1. Introduction
Unavoidable interference between SAR systems and imaging scenes, interruptions in SAR systems for different purposes, or new SAR mission requirements will result in the azimuth missing data (AMD) [
1,
2]. If the conventional SAR imaging algorithm is used directly in the AMD echo case, false targets or severe defocus will be produced in the final imaging results [
3].
In order to overcome the AMD-SAR imaging challenge, an auto-regressive linear prediction approach used initially in the discontinuous aperture SAR imaging [
4]. However, it only improves image quality if the Azimuth Missing Ratio (AMR) is below 30%. The equal-gap AMD-SAR imaging problem was solved by P. Stoica and J. Li. They proposed the Gapped-data APES (GAPES) algorithm based on the Amplitude and Phase EStimation (APES) algorithm [
5,
6,
7]. In order to enhance AMD-SAR imaging performance in random AMD conditions, they took advantage of the Expectation-Maximization algorithm and then further presented the Missing-data APES (MAPES) algorithm [
8]. However, its reliability decreases rapidly when the AMR increases, and the computational complexity is relatively expensive. To address this problem under the high AMR conditions, a random Missing-data Iterative Adaptive algorithm (MIAA) was proposed in [
9]. The maximum AMR threshold can achieve nearly 80%. Compared with the MAPES, the MIAA’s recovery performance is greatly improved when the AMR is higher than 60%. However, due to the fact that the MIAA involves numerous matrix inversions and iterations, its computing cost will be insurmountable for large-scene AMD-SAR imaging.
The partial data SAR imaging problem has been addressed from a new perspective since the Compressed Sensing (CS) technology was proposed [
10,
11]. Various CS-based sparse SAR imaging algorithms and methods have been proposed and improved in recent years [
12,
13,
14]. A segmented reconstruction algorithm for the large-scene sparse SAR imaging was proposed in [
15]. The whole scene is split into a set of small sub-scenes. With the appropriate increase in the segment number, the reconstruction time and running memory can be greatly reduced. Additionally, ref. [
16] proposed an improved method to speed up the sparse SAR imaging and reduce the memory requirement using the Non-Uniform Fast Fourier Transform (NUFFT).It applies interpolation coefficients instead of multiplication of observation matrices and vectors, leading to a smaller computational complexity and memory usage. Furthermore, since the strong scattering points are rebuilt directly, the imaging accuracy will be severely degraded under low echo signal-to-noise ratio (SNR).
For this problem, an Azimuth Missing Data Imaging Algorithm (AMDIA) was proposed in 2018 [
17]. It estimates and recovers the full echo of sparse targets from the AMD echo. The CS methods cannot reconstruct the complete SAR echo in the time domain because it is not sparse. Hence, influenced by PFA’s motion compensation approach [
18,
19], Literature [
17] discovered that multiplying the dense SAR echo with a Phase Compensation Function (PCF) in the range-frequency domain can yield a sparser signal in the Doppler domain. Next, a phase-compensated complete echo can be recovered from the phase-compensated AMD echo using the CS method. Then, by multiplying the phase-compensated complete echo with the conjugate of the previous PCF, the complete echo can be estimated. Lastly, using the traditional SAR imaging algorithms, the final image can be focused via the estimated full echo. Compared with the sparse SAR imaging algorithms, the AMDIA can obtain an excellent-focused image even at low SNR due to the two-dimensional Matched Filtering process [
3]. Its improved algorithms have developed rapidly in these years. K. Liu improved the imaging capabilities of the AMDIA by extending it into the spaceborne FMCW SAR system [
20,
21]. J. Wu suggested a sparsity adaptive StOMP algorithm for AMD-SAR imaging [
22]. It exhibits excellent recovery performance when the prior sparsity is unknown. In 2022, we proposed a Moving Target AMD-SAR Imaging (MTIm-AMD) method based on the AMDIA [
23]. Since the motion parameters are considered, the PCF is modified to be more efficient, and hence the moving target can be well-focused in the AMD case. Moreover, we proposed an Enhancement AMDIA (EnAMDIA) to improve the AMD-SAR imaging performance in [
24]. The EnAMDIA recovers the RCMC echo instead of the time domain echo. Therefore, it demonstrates a more accurate recovery and a more moderate computational burden.
However, the PCFs of all the above-mentioned AMD-SAR imaging algorithms are designed based on one reference point, which is generally regarded as the scene centroid. Therefore, the phase compensation error for the non-centered targets will increase significantly with the expansion of the imaging scene. Once the imaging scene is larger than the limit of the focusable region, the PCF of the State-Of-the-Art (SOA) AMDIA will result in unsatisfactory sparsity of the phase-compensated signal. Therefore, the estimation accuracy of the complete echo will decrease. It indicates that the imaging performance of SOA-AMDIA is unsatisfactory when the imaging scene is relatively large.
Therefore, to enlarge the maximum focusable region under the AMD conditions, an improved Sub-echo Segmentation and Reconstruction AMDIA (SSR-AMDIA) is proposed in this paper. We consider enhancing the phase-compensated signal’s sparsity and then enlarging the imaging scene limits in azimuth and range direction, respectively. First, we apply RCMC processing on the raw AMD echo and then design a new PCF for AMD-RCMC echo. Since the range migration is removed, a sparser signal can be obtained along the azimuth direction. Subsequently, the AMD-RCMC echo is split into a series of AMD-RCMC sub-echoes along the range direction. Each sub-scene’s centroid is regarded as the reference point for phase compensation. Instead of designing a PCF for the whole scene, many PCFs are redesigned for each sub-scene. Therefore, each phase-compensated RCMC (PC-RCMC) sub-echo is sparser, which implies that the complete RCMC sub-echoes can be estimated more precisely. Finally, by combining the reconstructed RCMC sub-echoes, a reliable complete RCMC echo is obtained. A superior imaging result of the edge targets can be obtained via azimuth compression.
The main innovations of the article consist in:
We first rebuilt the full RCMC echo rather than the full raw echo. The SOA-AMDIA only focuses on reconstructing the full raw echo before range compression and RCMC processing, resulting in an inaccurate reconstruction of azimuth far-field targets. Thus, the proposed algorithm first eliminates the negative effect of range migration on echo recovery. It significantly reduces the azimuth far-field target’s residual phase error, and expands the azimuth maximum Depth of Focus (DOF) of the sparse domain signal. Additionally, the computational cost can also be reduced if the range direction’s targets are adequately sparse.
We first exploited range segmentation to improve the SOA-AMDIA. Instead of using one PCF for the whole imaging scene, we redesigned a series of PCFs for each sub-scene. It ensures the significant reduction of the range far-field target’s residual phase error, and the imaging range limits can be eliminated with a reasonable segmentation strategy.
We also carried out the mathematical derivation for the two-dimensional maximum DOFs of the proposed algorithm. The advantage of the proposed SSR-AMDIA over SOA-AMDIA for the imaging scene scope is theoretically verified.
The rest of this article is organized as follows. In
Section 2, the SAR echo models are introduced. In
Section 3, the proposed SSR-AMDIA is derived in detail. In
Section 4, the azimuth maximum DOF, range segmentation strategy and the computational complexity of the proposed algorithm are analyzed and mathematically derived. The findings of the simulation and measured experiment are shown and discussed in
Section 5. Finally,
Section 6 serves as our conclusion.
3. Sub-Echo Segmentation and Reconstruction Azimuth Missing Data SAR Imaging Algorithm
The detailed steps of the proposed SSR-AMDIA are demonstrated in
Figure 2. First of all, the raw AMD echo is range compressed, and the range cells migration is corrected. Then, the entire AMD-RCMC echo is split into a series of AMD-RCMC sub-echoes along the range direction. Next, as the critical step, a set of PCFs are redesigned based on the RCMC sub-echoes. By multiplying each RCMC sub-echo with its corresponding PCF in the range-frequency domain, each complete PC-RCMC sub-echo can obtain a sparse representation in the Doppler domain. Subsequently, the accurate estimations of the complete PC-RCMC sub-echoes are recovered from the AMD-PC-RCMC sub-echoes using the CS method. In this article, we employ the Generalized Orthogonal Matching Pursuit (GOMP) algorithm [
26] to reconstruct the full PC-RCMC sub-echoes in order to remove any potential error impacts of various CS techniques, as in the SOA-AMDIA [
3,
17,
20,
27]. Then, the accuracy estimations of the complete RCMC sub-echoes can be obtained by multiplying the complete PC-RCMC sub-echoes with the conjugation of the previously mentioned PCFs. Finally, by combining each reconstructed RCMC sub-echo, the reconstructed RCMC echo of the entire imaging scene can be acquired, and then after azimuth compression, a satisfied imaging result can be obtained.
Next, we describe the specific derivation steps of the proposed SSR-AMDIA in detail.
3.1. Range Compression and Range Cell Migration Correction
First,
should accomplish the range Fourier transform to obtain the range-frequency domain signal
, which can be expressed as
where
denotes the range frequency.
Then, the range compressed signal
can be obtained after range compression, and the range Doppler domain signal
can be presented as
where
stands for the sinc function and
is the azimuth frequency. The Doppler instantaneous distance
can be written as
where
is shortest distance between the platform and the target.
is Doppler chirp rate. The second term of (
11) is the range cells migration term.
After RCMC and the azimuth Inverse Fourier Transform (AIFT), the AMD-RCMC signal
is demonstrated as
3.2. Reconstructing the Sub-Echoes
To reduce the residual phase error of the range far-field targets, the AMD-RCMC echo is split into
K sub-patches along the range direction. Since target range locations are determined after RCMC, the range segmentation will not distort the adjacent sub-scenes. To redesign the more effective PCFs,
k-th AMD-RCMC sub-echo
transforms to range-frequency domain, which is described as
Thus,
k-th redesigned PCF
is defined as
where
is the shortest slant range of
k-th reference point
. The
represents the slant range between
and the moving platform
. It can be expressed as
Next, the azimuth missing redesigned PCF
is acquired based on (
6), that is
The PCF designation is the key step of the proposed SSR-AMDIA. A sparse PC-RCMC sub-echo
that is likewise the waiting-recovering signal may be generated in the Doppler domain by multiplying
by
, which is represented by
where
and
are azimuth Fast Fourier Transform and range Inverse Fast Fourier Transform, respectively.
Since the main purpose of the proposed SSR-AMDIA is to reconstruct , its sparsity is vital for the signal reconstruction.
To evaluate the focusing performance of the redesigned PCFs,
results obtained by different methods are shown in
Figure 3. There are nine targets in the imaging scene, and the SOA-AMDIA’s
is shown in
Figure 3a. Obviously, only the center point is well-focused. The significant defocus can be easily found on the edge targets. Therefore, the effectiveness of SOA-AMDIA’s PCF for larger scenarios is limited. Moreover,
Figure 3b is imaged using the proposed SSR-AMDIA before the sub-echo segmentation. Compared with the
Figure 3a, the azimuth far-field targets at the center range are well-focused. However, it still cannot remove the residual phase error caused by the range differences. The range edge targets are still defocused. When a series of segmented
are used, the most sparse
can be obtained by observing
Figure 3c. The two-dimensional residual phase errors of the borderline targets are significantly reduced. The focusing performance of the redesigned PCFs is validated.
Next, the detailed reconstruction steps of the proposed SSR-AMDIA are introduced as follows. Firstly, the small size phase-compensated signal
can be demonstrated as the same as (
7)
must be segmented into several one-dimensional range signals in order to accommodate the one-dimensional signal recovery processing. The
q-th (
) range signal can be expressed as
, where
denotes the number of entire range gates. In the proposed SSR-AMDIA, the
q-th estimated range signal
is regarded as the signal
in CS method, while
is considered as the compressed signal vector
. Accordingly, since
is direct sparse,
is understood as being the sensing matrix
. First, the complete AIFT matrix
is illustrated as
Similar to the Equation (
6) and (
16), the partial missing AIFT matrix
is obtained by
Similar to (
7), the small size AIFT matrix
is gained by
Its size is equal to
. Consequently, the sub-echoes reconstructing process can be formulated as
where
denotes the threshold value.
To eliminate the possible error effects of different recovery methods, the estimated
is reconstructed using the GOMP algorithm in this article.
Table 1 or [
26] contains the specific steps for the GOMP algorithm.
Once the reconstructed one-dimensional signals
are combined, the
k-th segment sub-echo
can be acquired. It follows that the conjugation of
must be compensated in order to obtain the
k-th segment
, which is represented by
3.3. Combining the Sub-Echoes and Entire Scene Imaging
To obtain the reconstructed complete RCMC echo
of entire imaging scene, a series of
are combined in sequence, that is
where
denotes the combination operation.
The normalized recovery error results between
and
are illustrated in
Figure 4.
Figure 4a is obtained using the SOA-AMDIA and
Figure 4b is obtained using the proposed SSR-AMDIA. In order to quantitatively evaluate the reconstruction performance, the average normalized recovery error
is defined as
where
represents the range cells set corresponding to the presence of targets after RCMC, and
denotes the average function. According to
Figure 4,
of
Figure 4a is equal to 0.087, which is almost half to that of
Figure 4b, which is equal to 0.179. Therefore, the proposed SSR-AMDIA obviously has a better reconstruction performance.
Finally, since a satisfied complete RCMC echo of the entire scene is estimated, an excellent-focused image can be obtained via azimuth compression.
4. Parameter Analysis
As mentioned before, we consider extending the maximum DOFs of SOA-AMDIA in azimuth and range directions, respectively. Thus, the range segmentation is applied in the proposed SSR-AMDIA. However, the identical segmentation idea cannot be exploited to enlarge the azimuth imaging scope. Azimuth segmentation may result in too few azimuth samples available in some sub-apertures, especially in the case of random missing. Sub-apertures with a high AMR will be detrimental to complete sub-aperture reconstruction [
3]. Hence, although the proposed SSR-AMDIA extends the azimuth maximum DOF, it has a limitation. Moreover, there is no limit to the maximum range imaging scope of the proposed SSR-AMDIA under a proper segmentation. We next analyze the azimuth maximum DOF and range segmentation strategy of the proposed algorithm. The computational complexity advantage is also investigated.
4.1. Azimuth Maximum Depth of Focus
In 2005, B. Rigling analyzed the imaging scene size limits for the PFA in the monostatic SAR system situation [
28]. When the absolution value of residual Quadratic Phase Error after the PFA (PFA-QPE)
, the PFA’s maximum well-focused radius
can be expressed as
where
and
denote the azimuth resolution and wavelength, respectively. The prerequisite for applying (
26) is that there are no unknown motion measurement errors during the flight. Otherwise, the phase errors will lead to an irreparable defocus to the image long before the far-field approximation (
26) breaks down.
In 2016, L. Gorham and B. Riging further derived the imaging scene size limits for the PFA in the linear flight case [
18]. The residual PFA-QPE
can be written as
where
is the length of synthetic aperture and
can be calculated as
The value of
is related to the position of targets. Assume
GHz,
m,
result in different positions is shown in
Figure 5a. The maximum focus area is limited to a circle of which radius equals 148.3 m after the PFA imaging. However, since the SOA-AMDIA only utilizes the phase compensation process of PFA, the maximum DOF of
decreases rapidly [
29,
30].
To illustrate this phenomenon more clearly, we form
using the SOA-AMDIA with a regular grid in actual coordinates in
Figure 5b. Due to the impact of the range cells migration and the residual phase error, the azimuth edge targets of
Figure 5b are distorted and defocused severely while only the centroid is excellent-focused. Although it is clearly sparser than the dense time domain signal, the maximum DOF of
is only 25 m, much less than 148.3 m. The aforementioned conclusion has been verified in simulation. The low-quality
will impair full echo reconstruction. The applicable imaging scope of the SOA-AMDIA will be significantly reduced.
In response to this problem, the range cells migration effect of the raw echo is removed by the proposed SSR-AMDIA. Then, a new PCF is redesigned to enhance the sparsity of and to extend the maximum azimuth DOF.
First, the residual phase error
between the scattering point and the reference point can be represented as
where
is the differential slant range, and the expressions of
and
can be found in (
3) and (
15), respectively. The second-order Taylor series approximation of
is performed as
Let the monostatic SAR system moves at a constant speed in the azimuth direction (
direction), the first and second derivatives of
and
can be calculated as
where
and
.
The residual QPE after RCMC and phase compensation (PC-RCMC-QPE)
is thus computed as
Suppose
GHz,
m, PC-RCMC-QPE
result in different positions is shown in
Figure 6a and we form the calculated
by (
17) when
with a regular grid in actual coordinates in
Figure 6b.
Figure 6b is significantly sparser than
Figure 5b. The azimuth maximum DOF is substantially enlarged. Let
and
, the azimuth coordinate
can be calculated as
Substitute the above-mentioned simulation parameters into (
33), then two farthest points,
and
, that can achieve an excellent focus are obtained. Thus, the azimuth maximum DOF
m, which is much larger than that in
Figure 5b.
A strong sparsity of
will facilitate the signal reconstruction. Hence, the azimuth maximum DOF of
given by (
33) is conservative compared to that of the final image. We found that
in the final image can still be accurately focused using the proposed algorithm with the above-mentioned simulation parameters. Conversely, the azimuth maximum DOF using SOA-AMDIA only reaches about
m under identical simulation conditions.
4.2. Range Segmentation Strategy
Figure 6b exhibits the limited range maximum DOF of
as well. Motivated by [
15,
19], the range segmentation is applied to expand the imaging scene scope in range direction. By observing Equation (
32), the contour shape of
is a circle in the plane of
, of which radius equals
. Since the imaging scene is limited in the plane of
,
can only be obtained at two points
and
. Similarly, the contour shapes of
and
are two circles in the plane of
with different radius. It indicates that only four targets can ensure
in the
case.
Let
and
, the range coordinate
can be calculated as
Hence,
,
,
and
are obtained by substituting the above-applied simulation parameters into (
34). Obviously, only
and
are located in the imaging scene. The range maximum DOF is
m.
Obviously, when there is no target existing in the q-th range profile, the value of is equal to zero. The zero row vectors do not need to be reconstructed. We assume that the number of the range profiles that exist targets equals . Therefore, the range profiles of AMD-RCMC echo should be split into K sub-patches. The size of each sub-echo equals , where and denotes the interval of adjacent range cells.
Figure 7 demonstrates the simulated images of
based on the aforementioned range segmentation strategy. By adequately segmenting the imaging scene within the azimuth maximum DOF, all targets corresponding to
can be well-focused. The proposed SSR-AMDIA can guarantee the estimation accuracy of the complete echo in a larger imaging scene.
Moreover, since the proposed algorithm performs RCMC on the raw data before the range segmentation, the range position information of the sparse target is determined. Thus, it will not deteriorate the reconstruction error.
4.3. Computational Complexity
Assume that the number of entire range gates equals and sparse targets spread out along the range direction (). Since the range position information of the target is hidden in all range profiles before range compression and RCMC, no matter how many sparse targets exist, times reconstructions are required using the SOA-AMDIA. Contrarily, the proposed SSR-AMDIA only needs times reconstructions to complete echo reconstruction. Suppose the computational complexity of one-dimensional GOMP algorithm equals , then the computational complexity of the SOA-AMDIA is , while that of the proposed SSR-AMDIA is equal to . It implies that the computational complexity of the proposed SSR-AMDIA is times to that of SOA-AMDIA. Obviously, when fewer range profiles exist sparse targets, the computational complexity advantage of the proposed SSR-AMDIA is prominent.