You are on page 1of 21

GISC9216-Introduction to Supervised

Classification
GIS-Geospatial Management

Assignment 1

Prepared by:
Md Shakhin Uz Zaman
GIS-GM Program Candidates
Niagara College
135 Tylor Road, Niagara-on-the-lake, ON
L0S 1J0

Prepared for:
Ms. Janet Finlay
Professor and Program Coordinator
GIS-Geospatial Management
Niagara College
135 Tylor Road, Niagara-on-the-lake, ON
L0S 1J0

February 6, 2016
6 February 2016
GISC9216-D1
Ms. Janet Finlay
Professor and Program Coordinator
GIS-Geospatial Management
Niagara College
135 Tylor Road, Niagara-On-The-Lake, ON
L0S 1J0

Dear Ms. Finlay,

Re: Assignment GISC9216D1- Introduction to Supervised Classification.

Please accept this letter as my formal submission of Assignment#1: Digital image processing
for partial fulfillment of the course GISC9216-Digital Image Processing.

This assignment has a purpose to involve the creation of training area and the production of a
supervised classification. It involves to create a subset image, an unsupervised classification
image and a supervised classification image. Supervised classification has been performed
after training site creation and signature evaluation. All Image has been created using ERDAS
imaging 2014. Finally, a comparison has been presented in between supervised and un
supervised classification images. This classification process allowed me to categorize all the
pixel in a digital image into one of the several land cover classes or themes which are
ultimately used to create thematic maps of the land cover present in the image.

Should you have any question regarding the enclosed documents, or if there are technical
issues regarding the files please contact me at your convenience at (647) 739 8609. Thank
you for your time and attention. I look forward to your comments or suggestions.

Kindest Regards,

Md Shakhin Uz Zaman
GIS-GM graduate candidate
MZ

Enclosure: 1.) Supervised classification report.

2
TABLE OF CONTENTS

List of Figures .................................................................................................................................. 3


1. Introduction ........................................................................................................................ 4
2. Purpose ............................................................................................................................... 4
3. Procedure ............................................................................................................................ 4
3.1 Subset creation .................................................................................................................... 4
3.2 Unsupervised Classification ................................................................................................. 4
3.3 Supervised Classification...................................................................................................... 1
3.4 Signature Evaluation ............................................................................................................ 2
3.5 The differences between different types of classification images of Maximum likelihood,
Mahalanobis, Minimum distance and Spectral Angel Mapper Methods. .................................. 6
4. Classification comparison ................................................................................................. 11
5. Conclusion ......................................................................................................................... 13
6. Bibliography ...................................................................................................................... 14

LIST OF FIGURES

Figure 1: Subset Image.................................................................................................................... 1


Figure 2: Unsupervised classification Image ................................................................................... 2
Figure 3: Attribute table of Supervised Classification .................................................................... 1
Figure 4: Histogram of Supervised Classification ............................................................................ 3
Figure 5: Signature Seperability of Supervised Classification ........................................................ 4
Figure 6: Statistics Window............................................................................................................. 5
Figure 7: Mean Plot of Supervised Classification Classes ............................................................... 5
Figure 8: Supervised Classification, Maximum Likelihood Method ................................................ 7
Figure 9: Supervised Classification Image, Mahalanobis Distance Method ................................... 8
Figure 10: Supervised Classification Image, Minimum Distance Method ...................................... 9
Figure 11: Supervised Classification Image, Spectral Angel Mapper Method .............................. 10
Figure 12:Supervised vs Unsupervised Classification Images....................................................... 12

3
1. INTRODUCTION

The image classification process using ERDAS imaging software 2014 has been allowed to
categorized all the pixel in a digital image into one of the several land cover classes or themes
which are ultimately used to create thematic maps of the land cover present in the image
(source: Terms of reference for Deliverable GISC9216DI).

For this assignment, two image classification process has been used: supervised and
unsupervised classification. Supervised classification refers that the output dataset are
provided to train the software to get the desired outputs whereas in unsupervised
classification no data set is provided, instead the data is clustered in different classes based
on software analysis.

The process involves the creation of a subset image after joining individual files into a single
layers_stacked*.img file, a supervised classification image and an unsupervised classification
image. Supervised classification has been performed after training site creation and signature
evaluation. An unsupervised classification has been performed with post classification editing.
Finally, a comparison has been presented in between supervised and unsupervised
classification images. This classification process allowed me to categorize all the pixel in a
digital image into one of the several land cover classes or themes which are ultimately used
to create thematic maps of the land cover present in the image.

2. PURPOSE

To involve the creation of training areas and the production of a supervised classification.
To gain a basic practical familiarity and skills development with digital image processing.

3. PROCEDURE

3.1 Subset creation

A subset has been created from a large layer_stacked.img file using subset utility under the
Raster tab subset and chip. A subset area has been selected through the features that has been
identified within the area such as water, commercial area, residential area, vegetation, farmland,
and pasture. Thus, these features have been used to complete an unsupervised and supervised
classification. The subset presents a size of 1024 x 1024 pixels (figure-1).

3.2 Unsupervised Classification

Once the subset image has been produced, unsupervised classification has been performed.
Prior to unsupervised classification, some parameters such as number of classes, maximum
iterations, convergence threshold, and skip factor have been set within the unsupervised
classification dialogue window. 15 number of classes have been selected so that the features

4
noticed in image have been classified into 15 classes. They have been identified distinct with
appropriate colors within the attribute table.

ERDAS uses ISODATA algorithm for its unsupervised classification. The following parameters have
been used:

a. Number of classes = 15
b. Max Iterations = 10
c. Convergence threshold = 0.95
d. Skip factor (X & Y) = 1
a. Number of classes: It allows to define how the initial clustered are generated.
b. Max Iterations: This is the maximum number of times that the ISODATA utility re-clusters the
data. It prevents this utility from running too long, or from potentially getting stuck in a cycle
within reaching the convergence threshold.
c. Convergence threshold: it is the maximum percentage of pixels whose cluster assignments
can go unchanged between iteration. This threshold prevents the ISODATA utility from
running indefinitely. By specifying a convergence threshold of .95 means that 95% or more of
the pixels stay in the same cluster between one iteration.
d. Skip Factor: When Skip Factor setting increases, the classifying process runs faster. However.
The statistics are calculated on fewer number of pixels. This means areas of high density
differences may result in more pixels that are unsuitably classified.

Thus unsupervised classification image (thematic raster image) (figure-2) has been produced
with 15 separate classes with district colour. It is difficult to compare with the original false colour
image. The spectral classes has been created solely based on the numerical information such as
pixel values. Clustering algorithms has been used to group the pixels together based on their
spectral similarity. Thus, the spectral classes do not always correspond to informational classes.
Moreover, it has to spend time interpreting and label the classes following the classification.

Finally, the recode tool has been used to group classes that cover similar areas and assigned
proper land cover class names. Thus, 6 number of classes has been identified as water, residential
area, commercial area, vegetation, farmland, and pasture.

In my opinion, the increased number of classes, Iteration and Convergence Threshold and
decreased number of Skip Factor would improve the classification.

5
Figure 1: Subset Image

1
Figure 2: Unsupervised classification Image

2
3.3 Supervised Classification

To perform the supervised classification, the same subset has been used as for unsupervised
classification. The first step for this process is to create a training site. It is done by selecting the
Drawing tab, and using geometry tools. It allows to select training site into the image of interest
representing the district classes. Thus, the Signature Editor (located under Raster tab,
Classification and Supervised) has been opened and specific classes have been set in according
to their similarities. 6 signatures: water, residential area, commercial area, vegetation, farmland,
and pasture have been created (figure-3).

Figure 3: Attribute table of Supervised Classification

Once the signature file has been created, different supervised classification methods have been
applied. In Maximum Likelihood classification image (figure-7) and Mohalanobis Classification
image some (figure-8), some commercial class pixels overlap with water class pixels. In Spectral
Angle Mapper Classification (figure-10) some commercial class pixels overlap with water class
pixels and some vegetation class pixels overlap with pasture class pixels.

Among them Minimum Distance Classification Method is allowed the most accurate view of the
covered area in accordance with the classes created (figure-9).

The steps for supervised classification can be summarized as follows:

1. Locate representative examples of each cover type that can be identified in the image

2. Digitize polygons inside each training site, assigning a unique identifier to each cover
type.

3. Analyze the pixels within the training sites and create spectral signatures for each of
the cover types.

1
4. Classify the entire image by considering each pixel comparing its particular signature
with each of the known signatures.

3.4 Signature Evaluation

signatures have been created, they evaluated, deleted, renamed, and merged with other
signatures.

3.4.1 Signature Alarm

Signature Alarm utility highlight the pixels in the viewer that belongs to, or are estimated to
belong to the parallelpiped decision rule. the alarmed pixels are displayed in the viewer with the
corresponding class colour.

3.4.2 Histogram Plot

Histograms are useful to help determine if training areas that have been selected are
homogeneous and classes have good separability in different bands. The figure-4 shows the
histogram of Supervised Classification.

The Supervised Classification presents both x and y-axis. The x-axis presents the pixel values and
the y-axis presents the frequency that these pixels appear in the bands.

2
Figure 4: Histogram of Supervised Classification

3.4.3 Signature Seperability

Signature Separability is calculated as the statistical difference between pairs of spectral


signatures to monitor the quality of your training sites (figure-4). It uses to determine which
bands are the best for identifying features.

3
Figure 5: Signature Seperability of Supervised Classification

3.4.4 Signature Statistics

The Statistics utility allows you to analyze the statistics for the layers to make evaluations and
comparisons. It is possible to get the Statistics window (figure-6) and the Mean Plot of Supervised
classification (Figure-7) class through Signature Editor Window.

4
Figure 6: Statistics Window

Figure 7: Mean Plot of Supervised Classification Classes

5
3.5 The differences between different types of classification images of Maximum likelihood,
Mahalanobis, Minimum distance and Spectral Angel Mapper Methods.

3.4.1 Maximum Likelihood Classification Algorithm

Maximum Likelihood Classification Algorithm assigns each pixel having pattern measurement.
This classification can be considered the most accurate because it assumes that the training data
for each class in each band are normally distributed and it uses variance and covariance analysis.
However, Maximum Likelihood method seems to be computationally slower comparing to other
methods. The figure-8 shows the Supervised Classification image through the Maximum
Likelihood method.

3.4.1 Mahalanobis Distance Classification Algorithm

The Mahalanobis Distance Classifier is a direction distance where statistics are used for each
class. It assumes that all classes covariance is equal, and it makes a faster method compared with
Maximum Likelihood classification. Like Minimum Distance method assumes histograms have
normal distribution. The figure-9 shows the Supervised Classification image through the
Mahalanobis Distance method.

3.4.1 Minimum Distance Classification Algorithm

The Minimum Distance Classification Algorithm calculates the distance of a pixel's reflectance
values to the spectral mean of each signature file, and then assigns the pixel to the category with
the closest mean. There are two choices on how to calculate distance with this classifier. The first
one calculates the Euclidean, or raw, distance from the pixel's reflectance values to each
category's spectral mean. the second kind of distance calculation is normalized the distances. In
this case, the classifier will evaluate the standard deviations of reflectance values about the
meancreating contours of standard deviations. It then assigns a given pixel to the closest
category in terms of standard deviations. The figure-10 shows the Supervised Classification
image through of the Minimum Distance method.

3.4.1 Spectral Angle Mapper (SAM) Classification Algorithm

The Spectral Angle Mapper (SAM) algorithm is based on an ideal assumption that a single pixel
of remote sensing images represents one certain ground cover material and can be uniquely
assigned to only one ground cover class (Rashmi et. al.,2014). The SAM algorithm determines the
spectral similarity between two spectra as vector and calculate the angle between them. The
figure-11 shows the Supervised Classification image through of the method.

6
Figure 8: Supervised Classification, Maximum Likelihood Method
7
Figure 9: Supervised Classification Image, Mahalanobis Distance Method
8
Figure 10: Supervised Classification Image, Minimum Distance Method
9
Figure 11: Supervised Classification Image, Spectral Angel Mapper Method

10
4. CLASSIFICATION COMPARISON

After following all the processes, supervised and unsupervised classification images have
been produced (figure-12). Now it is possible to prepare a summary of some advantages and
disadvantages of both procedures. The two types of image classification present the same
goal and both showed to be effective; however, each one presents its own advantages and
disadvantages. It is essential that the analyst understands the alternative strategies for image
classification. Because, it is not possible to state that the given classifier is the best for all the
situation. The characteristics of each image and the circumstances for each study vary so
greatly.

In the supervised approach of classification, numerical description of the various land covers
present specified by an analyst using computer algorithm. To do this, representative sample
sites of known cover types called training sites are used as numerical interpretation key that
describes the spectral attributes for each feature type. Each pixel in the data set is then
compared numerically to each category in the interpretation key and labeled with the name
of the category it looks most like.

On the other hand, the unsupervised classification uses software to analyse an image. The
software identifies a number of classes and uses its own techniques to determine which pixels
are related and groups them into classes. It performs very well in cases where the values
within a given cover type are close together in the measurement space, data in different
classes are comparatively well separated.

In the supervised approach the analyst defines the classes and then examines their spectral
separability, whereas in the unsupervised approach, it first determines spectral separability
and then defines into the classes. However, the user should be able to determine labels to
each class. Consequently, the analyst needs to have a knowledge of the area to classify into
classes.

In areas of complex terrain, unsupervised approach is preferable to the supervised one. This
is because of difficulty in selecting training sites in supervise classification due to the
variability of spectral response within each class. Consequently, a prior ground data collection
is time consuming. Moreover, it requires less human intervention.

11
Figure 12:Supervised vs Unsupervised Classification Images
12
The advantages and disadvantages for both the classification is given below:

Supervised Classification method:


Advantages:
1. Faster method compared to Unsupervised Method
2. User has the control over the selected classes that allows getting the most
effective result
3. Presents statistical monitoring that includes the Signature Editor Statistics
window, the Signature Separability panel, and Histograms;
4. Training sites when developed properly can make the supervised classification
method more effective.
Disadvantages:
1. Can present human failures;
2. Training data selected may not be accurate;
3. Selection of training data can take a long time
Unsupervised Classification method:
Advantages:
1. The user does not need to necessarily present a prior knowledge of the area;
2. Human error is minimized.
Disadvantages:
1. Slower compared to Supervised Classification method;
2. The user has limited control of classes;

5. CONCLUSION

In this assignment both Supervised and unsupervised classification methods are important for
Digital Image Processing. Both methods present their own advantages and disadvantages. Thus,
the quality of each classification depends on of which situation it is going to be applied. Moreover
the user should present a background about Digital Image Processing as well as a familiarity with
the area of interest. All of these factors are fundamental to get an accurate process.

13
6. BIBLIOGRAPHY

1. Terms of Reference for deliverables 9216-D1


2. Lillesand et al., 2015: Remote Sensing and Image Interpretation, 6 th Edition.
3. Rashmi S , Swapna Addamani , Venkat1 and Ravikiran S.,: Spectral Angle Mapper
Algorithm for Remote Sensing Image Classification.

You might also like