A major potential problem in image processing is poor imagery quality due to noise effects induced by miscalibrated or malfunctioning equipment. Generally, systematic noise of this type produces striping or banding in the imagery. The resulting imagery is not only distracting to view but can also yield very inaccurate products. While attempting to complete the Idrisi Advanced Tutorial 7 - Vegetation Analysis in Arid Envronments, I encountered just such a dataset. The tutorial directions clearly state that the 1980 MSS imagery has horizontal striping effects but instructs the student to use it anyway since it is the best data available. However, by utilizing multiple modules in Idrisi, the datasets (M801W, M802W, M803W, and M804W) can be greatly enhanced.
This project will demonstrate how to enhance imagery by using a combination of Principal Components Analysis (PCA) and DESTRIPE. Once the datasets are corrected, I will illustrate the difference between the original dataset and the enhanced dataset by directly comparing the two within the context of an NDVI image.
The datasets used in this project are contained in the Idrisi Exercise files. The images were taken on October 10, 1980 by LANDSAT 4 Multi-Spectral Scanner (MSS). The images (M801W, M802W, M803W, and M804W) correspond to the four MSS bands which are visible green, visible red, and two wavelengths of near-infrared. The scenes depict an area in southern Mauritania near the Senegal border. The majority of the land-cover consists of riparian vegetation, poor grassland, and barren ground; a small a portion of the Gorgol River can be seen in the upper left corner. The image below is Band 1, or visible green, viewed with a Grey 256 pallette. This particular pallete highlights the striping resident in the imagery. In terms of striping, the other three bands are very similar to Band 1. Notice the dark heavy lines running horizontally across the image.
Not only is the image heavily striped, but it is also relatively
hazy. Apparently, atmospheric effects such as scattering and background
noise have negatively impacted on the data in addition to the satellite's
mechanical noise. Before this image is used for analyzing anything,
it needs to be enhanced through a series of steps centered around the Principal
Components Analysis and DESTRIPE modules within Idrisi.
Principal Components Analysis (PCA) Module
The PCA module will separate a collection of bands into statistically seperate components, each of which contains specific information about the dataset. In this dataset, four bands are input so four components are derived. The first two components contain nearly 100% of the critical data contained in the bands. The last two components contain mostly mechanical and atmospheric noise respectively. The four components are illustrated below with a NDVI 256 pallette.
By eliminating the noise, eg Components 3 and 4, the overall image can be greatly enhanced while maintaining over 99% of the critical data. When PCA is executed, it creates a text table that contains numerous values pertaining to the components and the original bands. The information in this table is used to reassemble the imagery into its enhanced format.
To reassemble, IMAGE CALCULATOR is used to multiply each component image by its corresponding eigenvector element (found in the table) for a particular band and summing the result. Only Component 1 and Component 2 are used to reassemble the image. Component 3 and Component 4 are simply dropped thus new bands are computed without any of the original noise. Below is the recalculated Band 1 (again, the other bands yielded similar results). The image on the left is in the Grey 256 pallette for comparison against the original Band 1 image. The image on the right is in the NDVI 256 pallette for future comparison.
By comparing the two Grey 256 images, one can tell that much, but not all, of the striping has been eliminated. More importantly, the image is much crisper and clearer than the original image. Now it is time to further enhance the image.
DESTRIPE works by calculating a mean and standard deviation for the entire image and then for each detector separately. Then the output from each detector is scaled to match the mean and standard deviation of the entire image. DESTRIPE only works on byte-binary data so each image must be reformated using the CONVERT module. Once converted to byte-binary, DESTRIPE can be run. There are several fields in the DESTRIPE menu; all can be left at the default value except for "Number of detectors". The default for this field, which is 16, consistently failed to remove all of the striping. By increasing the detectors to 256, all striping can be removed from the images. The image on the left is the original M801W data; the image in the center is the PCA enhanced image; and the image on the right is the PCA and DESTRIPE enhanced image. Note the increased clarity and lack of striping in the final Band 1 image.
Final NDVI Comparison
Lastly, to demonstrate the value of enhancing imagery, I have made two NDVI images of the same scene. Both were made using the VEGINDEX module and displayed with the NDVI 256 module. The image on the left is the original data, and the one on the left is the PCA/DESTRIPE enhanced data. The image constructed using the original image is heavily striped which leads to false reflectance values. The enhanced image provides much greater clarity. While enhancing each of the four bands in the dataset was a somewhat lengthy process, it yielded a much higher quality product in the end.
Aber, J. S. Course material from ES 771 Remote Sensing. Spring 1999.
Aber, J. S. Course material from ES 775 Advanced Image Processing. Spring 2000.
Eastman, Ronald J. Idrisi for Windows User's Guide Version 2.0 (Revision 5). December 1997.
Eastman, Ronald J. Idrisi for Windows Tutoral Exercises (Revision 5). December 1997.
Idrisi data files M801W, M802W, M803W, M804W
This page prepared for partial fulfillment of ES
775 Advanced Remote Sensing
taught by Dr. James Aber at the Emporia State University Earth Science Department
by Steve Tucker