Remote Sensing Lab 6

Goal and Background

The goal of this lab is to gain familiarity with geometric correction, a part of image preprocessing. In particular, the lab focuses on two types of geometric correction: image-to-map rectification and image-to-image registration.

Geometric correction is the removal of geometric distortion in an image so that individual pixels can be in their proper planimetric position. Internal geometric errors, those inherent in the satellite itself, and external errors, those not influenced by the sensor, are both corrected for by completing geometric correction. In both types of geometric correction ground control points (GCPs) must be obtained. These are identified on both the distorted image and the image or map used to correct the distorted image. GCPs can also be obtained using a GPS system or a total station.

Within the process of geometric correction, there are two basic operations: spatial interpolation and intensity interpolation. The former involves taking the GCPs to create a polynomial equation that rectifies the locations of the pixels in the distorted image. The latter involves relocating the brightness values in the distorted image to the new output image. Finally, the accuracy of geometric correction is determined using the root mean square error (RMS error). This number shows the difference between the distance of the distorted GCP and the same GCP in the output image. Ideally, this number is 0.5 pixels or lower.

Methods

Image-to-Map Rectification

A United States Geological Survey (USGS) 7.5 minute digital raster graphic (DRG) of the Chicago Metropolitan Statistical Area and adjacent regions was used as the reference map to correct a Landsat TM image of the same area. In Erdas Imagine, a first order polynomial geometric model was enabled with DRG set as the reference map.  The next step was to add GCP points (figure 1).

Figure 1. Geometric correction window showing GCPs added to both a distorted image and a reference map (DRG).

In this lab four GCP points were added. The GCPs then had to be repositioned to reduce the RMS error to 2.0 or lower (figure 2). This higher threshold was used because this was the first attempt at creating GCPs.

Figure 2. Geometric correction window showing a GCP on the distorted image (left) being repositioned to lower the overall RMS error.

After the GCPs were corrected, the image was corrected using the GCPs and a 1st order polynomial equation and resampled to produce an output image.

Image-to-Image Registration

For this section of the lab, an image of eastern Sierra Leone, the same view as the distorted image, was used as the reference for geometric correction. In Erdas, a third order polynomial geometric model was enabled, which required 10 GCPs instead of the 3 required in the image-to-map rectification. For good measure, 12 GCPs were recorded. These GCPs were then repositioned to reduce the RMS error to 1.0 or lower, although 0.5 RMS error is ideal.

After the RMS error was brought to an acceptable level, the image was corrected using the GCPs and the third order polynomial equation and then resampled using a bilinear resampling method to produce the corrected image.

Results

Figure 3. Geometrically corrected image using a first order polynomial equation.

In figure 3 the original image of The Chicago Metropolitan Statistical Area and the surrounding area has been corrected using a first order polynomial equation. The DRG for this image correction was a United States Geological Survey (USGS) 7.5 minute topographical map. The GCPs and the polynomial equation completed the spatial interpolation and the nearest neighbor resampling method was used to perform intensity interpolation. The first order polynomial equation was used because there was minimal distortion in the original image, so only small changes were needed. Although only three GCPs were needed, four were collected and dispersed throughout the image to ensure the whole area was geometrically corrected.

Figure 4. Geometrically corrected image using a third order polynomial equation.

In figure 4 the original image of eastern Sierra Leone has been geometrically corrected with a third order polynomial. The correction was done with an already corrected image of the same area and utilized 12 GCPs. The corrected image is fairly accurate. The upper right corner of the image and the whole right side in general is more accurate than the left side. This is because more GCPs were taken on the right side than the left side. The biggest difference between the two images in in the top left corner. This makes sense because that corner has the least amount of GCPs. The lower left-hand corner is also slightly off, but this is less severe than the top left-hand corner. Although the middle of the image does not have many GCPs either, it is corrected due to the presence of GCPs around itself. In this correction the bilinear transformation was used to perform intensity interpolation because it renders a more spatially accurate image.

Both images were geometrically corrected using ground control points (GCPs) but differed in the order of polynomial used to correct the images. The first order polynomial was used on an image with less distortion and the third order polynomial was used on an image with more distortion. Both images could have been corrected further with more GCPs and repositioning of the GCPs.  This lab successfully introduced the concept of geometric correction.

Sources

Satellite images from Earth Resources Observation and Science Center, United States Geological Survey
Digital raster graphic (DRG) from Illinois Geospatial Data Clearing House