Anda di halaman 1dari 11

ISODATA Unsupervised Classification of

the Mount Baker and Bellingham Bay


Region
Alix Hartmann
February 2016
ESCI 442

Abstract
In this analysis, I have classified land use and land cover types in the Mount
Baker and Bellingham Bay region. This process was split into two stages: initial
classification and refined classification. During the initial classification stage, I used the
ISODATA unsupervised classification technique in ENVI. During the refined
classification stage, I attempted to improve the results of this classification with a
series of conditional tools in ArcMap. Statistical analysis comparing the results of each
stage shows improved accuracy: initial classification of 14 information classes yielded
an overall accuracy of 44.21%, while refined classification using 11 information classes
yielded an overall accuracy of 63.35%. This analysis has highlighted the challenges of
accurately classifying land use and land cover in this region and may offer insight into
techniques to overcome these challenges.

Methods
This analysis is based on the procedures described in David Wallins Lab III:
Unsupervised Classification with ENVI (Wallin, 2006)
Imagery
The imagery used in this analysis is a subset of a Landsat TM image of the
Mount Baker and Bellingham Bay region in July 2011. The image is 1500 lines by 2500
columns, each pixel representing 25 square meters. The image spans approximately
2,378 square kilometers. Analysis of spectral classes was based on TM bands 1, 2, 3, 4,
5, and 7.

Initial Classification
This imagery is used as the input for the ISODATA unsupervised classification
tool in ENVI. This tool separates pixels into unique classes based on similar spectral
characteristics. The parameters of this tool may be changed to suit classification
needs. In this analysis, I have chosen to create a maximum of 50 unique classes.
Table 1: The chosen parameters for the ISODATA unsupervised classification tool.
Minimum Number of Classes

Maximum Number of Classes

50

Maximum Iterations

10

Change Threshold

5.0

Minimum Number of Pixels in Class

Maximum Class Standard Deviation

1.0

Minimum Class Distance

5.0

Maximum Number of Merge Pairs

Running this tool produces a new image containing 50 classes. I completed a


preliminary assignment of each of these classes to one of 14 information classes by
cross-referencing the classified image with a color-infrared image. These 14 classes
represent land use and land cover types. The 14 classes are as follows: residential,
urban, pasture, clearcuts between 1973 and 1992, clearcuts between 1992 and 2005,
clearcuts between 2005 and 2011, deciduous, conifer, water, soil and rock, alpine, snow
and ice, and clouds.
This first method of eyeballing each of the 50 classes to narrow the image into
14 LULC classes is fairly subjective. In order to improve my accuracy using a more
objective approach, I cross-referenced the classified image with groundtruth data. I
cross-tabulated my spectral class statistics with groundtruth information class
statistics. Using this table, I was able to clearly see which information class each
spectral class was most likely to belong in. Using this data, I revised my assignments of
the 50 spectral classes to the 14 LULC classes.

Next, the results of this initial classification are compared to groundtruth data in
order to assess accuracy. A confusion matrix was generated and analyzed in Excel. In
this matrix, my classifications are cross-tabulated with groundtruth classifications.
Calculations using this data produces several statistics used to assess the results of
my classification: producers accuracy, users accuracy, and overall accuracy.
Producers accuracy gives the probability that a pixel in a certain class is classified
correctly, assuming that the groundtruth data is correct. Users accuracy gives the
probability that a pixel labeled as a certain information class will also belong to that
same information class in the groundtruth data. Overall accuracy gives the percentage
of correctly classified pixels across all information classes. To improve classification
accuracy, I combined several information classes within the confusion matrix. The
following classes were merged:
urban, soil/rock
deciduous, clearcuts between 1973 and 1992, clearcuts between 1992 and 2005
pasture, crops

Refined Classification
Next, I refined the results of my initial classification using a series of conditional
tools in ArcMap. This process further improves classification accuracy, and allows me
to retain my original 14 classes. Adjustments to the classified image were made using a
series of conditional tools in ArcMap. These refinements are described in Table 2.

Table 2: Using a series of conditional tools, the classified image is refined to more
accurately represent real-world LULC types.
Input Condition

Reclassification

proximity to road

Residential pixels greater than 100 meters from a road are


reclassified as soil/rock.
Urban pixels greater than 100 meters from a road are reclassified
as soil/rock.
Soil/rock pixels that are within 100 meters of a road are
reclassified as urban.

slope

Water pixels located where slope does not equal 0 are reclassified
as shadow (this creates a new class)

elevation

Crop, clearcut, and alpine pixels located at low elevations ( less


than 200 m) are reclassified as crop.
Pasture, crop, and alpine pixels located at mid elevations (200 1500 m) are reclassified as clearcut (1992-2005)
Pasture, crop, and clearcut pixels at high elevations (greater than
1500 m) are reclassified as alpine.

Finally, this refined classification image was brought back into ENVI. I generated
a second confusion matrix and merged classes using the same process described in
the initial classification stage.

Results
Initial Classification
The results of the initial classification stage are shown in Figure 1.

Figure 1: The result of initial classification using 14 information classes. The overall
accuracy for this classification 44.21%.

Figure 2: The color legend corresponding to Figure 1.

Analysis of classification statistics from this image (Figure 1) shows that the
overall accuracy is 44.21%. The combination of classes increased this accuracy to
59.84%.

Table 3: Statistical analysis of all 14 classes in the initial classification image. This result
produced the least accurate classification of all the approaches.

Table 4: Statistical analysis of 10 classes, condensed from the original 14. While merging
classes sacrifices the breadth of land cover types, it significantly increases the accuracy
of classification. Overall accuracy has increased from 44.21% to 59.84% due to the
simplified classes. Producers and users accuracy is also greatly improved by merging
classes.

Refined Classification
Refinements to classification done in ArcMap significantly improved the accuracy of
land cover classification. The results of the refined classification stage are shown in
Figure 3.

Figure 3: The final result of refined classification. The overall accuracy of this
classification is 47.11%. Note the added Shadow class, which was not present in initial
classification.

Figure 4: The color legend corresponding to Figure 3.

The results of this refined classification show significant improvement in accuracy.


Shadows from mountain slopes and cloud cover are no longer classified as water.

Residential and urban cover types that clearly did not belong in rugged, mountainous
terrain have been appropriately replaced by soil/rock and alpine cover types. The overall
accuracy of the classified image containing all 15 classes is 47.11%. Condensing these
classes into 11 information classes results in an overall accuracy of 62.35%.

Table 5: individual classes, refined classification

Table 6: Statistical analysis of 11 classes, condensed from the previous 15. While
merging classes sacrifices the breadth of land cover types, it significantly increases the
accuracy of classification. Overall accuracy has increased from 47.11% to 62.35% due to
the simplified classes.

Discussion
Several issues are immediately evident in the results of the initial classification.
First, large areas of dark, north-facing slopes on Mount Baker are classified as water.
Second, rocky surfaces are classified as urban or residential. These errors occur
because unique information classes may share similar spectral properties. This is
corrected in the refinement stage of the analysis using the conditional tools in ArcMap. I
chose to use distance from roads, slope, and elevation to refine my results. These
choices (both the tools utilized and the thresholds used within the tools) were fairly
subjective and may be modified in further analysis.
The results of this analysis could be further improved by incorporating more
conditional tools. For example, a layer containing parcel data could be used to more
accurately assess where residential land use types are located.
One issue with this classification technique is the reliance on groundtruth data.
Analysts assigning information classes based on similarity to groundtruth data should
keep in mind that this data may be inaccurate or insufficient. Low numbers of

groundtruth data points may also affect results. For example, the alpine class in initial
classification yielded a 0% accuracy. This can be attributed to the fact that my
groundtruth data contained very few alpine reference points, resulting in a very limited
ability to cross-reference results. In addition, it was not possible to calculate accuracy
for cloud and shadow cover types due to the absence of groundtruth data.
While merging classes sacrifices the breadth of land cover types, it significantly
increases the accuracy of classification. All three measures of accuracy (overall,
producer, user) improve due to the merging of information classes. This balance
between accuracy and land cover type breadth is a choice the image analyst must
make.
My final classification using 11 information classes resulted in an overall
accuracy of 62.35%. While this figure improved over the course of my analysis, it may
still be considered poor. Further analysis using different LULC types, tool parameters,
and reference data may produce improved accuracy.

Literature Cited
Wallin, D. (2006). Lab III: Unsupervised Classification with ENVI.
http://faculty.wwu.edu/wallin/envr442/ENVI/442_unsup_class_ENVI.html

Anda mungkin juga menyukai