Determining the best possible solution for the math model is the foundation of your project. Therefore, it is important for you to know if your solution is good enough to achieve the results you expect. If it is not, you must also know what to do to adjust the model.
The residual errors will help you determine if the solution is good enough for your project. Residual errors are the difference between the coordinates that you entered for the ground control points (GCP) or tie points (TP) and where those points are according to the computed math model.
Residual errors do not necessarily reflect errors in the GCPs or TPs, but rather the overall quality of the math model. That is, residual errors are not necessarily mistakes that need to be corrected. They may indicate bad points, but, generally, they simply indicate how well the computed math model fits the ground control system.
By using the tools in the Residual Report window, you can test the quality of the math model and the effect various points have on the model by switching the points between active and inactive. You can do this for both GCPs and TPs. Inactive points are not used to compute the math model, so you can deactivate points and see whether or not it improves the model. With TPs, you can also deactivate the point for a single image to see if it improves the overall error values.
Note: In projects of types Rational Function, Polynomial, and Thin Plate Spline, images are not connected together with TPs. Therefore, the math model and the resulting residual errors are calculated for each image separately. If you selected the Thin Plate Spline math model for your project, the residual errors will always indicate zero. To check its accuracy, use check points.
Another way to verify the quality of the model is to collect some GCPs as check points. Check points are not used to compute the math model, but OrthoEngine calculates the difference between their position and the position determined by the model, and includes the error in the Residual Errors report. Therefore, check points provide an independent accuracy assessment of the math model.
In most projects, your objective should be for the residual errors to be one pixel or less. However, you should also consider how the resolution of the image, the accuracy of your ground control source, and the compatibility between your ground control source and the images can affect the residual errors.
- You may want to use a topographic map as a ground control source; however, features on topographic maps can sometimes, for aesthetics, be shifted several meters. This limits the accuracy of the coordinates that you can obtain from the map. Also, the detail visible on a 1:50,000 scale topographic map may not be compatible with the high resolution of an aerial photograph. For example, if you choose a road intersection in a topographic map as your coordinate, the same road intersection in the aerial photograph may consist of several pixels. Therefore, the residual error will likely be larger than a pixel.
- An existing LANDSAT orthorectified image may make a convenient ground control source for registering a new IKONOS image, but the resolution of the LANDSAT image is 30 meters and the resolution of the IKONOS raw image is 1 meter. Therefore, even if you could pick the right pixel in the IKONOS image, your GCP from the LANDSAT image is only accurate to 30 meters. You cannot achieve accuracy of 2 to 4 meters unless your ground control source is equally accurate.
- At first glance, a residual error of 250 meters in ground distance may appear too high. However, if your raw data has a resolution of 1000 meters, such as AVHRR, you have already achieved sub-pixel accuracy.