You are on page 1of 16

GISC9216 Deliverable 2 Principal Component Analysis

Shannon Graup

2/12/2014
Page | 0

Shannon Graup
47 Notley Place Toronto Ontario M4B 2M7 647-921-4930 shannon.graup@gmail.com

February 12th, 2014 GISC 9216-D2 Janet Finlay Program Coordinator Niagara College 135 Taylor Road Niagara-on-the-Lake, ON L0S 1J0

Dear Ms. Finlay, Re: GISC9216-D2 Principal Component Analysis Please accept this letter as my formal submission of deliverable 2, Principal Component Analysis for GISC9216 Digital Image Processing. The purpose of this deliverable was to compare two different unsupervised classification images of a Barrie, Ontario Region subset. A Principal Component Analysis was completed on the subset image and then an unsupervised classification of the PCA. The PCA classification was compared to an unsupervised classification of the original image subset. After contrasting the two images it was determined that the PCA classification better classified land areas into logical classes in comparison to the original classification. If there are any technical issues regarding the deliverable .img files or you have any questions regarding the assignment submission please feel free to contact me by phone (647-921-4930) or e-mail at your convenience. I look forward to receiving your feedback and suggestions. Sincerely,

Shannon Graup BAH GIS GM Candidate S.G/s.g Enclosures 1) GISC9216 Deliverable 2 Principal Component Analysis

Shannon Graup shannon.graup@gmail.com

Table of Contents_Toc379826994
1.0 Introduction .......................................................................................................................... 1 2.0 Methodology ........................................................................................................................ 1 2.1 Feature Space Image ....................................................................................................... 1 2.2 Principal Component Analysis .......................................................................................... 1 2.3 Unsupervised Classification of Principal Component Analysis .......................................... 1 3.0 Discussion............................................................................................................................ 2 3.1 Principal Component Analysis & Channel Variance .......................................................... 2 3.2 Subset Band Correlation................................................................................................... 2 3.3 Unsupervised Classification Comparison .......................................................................... 4 4.0 Conclusion ........................................................................................................................... 9 Bibliography ............................................................................................................................... 9

List of Figures
Figure 1: Feature Space Image Bands 1 & 2, Strong Correlation ............................................... 3 Figure 2: Feature Space Image Bands 3 & 6, Moderate Correlation .......................................... 3 Figure 3: Feature Space Image Bands 2 & 4, Weak Correlation ................................................ 3 Figure 4: Feature Space Image Bands 2 & 3, Strong Correlation ............................................... 3 Figure 5: Feature Space Image Bands 5 & 6, Moderate Correlation .......................................... 3 Figure 6: Feature Space Image Bands 4 & 6, Weak Correlation ................................................ 3 Figure 7: Feature Space Image Channels 1 and 2, Weak Correlation ....................................... 3 Figure 8: Feature Space Image Channels 1 and 3, Weak Correlation ........................................ 3 Figure 9: Feature Space Image Channels 2 and 3, Weak Correlation ........................................ 3 Figure 10: Barrie, Ontario Region Unsupervised Classification .................................................. 4 Figure 11: Barrie, Ontario Region Principal Component Analysis, Unsupervised Classification . 5 Figure 12: Unsupervised Classification, Bareground Agriculture Clip ......................................... 6 Figure 13: PCA Unsupervised Classification, Bareground Agriculture Clip ................................. 6 Figure 14: Barrie, Ontario Region Subset, Bareground Agriculture Clip ..................................... 6 Figure 15: Unsupervised Classification, bareground agriculture clip........................................... 7 Figure 16: PCA Unsupervised Classification, bareground agriculture clip .................................. 7 Figure 18: Unsupervised Classification, Residential Area .......................................................... 8 Figure 19: PCA Unsupervised Classification, Residential Area .................................................. 8

Page | i

Shannon Graup shannon.graup@gmail.com

List of Tables
Table 1: PCA Eigenvalues Variance .......................................................................................... 2 Table 2: Barrie, Ontario Region Subset Band Correlation .......................................................... 2

List of Appendices
Appendix 1: Barrie, Ontario Region Subset ..............................................................................10 Appendix 2: Barrie, Ontario Region Unsupervised Classification ..............................................11 Appendix 3: Barrie, Ontario Region Principal Component Analysis, Unsupervised Classification .................................................................................................................................................12

Shannon Graup shannon.graup@gmail.com

1.0 Introduction
Principal component analysis is a technique used to reduce redundancy in multispectral data (Lillesand, Kiefer, & Chipman, 2008). Correlation between bands is a problem in analysis of multispectral image data, as various wavelengths produce the same information (Lillesand, Kiefer, & Chipman, 2008). Using a subset of the Barrie, Ontario Region a principal component analysis (PCA) was completed to reduce the redundancy and band correlation. Once the PCA was completed an unsupervised classification was completed and compared to an unsupervised classification of the original Barrie, Ontario Region subset.

2.0 Methodology
To complete analysis on the Barrie, Ontario Region subset and principal component image, and to complete a comparison of the two unsupervised classifications a number of steps were completed to create image files and complementary statistics.

2.1 Feature Space Image


As principal component analysis works to reduce redundancy in multispectral data, it is useful to determine band correlation in an image. To determine which bands have a strong correlation feature space images were created in ERDAS.

2.2 Principal Component Analysis


A principal component analysis was completed in ERDAS to reduce redundancy of the bands. For this transformation three channels were selected, and the option to write the eigen matrix and eigenvalues to file was selected. The eigenvalues are used to analyze the variance in the PCA channels.

2.3 Unsupervised Classification of Principal Component Analysis


An unsupervised classification was completed on the principal component analysis in ERDAS. The original classification image had 10 classes, max iteration of 10 and a convergence threshold of 0.95. These settings were set to the same as the unsupervised classification of the original Barrie, Ontario Region subset shown in Appendix 2: Barrie, Ontario Region Unsupervised Classification. The principal component analysis classification was recoded into seven classes; agriculture, bareground, bareground agriculture, commercial, forest, residential and water.

Shannon Graup shannon.graup@gmail.com

3.0 Discussion
3.1 Principal Component Analysis & Channel Variance
The purpose of completing the principal component analysis on the original subset is to compress the information from the six bands into fewer bands. For this PCA the data was transformed into three channels. In a principal component analysis it can be noted that the first channel will have the largest percentage of scene variation (Lillesand, Kiefer, & Chipman, 2008). The scene variation percentage will decrease with succeeding components. Table 1 lists the eigenvalues for each PCA channel and their respective scene variation percentages. Channel 1 has 67.41% of the scene variation, channel 2 has 28.34% and channel 3 has 3.54%. The percentage of scene variation that was transformed onto the three channels is 99.29%, meaning that 0.71% of the entire scene variation was lost.
Table 1: PCA Eigenvalues Variance

PCA Channel 1 2 3 4 5 6 Sum

Eigenvalues 1795.9993 755.2083 94.2291 12.4023 4.8708 1.6578 2664.3675 PCA Channel 1, 2, 3 Lost 0.9929 0.0071 0.6741 0.2834 0.0354 0.0047 0.0018 0.0006

% 67.41 28.34 3.54 0.47 0.18 0.06

99.29 0.71

3.2 Subset Band Correlation


The reason for completing a principal component analysis is to try and eliminate redundancy between bands. To determine which bands were highly correlated in the original Barrie, Ontario Region subset feature space images were created. Table 2 identifies which bands have strong correlation, moderate correlation and which bands have a weak correlation. It can be seen that bands 1 and 2, 1 and 3, and 2 and 3 are the bands that have a strong correlation in the Barrie, Ontario region subset. Other band correlations are moderate or weak.
Table 2: Barrie, Ontario Region Subset Band Correlation

Band 1 2 3 4 5

2
Strongly correlated

3
Strongly correlated Strongly correlated

4
Weakly correlated Weakly correlated Weakly correlated

5
Weakly correlated Weakly correlated Weakly correlated Weakly correlated

6
Weakly correlated Moderately correlated Moderately correlated Weakly correlated Moderately correlated

Shannon Graup shannon.graup@gmail.com Figure 1 through 6 show examples of the feature space images that were produced by ERDAS to identify band correlation from the Barrie, Ontario Region Subset. Figure 1 and Figure 4 show feature space images of bands that have strong correlation. Figure 2 and Figure 5 show bands 3 and 6, and bands 5 and 6. These feature space images show moderate correlation. Lastly Figure 3 and Figure 6 show examples of feature space images with weak correlation between bands.

Figure 1: Feature Space Image Bands 1 & 2, Strong Correlation

Figure 2: Feature Space Image Bands 3 & 6, Moderate Correlation

Figure 3: Feature Space Image Bands 2 & 4, Weak Correlation

Figure 4: Feature Space Image Bands 2 & 3, Strong Correlation

Figure 5: Feature Space Image Bands 5 & 6, Moderate Correlation

Figure 6: Feature Space Image Bands 4 & 6, Weak Correlation

After completing the principal component analysis feature space images were created to identify if any band correlation exists. Figure 7, Figure 8 and Figure 9 show the band correlation between the principal component channels. It can be seen that completing the PCA has eliminated any band correlation in the image, therefore reducing the redundancy between channels.

Figure 7: Feature Space Image Channels 1 and 2, Weak Correlation

Figure 8: Feature Space Image Channels 1 and 3, Weak Correlation

Figure 9: Feature Space Image Channels 2 and 3, Weak Correlation

Shannon Graup shannon.graup@gmail.com

3.3 Unsupervised Classification Comparison


An unsupervised classification was completed on the original Barrie, Ontario Region subset, and on the principal component analysis of the subset. Figures of the classification images can be seen in Figure 10 (subset unsupervised classification) and Figure 11 (PCA unsupervised classification). Proper map layouts of the two classifications can be seen in Appendix 2: Barrie, Ontario Region Unsupervised Classification and Appendix 3: Barrie, Ontario Region Principal Component Analysis, Unsupervised Classification.

Figure 10: Barrie, Ontario Region Unsupervised Classification

The unsupervised classification of the Barrie, Ontario Region subset can be seen in Figure 10. This classification originally had 10 classes but was reclassed into six; agriculture, bareground,

Shannon Graup shannon.graup@gmail.com commercial, forest, residential and water. The original subset classification has strong forest and water classes. The main downfall of this classification is the improper classification of the roads class. Some roads in the subset were classified as agriculture, while others were classified as residential. Another issue with this classification is the improper classification of some bareground and bareground agricultural areas.

Figure 11: Barrie, Ontario Region Principal Component Analysis, Unsupervised Classification

The unsupervised classification of the PCA can be seen in Figure 11. This classification has seven classes; agriculture, bareground, bareground agriculture, commercial, forest, residential, and water. It can be noted that an extra class was added to this classification in comparison to

Shannon Graup shannon.graup@gmail.com the original classification, called bareground agriculture. This class has identified the agricultural areas that are not made of healthy vegetation and separated them from the bareground areas. This is a vast improvement from the original unsupervised classification of the Barrie, Ontario region subset where these agricultural bareground areas were classified as bareground. Figure 12, Figure 13 and Figure 14 show an area in Barrie, Ontario where there are bareground agriculture fields surrounded by forest. This can be identified in Figure 14, the false colour subset that shows the forest in dark red, and the bareground agricultural fields in green. In Figure 12 the bareground agriculture fields are classified as bareground, or residential while in Figure 13 the fields are classified as bareground agriculture and bareground, which is more correct than the original classification image.

Figure 12: Unsupervised Classification, Bareground Agriculture Clip

Figure 13: PCA Unsupervised Classification, Bareground Agriculture Clip

Figure 14: Barrie, Ontario Region Subset, Bareground Agriculture Clip

Shannon Graup shannon.graup@gmail.com Figure 15 and Figure 16 show a residential area in Barrie, Ontario. Figure 16 (PCA unsupervised) classified this residential area much better than Figure 15 (original unsupervised). The original unsupervised classification classed most of the residential area as forest and agriculture, while the PCA unsupervised classification has classified the area as residential and bareground (for the roads).

Figure 15: Unsupervised Classification, bareground agriculture clip

Figure 16: PCA Unsupervised Classification, bareground agriculture clip

Shannon Graup shannon.graup@gmail.com Although there is no roads class separated from bareground, the bareground class in the PCA unsupervised classification has identified the roads in Barrie, Ontario very well, in comparison to the original unsupervised classification. The roads in the urban/residential area have been classified very well as bareground in the PCA unsupervised classification. Figure 17 and Figure 18 identify a residential area in Barrie, Ontario. The PCA unsupervised classification classifies this area much better than the original unsupervised classification. Residential, bareground and commercial areas are clearly identified in the PCA classification, while these same areas in the original classification are mixed with agriculture, forest and residential.

Figure 17: Unsupervised Classification, Residential Area

Figure 18: PCA Unsupervised Classification, Residential Area

Completing the principal component analysis on the image and thereby reducing redundancy between bands has helped in creating an unsupervised classification image that has better identified urban/residential and agricultural areas. The reason for misclassification is similar spectral signatures between classes, meaning redundancy in the data and high band correlation. Bareground land areas and bareground agricultural areas have very similar spectral signatures and without training data an unsupervised classification can group these areas together. When a principal component analysis eliminates redundancy of multispectral data, an unsupervised classification can be more successful in separating classes with similar spectral signatures.

Shannon Graup shannon.graup@gmail.com

4.0 Conclusion
During this deliverable a principal component analysis was completed on a subset of the Barrie, Ontario Region. Once the PCA was created an unsupervised classification was completed on the image, and then compared to an unsupervised classification of the original subset image. After comparison of the two unsupervised classification images it was determined that the PCA unsupervised classification grouped areas into logical classes better than the original unsupervised classification.

Bibliography
Lillesand, T. M., Kiefer, R. W., & Chipman, J. W. (2008). Remote Sensing and Image Interpretation (Sixth ed.). John Wiley & Sons, Inc.

Shannon Graup shannon.graup@gmail.com

Appendix 1: Barrie, Ontario Region Subset

Shannon Graup shannon.graup@gmail.com

Appendix 2: Barrie, Ontario Region Unsupervised Classification

Shannon Graup shannon.graup@gmail.com

Appendix 3: Barrie, Ontario Region Principal Component Analysis, Unsupervised Classification

You might also like