Image Registration and Georeferencing in ENVI (Cavite, Philippines)

Georeferencing and Registration Methodology
The following steps were undertaken for the rectification or georeferencing of the image:

1. I have created a subset of the province of Cavite from the Landsat ETM+image that I have downloaded from GLCF . A topographic map of the equivalent subset has been obtained with the following basemap information.

Fig. 1: Basemap information

Scale 1: 50000..
Sheet name Cavite, Philippines
Sheet number 3163 II

2.  A scanned 1:50000 topographic map of Cavite appears below.
Fig. 2a: 1:50000 Topographic map of Cavite
3.  An Image-to-map registration was performed using the coordinates on the boundary of the map as ground control points. A total of 10 points are obtained with a total RMSE of 0.396853.

4.  An Image-to-Image registration was later performed, yielding 9 GCPs with a total RMS error of 0.970929. The Landsat ETM+ image that I have obtained has already undergone the Level 1G correction hence the warping from the originally downloaded image to the warped image was observed.
Fig. 3a: Distribution of ground control points within the Landsat ETM+ image of Cavite

Fig. 3b: Georeferenced image of Cavite, Philippines

Some Discussions

On topographic maps vs. satellite images
Topographic maps are already orthorectified hence it can be used to measure distances between two points. Name of features are also available on a topographic map. Depending on the level of correction contained on the satellite image, it may or may not be georeferenced or orthorectified. Satellite images show terrain of ground features. Moreover, satellite images are captured at different wavelength bands in the electromagnetic spectrum hence analysis and interpretation can be performed on more than three dimensions. Depending on the field of application or query, one can switch on different color composite diplays on the different bands contained in the satellite images to facilitate interpretation. Scanned topographic maps on the other hand utilizes only the visual RGB bands. Topographic maps are available on different scales depending on the area needed for interpretation. Satellite images, depending on the capabilities of the sensor, are available in different radiometric, spectral and spatial resolutions. One satellite image covers a very large area unlike topographic maps which contains relatively smaller areas hence producing finer details and information on features in it.

On the use of topographic maps for geometric rectification
Use of topographic maps poses some potential problems when used for geometric rectification. Depending on the date of execution of the topographic survey used for the map, information contained in it whether man-made or natural fatures may have already been altered through time. This is crucial for the identification and location of ground control points. Medium used for printing the topographic map is also a factor to be considered. Different materials expand and contract as the temperature of its environment changes. This expansion or contraction causes significant distortions on the features printed in the map. Paper when folded or crampled also causes geometric distortion of the features on the map. When scanned, printer resolution and other abberations caused by the glass on the scanner may also cause geometric distortion on the basemap. Lastly, when ground control points (GCP) are chosen, the interpolation of the corresponding coordinates of the GCPs may also be subjective depending on the discretion of the observer which may cause significant difference on the inetrpolated versus the true ground coordinates of the GCP.

On coordinate transformation
The RMS error measures the errors between the destination control points and the transformed locations of the source control points. A root mean square error is calculated for each transformation performed and indicates how good the derived transformation is. The transformation is derived using least squares, so more GCPs can be given than are necessary. Specifying a minimum of three GCPs are required to produce a transformation that results in an RMS error. The formula for the calculation of RMSE appears below.

To come up with an acceptable degree of spatial accuracy, RMSE should not be greater than 1.

Depending on the method of transformation, the number of minimum number of coordinates needed for transformation may vary. For 2-dimensional conformal coordinate transformation, which preserves the orthogonality of the coordiante systems, a minimum of 2 control points for scale change, rotation and translation are needed for transformation. On the other hand, 2D-Affine transformation which preserves the parallelism of parallel lines, 6 parameters are needed to facilitate transformation hence a minimum of three control points to produce 6 equations with 6 unknowns.

Computation for the parameters of each transformation method.

Using the following formula for conformal and affine transformation, a JAVA program was produced to compute the parameters for transformation.
    Fig 4a: Working equations for the formulation of matrices for 2D Conformal and Affine Transformations
The input file contains the coordinates from the GCP text file of Image-to-Image registration in ENVI. The output file produces the parameters for transformation.

The following table are the input coordinates in solving for the parameters for transformation.

Fig. 4b : Base and warp image coordinates as ground control points
Base Image Warp image
x y x y
2197 467 4288 3435.75
2101 915.25 4255.75 3581.5
1406.25 1293 4027 3709
757.5 1964.5 3806.25 3933.5
377 256.5 3705.16 3376.46
2831.5 1920 4491.25 3914
2869.25 638 4504.5 3487.25
1804 1168.25 4156.94 3667.62
1797.75 797.75 4157.56 3545.68

The following tables contain the resultant parameters for 2D Affine and and 2D Conformal Coordinate Transformation respectively.

Fig. 4c : Parameters for 2D Affine transformation
Parameter Value
a0 3579.92167
a1 0.32503
a2 -0.00779
b0 3291.26455
b1 -0.00444
b2 0.32919

Fig. 4d : Parameters for 2D Conformal transformation
Parameter Value
a1 0.32605
a2 -0.00021
a3 3569.72659
a4 3286.96612

On different transformation and resampling methods in ENVI
Upon application of the three different transformation methods , the resulting images are warped differently. The figure below compares the amount of warp present in the image after each transformation using different resampling methods.

    Fig. 5: Amounts of warping in the Cavite satellite image using RST, 2nd and 3rd order polynomial transformation
    After geometric corrections and translations, resampling is being performed to produce a better estimate of the DN values for individual pixels. In the nearest neighbor algorithm, the transformed pixel takes the value of the closest pixel in the pre-shifted array. In the bilinear interpolation, the average of the DN values of 4 surrounding pixels is used while cubic convolution averages the 16 closest input pixels.
    Images resampled using cubic convolution produces the sharpest image.


Hold-out validation (HOV) uses another set of GCPs for the same image dataset to verify the spatial accuracy of the georeferenced image.

    Fig. 6a: Hold-out Validation using a new set of ground control points of the Cavite Landsat ETM+ Image
Shown above are the relative loaction and ditribution of the new set of ground control points. Nine (9) GCPs which yield a total RMS error of 0.448321.

Leave-one-out Cross-Validation (LOOCV) - Cross-validation uses all of the data to estimate the trend and autocorrelation models. It removes each data location, one at a time, and predicts the associated data value.

Fig.6b : Sample Leave-one-out Cross Validation
The figure above shows how LOOCV works. After choosing the GCPs and minimizing the total RMSE to less than 1, the Image to Image GCP list was arranged such that the point with the largest RMS error appears on top of the list. This point was turned off and effectively the total RMS error of the GCPs lowered from 0.970929 to 0.818812. If we hit the predict button in the Ground Control Points Selection dialog,the cross hair on the zoom window will center on the point that will give the lowest total RMS based on the correlation of points of the image.

On Level 1G correction of Landsat images 
Upon rectification, distortions caused by platform and surface geometric characteristics can't be easily distingushed since the image obtained has already undergone level1G correction which is a format created by NASA to indicate imagery that is basically ready to use .L1G" is indicative of "Level 1G", meaning the data has been processed to level 1 and is radiometrically and geometrically corrected. 

On Image-to-image vs. Image-to-map registration
An image to image registration is a lot easier than image to map registration. You just have to scan the topographic map and georeference it using the graticule values given on the borders of the map. Bias on the map-scaling interpolation is removed because once you georeference the image, the coordinates on the pixel of the desired GCP within the georeferenced image is readily available. The drawback however is on the manner of scanning the topographic map. The resolution of the scanner that will produce the output topo map as well as the current state of topo map upon scanning greatly affects the quality of data one can extract in the scanned map. Folds and cramples or obliterations on the map produces significant distortions on the coordiantes derived from the map.

    1 comment:

    1. Very interesting article, i wonder if you found some similar idea about satellite image interpretation in here http://www.imagesatintl.com/services/satellite-image-analysis/