RE: Deliverable 1 - Introduction To Supervised Classification For GISC9216 Digital Image Processing
RE: Deliverable 1 - Introduction To Supervised Classification For GISC9216 Digital Image Processing
Deliverable 1 uses a classification process that allows one to classify pixels in a digital image into
several land cover classes: agriculture, commercial, residential, forest and water. Using ERDAS
Imagine, a subset image (512x512 pixels) of the Southern Ontario was chosen. Using this subset, a
supervised and unsupervised classification was performed. The purpose of the assignment is to
categorize and use pixel data to create thematic maps of land cover within an image. The deliverable
also will include the creation of training sites and the creation of a supervised classification.
Completion of the unsupervised and supervised classifications produced different results. Within this
document, both classifications are analyzed for procedure and quality of the output image. It was found
that the unsupervised classification was relatively easy to complete but provided the lowest quality of
results. Apart from this, the supervised classification took a longer amount of time but provided the
best quality thematic map. Each classification should be used on a project basis as each has its pros and
cons.
Should you have any concerns regarding the enclosed document or if there are any questions, please
contact me at your convenience via e-mail at [email protected] or phone at (905)932-4534. I
look forward to hearing from you.
Regards,
Introduction
Classifying an image allows a user to categorize pixels in a digital image into several land cover classes
such as residential, commercial, forest, water and agricultural land. From these, thematic maps can be
created to better display the land cover present within the image (Niagara College, 2018).
Supervised and unsupervised classification are two different types of classification processes that can be
used. Both classification processes have their own advantages and disadvantages associated with them. The
user can decide based on specific parameters which classification method they can choose.
Unsupervised classification is performed through software by clustering pixels into groups and using
specific algorithms to display them (GIS Geography, 2017). Unsupervised classification is very basic
because it is completed by the software and not the user. Based on these clusters and the algorithm defined,
the software separates the land mass into specific categories specified by the program or by the user (GIS
Geography, 2017).
Supervised classification is done through choosing representative samples of pixels to represent a specific
land mass. For example, a group of pixels within water will be selected to provide an overview of pixel
range (GIS Geography, 2017). These representative samples are called training sites which are used by the
software to define land cover classes (Niagara College, 2018). Different algorithms can be used to classify
the pixels including: maximum likelihood, Mahalanobis distance and minimum distance (Niagara College,
2018).
Unsupervised Classification
The unsupervised classification process was completed by inputting
the 5 classes with a maximum iteration of 10 and a convergence
threshold of 0.950, as depicted in Figure 1. After the unsupervised
classification software was completed, the attribute table was
observed showing each class with a corresponding color. Colours
were then chosen for each land cover class and named accordingly
(ex. Blue colour for water). Results of the colour changes made to
the attribute table are displayed in Figure 2.
2
Deliverable 1: Introduction to Supervised Classification | GISC9216 | Winona Drouin
Supervised Classification
To initially begin the supervised classification process, areas of interest or training sites were created using
the polygon tool. Areas that were in known land cover classes (Ex. Middle of Lake Ontario) were selected
with the tool and grouped together. This was completed for all land cover classes. Using the Signature
Editor, a new signature was created for each area of interest group or more specifically, for each individual
class. The resultant Signature Editor is show in Figure 4.
3
Deliverable 1: Introduction to Supervised Classification | GISC9216 | Winona Drouin
After creating each signature, the histogram was displayed for each
land cover class. The x-axis of the histogram is the range of data
values sampled and the y-axis is the frequency of pixels with that
data value. For example, Figure 5 shows the histogram for band 1
or the water classification. Since the band is somewhat normally
distributed we can conclude that the pixels that were chosen for this
band represent the water class.
Figure 5. Water class histogram; supervised
classification.
After the histograms were viewed for each class the best minimum separability was observed to solidify
the understanding of the classifications. The separability for each 6 land cover classes is shown in Figure
6. If the class pair displays a higher number of pixels than we can conclude that the pixels chosen for that
band represents the class accurately. Based on that information, Figure 6 shows a low number of pixels for
class 4 (residential) and class 6
(commercial). This indicates that the
software had a difficult time
distinguishing between the two classes
therefore, we can not say that those bands
accurately represent the pixels of that
class. When viewing the original subset
image, this can be confirmed. The
commercial and residential areas of
Figure 6. Best minimum separabilitiy; unsupervised classification.
interest were similar during the signature
editor creation.
4
Deliverable 1: Introduction to Supervised Classification | GISC9216 | Winona Drouin
each unknown pixel to the mean vector for each class (Humbolt State Unversity, 2018). Lastly, the
mahalnobis distance assumes all class covariances are equal and all pixels are classified to the closest
training data (Humbolt State University, 2018). When viewing the images, we notice a difference in the
commercial and residential classes. These two classifications alter significantly between the different
parametric rules. The other classes, such as water and forest remain somewhat constant. Maximum
likelihood was chosen to best represent the supervised classification.
Figure 8 shows the difference between the unsupervised classification and the supervised classification
(maximum likelihood). When comparing the two images the water class remains constant; from each
classification method, the water class stays the same. The forest class remains generally the same but is
more homogenous in the unsupervised classification and has some other classes speckled within it in the
supervised classification. It is important to note that there is a difference in class numbers for each
classification. The unsupervised class has 5 classes whereas the supervised classification has 6 (agriculture
is broken up into agriculture and fields). With that being said, there does appear to be a difference in
agriculture land cover when comparing the two images; the unsupervised classification displays more
agriculture on the southern portion. The main difference between the two images in the classification of the
commercial and residential land cover. The supervised classification has a higher concentration of
residential than the unsupervised classification.
Each classification method has its own advantages and disadvantages. Table 1 shows the advantages and
disadvantages for the unsupervised and supervised classification process.
5
Deliverable 1: Introduction to Supervised Classification | GISC9216 | Winona Drouin
Based on this information we can conclude that the best method for classifying land covers on an image is
the supervised classification. This method shows a better representation of land cover classes because of
the user manually choosing areas of interest. Lastly, the maximum likelihood is the best parametric rule for
the supervised classification.
References
GIS Geography. 2017. [Online]. Image classification techniques in remote sensing. Accessed on February 7, 2018.
Available from: http://gisgeography.com/image-classification-techniques-remote-sensing/
Humbolt State University. 2018. [Online]. GSP 216 Introduction to remote sensing; supervised classification.
Accessed on February 8, 2018. Available from:
http://gsp.humboldt.edu/olm_2015/Courses/GSP_216_Online/lesson6-1/supervised.html
Niagara College. 2018. [Online]. Assignment 1: introduction to supervised classification. GISC 9216 Digital Image
Processing. Accessed on Obtained from Niagara College Blackboard.