This work was supported in part by the national Natural
Science Foundation of China under Grant (NO.
61201055).
Application of Sparse Prior in Aperture Synthesis
Radiometric Imaging of Extended Radiation Source
He Fangmin, Wang Qian, Xiao Huan, Li Yi, Tang Jian, Meng Jin
National Key Laboratory for Vessel Integrated Power System Technology
Naval University of Engineering
Wuhan, China
hefangmin82@gmail.com
Abstract—Aimed at the extended source of earth thermal
radiation scene, the sparse prior is extracted from the transform
domain, and used in the statistical inversion approach (SIA) to
deal with the inverse problem in aperture synthesis radiometric
imaging of the extended source. As the transform basis, Laplace
basis, Fourier basis and Daubechies wavelet basis are proposed
to explore the implicit sparse prior about the extended source.
For the SIA, the image inversion of aperture synthesis
radiometers is recast as the statistical inference about the
hyperparameters based the sparse prior in the transform domain,
which can be automatically derived from an expectation
maximization (EM) algorithm. The simulations show that the
proposed SIA can improve the radiometric accuracy of the
reconstructed image by introducing the sparse prior as compared
to the traditional deterministic inversion approaches.
Index Terms—aperture synthesis radiometer, imaging, inverse
problem, sparse prior, statistical inversion approach.
I. INTRODUCTION
For passive microwave remote sensing from space, the
aperture synthesis radiometers (ASRs) are proposed in order to
provide high spatial resolution image when there is no need for
a very large and massive scanning antenna of the real-aperture
system [1]. The inversion of brightness temperature
distribution (BTD) from the visibilities measured by ASRs is
an ill-posed inverse problem [2], [3].
The approaches used in ASRs imaging are generally the
traditional deterministic inversion approaches (DIAs),
Tikhonov regularization, the truncated singular value
decomposition (TSVD) approach, and the band-limited
regularization [4]. DIAs do not explicitly refer to the statistical
properties of BTD prior. However, from the viewpoint of
statistics, the DIAs have implicit statistical properties with
regard to the BTD. According to the linear observation model,
the least square solution coincides with the maximum
likelihood estimate with white Gaussian noise. For example,
Tikhonov regularized solution can be interpreted statistically as
a Maximum A Posterior (MAP) estimate with a white noise
prior [5]. Given a nonuniform scene, the regularized solution
can be seen as a MAP estimate with erroneous BTD prior.
The statistical inversion approach (SIA) is aimed at
removing the ill-posedness about T and G by recasting the
inverse problem as a well-posed extension in a larger space of
probability distributions [5], [7]. At the same time, it allows us
to be explicit about the BTD prior that is often hidden in the
regularization approaches. The crucial task of ASRs inversion
is to extract and model the statistical properties of the BTD
priors as accurately as possible. In this paper, a combined
prior-based SIA has been presented. According to the SIA, the
inversion of ASRs is converted to a problem of hyperparameter
estimation. The effects of the BTD prior and the errors of
ASRs on the performance of the SIA have been analyzed.
II. ASR
S IMAGING BY STATISTICAL INVERSION APPROACH
The ASRs measure the correlation between the signals
collected by two spatially-separated antennas with overlapping
fields of view (FOV), yielding samples of complex visibilities
V(u), of the BTD T(ξ) of the scene under observation. The
relationship between V(u) and T(ξ) can be written as [3]
()
() () ()
()
ξξξuξξξu
ξ
ξu
defrFFTV
kj
j
kjjkkj
∫∫
≤
−
⎟
⎠
⎞
⎜
⎝
⎛
−−∝
1
2
2
0
*
2
1
~
π
. (1)
where u
kj
is the spatial frequency associated with the two
antennas k and j, the components ξ
1
= sinθcosφ and ξ
2
=
sinθsinφ of the angular position variable ξ are direction cosines
(θ and φ are the traditional spherical coordinates), F
k
(ξ) and
F
j
(ξ) are the normalized voltage patterns of the antennas k and j
(the superscript '*' indicates the complex conjugate),
kj
r
~
is the
so-called fringe-wash function, and f
0
= c/λ
0
is the central
frequency of observation.
Assuming M visibilities are measured, a series of
visibilities can be combined in the matrix equation V=GT+e,
where the M×P matrix G indicates the discrete model operator,
the column vector T denotes the discrete unknown BTD with P
pixels, and the P-dimensional column vector e denotes
visibility noise vector. Then, the corresponding inverse
problem aims at inverting the discrete version of Eq. (1) to
inverse the BTD T from the visibilities V with the disturbance
of noise.