Kalman
Kalman
Kalman
This paper is organized as follows. An overview of the ScanSAR mode and the modified
Kalman filer for scalloping removal are presented in Section II. Analysis of the Kalman
filter's limitation is analyzed in Section III, followed by the proposed method in Section
IV. Experimental results for scalloping suppression on GF-3 and TerraSAR images are
provided in Section V and conclusions are drawn in Section VI.
2. A. ScanSAR Model
In ScanSAR mode, azimuth resolution is traded off for swath coverage. To obtain a wide-
swath in range, the antenna elevation beam of ScanSAR mode is steered to different
contiguous range sub-swaths periodically, as shown in Fig. 2. As a result, for each sub-
swath, the dwell time is reduced compared to the stripmap mode, and the scene is
illuminated by the antenna beam for a short time interval, which is dependent on the
azimuth resolution and the number of sub-swaths [12]. Note that each target is
illuminated with only a small portion of the azimuth antenna pattern (AAP), and targets
located in different azimuth positions will be weighted by different portions of the AAP,
resulting in non-stationarity of amplitudes in azimuth. Consequently, the scalloping
effect appears after image formation processing, as illustrated in Fig. 3.
Fig. 2. ScanSAR acquisition geometry.
Fig. 3. Scalloping in a GF-3 ScanSAR image for one sub-swath.
3. B. Removal of Scalloping
Maximal Kalman filter may help reduce the noising effect during picture focusing,
however the noising effect still has a major impact on image quality because of
inaccuracy and error pattern. Therefore, a noisied picture model is provided by, to
characterise the noise-reduction effect.
The correction of error range is often carried out during the process of picture focusing,
resulting in little variance in the error range. In order to maintain simplicity and
preserve generality, this work assumes that the intensity of error remains constant
throughout the range direction. Hence, for each given frame, equation (1) may be
reformulated as
In the context of a discrete process, the estimation of the error value is often performed
using the stochastic difference equation shown below.
where x represents the linear state of PS , A is the linear data transition matrix, ω is the
processed noise.
z l=H x l+ v l (4)
In the given context, z is the observed state of PS, H represents the observation gain
matrix, and v represents the observation noise.
Typically, it is assumed that they are independent and follow a white Gaussian
distribution, that is,
p(ω )∼ N (0 , Q)
(5)
p(v)∼ N (0 , R)
where p(∙) is the probability density function is denoted as Q, which represents the
variance of system noise, whereas R represents the variance of observation noise.
Considering the information presented in (2), the variable PS is the quantity that has to
be approximated. Given that OS is a constant, it remains in a fixed state, resulting in
parameter A being equal to 1 in equation (3).The variable I 0 (b) ,b ∈[1 , B] , where b
belongs to the interval [1,B], represents the observation noise in equation (4). In this
estimate, the pixel value of the interfering picture serves as the measurement value of the
system, with no additional gain. As a result, the value of H in equation (4) is equal to 1.
Based on the aforementioned model, it is possible to express the prior estimate
covariance and posterior estimate covariance of the system.
¿ (6)
−¿¿
where x́ l and x́ l are the k th is the previous estimate and posterior estimate refer to the
initial and updated estimations of the system state, respectively. The time update
equations and state update equations of the maximal Kalman filter are derived as
follows.
Prediction equation:
¿ (7)
Update equation:
¿ (8)
where K l The blending factor that minimizes the covariance of the a posteriori error is
determined by the covariances of the process noise (Q) and observation noise (R), with
the unity matrix (I) also playing a role.
The maximal Kalman filter, when used iteratively, demonstrates the ability to effectively
estimate the intensity of error in each column. This estimation may then be utilized to
remove the estimated error from each column, so achieving error suppression.
It is known that the derivation of the Kalman filter is based on the assumption of an
independent white Gaussian noise. From the previous section, the real amplitude value
of image is considered as the observation noise when estimating the intensity of the
scalloping. However, the distribution of SAR image amplitude is subject to the Rayleigh
distribution, and the correlation between pixels makes the distribution not independent
anymore. Fortunately, experimental results show that although the statistical
distribution of the image does not satisfy the independent requirement or Gaussian
distribution, the Kalman filter can still achieve accurate results, as long as the statistical
distribution of each column does not deviate far from the Gaussian distribution.
J−B=¿
[ ]
n 3/ 2
1
∑
n i=1
❑ ( xi −x́ )
2
n
1
∑ ❑ ( x i−x́ )4
n i=1
K=
[ ]
n 2
1
∑
n i=1
❑ ( x i−x́ )
2