|
|
|
Image Processing
|
Geometric Registration Method For 10-Day Composite Avhrr Data For Asian Region
4.2 AVHRR image navigation
A final step of first stage correction is the conversion of each line and sample numbers of raw AVHRR imagery to corresponding earth locations. In this step, a set of GCPs determined by automated image matching, were used for adjustment of orbital parameters and satellite attitude correction. Then the conversion procedure is realized by indirect navigation method using PaNDA package (1,2). However, PaNDA package does not consider elevation effect in image navigation procedure because elevation data were not adapted to the package. The analysis showed that if an elevation is not considered for in the image navigation procedure, registration errors for off-nadir pixels can occur up to 4 pixels in high land areas. Therefore, elevation data created from GTOPO 30 were added to the PaNDA and earth location of each pixel of AVHRR imagery was calculated considering elevation effect in image navigation. Thus, first stage correction results in determining line and sample numbers of satellite image for given latitude and longitude of Plate Carree projection coordinate system. The navigation procedure is also used to compute sun-target geometry angles for each pixel.
5. Second-Stage Correction
Error evaluation of calculated ground locations of raw imagery showed that AVHRR image navigation does not work well over the whole image. Moreover, test composite data produced from the images corrected by only first-stage correction procedure showed that navigation accuracy does not satisfy compositing requirement and it was easily observed that small lakes and rivers were significant blurred in the composite image. Therefore, in order to achieve satisfactory accuracy for multitemporal compositing, it was necessary to register the each navigated image to fixed map overlay image. This reason motivates to develop second stage correction that is carried out in three steps: estimation of navigation accuracy, determination of mapping function and resampling.
5.1 Determination of navigation accuracy
A positional accuracy of the navigated image is determined by the correspondence between the reference overlay image and the navigated image. The correspondence is established by precise matching GCMPs as reference windows with searching windows in Plate Carree projection. The reference window in this step is the land and water surface pattern surrounding the center point of GCP. The GCMPs (Figure 2, A1~F1) described in section 3 are used as reference window in the pattern matching procedure. Simultaneously, an image window of 64 by 64 size centered at the same location as for the GCMP, is produced from raw imagery referring calculated ground location points. Then the image window is used to generate searching window of binary data of "1" for land or "0" for water (Figure 2, A3~F5). Then, the image correlation method described in section 4.1, is applied to match the reference and searching windows. The result of the matching procedure is a set of GCPs locations both in the navigated and reference overlay images. Thus, the set of GCPs selected allows estimating navigation errors and they are used for determination mapping function for final correction.
Figure 2. GCMP's and searching windows used in matching. 1-GCMP's, 2, 3- Ch2 and NDVI image windows extracted from raw AVHRR data, 4,5,6- searching windows extracted by NDVI, filtering and adaptive threshold value respectively.
5.2 Determination mapping function and resampling
The resampling process includes two steps. The first is determination of polynomial mapping function. Applying least square method to the set of GCPs selected in previous step, the polynomial mapping functions of degree two is established and they are formulated as following:
x i = a 0 + a 1*X i
+ a 2*Y i +
a 3*X i*Y i +
a 4* X i2 + a 5*Y i2
y i = b 0 + b 1*X i+
b 2*Y i + b 3*X i*Y i+
b 4* X i2 + b 5*Yi 2
where x i and y i correspond to a point in the navigated image,
X i and Y i correspond to a point in the corrected image. Using the set of GCMPs, the unknown coefficients a 0~a 5 and
b 0~b 5 were determined by minimizing the sum of squared errors using least-squares technique.
Using the mapping functions, the resampling procedure is carried out in the final step. The resampling is performed by transformation of pixel locations in the navigated image to corresponding locations in the resulting output image. In the transformation, nearest neighbour interpolation technique is used.
|
|
|
|
|
|
|
