Digital Image Mosaicing
R. Parvathi, V Jayaraman
Earth Observation System Indian Space Research Organisation
Bangalore-560094
S. Kannan
Regional Remote Sensing Service Centre
Introduction
A problem which frequently arises when two or more scenes are mosaiced is the creation of artificial edges at the seams between the images. These artificial edges occur where there are perceptible radiometric differences within the region of overlap. The radiometric differences between images are usually caused by changes in atmospheric and earth surface conditions. Changes in atmospheric transmittance will tend to transform the grey scale over the whole image. Illumination caused by different sun angle, changes caused by human activities, precipitation and seasonal changes of surface multitemporal images. All these factors make it difficult for manual or photographic methods to produce image mosaics without artificial edge.
The Remote Sensing images may be mosaiced to a map projection or a reference image. Inaccuracy and uncertainty in determination of data on spacecraft and position, orientation and dam era attitude data cause registration error. Thus the image should be geometrically corrected and registered prior to mosaicing. The image mosaic process consists of four steps, namely (i) geometric registration of the input images. (ii) a grey level adjustment for each image in order to match their grey level histograms in the overlap region (iii) a seam point identification, at which the corresponding lines in the two images will be jointed, (iv) averaging the degree of mismatch of the grey level in the overlap region over the neighbourhood of the seam point.
Geometric correction is carried out using higher order polynomial transformation model. Linear Histogram Analysis Operation (LHAO) is used to perform radiometric temporal scene normalization. A two dimensional seam point searching algorithm (Yang Shiren et al, 1989) is implemented to select the best seam points. Through nearest neighbourhood interpolation the degree of mismatch of the grey level in the overlap region is averaged over the neighbourhood of the seam point.
Grey – Level Normalisation
If A, B be are the images which are being mosaiced and WA & WB are their respective overlapping windows, ideally, WA and WB are identifical (radiometrically), but in general they will not be identical due to various resons including atmospheric and surface conditions. Atmospheric and sensor conditions tend to operate on the grey level distribution rather than on spatial features and it is appropriate to ask what grey level mapping should be applied to minimize these changes. More specifically, if HW
A, HW
B are the histrograms of W
A, W
B respectively. A and B are transformed (requantized) so that HW
AºHW
B.
In contrast to atmospheric and sensor effects, the ground changes, produce grey level differences that are irreconcilable by histogram remapping. Therefore to define the appropriate grey level transformation one should exclude the regions that occur in only one
of the overlapped images. Another case is where the grey level difference occurs due to ground reflectance changes. By masking the pockets where the changes occur the histogram should be calculated using rest of the image points (Milgram, 1975).
The radiation propagation to a Remote Sensing system may be modeled as: L =
a * R +
b where,
L is the radiance reaching the sensor, R Reflectance of the target
a is a factor including the solar irradiance, atmospheric transmittance and
b is the radiation contribution by the air column.
Each term in the above equation is dependent on the spectral bandwidth and the solar geometry. In terms of temporal imagery, and define the characteristic changes in the illumination and atmospheric conditions, while R defines the characteristic feature changes and L is directly related to the brightness of the features in the image. Given two temporal images, the brightness values of feature within each image are modeled as :
The approaches to temporal image normalization to date involve the independent normalization of the individual images for the time varying effects. These techniques essentially normalize each image by correcting for the specific atmospheric and illuminating conditions on the specific date. The approach is to determine the scene specific values of
ai and
b to perform an independent conversion of each image to values representing surface reflectances, Ri. Studies have shown that the normalization techniques are generally only successful under optimal conditions and are not easily implemented due to the dependence on ground truth and/or characterization of the atmospheric conditions at the time of imaging. Because of these reasons, a simpler and effective approache is established to determine the transformation parameters is established to determine the transformation parameters (
a1,
a2,
b1 and
b2)using brightness distribution from each temporal image.
The method used to derive the normalization transforms in this study is a Linear Histogram Analysis Operation (LHA0). This method assumes that the transformation function is a linear transform and acts to align the central tendency and dispersion at the test distribution with that of the target distribution. Hence, given test distribution with the mean m
2 and standard deviation s2, and a target distribution with mean m
1, and standard deviation, 01, a linear transform, T, is defined as,
If the digital count values are linearly related to radiance, above transformation coefficients are used to define a linear combination of the input values which is applied locally by the equation Y (i, j) = a * X (i, j) + b
