2 Method
2-1 Top-hat transformation
Because the remotely sensed image is the snapshot of ocean surface when the satellite passes over the sea, the wave pattern in the image is disturbed by the variations in scene illumination conditions, the change of water color and other effects. To correctly compute wave length and direction, the wave pattern must be separated from other phenomenon in the image. Top-hat transformation [1] is used to extract wave pattern from background, the original gray scale image f is first opened by a cylinder structuring element B with a radius of 10 pixels. The resultant image is then subtracted from the original, producing a different image Y which retains wave information. The process can be expressed as follow.
Y = f - (f° B) (1)
2.2 Probabilistic Relaxation
Since the wave information contained in the difference image is, to some extent, obscure, it is difficult to determine the wave length and direction from its spectrum. Thus, a probabilistic relaxation scheme is used to classify the wave information contained in the different image.
Probabilistic relaxation is an iterative algorithm to reduce the ambiguity in local pixel assignment by means of the contextual information. In probabilistic relaxation, must be defined first. Several methods have been reported for defining the compatibility coefficient process and parameter settings are based on the scheme proposed by Danker et al [4]. Since the relaxation process can converge to a good classification result at the early iteration of the process [5], the process is controlled to iterate ten times for the reason of efficiency.
2-2 Automatic thresholding
After the probabilistic relaxation classification process, the histogram of wave pattern image has been transformed to bimodal. It is easy to obtain a binary image by using a thresholding method. The algorithm based on the moment-preserving concept proposed by Tsai [6] was then applied to obtain a binary image.
2-3 Extend Image
Obviously, If the resolution in frequency domain is not high enough to discriminate the change of wave length and direction, it is impossible to identify the variation of these Parameters. To get higher resolution in frequency domain, and to compute wave parameters more a accurately, the binary wave-pattern image is placed is placed at the center of a 51.2 x 512 black image (a image with gray value 0 only). This extended image is then used as input of a two -dimensional fast fourier transformation algorithm. The reason behind this is that the frequency domain resolution
Df and the spatial domain resolution
Dx is related by (2)
Where N is the sample size per line.
2-4 Power spectrum rank filtering
Generally, the main surface waves must have large support in the frequency domain, while the support of noise in the frequency domain is small. Consequently, to reduce the impact of image extension, extension, a rank filter is designed to remove noise from the spectrum. The principle of designed rank filtering the fit of a probe in a shape, it is superior to the traditional morphological filter in the sense of noise insensitively. The rank filter in this research is represented as follows.
where Y(x) and X(x) represent the filtered and the original power of point x in the spectrum, respectively. L = ||A||, A= x|X(x-u) > mean of X,
"u i
eB. B is a disk mask and T is a threshold value.
2-5 Power spectrum band -pass filtering
For most surface waves, the value of wave length is always in the range of [0.05m, 500m] (excluding capillary wave)]. This suggests that a band-pass filtering may be used to remove the undesirable power in the frequency domain. Let x = (n
x,n
y) denote a point in the spectrum (where n
x and n
y represent x and y coordinate of x in the spectrum, respectively.), the band-pass filter used here can be represent as follows.
