Before analysis on remotely sensed images can begin,
the images need to be prepossessed to ensure planimetric accuracy of the image and thus the biophysical and sociocultural information which the image contains. There are two broad types of errors, Internal Errors and External
Errors. Internal errors are systematic and stem from an error with the sensor
itself. External errors are non-systematic and usually stem from the atmosphere.
Internal errors are categorized into three broad types, Image Offset, scanning
system induced variation, and scanning system one dimensional relief
displacement. Due to the systematic nature of internal errors, most of these
errors are predictable and can be corrected by the data vendor before the data is given to the
analyst.
However, external errors are categorized into two types, altitude changes and attitude changes. Altitude changes are caused by gradual
changes in altitude of a remote sensing platform as it collects data. Increases
in altitude cause a decrease in spatial resolution (smaller scale) in the
resulting image. While decreases in altitude cause an increase in spatial
resolution (larger scale) in the resulting image. Attitude changes are caused
by changes in roll, pitch and yaw.
Here is a short video explain Roll pitch and yaw in an
aircraft.
As a plane flies through the air, various atmospheric
conditions cause a plane to alter its roll, pitch and yaw so that the plan can
stay airborne. However, these changes alter the sensors orientation and cause
changes in data collection. While gyro stabilization is employed in aircraft
remote sensing, and can compensate for changes in roll and pitch a gyro
stabilizer does not account for errors in yaw. External Errors are
non-systematic and need to be corrected by tools or ancillary data to be in a
state where accurate information can be derived from the image.
The goal of this lab is to understand how to correct
for these external errors in remotely sensed images using ERDAS. The method employed will
be Image to Map Rectification and Image to Image Rectification, where the correction takes place via ground
control points and polynomial equations. Ground Control Points (GCPs) are
placed in both uncorrected image that has errors and simultaneously placed on
the reference image in the same location. A reference image may be a known
corrected image or an accurate map.
The GCPs are placed systematically using known
features for reference, such as intersections, airports, or road cross
sections. It is important to note that GCPs can never be placed on lakes or
water bodies, as well as vegetation, as these areas are too variable and change
frequently to be relied upon to be used as GCPs. The number of GCPs to use is
dependent on the size of the imagery as well as the extent of the distortion,
as well as the order of polynomial used in the correction process. A higher
order polynomial outputs a more robustly corrected image, however needs a
higher number of GCPs in order to compute the correction. Once the minimum GCPs
are used (it is recommended to use the number of GCPS at 1.5 times the minimum
for greater accuracy), we need to make sure that we have a very small Root Mean
Square Error (RMSE), which is a measure of accuracy in the placement of the
GCPs between the uncorrected image and the reference image. The RMSE varies
based on the order of polynomial used in the image correction, but generally
needs to be lower than 0.5.
The procedure for placing GCPs is as follows:
1. Locate
candidate points and collect GCPS
2. Compute
and test transformation
3. Create
an output image file with new coordinate information, pixels are resampled in
the process.
The first image that needs to be corrected is a Digital Raster Graphic of the Chicago area from the United State Geological Survey, and the reference image will be in
the form of a map of Chicago. A first order polynomial was used for this model,
so a minimum of 3 GCPs was needed. The total RMS error acceptable was anything
under 2.0 but ideally the RMS should be under 0.5.
 |
| Figure 1. Image to Map Rectification. The uncorrected image of Chicago is in the left window, while the accurate map is in the right window. Once the minimum number of GCPs are placed the total RMS control point error appears in the lower right hand corner of the screen (Control Point Error). |
Once the GCPs were placed the geometric correction
model was ran with a nearest neighbor resampling technique.
The next image that
needs to be corrected is of Sierra Leone, the photo is not in the correct planimetrically (X,Y) position as you can see in the picture below. The top image is of the
uncorrected photo, and the bottom images is of the reference photo, which is
the second method of geometric correction, or Image to Image rectification.
 |
| Figure 2. Image to Image Rectification. Here the uncorrected and correct images are stacked on top of each other. One can note the error in planimetric (XY) space that the uncorrected image (top) takes as it does not line up with the corrected image (bottom). |
For this model a third order polynomial was used,
which required a minimum of 10 GCP’s, again we used more than the minimum, and
used 12 GCPs. This time the RMS needed to be below 0.5, which was achieved by
fine tuning the placement of the GCPs used in the calculation.
 |
| Figure 3. Image to Image Rectification of Sierra Leone. When using a third order polynomial the minimum GCPs increase to 10. |
For the Chicago Image we can see both the uncorrected
and corrected images side by side (Figure 4). The uncorrected is on the left hand side and
the corrected is on the right hand side. While the images look similar, the corrected image has under gone the nearest neighbor interpolation method, which has changed the cell values in the corrected image to the closest cell value in the uncorrected image, if the two were to be overlapped. Its also important to note that a first order polynomial correction model is not as robust as higher order polynomial corrections and is used for images that have minimal inaccuracies, so we would not expect to see a drastic change in the Chicago DRG once it has been corrected.
 |
| Figure 4. Comparison of the the original uncorrected image of Chicago (left) and the corrected image (right). |
The Sierra Leone image on the other hand, had a much greater degree of inaccuracy than the Chicago Image (Figure 2). Therefore, the Sierra Leone image underwent a higher polynomial correction model in the form of a third order polynomial model, After the image to image geometric correction, Sierra Leone is now rectified (Figure 5), and in the correct planimetric position (x,y), and can
now be used for analysis. Bilinear Interpolation was used in this model which
did result in the output appearing to be hazy. While the haze may affect
analysis, the haze could be corrected by choosing a different interpolation model
or by using a tool to reduce haze. In this instance haze is not as important as
correcting the planimetric position of the original data.
 |
| Figure 5. Comparison of the reference image for Sierra Leone (left) and the corrected image for Sierra Leone (right). The corrected image now lines up planimetrically with reference image. |
Sources
Illinois Geospatial Data
Clearing House [USGS, 7.5 Minute Digital Raster Graphic (DRG)].
Retrieved November 16, 2016, from https://clearinghouse.isgs.illinois.edu/
USA | Earth Resources Observation and Science (EROS)
Center. (Landsat TM Image for Eastern Sierra Leone). Retrieved November 16,
2016, from http://eros.usgs.gov/usa
YouTube. Airplane control - Roll, Pitch, Yaw. Retrieved November 16, 2016, from http://youtu.be/pQ24NtnaLl8
