You are on page 1of 17

Chinese Journal of Electronics

Vol.32, No.5, Sept. 2023

A Novel Adaptive InSAR Phase Filtering


Method Based on Complexity Factors
XU Huaping1, WANG Yuan1, LI Chunsheng1, ZENG Guobing1,
LI Shuo2, LI Shuang3, and REN Chong4
(1. School of Electronic and Information Engineering, Beihang University, Beijing 100191, China)
(2. Nanjing Research Institute of Electronics Technology, Nanjing 211113, China)
(3. Beijing Institute of Radio Measurement, Beijing 100191, China)
(4. China Academy of Launch Vehicle Technology, Beijing 100076, China)

Abstract — Phase filtering is an essential step in in- to the existence of thermal noise and various decorrela-
terferometric synthetic aperture radar (InSAR) imaging.
tion factors, which increase the difficulty of phase un-
For interferograms of complicated and changeable terrain,
the increasing resolution of InSAR images makes it even
wrapping and ultimately affects the precision of DEM
more difficult. In this paper, a novel adaptive InSAR and deformation reconstruction [3]. Thus, phase filter-
phase filtering method based on complexity factors is pro- ing is crucial for improving the quality of SAR inter-
posed. Firstly, three complexity factors based on the noise ferograms before phase unwrapping [4]. The first and
distribution and terrain slope information of the interfero- foremost objective of InSAR phase filtering is suppress-
gram are selected. The complexity indicator composed of
ing the noise as much as possible while preserving fringe
three complexity factors is used to guide the adaptive se-
lection of the most suitable and effective filtering details adaptively.
strategies for different areas. Then, the complexity scalar With the rapid development of InSAR, a large
is calculated, which can guide the adaptive local fringe number of phase filtering methods have been proposed
frequency estimation and adaptive parameters calculation by researchers, which can be generally divided into spa-
in different filter methods. Finally, validations are per-
tial domain filtering and transform domain filtering.
formed on the simulated and real data. The performance
comparison between the other three representative phase
In the spatial domain filtering, the simplest filter-
filtering method and the proposed method have validated ing methods are the mean filter and the median filter.
the effectiveness and superiority of the proposed method. Early spatial domain filtering methods adopted fixed fil-
Key words — Interferometric synthetic aperture ter window [5], so they are poor in fringe detail preser-
radar, Adaptive, Phase filtering, Complexity Factors, vation. Lee et al. proposed a filtering method that ad-
Local fringe frequency (LFF).
aptively adjusts the size and orientation of the filter
window to maintain a balance between denoising and
fringe detail preservation [6]. However, its limited 16
window orientations may cause distortion to complic-
I. Introduction ated fringes. In view of the problems in Lee’s filtering
Synthetic aperture radar (SAR) is one of the most method, several papers propose improvements to the
vital technology in the field of microwave remote sens- Lee’s filter in terms of directional window [7], fringe fre-
ing. Interferometric synthetic aperture radar (InSAR) is quency [8], and phase gradient [9]. The aforementioned
an important branch of SAR. It provides the digital el- filtering methods are based on interferometric phase it-
evation model (DEM) and earth surface deformation self. In 2006, Vasile et al. proposed an intensity driven
[1], [2], which have been successfully and widely used in adaptive neighborhood (IDAN) method [10], which de-
civil and scientific research fields. tects homogeneous pixels in the neighborhood by SAR
Interferometric phase noise cannot be avoided due intensity information, to assist the interferometric phase

Manuscript Received Aug. 7, 2021; Accepted July 1, 2022. This work was supported by the Shanghai Aerospace Science and
Technology Innovation Fund (SAST2019-026).
© 2023 Chinese Institute of Electronics. DOI:10.23919/cje.2021.00.280
1090 Chinese Journal of Electronics 2023

filtering. However, its performance will be undermined and wavelet coefficients [24]. Bioucas-Dias et al. pro-
when there are not enough homogeneous pixels in the posed the phase estimation using adaptive regulariza-
neighborhood or heterogeneous pixels are included. tion based on local smoothing (PEARLS) algorithm
After that, a complex Markov random field (CMRF) fil- [25], which adaptively determines the polynomial fit-
ter, which estimates the phase by minimizing the local ting window size by the intersection of confidence inter-
energy function in the window, was proposed [11], [12]. vals (ICI) algorithm.
To adapt to the changes of fringe pattern, Li et al. pro- In Goldstein’s method, noise is suppressed based on
posed a variable window CMRF filter [13]. the different spectrum characteristics between useful
The spatial domain filtering methods introduced signal and the noise. However, the filter parameter α
above are all based on locally adjacent pixels. In the takes a fixed value from [0, 1], and α = 0 means no fil-
early days, due to the limitation of computing re- tering applied. With the increase of α, the smoothing ef-
sources, the efficiency of an algorithm was the most fect is enhanced but the fringe details may be damaged
concerned index. In recent years, with the improve- [26]. Therefore, to balance the fringe details preserva-
ment of computer technology, non-local phase filtering tion and noise suppression, efforts have been made by
methods have been developing rapidly [14], [15], so as researchers to improve the selection of α. Baran et al.
to overcome the constraint of estimating the phase in a associate α with the coherence, and at the same time
local window. With phase similarity calculated in a define the smoothing operator as the convolution with
matching window, weighted averaging of similar pixels the mean kernel function [27]. Li et al. proposed that
is performed, therefore non-local phase filtering can re- the noise standard deviation can be used for adaptive
duce noise while preserving structures [16], [17]. Buades selection of α [28]. These improved Goldstein filtering
et al. first proposed a non-local mean filtering method methods using coherence or noise standard deviation
[18]. Then, Deledalle et al. applied the non-local idea to are dependent on coherence, which is usually a biased
interferometric phase filtering based on the statistical estimation. In order to reduce the impact of biased co-
characteristics of InSAR data and proposed the non-loc- herence on the filtered phase, Zhao et al. proposed a
al InSAR (NL-InSAR) filtering method [19]. The patch pseudo coherence to determine α [29]. In addition, Sun
size is adaptively selected based on the heterogeneity of et al. proposed to use the signal-to-noise ratio defined
local scenes. Li et al. proposed an improved non-local by the noise variance as a criterion to selection α [30].
filter that uses a normalized probability density func- To overcome the subjectivity of the selection of α, Song
tion to measure the similarity between the center pixel et al. proposed two improved Goldstein filter methods
and the remaining pixels in the matching window [20]. based on empirical mode decomposition and adaptive
However, the interferometric fringes are dense in steep neighborhood respectively [31], [32]. Generally, there
terrain areas, it is difficult to select similar pixels, which tend to be more noise in the dense fringe area, which
in turn diminishes noise reduction. Therefore, introdu- increases the difficulty of filtering. Therefore, a Gold-
cing the fringe frequency compensation technology to stein filter method based on local fringe frequency com-
NL-InSAR can reduce fringe density and increase the pensation was proposed in [26]. However, fringe fre-
number of similar pixels in the search window, which is quency estimation in both noisy phase and residual
conducive to obtaining more reliable noise suppression phase makes it quite time-consuming.
results. It can be seen that neither non-local filtering nor
The transform domain filtering method rely on Goldstein filtering can perform well in dense fringes of
transforming the interferometric phase from the spatial steep terrain. Facing the tradeoff between noise sup-
domain to the frequency domain or the wavelet domain pression and fringe detail preservation, researchers com-
for filtering, and then transforming filtered results back bine classic filtering methods with the fringe frequency
to the spatial domain. In 1998, Goldstein et al. pro- compensation [26], [33]. The essence of local fringe fre-
posed a classic frequency domain filtering method [21], quency (LFF) compensation is to estimate the LFF of
which achieved noise reduction by smoothing the fre- the noisy phase [34], execute the filtering in residual
quency spectrum. Besides the frequency domain, also phase after removing LFF, and add back the removed
the wavelet domain has been considered for phase filter- phase to the filtered residual phase. Trouve et al. pro-
ing. A complex wavelet interferometric phase filter posed a filtering algorithm based on LFF estimation to
(WInPF) is implemented in [22]. The useful signals are enhance the fringe preserving ability [33], [35]. The lim-
extracted and amplified by utilizing discrete wavelet itation of this algorithm lies in its fixed window size
transform. In the wavelet domain, the phase informa- and linearity presumption of LFF. Aiming at this prob-
tion and noise are easier to separate [23], however, lem, Cai et al. proposed that the size and shape of LFF
which are highly depend on the wavelet decomposition estimation window should be adaptively determined by
A Novel Adaptive InSAR Phase Filtering Method Based on Complexity Factors 1091

the coherent accumulation principle, which is rather novel adaptive phase filtering method based on com-
time-consuming [36]. Suo et al. proposed a high-order plexity factors is presented in detail in Section II. The
LFF estimation method based on weighted least squares proposed method is tested on both simulated and real
phase unwrapping [37]. This algorithm breaks the lin- SAR data, where the experimental results are com-
earity presumption and behaves better in preserving pared with those of the slope adaptive filtering, im-
phase details. However, its pre-filtering and phase un- proved Goldstein filtering, and improved NL-InSAR fil-
wrapping process may introduce additional operation tering methods in Section III. The conclusions are
and errors. From the above analysis, the improved drawn in Section IV.
methods for LFF exhibit high time complexity and can-
not perform adaptive LFF estimation for fringe fre- II. Adaptive InSAR Phase Filtering
quency compensation. Method Based on Complexity Factors
For areas with sparse fringes in the interferogram, Each filtering method has certain limitations and
LFF compensation is unnecessary for noise suppression. unique applicability. Usually, a specific filtering meth-
However, the dense fringe area needs additional LFF es- od is selected according to the characteristics of the en-
timation operation, and the filtering methods based on tire image. However, different areas in the interferomet-
LFF compensation are time-expensive. In addition, the ric phase image usually exhibit different interferometric
main drawback of the existing LFF estimation methods fringe patterns due to different terrain scene. For ex-
is that the linear or nonlinear LFF cannot be estimated
ample, there are sparse and dense fringes in complic-
adaptively based on the terrain slope, resulting in large
ated and changeable terrain. For sparse fringe, most fil-
residual LFF estimation error or time consumption. On
tering methods can obtain satisfactory filtering result.
the other hand, there are heavy noises in low-correla-
So, the filtering methods with higher calculation effi-
tion area, where needs stronger filtering strength.
ciency is preferred. For the dense fringe, the filtering
However, the high-correlation area needs to reduce the
becomes more difficult, taking into account the fringe
filter strength to prevent the fringe details from being
detail preservation and noise suppression. It is difficult
damaged. For an interferogram, it is difficult to achieve
for a single filtering strategy to be the optimal filtering
high-performance noise suppression efficiently with a
for all areas when both filtering performance and calcu-
single phase filtering methods. The complicated and
lation efficiency are required at the same time.
changeable terrain in the interferogram will pose a huge
Therefore, a novel adaptive InSAR phase filtering
challenge on these non-adaptive phase filtering meth-
method based on complexity factors is proposed. The
ods. Therefore, to deal with these problems, it is neces-
flowchart is shown in Fig.1. Firstly, the three complex-
sary to balance noise suppression, fringe detail preserva-
tion and computational efficiency based on different im- ity factors related to the noise distribution and terrain
age characteristics. Factors like noise level and terrain slope are employed to adaptively select the filter
slope information should be taken into account to se- strategies. Based on the three complexity factors, the
lect the optimal filtering strategy and filtering paramet- complexity indicator CF1(m, n) of interferogram is con-
ers for different areas. structed to guide the adaptive selection of the most
To address the aforementioned problem, a novel suitable and effective filtering strategy for different
adaptive InSAR phase filtering method based on com- areas of the entire image. Then, the complexity scalar
plexity factors is proposed. Initially, the pseudo coher- CF2(m, n) is calculated according to the three complex-
ence, normalized maximum phase gradient (MPG), and ity factors, which can guide the adaptive LFF estima-
normalized phase derivative variance (PDV) are calcu- tion and compensation. The adaptive filter parameters
lated as the complexity factors. Based on the three are determined by CF2(m, n) of filter window in differ-
complexity factors, the complexity indicator is construc- ent filter methods for both residual phase and noisy
ted to guide the adaptive selection of the suitable and phase without removing LFF. In summary, the pro-
effective filtering strategy for different areas in the in- posed method adaptively selects the filtering strategy
terferogram. Then, the complexity scalar, calculated by and adjusts the filtering parameters based on the noise
the three complexity factors, is used to guide the adapt- and slope characteristics of the interferogram.
ive LFF estimation and adaptive filter parameters in 1. Complexity factors
different filtering methods. Experiments on simulated Many indexes, including coherence, pseudo coher-
and real data prove that for complicated and change- ence, PDV, MPG, and second-order phase gradient can
able terrain, the proposed method can not only effect- be used to describe the interferometric phase quality [38].
ively suppress noise and preserve phase fringe details, In this article, the pseudo coherence coefficient, normal-
but also increase the calculation efficiency. ized PDV and normalized MPG are selected as com-
The rest of this article is organized as follows. The plexity factors, which represent the noise level and slope
1092 Chinese Journal of Electronics 2023

Start

Interferogram

Pseudo Normalized Normalized


coherence PDV MPG

CF1 (m, n) of Adaptive selection


interferogram of filtering strategy

CF1 (m, n)=2 CF1 (m, n)=1 CF1 (m, n)=0


NL-InSAR filter Goldstein filter
Goldstein
based on LFF based on LFF
filter
compensation compensation

Adaptive LFF
estimation and
compensation
CF2 (m, n) of
filter wondow
Adaptive filter
Phase filtering
parameters
Complexity
factors
Add LFF to filtered
residual phase

Filterd phase

End

Fig. 1. Flowchart of the proposed method.

information. tion on (i, j) respectively, ∆¯xm,n and ∆¯ym,n are the local
1) Pseudo coherence coefficient mean values of the phase derivatives for pixel (m, n) in
The pseudo coherence coefficient γ(m, n) repres- row and column respectively.
ents the noise level of interferometric phase and can be Since γ(m, n) ∈ [0, 1], in order to ensure the con-
calculated with interferometric phase alone. The calcu- sistency for all complexity factors, the PDV′ (m, n) is
lation formula is expressed as [39] normalized to PDV(m, n) by
v 2  2 ( )
u
u ∑
u ∑ PDV′ (m, n) − min PDV′ (i, j)
t cos φi,j  +  sin φi,j  PDV(m, n) = ( ) ( )
(1) max PDV′ (i, j) − min PDV′ (i, j)
i,j i,j (3)
γ(m, n) =
k2
where i ∈ M, j ∈ N , M and N represent the number of
where γ(m, n) is calculated in a k × k window centered rows and columns of the interferogram respectively.
on (m, n). φi,j represents the interferometric phase on ( ) ( )
And min PDV′ (i, j) and max PDV′ (i, j) represent
pixel (i, j) in the local window. the minimum and maximum PDV of the interferogram
2) Normalized PDV respectively.
The PDV of an interferogram is defined as [38] 3) Normalized MPG
PDV′ (m, n) The MPG of an interferogram is defined as [39]
√∑ √∑
( )2 ( )2
∆xi,j − ∆¯xm,n + ∆yi,j − ∆¯ym,n MPG′ (m, n) = max(max( ∆xi,j ), max( ∆yi,j )) (4)
i,j i,j
=
k2 where MPG′ (m, n) is the maximum phase gradient of
(2) row and column direction in a k × k window centered
where ∆xi,j = wrap(φi+1,j −φi,j ) and ∆yi,j = wrap(φi,j+1− on (m, n).
φi,j ) are the phase derivatives in row and column direc- Similarly, MPG′ (m, n) is normalized to MPG(m, n)
A Novel Adaptive InSAR Phase Filtering Method Based on Complexity Factors 1093

by CF1(m, n) = γ1 (m, n) + PDV1 (m, n) × MPG1 (m, n)


( )
MPG′ (m, n) − min MPG′ (i, j) (9)
MPG(m, n) = ( ) ( )
max MPG′ (i, j) − min MPG′ (i, j)
(5) Only when PDV1 (m, n) and MPG1 (m, n) both
( ) ( ) equal to 1, the value of PDV1 (m, n) × MPG1 (m, n) is 1,
where min MPG′ (i, j) and max MPG′ (i, j) represent which makes the steepness indicator more robust and
the minimum and maximum MPG of the interferogram reliable. γ1 (m, n) = 1 means that the noise of the pixel
respectively. (m, n) is relatively large. Therefore, the three discrete
It is known that the pseudo coherence coefficient values 0, 1, and 2 of CF1(m, n) can be obtained by the
γ(m, n) is inversely proportional to the noise level. The formula (9), which guides the adaptive selection of the
PDV(m, n) and MPG(m, n) are related to the slope of filter strategy.
the terrain, and their values are proportional to the Moreover, the window size is very important for
steepness of the terrain. So pseudo coherence coefficient, the filtering. Here, the adaptive window size is determ-
normalized PDV, and normalized MPG represent the ined by the three complexity factors. Firstly, it is im-
noise level and terrain slope information and are selec- portant to set a basic filter window size window , then the
ted as complexity factors to guide the adaptive phase filter window size of different filter strategies is adjus-
filtering of the interferogram. ted based on window and CF1(m, n). window is calculated
2. Adaptive selection of filtering strategy by
For interferograms with complicated and change- window = 2 × (PDV_r + MPG_r + γ_r) + 1 (10)
able terrain background, the noise and terrain slope
changes greatly. It is necessary to select different filter- where
⌈ ⌉
ing strategies for different areas with different terrain max(abs(PDV(i, j) − PDVmean ))
PDV− r = (11)
characteristics. Hence, an adaptive selection of filtering PDVstd
strategy is proposed based on the complexity factors. ⌈ ⌉
γ(m, n) characterizes the phase noise. The pseudo max(abs(MPG(i, j) − MPGmean ))
MPG− r = (12)
coherence coefficient indicator γ1 (m, n) is calculated by MPGstd
{ 
0, γ(m, n) > γmean  −1, 0.8 < γmean ≤ 1
γ1 (m, n) = (6)
1, γ(m, n) ≤ γmean γ_r = 0, 0.4 < γmean ≤ 0.8 (13)

1, 0 ≤ γmean ≤ 0.4
where γmean is the mean value of pseudo coherence coef-
ficient of the interferogram. γ1 (m, n) = 0 means that PDVstd and MPGstd are the standard deviation of
the phase noise level of the pixel (m, n) is relatively low. the normalized PDV and normalized MPG of the inter-
Similarly, PDV indicator PDV1 (m, n) and MPG in- ferogram respectively. ⌈·⌉ rounds up its arguments to
dicator MPG1 (m, n) of pixel (m, n) are calculated by the nearest integer.
{ The PDV_r and MPG_r are the maximum value
0, PDV(m, n) ≤ PDVmean of Z-score normalized PDV and Z-score normalized
PDV1 (m, n) = (7)
1, PDV(m, n) > PDVmean MPG respectively. Z-score normalization [40] could con-
{ vert PDV and MPG to the same magnitude. The
0, MPG(m, n) ≤ MPGmean
MPG1 (m, n) = (8) PDV_r and MPG_r are related with the largest ter-
1, MPG(m, n) > MPGmean
rain slope pixel of the interferogram. Sun et al. state
where PDVmean and MPGmean are the mean value of the that larger terrain slopes lead to more severe baseline
normalized PDV and normalized MPG of an interfero- decorrelation and lower signal noise ratio (SNR) in the
gram. PDV1 (m, n) = 1 and MPG1 (m, n) = 1 mean that interferogram [41]. The larger window size should be set
the slope of the terrain near the pixel (m, n) is relat- for the lower SNR areas. Therefore, the PDV_r and
ively steep. MPG_r, determined by the largest terrain slope, are
According to the relationship between the three used to jointly set the basic radius of filter window size.
complexity factors γ(m, n), PDV(m, n), and MPG(m, n) Correlation is inversely proportional to the noise
of each pixel (m, n) and the corresponding mean value level, therefore γ_r, determined by correlation coeffi-
γmean , PDVmean, and MPGmean, γ1 (m, n), PDV1 (m, n), cient, is used to adjust the filter window size. Here, 0.8
and MPG1 (m, n) are calculated, which are binary para- and 0.4 are chosen experimentally [1]. The quality of
meters of 0 or 1 with no dimension. the interferimetric phase is good if 0.8 < γmean ≤ 1. On
The complexity indicator CF1(m, n) can be calcu- the contrary, 0 ≤ γmean ≤ 0.4 means that the noise is
lated by heavy and the filter window size needs to be increased
1094 Chinese Journal of Electronics 2023

to suppress noise. With the decrease of γmean , the filter the fringe details. Therefore, the Goldstein filtering
window size will increase accordingly, to enhance noise method is chosen for its high calculation efficiency, and
suppression. the filter window size can be slightly increased. So, the
According to formula (9), CF1(m, n) has three dis- filter window size is
crete values, 0, 1, and 2, which represents the filtering max_w = window + 1, CF1(m, n) = 0 (16)
difficulty of filter window centered on pixel (m, n).
Therefore, different filtering strategies and different fil- 3. Adaptive LFF estimation and compensa-
tering window size centered on (m, n) are chosen for tion
each pixel according to CF1(m, n). In the case of CF1(m, n) = 2 and CF1(m, n) = 1,
1) For CF1(m, n) = 1, Goldstein filter based on the LFF is estimated and removed before filtering, so
LFF compensation the noise can be effectively filtered out with little loss of
CF1(m, n) = 1 means that the noise level and ter- fringes detail. In the traditional LFF estimation meth-
rain slope in the pixel (m, n) are moderate. By tradeoff od [33], the size of the estimation window is fixed and
between noise suppression and computation burden, the LFF is assumed to be linear. However, for complicated
LFF compensation is needed, so the Goldstein filter and changeable terrain, LFF is generally nonlinear, in
method based on LFF compensation is adopted, and which case linear LFF compensation may cause terrain
the LFF estimation window size is equal to filter win- information loss during filtering. An adaptive LFF es-
dow size according to formula (10). Considering that timation and compensation method is proposed accord-
the noise and slope information are both moderate in ing to the complexity factors. To deal with the adverse
this case, the filter window size of the Goldstein filter impact of phase noise on LFF, prefiltering is employed
method based on LFF compensation equals to the ba- before LFF estimation, then nonlinear or linear LFF es-
sic filter window size window , which means timation is performed adaptively to the terrain.
The complexity scalar CF2(m, n), which reflects
mid_w = window , CF1(m, n) = 1 (14) the noise level and terrain slope information, is thus
2) For CF1(m, n) = 2, NL-InSAR filter based on defined as
LFF compensation CF2(m, n)
CF1(m, n) = 2 means that the noise level and ter- 1 − γ2 (m, n) + PDV2 (m, n) + MPG2 (m, n)
rain slope in the pixel (m, n) are relatively higher. Usu- = (17)
3
ally, this means dense fringes and heavy noise, which
increases the filtering difficulty compared with the first where
1 ∑∑
k k
case CF1(m, n) = 1. In this case, an NL-InSAR filter- γ2 (m, n) = γ(i, j) (18)
ing method based on LFF compensation is adopted, k × k i=1 j=1
which has great fringe detail preservation and noise
1 ∑∑
k k
suppression capabilities while heavy calculation burden.
PDV2 (m, n) = PDV(i, j) (19)
For the NL-InSAR filtering method based on LFF k × k i=1 j=1
compensation, the LFF estimation window size is also
equal to window . To ensure the fringe continuity of adja- 1 ∑∑
k k

cent pixel, the filter window size span of different filter MPG2 (m, n) = MPG(i, j) (20)
k × k i=1 j=1
methods should not be too large, otherwise it will bring
additional LFF estimation error. In the case CF1(m,
where γ2 (m, n), PDV2 (m, n) and MPG2 (m, n) repres-
n) = 2, due to the fringe is denser than that in the case
ent the mean value of γ(i, j), PDV(i, j), and MPG(i, j)
CF1(m, n) = 1, to ensure the accuracy of LFF estima-
in the k × k filter window centered on pixel (m, n) re-
tion, the filter window size needs to be slightly reduced.
spectively. And the values of γ2 (m, n), PDV2 (m, n), and
Therefore, the filter window size is calculated by
MPG2 (m, n) are within [0, 1]. Therefore, CF2(m, n) is a
min_w = window − 1, CF1(m, n) = 2 (15) continuous value within [0, 1], which is used to instruct
the adaptive LFF estimation and calculate the filter
3) For CF1(m, n) = 0, Goldstein filter parameters.
CF1(m, n) = 0 means that the noise and terrain 1) Prefilter with variable window size before LFF
slope are lower. Compared with CF1(m, n) = 2 and estimation
CF1(m, n) = 1, the filtering difficulty is greatly re- When CF1(m, n) = 2 and CF1(m, n) = 1, it is
duced. In this case, most filtering methods can achieve known that the fringes are relatively dense and LFF
satisfactory noise suppression effects while preserving compensation needs to be performed in the filter win-
A Novel Adaptive InSAR Phase Filtering Method Based on Complexity Factors 1095

dow. To improve the accuracy of LFF estimation, ad- phase. The components with the amplitude of the spec-
aptive mean filter is implemented as prefilter, whose trum greater than the threshold are retained, and lower
window size is determined by CF2(m, n) as than the threshold is set to zero.
 {
 7 × 7, 0.6 < CF2(m, n) ≤ 1 ′ S(u, v), |S(u, v)| ≥ b
S (u, v) = (23)
prefilter_win = 5 × 5, 0.2 < CF2(m, n) ≤ 0.6 0, |S(u, v)| < b

3 × 3, 0 ≤ CF2(m, n) ≤ 0.2
(21) where b ∈ (0, max (|S (u, v) |)]. b = max (|S (u, v) |) means
extracting the linear phase. The smaller the value of b ,
Since prefilter is performed within the filter win- the richer the nonlinear phase details. But if the value
dow mid_w or min_w , therefore, it is necessary to en- of b is too small, the extracted phase will contain part
sure that the prefilter window size prefilter_win is of the phase noise. Therefore, it is necessary to select a
smaller than mid_w or min_w . Also prefilter_win can- reasonable threshold to effectively distinguish the prom-
not be too large, otherwise it will reduce the estimation inent fringe spectrum from the noise spectrum.
sensitivity. In addition, prefilter_win should be small to It is proposed that b = max(|S(u, v)|) × 97% is be-
ensure the continuity of the LFF estimation value [26]. neficial to estimate the nonlinear fringes more precisely
So prefilter_win varies from 3 × 3 to 7 × 7, which is in [42]. When CF1(m, n) = 2 and CF1(m, n) = 1, the
small enough for most interferogram. interferometric fringes are complicated and changeable,
CF2(m, n) is proportional to the noise level and therefore, in our methods, the threshold is set to
terrain slope in the filter window. Conversely, correla-
tion coefficient is inversely proportional to the noise b = max(|S(u, v)|) × (100−X)%, X ∈ [1, 3] (24)
level. The boundary values 0.2 and 0.6 of CF2(m, n) are
Sorting the spectrum amplitude in descending or-
set according to the boundary values 0.8 and 0.4 of cor-
der, the value of X is determined by the complexity of
relation coefficient in formula (13) respectively. There-
the terrain, which is more adaptable to different scen-
fore, the areas of heavy noise will be prefiltered with
arios. When 0 ≤ CF2(m, n) ≤ 0.2, the prominent fringe
larger windows.
is approximately equal to the linear fringe, so we set
It should be noted that prefilter with variable win-
X = 1, that is b = max(|S(u, v)|) × 99% , so as to avoid
dow size will improve the LFF estimation accuracy
the noise spectrum. In the case of more complicated ter-
without losing details of the interferogram as it is not
rain, 0.6 < CF2(m, n) ≤ 1, the nonlinear phase details
used for phase filter but only for LFF estimation.
are richer, so the prominent fringe can be better extrac-
2) Adaptive prominent fringe components estima-
ted by taking X = 3, which means b = max(|S(u, v)|)×
tion
The fringes are often nonlinear in complicated ter- 97%.
rain. If only the linear fringe is removed, the residual The complex form of the prominent phase compon-
fringe will hamper the phase filtering. Several linear or ent in the local window is shown as
nonlinear LFF estimation methods have been proposed I ′ m (m, n) = FFT−1 (S ′ (u, v)) (25)
over the past years. However, the linear LFF estima-
tion methods cannot accurately compensate the fringe φm (m, n) = arg (exp(I ′ m (m, n))) (26)
frequency in areas with complicated terrain, which lim-
its the ability of noise suppression in dense fringe areas. The residual phase φr is obtained by removing the
The nonlinear LFF estimation method [37] breaks prominent phase φm from the noisy phase φn , which
through the first-order limitation of fringe frequency by can be expressed as
performing weighted least squares phase unwrapping, φr = arg (exp(j(φn − φm ))) (27)
which may affect the filtering result. Here, prominent
fringe components estimation method is presented and When CF1(m, n) = 2 or CF1(m, n) = 1, through
the linear or nonlinear fringe is compensated adaptively. the above steps, the adaptive LFF compensation are
The prominent fringe component is estimated by performed in the filter window centered on pixel (m, n).
extracting the prominent frequency components of Finally, the filtered interferometric phase φf consists of
phase in the frequency domain. Implement 2-D FFT in the removed phase φm and the filtered residual phase
local window to obtain the interferometric phase spec- φrf .
trum as follows: 4. Adaptive filter parameters
S (u, v) = FFT (In (m, n)) (22) Complexity scalar CF2(m, n) is also used to calcu-
late the adaptive filter parameters in the different filter-
where In = exp (jφn ) is the complex form of the noisy ing methods.
1096 Chinese Journal of Electronics 2023

1) For CF1(m, n) = 2, residual phase NL-InSAR ter parameter is usually set to 0.5. Therefore, in our
filtering based on LFF compensation method, αr (m, n) is adaptive to CF2(m, n) by
For CF1(m, n) = 2, NL-InSAR filtering is per-
formed on the residual phase. In the NL-InSAR filter αr (m, n) = αrmin + (1 − αrmin ) × (1 − CF2(m, n)) (29)
method [19], the weights are determined by the similar-
ity between the neighboring pixel blocks, and the calcu- Given that the area with CF1(m, n) = 1 has mod-
lation is related to the Euclidean distance of neighbor- erate noise and slope, The LFF compensation is per-
ing pixel blocks and smoothing parameter h(m, n). The formed before residual phase filter and the density of
filtering effect of NL-InSAR filter method is greatly af- fringes has decreased, so the filter parameter αr (m, n)
fected by the smoothing parameter h(m, n), which is can be a little larger. αrmin = 0.5 is set to ensure the fil-
closely related to the decay speed of the weight function. ter intensity, and 1 − CF2(m, n) is also used to adjust
Usually a smaller h(m, n) is beneficial to the preser- the fringe detail preservation
( ) ability. The filter para-
vation of fringes details, and a larger h(m, n) can im- meter αr (m, n) ∈ αrmin , 1 , therefore the filter intensity
prove the noise suppression ability. In [42], the smooth- can be adaptively adjusted according to CF2(m, n).
ing parameter h(m, n) = 10σn · γ · (1 + f ′ x2 +f ′ y 2 )−1/2 , 3) For CF1(m, n) = 0, noisy phase Goldstein filter-
which is jointly defined by the noise standard deviation ing
σn , the correlation coefficient γ and the residual phase For CF1(m, n) = 0, the noise level and terrain
( slope of the area are low. To improve the calculation ef-
fringe frequency (f x, f y). The value of γ · 1 + f ′ x2 +
′ ′
)−1/2 ficiency, Goldstein filter is adopted in noisy phase
f ′ y2 is within (0,1). Since the residual phase fringe
without removing LFF, its filter parameter αn (m, n) is
frequency (f ′ x, f ′ y) need to be estimated, this method
adaptively determined by
is time-consuming, and may introduce fringe frequency
estimation errors. win
αn (m, n) = 1 − γmean (m, n) × (1 − CF2(m, n)) (30)
Therefore, to ensure the fringe detail preservation
and noise suppression effect at the same time, the win
where γmean (m, n) is the mean value of pseudo coher-
smoothing parameter h(m, n) is adaptively calculated by ence coefficient in the filter window centered on pixel
(m, n). The Goldstein filter parameter in [27] is only
h(m, n) = 10σn × h′ (m, n)
changed with the noise level. To make the adaptive fil-
h′ (m, n) = 0.7 + 0.3 × (1 − CF2(m, n)) (28) ter parameter more robust, in our method, the filter
The larger CF2(m, n) means denser fringes and parameter αn (m, n) changes with the noise level and
win
heavier noise. CF2(m, n) varies from 0 to 1, so the value terrain slope simultaneously because of γmean (m, n)×
range of h′ (m, n) is within (0.7,1). Here the minimum (1−CF2(m, n)). In the case CF1(m, n) = 0, γmean win
(m, n)
value of h′ (m, n) is set as 0.7 because the residual phase is larger and CF2(m, n) is smaller. The minimum value
after LFF compensation is so sparse that the smooth- of αn (m, n) is 1 − γmean
win
(m, n), which can prevent over-
ing parameter h(m, n) can be larger. For the areas with filtering.
larger CF2(m, n), 1 − CF2(m, n) is smaller, which is
more conducive to preserving fringe details. Therefore,
III. Results and Analysis
CF2(m, n) is used to adjust the smoothing parameter In this section, to validate the proposed method,
h(m, n) to gain a tradeoff between fringe preservation experiments are performed on both simulated and real
and noise suppression. interferograms, and results are compared with those of
2) For CF1(m, n) = 1, residual phase Goldstein fil- several recognized and representative methods. The ef-
tering based on LFF compensation fectiveness of the three proposed adaptive strategies
Goldstein filtering is performed on the residual based on the complexity factors is demonstrated by the
phase after removing the LFF. Goldstein filter method first experiment with simulated data. In the second ex-
[21] converts the interferometric phase from the spatial periment, the superiority of the proposed method in
domain to the frequency domain, and then smooths the terms of noise suppression and fringe detail preserva-
frequency spectrum. The filter parameter αr (m, n) in- tion is verified by comparison with the three adaptive
dicates the extent to which the spectrum is smoothed. filter methods —The slope adaptive filter [33] and the
For Goldstein filter, the smoothing effect becomes improved Goldstein filter [27] can implement adaptive
more intense with the increase of the filter parameter phase filter based on fringe frequency and correlation
αr (m, n), whereas the ability of fringe detail preserva- coefficient, respectively. The improved NL-InSAR filter
tion reduces. In order to balance the details loss and [42] is one of the latest adaptive filter method proposed
noise suppression ability, the traditional Goldstein fil- in 2021. Finally, the real data are processed to further
A Novel Adaptive InSAR Phase Filtering Method Based on Complexity Factors 1097

verify the robustness and superiority of the proposed data [43]. The noise flattened phase and corresponding
method in comparison with three existing methods. real phase are shown in Fig.2. In order to verify the ef-
1. The first experiment to validate the ad- fectiveness of the three adaptive strategies in the pro-
aptive strategies posed method, the three adaptive strategies are sequen-
Two SAR single-look complex images are simu- tially replaced with fixed strategies, results are shown in
lated according to certain SAR geometry and DEM the Fig.3.

(rad) (rad)
3 3
20 20
2 2
40 40
1 1
60 60

80 0 80 0

100 −1 100 −1

120 −2 120 −2
140 140
−3 −3
20 40 60 80 100 120 140 20 40 60 80 100 120 140
(a) Real phase (b) Noisy flattened phase

Fig. 2. Simulated data 1.

(rad) (rad)
3 3
20 20
2 2
40 40
1 1
60 60
0 0
80 80

100 −1 100 −1

120 −2 120 −2
140 140
−3 −3
20 40 60 80 100 120 140 20 40 60 80 100 120 140
(a) The first fixed strategy (b) The second fixed strategy
(rad) (rad)
3 3
20 20
2 2
40 40
1 1
60 60
0 0
80 80

100 −1 100 −1

120 −2 120 −2
140 140
−3 −3
20 40 60 80 100 120 140 20 40 60 80 100 120 140
(c) The third fixed strategy (d) Proposed method

Fig. 3. Filtered results using different adaptive strategies.

Clearly, the filter result of proposed method shown opt NL-InSAR filtering method based on LFF compens-
in Fig.3(d) contains more fringe detail than those in ation for the interferogram, but the LFF estimation and
Fig.3(a), (b), and (c). The first fixed strategy is to ad- filter parameters are still adaptive. It can be seen that
1098 Chinese Journal of Electronics 2023

without adaptive selection of filtering method, result in filter, the Improved Goldstein filter and the Improved
Fig.3(a) presents more phase residues. The second fixed NL-InSAR filter are implemented as comparison. The
strategy is that the LFF estimation window and the noisy flattened phase and corresponding real phase and
prefilter window are fixed to 8 × 8 and 5 × 5 respect- are shown in Fig.4.
ively, X = 1 is adopted to extract prominent fringe, the
(rad)
filter results is shown in Fig.3(b). The third fixed 3
strategy is that the filter parameters h(m, n), αr (m, n), 20
αn (m, n) are fixed to 0.5, and Fig.3(c) shows the filter 40
2
results. Without adaptive LFF estimation and adapt-
60 1
ive filter parameters, the fringe detail in Fig.3(b) and
80
Fig.3(c) are preserved not as good as in Fig.3(a) and 0
Fig.3(d). 100
In order to evaluate the filtered results, the per- 120 −1
formance of each filter is assessed by the number of 140
−2
phase residues, the edge preservation index (EPI) [31] 160
and the root mean square errors (RMSE) [26]. The EPI −3
180
and RMSE are calculated by 20 40 60 80 100 120 140 160 180

∑ ( |φf (m, n) − φf (m + 1, n)| )


(a) Real phase
(rad)
+ |φf (m, n) − φf (m, n + 1)| 3
EPI = ∑ ( ) (31) 20
|φreal (m, n) − φreal (m + 1, n)|
2
+ |φreal (m, n) − φreal (m, n + 1)| 40
60 1
RMSE
√∑ ∑ 80
2
|arg (exp (j (φf (m, n) − φreal (m, n))))| 0
= 100
M ×N 120 −1
(32)
140
−2
where φf (m, n) and φreal (m, n) represent the filtered 160
phase and the real phase respectively. M × N repres- −3
180
ents the total number of pixels in the interferogram. 20 40 60 80 100 120 140 160 180
An EPI closer to 1 means better fringe and edge (b) Noisy flattened phase
preservation. EPI > 1 means that fringe details are false Fig. 4. Simulated data 2.
or more noise on the filtered phase, whereas EPI < 1
As can be seen in Fig.4, the fringe density is vari-
means the fringe details are damaged. Smaller RMSE
able, which is sparse on the right side and dense on the
represents a better noise suppression effect.
left side. The terrain is relatively flat in sparse fringe
The evaluation results are shown in Table 1. It is
areas and the terrain is more complicated in dense
obvious that the three adaptive strategies in the pro-
fringe area, which increases the difficulty of filtering.
posed method have improved fringe details preserva-
In order to facilitate comparison, the window size
tion and noise suppression.
of the other three filtering methods is same with that of
Table 1. Evaluation results of simulated data1 CF1(m, n) = 1 in the proposed method, i.e., win_size =
mid_w . In Fig.5, we show the corresponding filtered
Interferogram Residues EPI RMSE (rad)
Real phase 1 1 0 results of different methods. In each group, the left im-
Noisy flattened phase 3270 7.8684 1.1425 age is the filtered phase, and the right one is the corres-
The first fixed strategy 4 1.0620 0.1078 ponding phase error between the filtered and real phase.
The second fixed strategy 4 1.0995 0.1096 As shown in Fig.5(a) and Fig.5(b), in the flat area
The third fixed strategy 3 1.1029 0.1068 on the right, the slope adaptive filter has a better noise
Proposed method 1 1.0165 0.1011
suppression effect, but in the dense fringe area on the
left, the fringe details are damaged, so the phase error
2. The second experiment to validate su- is larger. In Fig.5(c) and Fig.5(d), although the overall
periority phase error of the improved Goldstein filter is smaller,
In this part, the superiority of the proposed meth- it can be clearly seen that the fringe edge preservation
od is evaluated on simulated data. The slope adaptive ability is poor, and the noise in some areas is still large.
A Novel Adaptive InSAR Phase Filtering Method Based on Complexity Factors 1099

(rad) (rad)
3 3
20 20
2 2
40 40
60 1 60 1
80 80
0 0
100 100
120 −1 120 −1
140 140
−2 −2
160 160
180 −3 180 −3
20 40 60 80 100 120 140 160 180 20 40 60 80 100 120 140 160 180
(a) Slope adaptive filtered result (b) Slope adaptive filtered phase error
(rad) (rad)
3 3
20 20
2 2
40 40
60 1 60 1
80 80
0 0
100 100
120 −1 120 −1
140 140
−2 −2
160 160
180 −3 180 −3
20 40 60 80 100 120 140 160 180 20 40 60 80 100 120 140 160 180
(c) Improved Goldstein filtered result (d) Improved Goldstein filtered phase error
(rad) (rad)
3 3
20 20
2 2
40 40
60 1 60 1
80 80
0 0
100 100
120 −1 120 −1
140 140
−2 −2
160 160
180 −3 180 −3
20 40 60 80 100 120 140 160 180 20 40 60 80 100 120 140 160 180
(e) Improved NL-InSAR filtered result (f) Improved NL-InSAR filtered phase error
(rad) (rad)
3 3
20 20
2 2
40 40
60 1 60 1
80 80
0 0
100 100
120 −1 120 −1
140 140
−2 −2
160 160
180 −3 180 −3
20 40 60 80 100 120 140 160 180 20 40 60 80 100 120 140 160 180
(g) Proposed method filtered result (h) Proposed method filtered phase error

Fig. 5. Filtered result and corresponding phase error.


1100 Chinese Journal of Electronics 2023

For the improved NL-InSAR filter, as it can be ob- 4


Real phase
served in Fig.5(e) and Fig.5(f), there are some error 3
Noisy phase
points at the edges of the image. Comparing the result 2 Slope

Phase error (rad)


Goldstein
of the proposed method in Fig.5(g) and Fig.5(h) with 1
the above three filter methods, the proposed methods 0
show a good performance in noise suppression and −1
fringe details preservation. The phase error diagrams −2
clearly show that the proposed method outperforms NL-InSAR
−3 Proposed
other filter methods.
−4
The quantitative evaluation results are shown in 0 5 10 15 20 25 30 35 40
Table 2. The slope adaptive filter has obvious over- Range (pixels)
filter phenomenon, and the EPI is far less than 1. Al- Fig. 6. Phase error of different filter methods in cross-sec-
though its residue number is close to that of the pro- tion A.
posed method, it is at the cost of fringe detail loss, res-
in the white rectangle are shown in Fig.7. It can be seen
ulting in a larger RMSE. The RMSE of the improved
that the fringe in Fig.7(b), contaminated with heavy
Goldstein filter has been reduced, but there are more
noise, represents the complicated terrain of ETNA Vol-
residues and poor fringe detail preservation in areas
cano. And the mean pseudo coherence coefficient is only
with steep terrain and low coherence. For the improved
0.5152.
NL-InSAR filter, the fringe details are damaged in the
(rad)
dense fringe area, resulting in residual points. Com-
3
pared with the other three methods, the EPI of the pro-
50
posed method is closer to 1, hence, the ability of fringe 2
details preservation is much better. Moreover, the 100
residues and RMSE of the proposed method are the 1
150
smallest because of an excellent noise suppression per-
200 0
formance, proving that the proposed method achieves
the best balance between noise suppression and fringe 250
−1
preservation compared with the other three methods. 300
−2
Table 2. Evaluation results of simulated data 2 350

400 −3
Interferogram Residues EPI RMSE (rad) Time (s)
50 100 150 200 250 300 350 400
Real phase 1 1 0 –
(a) Noisy phase
Noisy flattened phase 7258 7.9945 1.0807 – (rad)
Slope adaptive filter 1 0.9328 0.3020 36
3
Improved Goldstein filter 6 1.1004 0.2069 9 20
Improved NL-InSAR filter 2 1.0684 0.1946 45 2
40
Proposed method 0 1.0410 0.1750 30
60
1
80
As shown in Fig.4(b), “A” represents the area
100 0
where phase distortion often occurs. The cross-section
of phase error in “A” is extracted to validate the ro- 120
−1
bustness of the proposed method in filtering the steep 140
terrain region. As clearly shown in Fig.6, the filtered 160 −2
phase error of the proposed method is much closer to 180
zero than the other filter methods, which proves that 200 −3
20 40 60 80 100 120 140 160 180 200
the proposed method has a better performance on the (b) Enlarged areas
edge preservation than the other three methods.
Fig. 7. ETNA Volcano.
3. Experiments with real data
In this part, two sets of real data are employed to The filtering results of Fig.7(d) with the slope ad-
investigate the performance of the proposed method. aptive filter, the improved Goldstein filter, the im-
1) ERS SAR data proved NL-InSAR filter, and the proposed method are
ERS SAR images over the ETNA Volcano in shown in Fig.8. In Fig.8(a), due to the LFF estimated
September and October 2000 is used as the test data. by the slope adaptive method is not accurate enough,
The interferometric noisy phase and the enlarged area resulting in damage of the fringes edges, which causes
A Novel Adaptive InSAR Phase Filtering Method Based on Complexity Factors 1101

more residues. As can be seen in Fig.8(b) for the im- InSAR filter is much better than the slope adaptive fil-
proved Goldstein filter, the fringes in the dense fringe ter and the improved Goldstein filter in dense fringe
area are ambiguous, especially in areas with a heavy areas. In Fig.8(d), the proposed method shows a better
noise. Comparing Fig.8(c) with Fig.8(a) and Fig.8(b), it performance in fringe detail preservation and the fringe
is seen that the fringe preservation of the improved NL- in steep terrain is the most continuous.

(rad) (rad)
3 3
20 20
40 2 40 2
60 60
1 1
80 80
100 0 100 0
120 120
−1 −1
140 140
160 −2 160 −2
180 180
200 −3 200 −3
20 40 60 80 100 120 140 160 180 200 20 40 60 80 100 120 140 160 180 200
(a) Slope adaptive filter (b) Improved Goldstein filter
(rad) (rad)
3 3
20 20
40 2 40 2
60 60
1 1
80 80
100 0 100 0
120 120
−1 −1
140 140
160 −2 160 −2
180 180
200 −3 200 −3
20 40 60 80 100 120 140 160 180 200 20 40 60 80 100 120 140 160 180 200
(c) Improved NL-InSAR filter (d) Proposed method

Fig. 8. Filtered results of Fig.7(b) with four methods.

A quantitative evaluation is also performed to com- where φf (m, n) represents the filtered phase, and
pare the filtered results. The number of residues, the φf (m + i, n + j) is the filtered phase of eight adjacent
sum of phase difference (SPD) [44] and the phase stand- pixels.
ard deviation (PSD) [45] are employed as metrics. Com- The PSD of the interferogram is calculated by
pared with residues, the SPD and PSD can more accur- v
ately reflect the smoothness of the filtered phase. It is u M N
u∑ ∑
generally believed that smaller number of residues, SPD u (φf (m, n) − φ̄f (m, n))
2
u (35)
and PSD indicate a smoother phase with less noise. t
PSD = m=1 n=1
The SPD for the interferogram is the sum of M ×N −1
APD(m, n), which is expressed by
where φ̄f (m, n) is the linear phase in a window of size

M ∑
N
3 × 3, and M × N represents the size of interferogram.
SPD = APD(m, n) (33)
m=1 n=1
The evaluation results are shown in Table 3. As
can be seen from Table 3, the number of residues are
where APD(m, n) is calculated as
reduce by all methods. However, because of the fixed
1 ∑
1 ∑1 window used in slope adaptive filter, the loss of detail is
APD(m, n) = × |φf (m, n) severe in the dense fringe area, so its residues are more
8 i=−1 j=−1
than the proposed method. The improved Goldstein fil-
− φf (m + i, n + j)| (34) ter and the improved NL-InSAR filter have the prob-
1102 Chinese Journal of Electronics 2023

lem of under-filtering in the low-coherence area, result- posed method. Again, the proposed method shows a
ing in a large SPD and PSD, and the residues of im- good improvement in terms of noise suppression, the
proved Goldstein filter is far more than that of the pro- number of residues, SPD and PSD are greatly reduced.

Table 3. Evaluation results of real data1

Interferogram Residues SPD × 104 (rad) PSD (rad) Time (s)


Noisy phase 7455 6.3222 1.4953 –
Slope adaptive filter 19 1.8761 0.3273 40
Improved Goldstein filter 41 1.9855 0.4592 12
Improved NL-InSAR filter 13 1.9320 0.4289 51
Proposed method 4 1.8317 0.3109 25

2) Millimeter-wave airborne InSAR data Figs.10(b) and (c), the dense fringes in the upper right
The millimeter-wave airborne InSAR data are em- corner where there is a mountainous shaded area, is
ployed to conduct another experiment. The interfero- severely damaged because the improved Goldstein filter
grams are provided by Beijing Institude of Radio Meas- and the improved NL-InSAR filter do not perform LFF
urement, and the test site is situated in Zhaotong, Yun- compensation before filtering. The filtered result in
nan Province, Midwest China. The lower left corner of Fig.10(d) and its quantitative evaluation in Table 4
the interferogram is a residential area containing a lot show that the proposed method not only effectively
of architectural details, and the upper right corner is a suppresses noise, but also has the best performance in
mountainous area, with partial shadows and layovers. preserving fringe details.
The interferometric noisy phase and the area bounded It can be seen from the calculation efficiency in the
by rectangle are shown in Fig.9. tables, the filter speed of the improved Goldstein filter
The slope adaptive filter, the improved Goldstein is always the fastest due to the frequency domain filter.
filter, the improved NL-InSAR filter and the proposed The improved NL-InSAR filter has the slowest filter
method are performed on Fig.9(b), and the filtered speed because twice LFF estimation is performed. The
phase of all four methods are shown in Fig.10. The slope adaptive filter also requires LFF estimation for
quantitative evaluation is given in Table 4. the entire interferogram. The proposed method adopts
Since the airborne data have relatively high SNR adaptive filter strategy, LFF estimation is not needed
and sparse fringes, all four methods have achieved good when CF 1(m, n) = 0, which represents the noise level
noise suppression effect. From Fig.10 and Table 4, the and fringe density are small. Therefore, the calculation
fixed window in the slope adaptive filter causes the efficiency is improved compared with the slope adapt-
fringe details of the residential area to be completely ive filter and the improved NL-InSAR filter.
filtered out. And the fringe edges are destroyed, result-
IV. Conclusions
ing in residues. Compared with the slope adaptive filter,
the SPD and PSD of the improved Goldstein filter and With the increasing resolution of SAR imaging,
the improved NL-InSAR filter are reduced. However, in there will be much more rich terrain types and detailed

(rad)
3
20
40 2
60
1
80
100 0
120
−1
140
160 −2
180
200 −3
20 40 60 80 100 120 140 160 180 200
(a) Noisy phase (b) Enlarged areas

Fig. 9. Millimeter-wave SAR data.


A Novel Adaptive InSAR Phase Filtering Method Based on Complexity Factors 1103

(rad) (rad)
3 3
20 20
40 2 40 2
60 60
1 1
80 80
100 0 100 0
120 120
−1 −1
140 140
160 −2 160 −2
180 180
200 −3 200 −3
20 40 60 80 100 120 140 160 180 200 20 40 60 80 100 120 140 160 180 200
(a) Slope adaptive filter (b) Improved Goldstein filter
(rad) (rad)
3 3
20 20
40 2 40 2
60 60
1 1
80 80
100 0 100 0
120 120
−1 −1
140 140
160 −2 160 −2
180 180
200 −3 200 −3
20 40 60 80 100 120 140 160 180 200 20 40 60 80 100 120 140 160 180 200
(c) Improved NL-InSAR filter (d) Proposed method

Fig. 10. Filtered results of Fig.9(b) with four methods.

Table 4. Evaluation results of real data 2

Interferogram Residues SPD × 104 (rad) PSD (rad) Time (s)


Noisy phase 6619 4.4042 1.4268 –
Slope adaptive filter 19 0.8478 0.3624 32
Improved Goldstein filter 20 0.7319 0.2900 12
Improved NL-InSAR filter 21 0.8113 0.3243 38
Proposed method 5 0.7286 0.2858 29

terrain information in the interferogram than ever. and after that, the proposed method is tested on simu-
What is more, heavy noise caused by SAR imaging geo- lated data and real data sets from ETNA Volcano and
metry and complicated terrain background brings a Yunnan Province, mountainous area in western China.
huge challenge to phase filtering. However, most of the By comparing its performance with the other three re-
current phase filtering methods cannot simultaneously cognized and representative phase filtering method, it
take into account the three aspects of suppressing noise has been demonstrated that the proposed method of-
suppression effectively, preserving terrain details adapt- fers the best filtering results. The adaptive selection of
ively and improving calculation efficiency. Thus, a nov- filtering strategy could improve calculation efficiency.
el adaptive InSAR phase filtering method based on Moreover, the adaptive LFF estimation and adaptive
filter parameters based on complexity factors not only
complexity factors is proposed in this paper. The com-
can effectively suppress noise, but also have excellent
plexity factors can characterize the noise level and ter-
performance in preserving fringe details, the effective-
rain slope information of the interferogram effectively,
ness and superiority of proposed method are validated.
which are used to guide the adaptive selection of suit-
able and effective filtering strategies for different areas
in the interferogram. References
The proposed method is presented in detail firstly, [1] R. F. Hanssen, Radar Interferometry: Data Interpretation
1104 Chinese Journal of Electronics 2023

and Error Analysis, Springer, Dordrecht, pp.16–23, 2001. Simulation, vol.4, no.2, pp.490–530, 2005.
[2] H. A. Zebker and R. M. Goldstein, “Topographic mapping [19] C. A. Deledalle, L. Denis, and F. Tupin, “NL-InSAR: Non-
from interferometric synthetic aperture radar observations,” local interferogram estimation,” IEEE Transactions on
Journal of Geophysical Research:Solid Earth, vol.91, no.B5, Geoscience and Remote Sensing, vol.49, no.4, pp.1441–452,
pp.4993–4999, 1986. 2011.
[3] D. C. Ghiglia and L. A. Romero, “Minimum Lp-norm two- [20] J. W. Li, Z. F. Li, Z. Bao, et al., “Noise filtering of high-res-
dimensional phase unwrapping,” Journal of the Optical So- olution interferograms over vegetation and urban areas with
ciety of America A, vol.13, no.10, pp.1999–2013, 1996. a refined nonlocal filter,” IEEE Geoscience and Remote
[4] P. A. Rosen, S. Hensley, I. R. Joughin, et al., “Synthetic Sensing Letters, vol.12, no.1, pp.77–81, 2015.
aperture radar interferometry,” Proceedings of the IEEE, [21] R. M. Goldstein and C. L. Werner, “Radar interferogram
vol.88, no.3, pp.333–382, 2000. filtering for geophysical applications,” Geophysical Re-
[5] P. H. Eichel, D. C. Ghiglia, and C. V. Jakowatz Jr, Spot- search Letters, vol.25, no.21, pp.4035–4038, 1998.
light SAR Interferometry for Terrain Elevation Mapping [22] C. Lopez-Martinez and X. Fabregas, “Modeling and reduc-
and Interferometric Change Detection. Albuquerque, NM, tion of SAR interferometric phase noise in the wavelet do-
USA: Sandia National Lab, 1996, doi: 10.2172/211364. main,” IEEE Transactions on Geoscience and Remote
[6] J. S. Lee, K. P. Papathanassiou, T. L. Ainsworth, et al., “A Sensing, vol.40, no.12, pp.2553–2566, 2002.
new technique for noise filtering of SAR interferometric [23] Y. Bian and B. Mercer, “Interferometric SAR phase filter-
phase images,” IEEE Transactions on Geoscience and Re- ing in the wavelet domain using simultaneous detection and
mote Sensing, vol.36, no.5, pp.1456–1465, 1998. estimation,” IEEE Transactions on Geoscience and Re-
[7] C. F. Chao, K. S. Chen, J. S. Lee, et al., “Refined filtering mote Sensing, vol.49, no.4, pp.1396–1416, 2011.
of interferometric phase from INSAR data,” in Proceedings [24] X. Zha, R. S. Fu, Z. Y. Dai, et al., “Noise reduction in in-
of 2012 IEEE International Geoscience and Remote Sens- terferograms using the wavelet packet transform and wien-
ing Symposium, Munich, Germany, pp.1821–1824, 2012. er filtering,” IEEE Geoscience and Remote Sensing Letters,
[8] N. Wu, D. Z. Feng, and J. X. Li, “A locally adaptive filter vol.5, no.3, pp.404–408, 2008.
of interferometric phase images,” IEEE Geoscience and Re- [25] J. Bioucas-Dias, V. Katkovnik, J. Astola, et al., “Absolute
mote Sensing Letters, vol.3, no.1, pp.73–77, 2006. phase estimation: Adaptive local denoising and global un-
[9] Q. F. Yu, X. Yang, S. H. Fu, et al., “An adaptive con- wrapping,” Applied Optics, vol.47, no.29, pp.5358–5369,
toured window filter for interferometric synthetic aperture 2008.
radar,” IEEE Geoscience and Remote Sensing Letters, [26] Q. Q. Feng, H. P. Xu, Z. F. Wu, et al., “Improved Gold-
vol.4, no.1, pp.23–26, 2007. stein interferogram filter based on local fringe frequency es-
[10] G. Vasile, E. Trouve, J. S. Lee, et al., “Intensity-driven ad- timation,” Sensors, vol.16, no.11, article no.1976, 2016.
aptive-neighborhood technique for polarimetric and interfer- [27] I. Baran, M. P. Stewart, B. M. Kampes, et al., “A modifica-
ometric SAR parameters estimation,” IEEE Transactions tion to the Goldstein radar interferogram filter,” IEEE
on Geoscience and Remote Sensing, vol.44, no.6, Transactions on Geoscience and Remote Sensing, vol.41,
pp.1609–1621, 2006. no.9, pp.2114–2118, 2003.
[11] L. Denis, F. Tupin, J. Darbon, et al., “Joint regularization [28] Z. W. Li, X. L. Ding, C. Huang, et al., “Improved filtering
of phase and amplitude of InSAR data: Application to 3-D parameter determination for the Goldstein radar interfero-
reconstruction,” IEEE Transactions on Geoscience and Re- gram filter,” ISPRS Journal of Photogrammetry and Re-
mote Sensing, vol.47, no.11, pp.3774–3785, 2009. mote Sensing, vol.63, no.6, pp.621–634, 2008.
[12] W. B. Abdallah and R. Abdelfattah, “A joint Markov ran- [29] C. Y. Zhao, Q. Zhang, X. L. Ding, et al., “An iterative
dom field approach for SAR interferogram filtering and un- Goldstein SAR interferogram filter,” International Journal
wrapping,” IEEE Journal of Selected Topics in Applied of Remote Sensing, vol.33, no.11, pp.3443–3455, 2012.
Earth Observations and Remote Sensing, vol.9, no.7, [30] Q. Sun, Z. W. Li, J. J. Zhu, et al., “Improved Goldstein fil-
pp.3016–3025, 2016. ter for InSAR noise reduction based on local SNR,” Journ-
[13] H. Y. Li, H. J. Song, R. Wang, et al., “A modification to al of Central South University, vol.20, no.7, pp.1896–1903,
the complex-valued MRF modeling filter of interferometric 2013.
SAR phase,” IEEE Geoscience and Remote Sensing Let- [31] R. Song, H. D. Guo, G. Liu, et al., “Improved Goldstein
ters, vol.12, no.3, pp.681–685, 2015. SAR interferogram filter based on empirical mode decom-
[14] A. Buades, B. Coll, and J. M. Morel, “A non-local al- position,” IEEE Geoscience and Remote Sensing Letters,
gorithm for image denoising,” in Proceedings of 2005 IEEE vol.11, no.2, pp.399–403, 2014.
Computer Society Conference on Computer Vision and [32] R. Song, H. D. Guo, G. Liu, et al., “Improved Goldstein
Pattern Recognition (CVPR’05), San Diego, CA, USA, SAR interferogram filter based on adaptive-neighborhood
pp.60–65, 2005. technique,” IEEE Geoscience and Remote Sensing Letters,
[15] D. S. Fang, X. L. Lv, and B. Lei, “A novel InSAR phase de- vol.12, no.1, pp.140–144, 2015.
noising method via nonlocal wavelet shrinkage,” in Proceed- [33] E. Trouve, J. M. Nicolas, and H. Maitre, “Improving phase
ings of 2016 IEEE International Geoscience and Remote unwrapping techniques by the use of local frequency estim-
Sensing Symposium (IGARSS), Beijing, China, ates,” IEEE Transactions on Geoscience and Remote Sens-
pp.6429–6432, 2016. ing, vol.36, no.6, pp.1963–1972, 1998.
[16] X. Lin, F. F. Li, D. D. Meng, et al., “Nonlocal SAR interfer- [34] U. Spagnolini, “2-D phase unwrapping and instantaneous
ometric phase filtering through higher order singular value frequency estimation,” IEEE Transactions on Geoscience
decomposition,” IEEE Geoscience and Remote Sensing Let- and Remote Sensing, vol.33, no.3, pp.579–589, 1995.
ters, vol.12, no.4, pp.806–810, 2015. [35] E. Trouvé, M. Caramma, and H. Maître, “Fringe detection
[17] G. Baier, C. Rossi, M. Lachaise, et al., “A nonlocal InSAR in noisy complex interferograms,” Applied Optics, vol.35,
filter for high-resolution DEM generation from TanDEM-X no.20, pp.3799–3806, 1996.
interferograms,” IEEE Transactions on Geoscience and Re- [36] B. Cai, D. N. Liang, and Z. Dong, “A new adaptive mul-
mote Sensing, vol.56, no.11, pp.6469–6483, 2018. tiresolution noise-filtering approach for SAR interferometric
[18] A. Buades, B. Coll, and J. M. Morel, “A review of image de- phase images,” IEEE Geoscience and Remote Sensing Let-
noising algorithms, with a new one,” Multiscale Modeling & ters, vol.5, no.2, pp.266–270, 2008.
A Novel Adaptive InSAR Phase Filtering Method Based on Complexity Factors 1105

[37] Z. Y. Suo, Z. F. Li, and Z. Bao, “A new strategy to estim- LI Chunsheng received the
ate local fringe frequencies for InSAR phase noise Ph.D. degree in signal and information
reduction,” IEEE Geoscience and Remote Sensing Letters, processing from Beihang University in
vol.7, no.4, pp.771–775, 2010. 1998. Since 2005, he is a Professor with
[38] H. W. Yu, Y. Lan, Z. H. Yuan, et al., “Phase unwrapping in the School of Electronics and Informa-
InSAR: a review,” IEEE Geoscience and Remote Sensing tion Engineering, Beihang University. He
Magazine, vol.7, no.1, pp.40–58, 2019. has authored more than 100 journal and
[39] D. C. Ghiglia and M. D. Pritt, Two-Dimensional Phase Un- conference papers and four books. His re-
wrapping: Theory, Algorithms, and Software. Wiley-Inter- search interests include analysis and sim-
science, New York, NY, USA, 1998.
ulation of SAR satellite, highresolution image formation, and mul-
[40] E. Kreyszig, Advanced Engineering Mathematics, 4th ed.,
John Wiley & Sons, New York, NY, USA, pp.880, 1979. timodal remote sensing data fusion. (Email: lics@buaa.edu.cn)
[41] L. Sun, C. Y. Zhang, and M. L. Hu, “System design and
performance analysis of spatial baseline in spaceborne In- ZENG Guobing received B.S.
SAR,” Radar Science and Technology, vol.5, no.2, degree in aircraft engineering from Bei-
pp.133–138, 2007. hang University in 2019. He is currently
[42] H. P. Xu, Z. H. Li, S. Li, et al., “A nonlocal noise reduction pursuing the Ph.D. degree in signal and
method based on fringe frequency compensation for SAR in- information processing in the School of
terferogram,” IEEE Journal of Selected Topics in Applied Electronic and Information Engineering,
Earth Observations and Remote Sensing, vol.14, Beihang University. His current research
pp.9756–9767, 2021. interests include SAR interferometry and
[43] G. Franceschetti, A. Iodice, M. Migliaccio, et al., “A novel Differential SAR interferometry.
across-track SAR interferometry simulator,” IEEE Transac-
(Email: zengguobing@buaa.edu.cn)
tions on Geoscience and Remote Sensing, vol.36, no.3,
pp.950–962, 1998.
[44] Z. L. Li, W. B. Zou, X. L. Ding, et al., “A quantitative LI Shuo received the M.S. de-
measure for the quality of InSAR interferograms based on gree from China University of Mining
phase differences,” Photogrammetric Engineering & Re- and Technologyì in 2015 and Ph.D. de-
mote Sensing, vol.70, no.10, pp.1131–1137, 2004. gree from the School of Electronic and
[45] S. Li, H. P. Xu, S. Gao, et al., “An interferometric phase Information Engineering, Beihang Uni-
noise reduction method based on modified denoising convo- versity in 2021. He is currently working
lutional neural network,” IEEE Journal of Selected Topics in the Nanjing Research Institute of Elec-
in Applied Earth Observations and Remote Sensing, vol.13, tronics Technology, and is mainly en-
pp.4947–4959, 2020. gaged in the design of space-based inter-
ferometric SAR system. (Email: shuo201@buaa.edu.cn)
XU Huaping received the B.S.
degree in electronic engineering in LI Shuang received the Ph.D.
1998 and the Ph.D. degree in communic- degree in communication and informa-
ation and information system in 2003 tion system from Beihang University in
both from Beihang University. She is 2013. She is currently a Researcher in
currently a Professor with the School of Beijing Institute of Radio Measurement.
Electronic and Information Engineering, She has published more than 20 academ-
Beihang University. She has published ic papers and applied for 5 patents. Her
more than 100 journal and conference pa- current research interests include space-
pers, and a research monograph about signal processing. Her cur- based interferometric SAR system, data
rent research interests include SAR interferometry, differential
processing and high-precision 3D information application techno-
SAR interferometry, image processing, and radar waveform
design. (Email: xuhuaping@buaa.edu.cn) logy. (Email: lishuang0108@sohu.com)

WANG Yuan (corresponding REN Chong recieved the B.S.


author) received the B.S. degree in degree in materials science and engineer-
School of Information and Communica- ing from University of Science and Tech-
tion Engineering from Communication nology Beijing in 2003 and Ph.D. degree
University of China, Beijing, China, in in materials science and engineering from
2019. She is currently working toward Tsinghua University in 2012. She is cur-
the Ph.D. degree with the School of Elec- rently a Deputy Chief Designer in the
tronic and Information Engineering, Bei- China Academy of Launch Vehicle Tech-
hang University. Her current research in- nology. Her current research interests fo-
terests include SAR interferometry, and interferometric SAR im- cus on thermal protection design for reusable launch verhicle.
age processing. (Email: wyuan@buaa.edu.cn) (Email: 674686864@qq.com)

You might also like