifier, similar to maximum likelihood, also assumes normal
distribution and tends to overclassify signatures with large values in the
covariance matrix. It considers variability during processing (Janet Finlay, 2015). It
is an exacerbated version of the maximum likelihood classifier in this case, with
agriculture (orange) swallowing even more forest.; the roads in the urban feature
(yellow) looking more like a solid polygon of yellow and more fireworks from the
shallow water (cyan) feature exploding into the deep water (dark blue) feature.
The minimum distance classifier is the most efficient classifier and assigns each
unknown pixel to the closest category mean, but does not take into account
variability or dispersion (Janet Finlay, 2015). It was determined that minimum
distance was the best representation of the results from a supervised classification
because it was much cleaner than the other classifiers and had more features
categorized correctly, as well, the features were much more distinguishable. The
spectral distributions of forest and agriculture seemed more viable. The roads
(yellow), even in the middle of the city, were distinguishable from each other. The
shallow coast (cyan) feature avoided spilling an unreasonable amount of pixels into
the deep water (dark blue) feature, making it much more representative.

January 30th, 2015

P a g e | 10

4.3 Differences
The results of both the unsupervised and supervised classification methods are
displayed below in Table 1. A closer look at the unsupervised classification, located
in the Appendix, is displayed in Figure 13. A closer look at the supervised
classification, located in the Appendix, is displayed in Figure 14.
The first notable difference in the colour scheme is that the urban area for the
unsupervised classification is divided into two colours: red and yellow; whereas the
supervised classification represents that same urban area with one colour: yellow.
The unsupervised classification could not identify a generalized feature such as
urban area, therefore dividing roads from buildings. So, the roads were assigned
the red colour and the buildings were categorized as urban area and assigned
the yellow colour, as shown by the black indicator arrow in Table 1. The training
sites for the supervised classification were sufficient enough to categorize the
roads and buildings into one class. The roads class, for the unsupervised
classification, included some blocky, farmland shapes, hence the name Roads /
Some Agriculture. The black circle indicator in Table 1 shows the difference in
pixel density for the shallow coast feature. The unsupervised classification hardly
represented the amount of visible sediment seen in the original true colour subset
raster image (Figure 15). The light blue circle indicator in Table 1 shows the
difference in consistency between agriculture and roads. The unsupervised
classification demonstrates difficulty in identifying the roadways that run through
that area and there is hardly any consistency in features; almost looking like a
random assortment of red and orange starburst candy had been tossed into a bowl.
The supervised classification demonstrates an easy to follow roadway and its
surrounding agriculture and forest features. The yellow indicator arrow in Table 1
shows the difference in pixel density for the shadow coast feature. The supervised
classification demonstrates a realistic extent of the shallow coast, again, evident in
the original true colour subset raster image (Figure 15).
Table 1 - Resultant Images for both Classification Types
Unsupervised Classification

Supervised Classification

January 30th, 2015

P a g e | 11

January 30th, 2015

P a g e | 12

5.0 Conclusions
After following workshop and deliverable procedures, it can be concluded that
ERDAS Imagine excels at manipulating and analyzing digital imagery. It was
determined that minimum distance was the best representation of the results from
a supervised classification because it was much cleaner than the other classifiers
and had more features categorized correctly, as well, the features were much more
distinguishable. The supervised classification is a more accurate representation of
features on an image, but an unsupervised classification has its benefits as well.

January 30th, 2015

P a g e | 13

Bibliography
1) Tutorialspoint. Digital Image Processing. (2014). Retrieved on January 28th
from www.tutorialspoint.com/dip/.
2) Niagara College. Janet Finlays Digital Image Processing (DIP) Lecture on
Supervised Classification. (2015).

January 30th, 2015

P a g e | 14

Appendix

January 30th, 2015

P a g e | 15
Figure 13 - Supervised Classification (Minimum Distance) of the AOI Using
ERDAS Imagine 2014

January 30th, 2015

P a g e | 16

Figure 14 - Unsupervised Classification of the AOI Using ERDAS Imagine

2014

January 30th, 2015

P a g e | 17

Figure 15 - Subset Raster Imagery for the AOI

January 30th, 2015