Technical Report Exploring Classification Methods

Exploring Supervised Classification Through Maximum Likelihood, Mahalanobis and K Means Methods 12 January 2014
1

Introduction to Supervised Classification

Introduction

The process of land classification involves analyzing a raster image to categorize
common areas of land into distinct land types. This is accomplished by assigning each pixel in
the image to a specific land class. This process can be undertaken using two techniques:
unsupervised and supervised classification. Both classification techniques were completed and
compared for a subsection of the Niagara Region in Ontario, Canada. Supervised classification
involves choosing from three methods: maximum likelihood, minimum distance, or
mahalanobis and involves the most user input through the use of training sites. Unsupervised
classification relies on computer algorithms to find spectral patterns in the data.


Procedures

Unsupervised classification involves the automatic classification of all pixels in an image
with limited user input. The user has the ability to specify number of classes, number of
iterations, but no prior knowledge of the study area is necessary. The algorithms within the GIS
find natural groupings of the pixels based on the different values throughout the spectral
bands. Spectral signatures are created automatically by the GIS to define the uniqueness of
each class. These signatures show spectral reflectance across the different bands of the raster
image (Figure 1).

Technical Report Exploring Classification Methods
Exploring Supervised Classification Through Maximum Likelihood, Mahalanobis and K Means Methods 12 January 2014
2


Figure 1: Spectral Signatures for 8 classes
Analyzing the spectral signatures will help in identifying the characteristics of each land class.
For example, vegetation layers emit high levels of IR radiation so those spectral signatures that
have high values at Band 4- the IR band- will indicate vegetation present on the surface.
Unsupervised land classification was undertaken for the Niagara Region using the
following parameters: Method: K-Means; # of Classes: 8; Maximum Iterations: 10; Convergence
Threshold: 95%. After experimentation with various parameters, the above set was determined
to be the most clear and representative of the study area as will be described in the Discussion
section.
The 8 classes were then inspected and named based on their spectral characteristics
and previous information about the area. Signature Separability was inspected to identify
classes that were similar to each other. The final step was to recode the 8 classes into 5 distinct
classes (Figure 2). There were three Forest classes that had to be merged into one and two
Agricultural classes were merged into one. This left the 5 classes: Forest, Water, Agriculture,
Bare Agriculture, and Urban.

Technical Report Exploring Classification Methods
Exploring Supervised Classification Through Maximum Likelihood, Mahalanobis and K Means Methods 12 January 2014
3


Figure 2: Unsupervised classification result with 8 classes before recode
Unsupervised classification does not allow significant user control over the output
classes. This can be most beneficial when minimal knowledge is known about the study area,
unlike in this case. The use of training sites in supervised classification allows for more certainty
in land classes. An in-depth analysis of supervised versus unsupervised classification will take
place in the Discussion section.
Supervised classification was also undertaken for the Niagara study area. This technique
uses known information about the study area to help in land classification. This is done through
the delineation of training sites. The number of land classes is determined before the training
stage is started. Training sites are then digitized for each class. A typical rule for training sites is
that each class should contain at least 10*n (n being the number of classes) pixels and should
Technical Report Exploring Classification Methods
Exploring Supervised Classification Through Maximum Likelihood, Mahalanobis and K Means Methods 12 January 2014
4

be representative of the class throughout the whole study area. In this case the number of
training sites for each class was approximately 5 to 10. Once all the training sites have been
selected, a spectral signature is created for the respective class. The spectral signature shows
the amount of radiation that is reflected across the different wavelengths of light (or spectral
bands). Furthermore, signature separability can be viewed to see how each spectral class varies
from one another. Through the inspection of these signatures and associated statistics, classes
can be merged, omitted, edited and land characteristics can be determined. Additionally,
histograms can be viewed for each class regarding each band of the image (Figure 3). The x axis
indicates the value of the pixel and the y axis shows the frequency of that pixel value. It is
important to view the range of the x axis to realize how high a pixel value is and in which band.
Again, the forest class should have high IR values so band 4 should show higher values than
other bands. The range here is 72 to 189, much higher than the next highest band 30 to 98.
Histograms can be simultaneously compared over all classes for each band to obtain a valuable
comparison between classes.


Figure 3: Histograms for Class 2- Forest

Once the spectral classes are set, the classification stage can be run. The three methods
of supervised classification are: maximum likelihood, minimum distance to means, and
mahalanobis distance.
Technical Report Exploring Classification Methods
Exploring Supervised Classification Through Maximum Likelihood, Mahalanobis and K Means Methods 12 January 2014
5

The maximum likelihood classifier takes into account the variance and covariance of the
spectral response patterns when classifying a pixel. Each spectral class is fitted with a
probability density function which calculates the chance of a given pixel belonging to a certain
class. The pixel gets classified into the class which contains the highest probability.
The minimum distance to means classifier classifies unknown pixels based on their
distance from each of the class means. A pixel will be classified to a certain class if its value is
closest to the mean value of the entire class. The mean values are gathered from classes of the
training sites. It does not use the variance or covariance functions.
The mahalanobis distance classifier is very similar to the minimum distance classifier but
also takes into account variance and covariance. Therefore, this method can be slightly more
effective as it takes into account the spread of the data and the characteristics of other classes
involved.
After running classification through all three methods, the maximum likelihood method
showed the clearest results (Figure 4).












Technical Report Exploring Classification Methods
Exploring Supervised Classification Through Maximum Likelihood, Mahalanobis and K Means Methods 12 January 2014
6




Figure 4: Comparison of 3 classification methods (L-R): Maximum Likelihood, Mahalanobis and
Minimum Distance to Means




Technical Report Exploring Classification Methods
Exploring Supervised Classification Through Maximum Likelihood, Mahalanobis and K Means Methods 12 January 2014
7

Analysis

The final supervised and unsupervised classification results can be compared in
Appendix 1 and 2. The unsupervised classification was able to classify water very well, whereas
supervised classification had some difficulties due to the variation in bodies of water in Niagara
(Canals, ponds, lakes, creeks).
The urban land class was identified correctly for unsupervised classification although it
could not tell the difference between residential and commercial classes. This can be attributed
to the mixture of land classes in residential settings- trees, roads, roofs, houses, open areas,
etc. Training sites would be beneficial in this case to inform the GIS of this intricate land class.
Supervised classification was able to produce an effective result between dense urban and
residential areas by identifying explicit parcels of land as residential and dense urban through
training sites. The cities of St.Catharines, Niagara Falls, and Welland are clearly visible in the
supervised classification. In the unsupervised classification, the cities are covered with forest
and agricultural land classes.
The forest land class was much less dominant in the supervised classification.
Unsupervised classified approximately one-third of the image as forest compared to supervised
at 13% (Table 1). The addition of a residential class in supervised classification reduced the
forest class greatly. Both agriculture classes also seem to have affected the forest class as is
most evident in the northeastern section of the study area. This is due to the vegetation
present on the agricultural surfaces.

Table 1: Percentage of pixels in each class for each method of classification


Water 164,811 16% Water 161,319 15%
Forest 369,430 35% Forest 131,457 13%
Agriculture 344,385 33% Agriculture 220,306 21%
Bare Agriculture 92,456 9% Bare Agriculture 174,157 17%
Urban 77,494 7% Urban 53,348 5%
Residential 307,989 29%
UNSUPERVISED SUPERVISED
Total Pixels: 1,048,576
Technical Report Exploring Classification Methods
Exploring Supervised Classification Through Maximum Likelihood, Mahalanobis and K Means Methods 12 January 2014
8

The two agricultural classes combined, covered a similar amount of total area in each
classification but individually were quite different from each method. This was affected mainly
by the selection of training sites.
For supervised classification, the inspection of signature separabilities shows that the
residential class is very similar to all classes except for water (Figure 5).

Figure 5: Signature separability for supervised classification. Red numbers indicate signature
similarities

Since the residential class is very different from the others, it cannot be merged. The forest and
agriculture signatures (39) are the most similar due to the presence of vegetation. After using
the Image Alarm to view individual classification results, the Agriculture land was distinct from
the forest regions. The Image Alarm was helpful in visualizing how the classes would turn out
before the classification stage was completed. Bare agriculture and Urban signatures (41) were
similar due to the lack of vegetation present. Again, these classes had to be kept separate
because they showed distinctly different areas.

Discussion

The supervised and unsupervised classification results had some significant differences.
Overall, the unsupervised classification created a more pixelated image which is most notable
in the cities of St.Catharines, Niagara Falls, and Welland. This can be attributed to the mixture
Technical Report Exploring Classification Methods
Exploring Supervised Classification Through Maximum Likelihood, Mahalanobis and K Means Methods 12 January 2014
9

of land uses in cities. Unsupervised classification classified water more effectively than
supervised classification. It also classified much more of the region in the Forest land use. It had
difficulty classifying roads which could be a result of the number of classes chosen. Eight classes
were used as visual inspection proved 10 and 12 classes did not add any significant information
to the classification result. The number of maximum iterations and the convergence threshold
work together and 10 and 95% were most suitable. Most of the classification results made it to
only 8 iterations, regardless of maximum iterations, as there were no significant changes in the
means for the spectral classes at the 95% threshold.

Supervised classification created similar signatures between agriculture and forest
classes. This, along with the addition of the residential class greatly reduced the amount of
forest pixels in the image. Supervised classification works best when the user knows the study
area or has significant reference information for it. The use of training sites allows unique
signatures to be created for known land classes. This method was able to classify roads fairly
well under the residential land class. The output in Appendix 2 shows road networks
throughout most of the rural areas of Niagara. It did not classify the QEW highway well,
compared to the unsupervised result.

Conclusion

Supervised and unsupervised classifications each have their advantages. When little
information is known about the study area, unsupervised classification can be very effective in
creating spectral signatures as it uses inherent patterns in the pixel values. A disadvantage of
this method is classifying cities with a mixture of land uses. Supervised classification can
overcome this problem through the use of training sites. Training sites allow the user a degree
of control over the classification results. Though this process can be time consuming, it can be
more effective than unsupervised classification if done thoroughly. Unsupervised classification
has the advantages of speed. The GIS can run the procedure with limited user input and create
meaningful spectral classes. Another disadvantage of supervised classification is that the
Technical Report Exploring Classification Methods
Exploring Supervised Classification Through Maximum Likelihood, Mahalanobis and K Means Methods 12 January 2014
10

spectral signatures are created solely on the training sites opposed to the inherent patterns of
pixel values. The best approach to classification would be a hybrid method which uses results
from both techniques. Again, this would be most suitable if time is available.
Overall, the supervised method produced the more effective result though it may have
under-classified the forest land use. Cities stand out distinctly from the other classes and roads
are more visible in this result. The unsupervised result classified water more effectively due to
the variance of water bodies in Niagara. This could have been improved in supervised
classification if the training sites were more carefully chosen to represent each different type of
water body.




















Technical Report Exploring Classification Methods
Exploring Supervised Classification Through Maximum Likelihood, Mahalanobis and K Means Methods 12 January 2014
11

Appendix 1: Unsupervised Classification for Niagara With 5 Classes



Technical Report Exploring Classification Methods
Exploring Supervised Classification Through Maximum Likelihood, Mahalanobis and K Means Methods 12 January 2014
12

Appendix 2: Supervised Classification for Niagara With 6 Classes