Sie sind auf Seite 1von 15

Principal Component Analysis

By: Kirsten Anderson For: Janet Finlay

February 12

2014
GISC9216 Digital Image Processing Assignment 2

Kirsten Anderson 1208 Lansdown Drive Oakville, Ontario L6J 7N6 Phone (289) 259-2790 E-mail kirstenpatrice@gmail.com

March 4, 2014 Deliverable 2 Principal Component Analysis Janet Finlay Instructor and Thesis Project Co-ordinator Niagara College, NOTL Campus 135 Taylor Rd. Niagara-on-the-Lake, ON, L0S 1J0 Dear Mrs. Finlay, RE: GISC9216 Deliverable 2 Principal Component Analysis Please accept this letter as a formal submission of Deliverable 2 Principal Component Analysis for GISC9216, Digital Image Processing. The purpose of this deliverable is to become familiar with Principal Component Analyses. This deliverable begins by explaining the purpose of carrying out a Principal Component Analysis (PCA) by looking at the correlation between the bands of the original subset image, the eigenvalues for the bands of the PCA, as well as the correlation between the bands of the PCA. An unsupervised classification was then done on the PCA image and compared to the unsupervised classification of the original subset image. This comparison focused on the urban and agriculture land classes. If there are technical difficulties regarding the documents or necessary questions regarding the documents, please contact me at your convenience at (289) 259-2790. Thank you very much for your time and I look forward to your comments and feedback regarding this deliverable.

Sincerely,

Kirsten Anderson
Kirsten Anderson BAH GIS-GM Certificate Candidate Enclosures: i. Deliverable 2 Principal Component Analysis

Table of Contents
1.0 2.0 Introduction ...................................................................................................................................... 1 Analysis ............................................................................................................................................. 1

2.1 Purpose of a Principal Component Analysis ....................................................................................... 1 2.2 Unsupervised Classification of a Principal Component Image............................................................ 4 2.2.1 Urban Area Comparisons ............................................................................................................. 4 2.2.2 Agricultural Land Comparisons .................................................................................................... 5 3.0 Conclusion ......................................................................................................................................... 6

4.0 References .............................................................................................................................................. 7 Appendix 1 .................................................................................................................................................. 15 Appendix 2 .................................................................................................................................................. 17

List of Tables
Table 1: Correlation Between Bands ............................................................................................................ 2 Table 2: PCA Result Eigenvalues ................................................................................................................... 3

Kirsten Anderson 1208 Lansdown Drive Oakville, Ontario L6J 7N6 Phone (289) 259-2790 E-mail kirstenpatrice@gmail.com

ii

List of Figures
Figure 1: Feature Space Image of Bands 1 and 2 .......................................................................................... 2 Figure 2: Feature Space Image of Bands 1 and 4 .......................................................................................... 2 Figure 3: Feature Space Image of Bands 2 and 5 .......................................................................................... 2 Figure 4: Feature Space Image of Bands 3 and 6 .......................................................................................... 2 Figure 5: Feature Space Image of Bands 4 and 5 .......................................................................................... 2 Figure 6: Feature Space Image of Bands 5 and 6 .......................................................................................... 2 Figure 7: PCA Feature Space Image of Bands 1 and 2................................................................................... 3 Figure 8: PCA Feature Space Image of Bands 1 and 3................................................................................... 3 Figure 9: PCA Feature Space Image of Bands 2 and 3................................................................................... 3 Figure 10: Urban Area in Unsupervised Classification for Principal Component Image............................... 5 Figure 11: Urban Area in Unsupervised Classification for the Original Subset Image .................................. 5 Figure 12: Urban Area in Subset Image ........................................................................................................ 5 Figure 13: Agricultural Land in Unsupervised Classification for Principal Component Image ...................... 6 Figure 14: Agricultural Land in Unsupervised Classification for the Original Subset Image ......................... 6 Figure 15: Agricultural Land in Subset Image ............................................................................................... 6

1.0 Introduction
The purpose of performing a Principal Component Analysis (PCA) is to disregard redundant information in the multispectral image data. The redundancy is eliminated by compressing the information held in the bands into fewer bands (Lillesand, Keifer, & Chipman, 2008). The first Principal Component image usually has the largest percentage of the variance of the image, as presented by the Eigenvalue for each of the images (Lillesand, Keifer, & Chipman, 2008). For this particular assessment, a subset image of Midland, Ontario was classified using an unsupervised classification (Appendix 1). The original subset image was then converted to a Principal Component image and then classified using an unsupervised classification (Appendix 2).

2.0 Analysis
The following is the completed analysis undertaken to compare the original subset image to the Principal Component image, as well as the unsupervised classification of the Principal Component image (Appendix 1), to the unsupervised classification of the Principal Component image (Appendix 2).

2.1 Purpose of a Principal Component Analysis


The original bands of the subset image were transformed to the Principal Components in order to eliminate the correlation between the bands by compressing the information from the 6 bands into 3 bands (Lillesand, Keifer, & Chipman, 2008), which is why this process was completed on the original subset image. In the original subset image bands, there are 3 pairs of bands that are strongly correlated ( Table 1). These pairs are; bands 1 and 2, 1 and 3, and 2 and 3.

Kirsten Anderson 1208 Lansdown Drive Oakville, Ontario L6J 7N6 Phone (289) 259-2790 E-mail kirstenpatrice@gmail.com

2
Table 1: Correlation Between Bands

Band # 1

1 -

2 Strongly Correlated (Figure 1) -

3 Strongly Correlated Strongly Correlated -

4 Weakly Correlated (Figure 2) Weakly Correlated Weakly Correlated

5 Weakly Correlated (Figure 3) Weakly Correlated Weakly Correlated Weakly Correlated (Figure 5) -

6 Weakly Correlated Weakly Correlated Moderately Correlated (Figure 4) Weakly Correlated Moderately Correlated (Figure 6)

2 3

4 5 -

Figure 1: Feature Space Image of Bands 1 and 2

Figure 2: Feature Space Image of Bands 1 and 4

Figure 3: Feature Space Image of Bands 2 and 5

Figure 4: Feature Space Image of Bands 3 and 6

Figure 5: Feature Space Image of Bands 4 and 5

Figure 6: Feature Space Image of Bands 5 and 6

Kirsten Anderson 1208 Lansdown Drive Oakville, Ontario L6J 7N6 Phone (289) 259-2790 E-mail kirstenpatrice@gmail.com

The variance on the first, second and third channel of the PCA result is demonstrated by the Eigenvalues (Table 2).
Table 2: PCA Result Eigenvalues

Channel 1 2 3 4 5 6

Eigenvalues 2849.466589 476.3079803 35.89556355 5.925213718 2.673038666 1.379185075

Percentage 84.51 14.13 1.06 N/A N/A N/A

Most of the scene variance was in channel 1, as expected, with 84.51% of the scene variance. The scene variance decreases as the channel number increases. The channels after channel 3 are insignificant, because they display relatively no variance, which is why the PCA displays the first 3 channels.

Figure 8: PCA Feature Space Image of Bands 1 and 3

Figure 9: PCA Feature Space Image of Bands 2 and 3

Figure 7: PCA Feature Space Image of Bands 1 and 2

Figure 7 shows the relationship between bands 1 and 2 of the Principal Component image, and they have no correlation. Comparing this to Figure 1, which shows the relationship between bands 1 and 2 of the subset image, it shows that the Principal Component image eliminates redundancy. This is also relevant when comparing Figure 8, which show the weak correlation between bands 1 and 3, to the correlation of bands 1 and 3 in the subset image, indicated in Table 1, indicating that bands 1 and 3 are strongly correlated. Similarly, unlike bands 2 and 3 in the original subset image (Table 1), bands 2 and 3 of the Principal Component image are not correlated (Figure 9).

Kirsten Anderson 1208 Lansdown Drive Oakville, Ontario L6J 7N6 Phone (289) 259-2790 E-mail kirstenpatrice@gmail.com

2.2 Unsupervised Classification of a Principal Component Image


Appendix 2 shows the unsupervised classification of the Principal Component image. In order to successfully compare the unsupervised classifications for the subset image and Principal Component image, the same parameters were used to create each of the unsupervised classifications.
2.2.1 Urban Area Comparisons

Figure 12 shows an urban area in the subset image. Looking at the urban regions for the first unsupervised classification of the subset image, they are very well defined (Figure 11). In the Principal Component unsupervised classification image (Figure 10), the urban areas are mainly classified as bare ground, likely because these pixels are light in colour.

Kirsten Anderson 1208 Lansdown Drive Oakville, Ontario L6J 7N6 Phone (289) 259-2790 E-mail kirstenpatrice@gmail.com

Figure 10: Urban Area in Unsupervised Classification for Principal Component Image

Figure 11: Urban Area in Unsupervised Classification for the Original Subset Image

Figure 12: Urban Area in Subset Image

2.2.2 Agricultural Land Comparisons Figure 15 shows a plot of agricultural land on the original subset image. There was difficulty recognizing agricultural land on the unsupervised classification of the subset image because much of the agricultural land was classified as dense forest (Figure 14). Much of the agricultural land was properly classified in the unsupervised classification for the Principal Component Image (Figure 13).

Kirsten Anderson 1208 Lansdown Drive Oakville, Ontario L6J 7N6 Phone (289) 259-2790 E-mail kirstenpatrice@gmail.com

Figure 13: Agricultural Land in Unsupervised Classification for Principal Component Image

Figure 14: Agricultural Land in Unsupervised Classification for the Original Subset Image

Figure 15: Agricultural Land in Subset Image

3.0 Conclusion
Creating a Principal Component image simplified the image by reducing the number of bands the image is comprised. This reduction of bands eliminates redundancy between the bands, demonstrated by the weak correlation of the bands in the PCA. As shown in the comparison between the unsupervised classification of the original subset image and the Principal Component Image, a PCA can be useful for analyzing different land classes. However, in some cases a PCA may not be helpful for classifying pixels. Whether or not this analysis is carried out is dependent on the individual subset image, as well as what the user wishes to accomplish.

Kirsten Anderson 1208 Lansdown Drive Oakville, Ontario L6J 7N6 Phone (289) 259-2790 E-mail kirstenpatrice@gmail.com

4.0 References
Lillesand, T. M., Keifer, R. W., & Chipman, J. W. (2008). Remote Sensing and Image Interpretation. New Delhi, India: John Wiley & Sons Inc.

Kirsten Anderson 1208 Lansdown Drive Oakville, Ontario L6J 7N6 Phone (289) 259-2790 E-mail kirstenpatrice@gmail.com

15

Appendix 1 Subset Image of Midland, Ontario Unsupervised Classification

16

Kirsten Anderson 1208 Lansdown Drive Oakville, Ontario L6J 7N6 Phone (289) 259-2790 E-mail kirstenpatrice@gmail.com

17

Appendix 2 Niagara Subset Image of Midland, Ontario Principal Component Analysis Unsupervised Classification

Kirsten Anderson 1208 Lansdown Drive Oakville, Ontario L6J 7N6 Phone (289) 259-2790 E-mail kirstenpatrice@gmail.com

18

Kirsten Anderson 1208 Lansdown Drive Oakville, Ontario L6J 7N6 Phone (289) 259-2790 E-mail kirstenpatrice@gmail.com

Das könnte Ihnen auch gefallen