COMPUTER MODELLING & NEW TECHNOLOGIES 2014 18(12A) 240-244 Yan Man, Hu Defa
240
Nonlinear Kalman filter phase unwrapping algorithm based on
the terrain
Man Yan
1,2
, Defa Hu
3
1
School of Computer Engineering, Weifang University, Weifang 261061, Shandong, China
2
School of Information Science and Engineering, Shandong University of Science and Technology, Qingdao 266590, Shandong, China
3
School of Computer and Information Engineering, Hunan University of Commerce, Changsha 410205, Hunan, China
Received 1 June 2014, www.cmnt.lv
Abstract
Phase unwrapping is one of the key steps in InSAR data processing. In condition of steep terrain, excessive stripes reduce the
conventional phase unwrapping algorithms’ implement, resulting in error propagation of unwrapping phase. Therefore, this paper
presented a nonlinear Kalman filter phase unwrapping algorithm based on terrain. This algorithm uses local fringe frequency estimation
as input control variable. Chirp-Z transform is added into Fourier transform in local frequency estimation, and this improves the
unwrapping result accuracy. Results obtained with simulated and real data validate effectiveness of proposed method through analyzing
and comparing with Kalman filter method, nonlinear Kalman filter method and quality map guiding method.
Keywords: InSAR, nonlinear Kalman filter, phase unwrapping, Chirp-Z transform, terrain
1 Introduction
One of the important steps in InSAR data processing is
phase unwrapping, its accuracy will directly influence
height measurement. We refer to the book by Ghiglia and
Pritt [1] for an excellent overview. Traditionally, there
have been two general types of conventional phase
unwrapping methods. The algorithms of the first group,
generally named path-following algorithms [2-4]. Those
advantages are that computing speed is fast and demanding
memory is less. These algorithms isolate problematic
zones containing residues and can be unwrapped the
interferogram by avoiding these zones. The techniques of
the second group provide a global solution which
minimizes a cost function over the whole interferogram [5-
8]. Independently from this traditional classification, some
techniques make use of a prefiltering stage before starting
unwrapping procedure with filtered phase, for instance [9].
Conventional Kalman filter algorithm transforms phase
unwrapping into state estimation problem, through the
establishment of phase state space model and vector
observation equation, and Kalman filter is used to unwrap
and filter simultaneously. However, the layover
phenomenon makes discontinuous streaks appear in course
of phase unwrapping; coherence reduction also affects
reliability of DEM, which occurs in the urban or dried area;
in steep terrain, too many stripes reduce phase unwrapping
algorithm’s execution, and this makes error propagation
phenomenon appear in unwrapping results. These
disadvantages will affect accuracy and resolution of image
which generated. But the previous two shortcomings are
difficult to overcome, in this paper nonlinear Kalman filter
Corresponding author’s e-mail: yanmansdust@126.com
phase unwrapping algorithm based on terrain is proposed.
It can be implemented through introduction of input
control variable associated with terrain to the state space
model, then in iterative formula Chirp-Z transform is
added into Fourier transform to estimate local frequency,
accuracy of phase unwrapping is improved.
2 Nonlinear Kalman filter phase unwrapping
algorithm based on terrain
2.1 OBSERVATION EQUATION
Inphase and quadrature components of complex
interferogram are consisted as two noisy observations of
true interferometric phase. Here, substituting again m, n
pixel by k pixel, ϕ(k) represents real value of interfergram
phase at k pixel, observation equation can be written as
follows [10]:
,
)(
)(
))(cos(
))(sin(
)(
)](Re[
)(
)](Im[
)()]([)(
2
1
kv
kv
k
k
ka
kz
ka
kz
kvkhky
(1)
where, z(k) is complex interferogram, a(k) represents
amplitude of interferogram, h(·) represents non-linearized
mapping between y(k) and state vector ϕ(k). v
1
(k) and v
2
(k)
are assumed to be white Gaussian noise, then: