Short URL

Image Preprocessing

 

Other Methodology areas:

Loading, please wait...

Chapter 4 - Methodology

Data Acquisition  |  GPS Ground Truth Data Acquisition  |  Image Preprocessing  |  Image Classification  |  Change Detection  |  1984 Land Use & Land Cover Map  |  Data Conversion  |  Land Assessment  |  Hardcopy Map Production

Image Preprocessing

Preprocessing of satellite images prior to image classification and change detection is essential. Preprocessing commonly comprises a series of sequential operations, including atmospheric correction or normalization, image registration, geometric correction, and masking (e.g., for clouds, water, irrelevant features) (Coppin & Bauer, 1996).

The normalization of satellite imagery takes into account the combined, measurable reflectances of the atmosphere, aerosol scattering and absorption, and the earth’s surface (Kim & Elman, 1990). It is the volatility of the atmosphere which can introduce variation between the reflectance values or digital numbers (DN’s) of satellite images acquired at different times Although the effects of the atmosphere upon remotely sensed data are not considered errors, since they are part of the signal received by the sensing device (Bernstein et al, 1983), consideration of these effects is important. The goal aptly stated by Hall et al (1991), should be that following image preprocessing, all images should appear as if they were acquired from the same sensor. The method employed to resolve the issue of satellite image normalization is discussed further in succeeding pages of this chapter. (See Change Detection)

Geometric rectification of the imagery resamples or changes the pixel grid to fit that of a map projection or another reference image. This becomes especially important when scene to scene comparisons of individual pixels in applications such as change detection are being sought (ERDAS, 1999).

Geometric Registration

To conform the pixel grids and remove any geometric distortions in the TM imagery, the first TM image, September 18, 1984, was registered to the UTM Zone 15, NAD27 coordinate system, based upon ground control points collected from overlain Arkansas Highway and Transportation Department (AHTD) road vector data. Each of the three 1999 scenes were registered to the September 18, 1984 image utilizing similar sets of ground control points (GCP’s).

GCP collection, first order transformation, and nearest neighbor resampling of the uncorrected imagery was performed. First order transformation is also known as a linear transformation which applies the standard linear equation (y = mx + b) to the X and Y coordinates of the GCP’s. The nearest neighbor resampling method uses the value of the closest pixel to assign to the output pixel value and thus transfers original data values without averaging them as other methods do, therefore, the extremes and subtleties of the data values are not lost (ERDAS, 1999).

Registered vectors (AHTD road data for Carroll County) were overlaid upon the 1984 uncorrected imagery. Image identifiable points were selected and matched to vectors until a semi-regular grid of GCP's covered the entire scene. Those GCP's were then used to project the uncorrected imagery to a UTM coordinate system, Zone 15, NAD27. Each GCP was ordered by the residual error it contributed to the polynomial fit. Points with high error were discarded before registration. Image fit was considered acceptable if the RMS error was < 15 m or one-half pixel wide (RMS = 0.5). Overall, RMS errors of less than 0.5 pixel were achieved for each transformation. RMS error is the distance between the input (source) location of a GCP, and the resampled location of the same GCP (ERDAS, 1999).

Additional research by Dai and Khorram (1998) obtained after the preprocessing phase was complete, indicate that due to misregistration the accuracy of remotely sensed change detection can be substantially degraded. Results of their analysis on Landsat TM data indicated that a registration accuracy of less than one-fifth of a pixel (0.2) is required to achieve a change detection error of less than 10%. However, Dai and Khorram (1998) also suggest that there are inherent differences between TM image pairs which may be more or less sensitive to image misregistration than other pairs. It should be noted that the registration of all four TM images did not coincide with such rigorous control; however, the two images incorporated in the change detection process, September 18, 1984 and October 6, 1999, were registered to an RMS error of less than 0.5 pixel. This RMS error was deemed acceptable at the time of registration.

Even the smallest amount of RMS error has the potential to introduce some degradation to the change detection accuracy. This degradation has the potential to affect the boundaries of the land classes incorporated in the study as spurious differences can be detected because the land surface properties at wrong locations are evaluated instead of the real changes at the same location between one time and another. The spatial resolution of the imagery becomes an important factor in this assessment.

Possible implications of the RMS error obtained for the registration of the imagery employed in the change detection could include land class and change area boundary discrepancies up to approximately 15 meters (0.5 pixels). The exact percentage of change detection error due to misregistration is unknown. The topic of image misregistration for change detection is still a topic of current research.

Transformation was previewed prior to resampling. All 1999 images were resampled to 30 meters to coincide with the 1984 imagery. To verify co-registration, bands 7 of each image date were displayed together in different image planes. Nearest neighbor resampling was selected due to quicker computer processing time as compared to other interpolation methods. In addition, nearest neighbor interpolation better maintains original reflectance values thus is the only resampling method which should be performed before image classification. The first order transformation provided sufficient accuracy and reduced potential introduction of unwanted geometric distortions in areas with no ground control points to provide precise control. Figure 4.1 portrays how the pixel grids in each Landsat image are shifted during resampling procedures.

 Figure 4.1

Subset of Study Area

In some cases, Landsat TM scenes are much larger than a project study area. In these instances it is beneficial to reduce the size of the image file to include only the area of interest. This not only eliminates the extraneous data in the file, but it speeds up processing due to the smaller amount of data to process. This is important when utilizing multiband data such as Landsat TM imagery. This reduction of data is known as subsetting. This process cuts out the preferred study area from the image scene into a smaller more manageable file (ERDAS, 1999). (Figure 4.2)

 Figure 4.2

A Landsat TM image is 115 miles (185 kilometers) wide by 106 miles (170 kilometers) long and has a total area of 12,190 square miles or 31,450 square kilometers.

Carroll County has an area of approximately 641 square miles. In order to subset the study area from each of the four Landsat scenes, a vector file defining the county boundary with the same georeferenced coordinates as the Landsat images, UTM Zone 15, NAD27, was imported into PCI Imageworks. The county boundary vector file was converted to a binary bitmap mask and overlaid on to each of the TM scenes. The county mask acts as a virtual cookie-cutter and subsets the study area similar to Figure 4.2.