Hao-Hsiung Huang and Chiao-Ju Hsiao
Associate Professor and Graduate Student
Department of Land Economics
National Cheng-Chi University
64, Sec. 2, Tzu-Nan Road, Wensan, Taipei, Taiwan
Tel: (886)-2-29379261, Fax: (886)-2-2939-0251
E-mail:[email protected] and [email protected]
Keywords: Supervised Classification, Change Detection
Natural healthy grass and varnished grass all appear green under a visible band, and generally not to be distinguishable. But this does not apply to the near infrared wave testing which is useful in determining the health conditions of plants. Experiments have been designed and executed, therefore, using both normal color slides and color infrared slides to take both natural health and varnished grass at a close distance. Then, the slides were scanned and transformed to obtain digital images for comparison. Observing the changes on the brightness value, determining the differences in healthy plants and varnished grass, and then studying all the above results for reference in order to improve the credibility on plant testing in the future. Variables are controlled and fixed in the whole process to provide an ideal remote sensing environment and to increase the accuracy of the experiment.
When investigating the characteristic of the ground area with remote sensed images, one should keep in mind that to minimize unwanted spectral variability as well as to maximize this variability when the specific application requires it (Lillesand, 2000). According to this, in order to test the accuracy of change detection using postclassification comparison, such as temporal effects and spatial effects are minimized in this research. The methods, experimental design, results and their analyses are discussed respectively in the following sections.
2.1. Image Classification
In general, image classification involves three procedures, supervised classification, unsupervised classification, and hybrid classification. Supervised classification with Gaussian maximum likelihood classifier has been used to simplify the study and improve the accuracy.
Base on the maps, images, pictures that could represent the particle category are taken as samples for training sites. To calculate the statistics of the mean, variance, variance-covariance matrix on every training sites, and then come up with the classification according to the produced equation and appropriate wave band formula.
Gaussian maximum likelihood classifier is one of the most used formulas in supervised classification. First, the brightness value of every category, wave band from the image is set as norm and then the operator determine the numbers of categories in the picture. And training sites are picked and selected in each picture. Using the brightness value of the picture in training sites, calculating the mean and covariance matrix within each wave band in each category. And then apply the result to the equation of possibility, and calculus then figure out the possibilities of the item to each category. Finally, classify according to the maximum likelihood.
2.2. Change Detection
To detect the changes on the ground area or gather the changes in a short period of time can only rely on the remote sensing image data. Change detection is to compare and contrast the two images with symmetrical positions, and use image-handling technique to analyze the reformed area. There are many methods to detect the reformation, such as Image Differencing Method, Multi-Date Composite Image Change Detection, and Post-Classification Comparison Change Detection…etc. The Post-Classification Comparison Change Detection is to classify the rectified images separately from two periods of time, giving appropriate marks to different particles on the surface of the ground. Then, compare and analyze the classified images from the two periods to figure out the change-detecting matrix, and finally construct the change map.
3. Experimental Design
This research is to investigate the light wave reaction on the real green grass. Normal color slides and color infrared slides have been used to closely take the same area. Flow chart of the experiment is shown in Fig.1.
Figure 1 Flow chart of the experiment
3.1. Image Acquisition
Healthy green grass has a strong reflection in the near infrared region. Therefore, they appear bright red on color infrared images. However, varnished grass often shows low reflection rate in the near infrared area, thus it appears dark red on the color infrared image. According to this, change of the environment has been simulated. A normal color film has been put in one 50mm single lens camera, and a color infrared film has been put in another 50 mm single lens camera. Control points have also been set up around the experimental area. To observe the reflection difference on both healthy and false green plants, four small areas are varnished in green. A pixel size of 400dpi was used for scanning. Four digital color images were generated in the size of 450 x 210, and processed as described in the succeeding sections.
3.2. Image Registration and Image Classification
Four images were registered by the use of ground control points. One of the four images has been used as the master image. The other images were then geometric registered to each other. ER Mapper software was used to do the supervised classification with Gaussian maximum likelihood algorithm.
3.3. Change Detection
Following registration and classification, two classified images were produced. One image was classified healthy grass without any change, and the other one was classified varnished grass. Post-Classification Comparison was then employed to detect the differences between the two images. Change maps have been complied to display the specific nature of the changes between the two classified images.
4. Outcome Analysis
4.1. Color Infrared Image
Images of healthy grass without any change and of varnished grass are shown respectively in Fig. 2 and Fig. 3. The white circles shown in these images are control points used in image registration. Healthy vegetation shown in Fig.2 reflects strong infrared wave before change. Some grass have been painted and then exposed dark red as shown in Fig.3. Because of the painted grass have low reflection in the IR region.
4.2. Supervised Classification
Classified images of healthy grass without any change and of varnished grass are shown respectively in Fig. 4 and Fig. 5. Four blue spot with red ring shown in these images are control points. Healthy vegetation appears to be green tone, painted grass shown in Fig.5 appears red tone, and unclassified pixels shown in both Fig.4 and Fig.5 appear to be black tone.
4.3. Change Detection Matrix
Post-classification change detection used in this research provides “from-to” information. A change detection matrix is shown as below. Note in the table that the pixels without change are located along the major diagonal of this matrix. Note also in the table that the producer’s accuracy of painted grass class is only 20.29%, the omission error is 79.71%. The user’s accuracy of painted grass class is 15.17%, the commission error is 84.83%. The overall classification accuracy of this error matrix is 62.82%. Kappa analysis is a discrete multivariate technique of use in accuracy assessment (Congalton and Mead, 1983). Khat computation incorporates the of-diagonal elements as a product of the row and column marginal. Computation of the Khat statistic may also be used to compare two similar matrices(consisting of identical categories) to determine if they are significantly different. The Khat computation of the error matrix is 17.34%, and it means that the two classified images have significantly difference.
|Categories||Image categories without change (from)|
sified pixel (0)
|Healthy grass (1)||Painted grass (2)||Control point (3)||Total|
sified pixel (0)
|Healthy grass (1)||5018||6924||755||1||12698|
|Painted grass (2)||2215||1931||755||76||4977|
|Producer’s Accuracy||User’s Accuracy|
|Unclassified pixel = 51461/58694 = 86.21%||Unclassified pixel = 51461/76529 = 67.24%|
|Healthy grass = 6924/31780 = 21.79%||Healthy grass = 6924/12698 = 54.53%|
|Painted grass = 755/3721 = 20.29%||Painted grass = 755 / 4977 = 15.17%|
|Control point = 228/305 = 74.75%||Control point = 228/296 = 77.03%|
|where||r is the number of rows in the matrix|
|Xii is the number of observations in row i and column i|
|Xi+ X+i are the marginal totals for row i and column i|
|N is the total number of observations|
=76529*58694+12698*31780+4977*3721+296*305 = 4913945263
(94500*59368-4913945263)/(945002-4913945263) = 17.34%
Figure 2 Color IR image without change
Figure 3 Color IR image with change
Figure 4 Classified image without change
Figure 5 Classified image with change
Figure 6 Change Map
Spectral variability has been controlled as possible as in this research to simulate an ideal remote sensing environment. The visual appearance of the change map was so good that one can easily distinguish any change by the use of post-classification change detection method. Furthermore, utilizing such a simple camera has provided considerable convenience for remote sensing experiment. On the other hand, the accuracy of registration influences the quality of classification. As the same reason, a classification with high quality indicates that change detection has excellent results.
- Augenstein, E., D. Stow, and A. Hope, 1991, “Evaluation of SPOT HRV-XS Data for Kelp Resource Inventories,” Photogrammetric Engineering & Remote Sensing,57(5):501-509
- Chavez, P. S., Jr., 1986. Digital merging of Landsat TM and digitized NHAP data for 1:24,000 scale image mapping, Photogrammetric Engineering & Remote Sensing, 52(10), pp. 1637-1646.
- Jensen, J. R. and S. Narumalani, 1992,”Improve Remote Sensing and GIS Reliability Diagrams, Image Genealogy Diagrams, and Thematic Map Legends to Enhance Communication,” International Archives of Photogrammetry and Remote Sensing,6(B6):125-132.
- Jensen, J. R., D. J. Cowen, S Narumalani, J. D.Althausen, and O.Weatherbee,1993a,”An Evaluation of Coast Watch Change Detection Protocol in South Carolina,” Photogrammetric Engineering & Remote Sensing,59(6):1039-1046.
- Jensen, J. R. 1996. Introductory digital image processing: A remote sensing perspective, Prentice Hall.
- Jensen, J. R., E. W. Ramsey, J. M. Holmes, J. E. Michel, B. Savitsky, and B. A. Davis 1990. Environmental Sensitivity Index (ESI) Mapping for Oil Spills Using Remote Sensing and Geographic Information System Technology, International Journal of Geographical Information Systems, 4(2), pp. 181-201.
- Lillesand, T. M., and R. W. Kiefer. 2000. Remote Sensing And Image Interpretation, 4th ed., John Wiley & Sons.
- Rutchey, K. and L.Velcheck, 1994,”Development of an Everglades Vegetation Map Using a SPOT Image and the Global Positioning System,” Photogrammetric Engineering & Remote Sensing,60(6):767-775.
- Walkey, J. A., “Development of a Change Detection Tool for Image Analysts,” M.S. Thesis, University of Wisconsin-Madison, 1997.