Image Registration

Image registration is a geometrical transformation that creates a correspondence between image coordinates (row and column) and a reference coordinate system. In the SPRING this reference coordinate system is, in the worst case, the planar coordinate system for a given cartographical projection. As any cartographical projection has a well defined relationship with a geographical coordinate system, it is possible to say that the registration establish a correspondence between image coordinates and geographical coordinates.

Other common terms used for the registration procedure are geocodification and georeferencing. It is important, however, to make a clear distinction between registration and geometric correction. The image geometric correction procedure removes systematical geometrical distortions introduced in the image acquisition step, while the registration just uses simple geometrical transformations - usually polynomial transformations - to establish a mapping between the image coordinates and geographical coordinates. So it is suggested to use the registration as a technique to refine the geometrical quality of images with system geometrical correction.

The registration is a required operation to make the integration of an image to the existing database in a GIS. For years the remote sensing projects assume that the images could be integrated to data extracted from existing maps or to measurements performed directly on the terrain. The registration is also important to combine images from the same area taken with different sensors or to multi-temporal studies, in case the images are taken in distinct periods of time.

Operations in the SPRING registration window


See also:
About Digital Images
About Remote Sensing
Digital Image Processing


Polynomial Transformations - control points

The usage of polynomial transformations is very common for image registration. The polynomial transformations make the relationship between image coordinates and coordinates in the reference system through the control points. Control points are features subject to identification in the image and in the reference system. Road crossings, airport runways and river junctions are natural candidates for control points.

The parameters determination of the polynomial transformation selected is made through the solution of an equation system. In order to build this equation system the coordinates of the control points have to be known in the image and in the reference system. The image coordinates (row, column) are obtained when the user clicks on the image features. The reference coordinates are usually obtained through reliable maps which have the corresponding features used as control points (Table mode in the registration window). The SPRING also accepts measurements performed directly in the field using a GPS (keyboard mode). Existing vectorial data and georeferrenced images can also be used as sources for referenced coordinates (Window mode).

Once the n control points are determined and the polynomial transformation is selected, a system with 2n equations is built to solve 6, 12 or 20 parameters, depending on the polynomial degree (1o, 2o or 3o degree). Thus, it is possible to conclude that the minimum number of control points is 3 for the 1o degree polynomial, 6 for the 2o degree polynomial and 10 for the 3o degree polynomial (see the equations for the 1o and 2o degree polynomials in the figure below).

registro_esq.gif - 6063 Bytes


The minimum number of control points represents the situation of a certain equation system, in which the number of equations is equal the number of unknown variables. However, as the measured coordinates of the control points are subject to errors, it is suggested to use a number of control points higher than the minimum. In this case we have an over-determined equation system, which has more equations than unknown variables and allows to distribute the control points measurements errors. In practical terms it is advised to use 6 control points for the 1o degree polynomial, 10 control points for the 2o degree polynomial and 14 control points for the 3o degree polynomial.

One should be aware that the control points distribution in the area to be registered is very important, because the polynomial transformations tend to behave adequately only inside the region where the control points are defined.

Registration


Images with system geometrical correction

The SPRING can recognize satellite images with system geometrical correction and treat them in a special way during registration. It is the case for TM images from LANDSAT-5, ETM from LANDSAT-7, CCD from CBERS-2 and from high resolution images available in the GEOTIFF format. The system correction is based on physical parameters usage inherent to the acquisition situation for each image (ephemerid and platform attitude, sensor imaging system and Earth reference model). As a result an image is obtained where the pixels already observe a relative positioning related to a certain cartographical projection system, where, usually, a residual translation is required because of the uncertainty in the satellite positioning. In this kind of image the internal geometry is almost always well solved and it does not require being modeled using polynomial transformations.

A potential advantage when treating specially images with system correction is to use fewer control points (in this case a single point would be enough) to refine the residual translation. Another advantage is related to the fact that the control points do not need to be well distributed by the whole area to be registered.

When an image with system correction is read by the IMPIMA module, it is generated a file in the GRIB format which allows the system correction option to be enabled in the registration step. Thus, the user can treat in a special mode an image with a correction system (system correction button activated), but it is also possible to treat an image in the conventional way, if desired (system correction button disabled). This functionality is present in the Importing GRIB files interface.

With the system correction option activated, the SPRING access, in the GRIB file, the navigation equation (relationship between image coordinates and projection coordinates) from the system correction and refines the translations (in X and Y) of this equation while the control points are acquired.

NOTICE: It is emphasized here that besides the fact that this option works with a single control point, it is possible to use all the good control points acquired.

Registration

 


CBERS-2 Images with system geometrical correction

The CBERS-2 images selected through the INPE images catalog (www.dgi.inpe.br/CDSR) has a system geometrical correction and then, subject to uncertainties from the ephemerid data and attitude used in the geometrical correction process. As the CBERS-2 images have been distributed freely to the Brazilian users community, it is suitable to emphasize the actual state of the geometrical quality of these images and the description of the best way to treat them through the SPRING registration module with the purpose of eliminating the positioning error and refine the internal error.

The precision in the positioning defines how much an image with system correction is out of its correct geographical position. The CBERS-2 images can present positioning errors up to 10 Km. Imprecise ephemerid data and proximity in the attitude data integration made on board of the satellite are responsible for the positioning error of images with system correction. The image registration removes the positioning error. However, even if the positioning error was just a few hundred meters the registration would be required.

The internal precision establish the integration possibility from an image with system correction to maps and other georeferrenced data. In other words, the internal error is the residual error, the one which is not completely removed when an image is superposed on a map, that is, a small internal error means a good superposition. The internal error is about 90 m (4.5 pixels) for the CCD images, 250 m (3.125 pixels) for the IRMSS and 700 m (2.7 pixels) for the WFI images. Consequently, if the system correction option is active, this internal error values remain in the registered image. In the other hand, tests performed at INPE shows that image registration through an affine transformation (1o degree polynomial) allows the internal error refinement, which is reduced to 24 m (1.2 pixels) for CCD images, 112 m (1.4 pixels) for IRMSS images and 416 m (1.6 pixels) for WFI images. Thus, if the registration is performed with the system correction button deactivated and there are more than three control points for using a 1o degree polynomial, it is possible to get a good result. Thus, it is suggested, as a general rule for CBERS-2 images, that the Registration and the Importing of GRIB Files is performed with the system correction button deactivated.

 

Registration

 


Resampling by interpolation

To compute the new gray level value in the image to be registered, the SPRING adopts as interpolation method the techniques known as bilinear interpolation and nearest neighbor allocation.

The nearest neighbor allocation interpolator attributes a gray level value to the pixel of the corrected image, equals the gray level value of the closest pixel to that position. There is no alteration in the gray level value. Because of this characteristic, it can be applied in images where there is not too much heterogeneity in the gray level values.

The bilinear interpolator computes the gray level of the pixel in the corrected image using the 4 pixel neighbors. As a result, the gray level of the pixel is changed by the values of its neighbors. It can be applied in images where there is a considerable heterogeneity in the gray level values of the pixels.

Check the procedures to make a georeferrencing (registration) of an image. The image has to be in the Grid Binary (GRIB) format and stored in the disk. The GRIB format is generated by the "Impima", the module for reading images.

NOTICE: The registration of images stored in other formats, such as TIFF, GEOTIFF and RAW, requires first a conversion to the GRIB format, which can be done using the "Impima" module.

Registration


Registration in the SPRING

The georeferrencing can be performed using the cartographical parameters in an active project (Window mode). If there is no active project the control points coordinates can be informed using the digitizing table (Table mode) or by computer keyboard (keyboard mode). See details for each mode:

TABLE: In the table mode the user needs to have a map (topographical chart of the same image area being considered). This map has to be calibrated in a digitizing table. As it is not required to activate a project (it is just enough to have an active database), the system requests the type of projection to be used in the registration.

WINDOW: In this mode the user can use an Information Layer in an active project. This IL can be an image that is already georeferrenced or any other map, as a thematic IL that has some identifiable features in the image.

KEYBOARD: In the keyboard case it is not required to have an active project (it is just enough to have an active database). Because of this, it is required to inform using planar or geographical coordinates, which can be directly measured in the terrain using, for instance, a GPS.

See all the procedures to register an image.

Registration


Image Registration and Mosaicking

When the area covered by an image is smaller than the project area, it is required to make a mosaicking with adjacent images.

An image mosaic is just a collage process of adjacent images to have a larger coverage of a given area.

The registration between these two images compounding a single information layer in the mosaicking process has to be precise enough so that in the superposed region there are no gaps or distortions in the continuity of geographical features. The error in pixels has to be the smallest possible.

Mosaicking with TM-Landsat images

The SPRING can make an automatic cut in TM-Landsat images, for complete scenes (full frame), not displaced from its orbit. This cut allows the maximum reduction of the superposed area among neighbor scenes, using the nominal area from a WRS (Worldwide Reference System). Just click on the Cut WRS button in the "Importing GRIB Files" window.

IMPORTANT: This resource only works for whole scenes of the TM-Landsat.


See how to execute mosaicking with two images.

About Registration


Logical sequence to perform a registration

Next it is presented an example for a logical sequence that the user has to follow to georeferrence an image using a map placed in a digitizing table. The example assumes there is a Database and a Project.

Step

Dialog Box or Menu

Result

1 - Image Acquisition:
     Read image or
     Import image (Tiff)


Module IMPIMA

Import...


GRIB format image

2 - Activate the Database and the Project that will receive an image

Database and

Projects

Active Database and Project.

3 - Map calibration at the digitizing table

Calibration

Calibrated map

4 - Load image in Window 5.

Select an image

Image for registration in the window.

5 - Get control points

Image Registration

Several control points.

6 - Select points for registration and error analysis.

Image Registration

The best selected points.

7 - Save image for registration.

Image Registration

Image in the GRIB format with registration parameters

8 - Import image for the project.

Import GRIB Files.

Registered Image in the GRIB format

NOTE: If the user gets control points using the Window mode, substitute step 3 by "Select an information layer as a reference in the main Window".

NOTE: If the user gets control points using the Keyboard mode, do not consider step 3.

About Registration



See also:
About Remote Sensing
Digital Image Processing