Deliverable 1 Unsupervised and Supervised

Classification

Tetiana Yurchenko
NIAGARA COLLEGE
 192 Queenston Street, St. Catharines, Ontario, L2R 2Z7
(905) 324-6825 • tatianayu20@gmail.com

February 13, 2019

GISC9217-D1
Mark Wilkinson
GISC9231 Course Instructor
Niagara College
135 Taylor Road,
Niagara-on-the-Lake, ON
L0S 1J0

Dear Mr. Wilkinson,

RE: GISC9217 Deliverable 1 Unsupervised and Supervised Classification

Please accept this document as my official submission for Deliverable 1: Unsupervised and
Supervised Classification for GISC9217 – Digital Image Processing.
The following document contains formal report that consists of introduction and procedure described
in Terms of Reference (Terms of Reference, Retrieved 2019). To complete this assignment ERDAS
IMAGINE 2018, lecture notes, workshops notes, book (Thomas Lillesand, Ralph W. Kiefer, Jonathan
W. Chipman, 2015) and several internet resources were used.
The idea of this assignment is creating unsupervised and supervised classifications with different
algorithms (Maximum Likelihood, Minimum Distance, Mahalanobis) of the tiff image and compare
the results. According to Terms of Reference, these classifications allow to categorize the pixels in a
digital image into one of several land cover classes (Terms of Reference, Retrieved 2019).
If you will have any questions regarding to the Deliverable 1, please contact me at your convenience
at (905)324-6825 or by e-mail at tatianayu20@gmail.com. I look forward to your comments and
suggestions.

Regards,

Tetiana Yurchenko
GIS-GM Student
T.Y.
 GISC9217-D1
February 13, 2019

TABLE OF CONTENTS

Table of figures ................................................................................................................................................................... ii

Introduction ......................................................................................................................................................................... 1
Procedure ............................................................................................................................................................................ 1
Section 1: Training Site Creation .................................................................................................................................... 1
Section 2: Signature evaluation ....................................................................................................................................... 2
Section 3: Perform a Supervised Classification .............................................................................................................. 5
Section 4: Classifications Comparison ............................................................................................................................ 8
Bibliography ..................................................................................................................................................................... 10

TABLE OF FIGURES

Figure 1 Signature editor ..................................................................................................................................................... 2

Figure 2 Panchromatic unsupervised classification ............................................................................................................ 2
Figure 3 Band 1 Histogram ................................................................................................................................................. 3
Figure 4 Band 2 Histogram ................................................................................................................................................ 3
Figure 5 Band 3 Histogram ................................................................................................................................................ 4
Figure 6 Band 4 Histogram ................................................................................................................................................ 4
Figure 7 Band 5 Histogram ................................................................................................................................................ 4
Figure 8 Band 6 Histogram ................................................................................................................................................. 5
Figure 9 Band 7 Histogram ................................................................................................................................................. 5
Figure 10 Band 8 Histogram ............................................................................................................................................... 5
Figure 11 Parametric Decision Rules (Maximum Likelihood, Mahalanobis Distance, Minimum Distance) ..................... 6
Figure 12 Parametric Decision Rules (Maximum Likelihood, Mahalanobis Distance, Minimum Distance) – NC Campus
area ...................................................................................................................................................................................... 7
Figure 13 Parametric Decision Rules (Maximum Likelihood, Mahalanobis Distance, Minimum Distance) - Landfill area
............................................................................................................................................................................................ 8
Figure 14 Comparison of supervised and unsupervised classification ................................................................................ 9
 GISC9217-D1
February 13, 2019

INTRODUCTION

There are two major categories of image classification techniques: unsupervised (calculated by
software) and supervised (human-guided) classification.
Unsupervised classification is based on the software analysis of an image without the user providing
sample classes. The computer uses techniques to determine which pixels are related and groups them
into classes. The user can specify which algorism the software will use and the desired number of
output classes but otherwise does not aid in the classification process. However, the user must have
knowledge of the area being classified when the groupings of pixels with common characteristics
produced by the computer have to be related to actual features on the ground (What's the difference
between a supervised and unsupervised image classification?, 2013).
Supervised classification is based on the idea that a user can select sample pixels in an image that
are representative of specific classes and then direct the image processing software to use these
training sites as references for the classification of all other pixels in the image. Training sites (also
known as testing sets or input classes) are selected based on the knowledge of the user. The user also
sets the bounds for how similar other pixels must be to group them together. These bounds are often
set based on the spectral characteristics of the training area, plus or minus a certain increment (often
based on "brightness" or strength of reflection in specific spectral bands). The user also designates
the number of classes that the image is classified into (What's the difference between a supervised
and unsupervised image classification?, 2013).

PROCEDURE

SECTION 1: TRAINING SITE CREATION

Firstly, to perform supervised classification in ERDAS training sites should be created. To do this
Area of interest needs to be created for each class separately, so it can be easily changed later if
needed. Carefully created polygons (at least 6) as training sites should be grouped into one, after that
it can be brought to Signature Editor as one single signature (class) (Figure 1). User assigns colour
for each signature before classification, and finally, in Signature Editor clicks Classify – Supervised
classification. To get a proper result 5 attempts to classify were made. As a personal observation, it
is better to select more than 10 polygons (training sites) for each class, while 6 is recommended.

1
 GISC9217-D1
February 13, 2019

Figure 1 Signature editor

To perform unsupervised classification tiff image is selected as an input. K Mean method, 6 classes
with maximum iteration 10, and 95% of accuracy are selected. Panchromatic unsupervised
classification is shown below (Figure 2). After that, we can assign a colour for each class in the
attribute table.

Figure 2 Panchromatic unsupervised classification

Both supervised and unsupervised classifications (coloured) are presented separately as .img files
(Disc image files) that can be opened in ERDAS.

SECTION 2: SIGNATURE EVALUATION

There are few ways to check signatures before classifying. One of them is histogram.
The histograms below represent 8 spectral bands for 6 classes: Forest (green), Buildings (orange),
Water (blue), Grass (dark green), Roads and Parking Lots (pink), and Soil (tane). The x-axis

2
 GISC9217-D1
February 13, 2019

represents the digital number and the y-axis shows how many pixels in the image have that particular
digital numb (How to work on histograms in ERDAS Imagine?, Retrieved 2019).
First histogram (Figure 3) shows how many pixels were recognized in Band 1 (Blue) for each class
and its RGB value from the lightest to darkest. It can be seen that the water plot has the biggest
number of pixels, that means it is homogenous in pixel value range. Roads and Parking Lots and
Forest classes have bell-shaped form, while Buildings class has fewer pixels and it is spread on x-
axis, which means that it is heterogeneous (has a variety of colours).
The analogic statistics are observed in other bands (Figure 4, Figure 5, Figure 6, Figure 7, Figure 8,
Figure 9, Figure 10).

Figure 3 Band 1 Histogram

Figure 4 Band 2 Histogram

3
 GISC9217-D1
February 13, 2019

Figure 5 Band 3 Histogram

Figure 6 Band 4 Histogram

Figure 7 Band 5 Histogram

4
 GISC9217-D1
February 13, 2019

Figure 8 Band 6 Histogram

Bands 7 and 8 are close to Near Infrared, therefore Forest class has big number of pixels in these
bands and it is homogenous in pixel value range.

Figure 9 Band 7 Histogram

Figure 10 Band 8 Histogram

SECTION 3: PERFORM A SUPERVISED CLASSIFICATION

5
 GISC9217-D1
February 13, 2019

As part of the assignment, three algorithms of classification were tested: Maximum Likelihood,
Mahalanobis Distance, Minimum Distance (Figure 11).
Gaussian Maximum Likelihood classifier quantitatively evaluates both the variance and covariance
of the category spectral response patterns when classifying an unknown pixel. It assumes Gaussian
(normal) distribution. The principal drawback of maximum likelihood classification is the large
number of computations required to classify each pixel (Thomas Lillesand, Ralph W. Kiefer,
Jonathan W. Chipman, 2015).

Mahalanobis Distance takes variability of classes into account and assumes normal distribution as
well. Covariance matrix is used in the equation. Minimum distance classifier is one of the simplest.
It assigns each unknown pixel to the closest category mean and doesn’t take variability into account
(Wilkinson, 2019).

Figure 11 Parametric Decision Rules (Maximum Likelihood, Mahalanobis Distance,

Minimum Distance)

Niagara College area was the most problematic because all types of parametric decision rules weren’t
accurate. Maximum Likelihood rule classifies the campus and part of the parking lot as soil, while
greenhouse is partly water object. Mahalanobis Distance rule recognizes campus as a building, but
most of the parking lots, roads, and vineyard (partially) as well. Minimum Distance rule classifies
campus building and lagoon as roads/parking, whereas shadow of the campus, greenhouse, residence
building, and Outlet Mall are classified as water sites, upfront parking lot – as building (Figure 12).

6
 GISC9217-D1
February 13, 2019

Figure 12 Parametric Decision Rules (Maximum Likelihood, Mahalanobis Distance,

Minimum Distance) – NC Campus area

If we look at the landfill in the East part of the subset, we can see the big difference in classification
using these three algorithms. While on the left image (Maximum Likelihood method) landfill
displayed most accurate – as a soil site and roads site where appropriate, the Mahalanobis Distance
rule classifies it as buildings and Minimum Distance rule – like a combination of roads/parking lots,
buildings and soil (Figure 13).

7
 GISC9217-D1
February 13, 2019

Figure 13 Parametric Decision Rules (Maximum Likelihood, Mahalanobis Distance,

Minimum Distance) - Landfill area

Based on the above, all algorithms have advantages and disadvantages, but Gaussian Maximum
Likelihood method works better in this case.

SECTION 4: CLASSIFICATIONS COMPARISON

From the result and the procedure above can be seen that supervised classification is much more
accurate. But it depends on skills of the person who perform classification, size and placement of the
training sites (better to have more small sites then one big), distance from the edges of the site, number
of pixels, etc. (Wilkinson, 2019). Supervised classification uses the spectral signature defined in the
training set. For example, it determines each class on what it resembles most in the training set (Image
Classification Techniques in Remote Sensing, 2018).

Unsupervised classification groups pixels into “clusters” based on their properties. In order to create
“clusters”, analysts use image clustering algorithms such as K-means and ISODATA. After picking
a clustering algorithm, you identify the number of groups you want to generate. In this case 6 clusters
(classes) were created (Image Classification Techniques in Remote Sensing, 2018).

Since unsupervised classification automatically assigns colours without creating training sites and
signatures manually, it has poor representation. For example, in black & white version Welland
Canal, landfill and Outlet Mall are very bright, and unsupervised classification considers these as one
class therefore in coloured image they are all blue (Figure 14).

8
 GISC9217-D1
February 13, 2019

Figure 14 Comparison of supervised and unsupervised classification

Looking at these two images it is obvious even for non-GIS person, that supervised classification
displays the classes better. On the unsupervised classification image there are no completely forest
areas, it is mixed with grass class. It is hard to recognise the roads class as most of the roads and
parking lots are classified as buildings, as well as some soil areas and crop fields.
Advantages of supervised classification. Supervised classification can be very accurate depending on
the spectral distinctness of the classes and proficiency of the user. It involves much more complicated
analysis. The software analyzes every pixel and trying to process it and find similarity with created
training sites using different algorithms (Maximum Likelihood, Minimum Distance, Mahalanobis
Distance, Parallelepiped, Stepped Parallelepiped, etc.).
Disadvantages of supervised classification. It requires more time and attention to details. If two (or
more) training sites of different classes are similar in spectral reflectance, it may lead to
misclassification (Supervised Classification, Retrieved 2019). This is what happened in the area of
Niagara College campus, when campus was classified as Roads and Parking Lots class.
Advantages of unsupervised classification. It doesn’t require any special skills of digital image
processing and doesn’t take a lot of time. The software automatically assigns classes based on spectral
reflectance.

9
 GISC9217-D1
February 13, 2019

While the process is basically automated, the user has control over certain inputs. This includes the
Number of Classes, the Maximum Iterations, (which is how many times the classification algorithm
runs) and the Change Threshold %, which specifies when to end the classification procedure
(Unsupervised Classification, Retrieved 2019).
Disadvantages of unsupervised classification. Can be confusing when user is trying to determine the
class name because it may merge few classes together. The classification carried out for this tiff
image cannot be considered successful because it has too many errors.
Overall, supervised classification is more precise comparing to unsupervised classifications based on
the above.

BIBLIOGRAPHY
How to work on histograms in ERDAS Imagine? (Retrieved 2019). Retrieved from StackExchange:
https://gis.stackexchange.com/questions/293132/how-to-work-on-histograms-in-erdas-imagine

Image Classification Techniques in Remote Sensing. (2018, December 14). Retrieved from GISGeography:
https://gisgeography.com/image-classification-techniques-remote-sensing/

Supervised Classification. (Retrieved 2019). Retrieved from Humboldt State Geospatial Online:
http://gsp.humboldt.edu/olm_2015/courses/gsp_216_online/lesson6-1/supervised.html

(Retrieved 2019). Terms of Reference. Niagara College.

Thomas Lillesand, Ralph W. Kiefer, Jonathan W. Chipman. (2015). Remote Sensing and Image Interpretation 7th
edition. Wiley.

Unsupervised Classification. (Retrieved 2019). Retrieved from Humboldt State Geospatial Online:
http://gsp.humboldt.edu/olm_2015/courses/gsp_216_online/lesson6-1/unsupervised.html

What's the difference between a supervised and unsupervised image classification? (2013, October 17). Retrieved
from Extension: https://articles.extension.org/pages/40214/whats-the-difference-between-a-supervised-
and-unsupervised-image-classification

Wilkinson, M. (2019). Supervised classification Week 3 GISC9217 – Digital Image Processing. Presentation. Niagara-on-
the-Lake.

10