Exploring the Spatial Correlation in Radio Tomographic Imaging by Block‐structured Sparse Bayesian Learning
Jiaju Tan,Xin Zhao,Xuemei Guo,Guoli Wang
DOI: https://doi.org/10.1049/sil2.12185
IF: 1.819
2023-01-01
IET Signal Processing
Abstract:IET Signal ProcessingVolume 17, Issue 2 e12185 ORIGINAL RESEARCHOpen Access Exploring the spatial correlation in radio tomographic imaging by block-structured sparse Bayesian learning Jiaju Tan, Corresponding Author Jiaju Tan judgetan7@163.com orcid.org/0000-0003-4322-4306 School of Mathematics and Computational Science, Wuyi University, Jiangmen, China Correspondence Jiaju Tan, School of Mathematics and Computational Science, Wuyi University, Jiangmen, China. Email: judgetan7@163.comSearch for more papers by this authorXin Zhao, Xin Zhao Institute of Robotics and Automatic Information System (IRAIS), Nankai University, Tianjin, China Tianjin Key Laboratory of Intelligent Robotic (tjKLIR), Nankai University, Tianjin, ChinaSearch for more papers by this authorXuemei Guo, Xuemei Guo School of Data and Computer Science, Sun Yat-sen University, Guangzhou, China Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, Sun Yat-sen University, Guangzhou, ChinaSearch for more papers by this authorGuoli Wang, Guoli Wang School of Data and Computer Science, Sun Yat-sen University, Guangzhou, China Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, Sun Yat-sen University, Guangzhou, ChinaSearch for more papers by this author Jiaju Tan, Corresponding Author Jiaju Tan judgetan7@163.com orcid.org/0000-0003-4322-4306 School of Mathematics and Computational Science, Wuyi University, Jiangmen, China Correspondence Jiaju Tan, School of Mathematics and Computational Science, Wuyi University, Jiangmen, China. Email: judgetan7@163.comSearch for more papers by this authorXin Zhao, Xin Zhao Institute of Robotics and Automatic Information System (IRAIS), Nankai University, Tianjin, China Tianjin Key Laboratory of Intelligent Robotic (tjKLIR), Nankai University, Tianjin, ChinaSearch for more papers by this authorXuemei Guo, Xuemei Guo School of Data and Computer Science, Sun Yat-sen University, Guangzhou, China Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, Sun Yat-sen University, Guangzhou, ChinaSearch for more papers by this authorGuoli Wang, Guoli Wang School of Data and Computer Science, Sun Yat-sen University, Guangzhou, China Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, Sun Yat-sen University, Guangzhou, ChinaSearch for more papers by this author First published: 22 February 2023 https://doi.org/10.1049/sil2.12185AboutSectionsPDF ToolsRequest permissionExport citationAdd to favoritesTrack citation ShareShare Give accessShare full text accessShare full-text accessPlease review our Terms and Conditions of Use and check box below to share full-text version of article.I have read and accept the Wiley Online Library Terms and Conditions of UseShareable LinkUse the link below to share a full-text version of this article with your friends and colleagues. Learn more.Copy URL Share a linkShare onFacebookTwitterLinkedInRedditWechat Abstract Radio Tomographic Imaging (RTI) is a low-cost computational imaging method realised by the Radio Frequency (RF) signal sensing. The target-induced shadowing effect in the RF sensing network is reconstructed as a probability image to estimate the target's position. Then, the RTI-based Device-free Localization (DFL) is becoming a promising research topic in the Location-based Services applications by the Internet of Things (IoT). However, the multipath interference in the RF sensing network often induces the imaging degradation and decreases the DFL accuracy. To deal with the multipath-induced imaging degradation, considering that the target's shadowing occupies a small spatial range in the RF network and expresses some spatial structure, this article explores the spatial correlation in the target's shadowing. Then, a new RTI reconstruction method based on the Structured Sparse Bayesian Learning is proposed to model the spatial correlation implied in the sparse target's shadowing image. Further, the localisation experiments in actual scenes are conducted to validate the utilisation of the spatial correlation in target's shadowing is able to improve the imaging quality of the RTI system by enhancing the robustness towards the multipath-induced imaging degradation. 1 INTRODUCTION Radio Tomographic Imaging (RTI), is a novel computational imaging technology to sense the environment through the wireless Radio-Frequency (RF) signal [1]. When the target enters the RF sensing network, which is covered by the RF-signal transmission defined as RF links, the target may induce the RF-link shadowing [2]. Then, the RTI system is to reconstruct the target's shadowing image from some quantitative measurements of the variation in RF-link propagation, such as the Received Signal Strength (RSS) [3]. As only the RSS measurements is use, the RTI system is not only low-powered but also provides good protection to the target's privacy [4]. Further, because the RF signal can pass through opaque obstacles or structures and penetrate the illumination or smokes, RTI is practical even in obstructed surroundings and through-wall scenes [5]. Thus, RTI is serviced in various Device-Free Localization (DFL) applications, such as emergency response [6], traffic and roadside surveillance [7], military reconnaissance [8] and healthcare for the elderly [9]. Due to the limited bandwidth of the RF signal, unexpected multipath propagation, such as the target's scattering and the obstacle-induced reflection that often causes multipath interference in RSS measurements, is an important problem in the RTI system [10]. These unreliable RSS measurements may induce the imaging degradation in the reconstructed shadowing image [11]. Thus, how to reduce the multipath-induced imaging degradation poses a significant challenge in the RTI system [12]. Addressing the above problems, many methods investigate how the RSS measurements are varied by the target's shadowing [13] and different multipath components, such as the electronic noise [14], scattering [15] and reflection [16]. Further, the RF-link fade level [17] and the multi-scale models [18] have proposed to precisely quantify the target's shadowing. By these models, the characteristics of the target's shadowing and multipath components have been deeply analysed to improve the RTI reconstruction performance. However, some detail problems such as the parameter pre-estimation for each concrete models in different scenes usually limit the universal application of these model-based methods [19]. In comparison to the above mentioned model-based methods, the reconstruction-based methods usually have a better universality as they reconstruct the target's shadowing from the RSS measurements by using the general models. Actually, the target may cause the shadowing of merely a few RF-links which are close to the target [20], while the spatial distribution of multipath noise is often unpredictable [21]. As the target's shadowing often occupies a small spatial area around the target [22], the RTI reconstruction is transformed into the sparse recovery [23]. However, the isolated multipath interference in the RF sensing network also induces the sparse imaging degradation [24]. Thus, an enhanced sparse RTI reconstruction method is needed for better dealing with the multipath-induced imaging degradation [25]. To decrease the multipath-induced imaging degradation in the reconstructed shadowing image, the sparse reconstruction methods with consideration to more information of the target's shadowing will be helpful. By utilising the prior distribution information to model the sparse signal and the measurement noise, Sparse Bayesian Learning (SBL) is an adaptive sparse-reconstruction method with the stronger sparsity-promoting ability [26], which is introduced into the RTI system [27]. Besides, many articles model the noise level of the environmental multipath to deal with the multipath-induced imaging degradation [28, 29]. However, in the actual RTI-based DFL applications, especially in the indoor and the multi-target scenes, the complex multipath interference is so unpredictable that the prior distribution of multipath interference is insufficient to quantify by using the simple noise models [30], and the distribution modelling of the multipath noise may cause amounts of calculation [31]. On the other hand, there exist some measurable characteristics and typical information in the distribution of the target's shadowing [32]. Moreover, SBL is a collectively learning method of the prior distribution of both the sparse signal and the isolated measurement noise [33]. Then, if the reasonable prior distributional information of the target's sparse shadowing in the RTI system is utilised under the framework of SBL, even the simple noise models are enough to quantify the multipath noise [34]. After that, the robustness of the RTI system towards the multipath-induced imaging degradation is also enhanced [35]. For the better reconstructed shadowing image by using the SBL methods, it will be helpful to explore the typical characteristics and prior information of the target's shadowing in the RTI system [36]. Based on the propagation principle of the RF signal [37], the target's shadowing merely affects the RF electromagnetic field which is spatially around the target [38]. Then, the spatial neighbourhood of the target's shadowing often contains some spatial distributional information of the target's shadowing [39, 40]. Besides, the spatial neighbourhood of the target's shadowing image often exhibits the spatial structural correlation [41], which is an important property to distinguish the target's sparse shadowing image with the multipath-induced isolated imaging degradation [42]. To analyse and model the spatial structural correlation in the sparse signal, Structured Sparse Bayesian Learning (SSBL) [43-45] was used to reconstruct the sparser results than the preliminary SBL method proposed in Ref. [27]. However, these SSBL methods have never been used in the RTI system to analyse the spatial correlation information of the target's sparse shadowing image. Besides, the SSBL methods are merely used to analyse the spatial correlation information in the one-dimensional sparse signal, while the spatial correlation of the target's shadowing image in the RTI system is two-dimensional. Therefore, to analyse the spatial correlation information implied in the target's shadowing image based on the framework of SSBL, the two-dimensional spatial correlation in the target's shadowing image should be explored. Based on the analysis above, this article tends to explore the spatial correlation of the target's shadowing in the RTI system. Then a spatial-correlation-based RTI reconstruction method under the framework of SSBL [43-45] is proposed. Further, the actual RTI-based DFL experiments that are conducted to evaluate the utilisation of the spatial correlation can deal with the multipath-induced imaging degradation. The major contributions of this article are summarised as follows: First, the spatial correlation of the target's shadowing in the RTI system was explored by the prior modelling of the two-dimensional block-structured sparsity in the target-induced shadowing image. Second, based on the above modelling of the structured spatial correlation, the RTI reconstruction algorithm under the framework of SSBL was proposed. It effectively reduced the multipath-induced imaging degradation and promote the imaging quality of the RTI system. Third, the real RTI-based localisation experiments, including the multi-target tests in cluttered indoor scenes, were carried out to evaluate the multipath-induced imaging degradation, which was effectively reduced by modelling the spatial correlation in the target's shadowing. This paper is organised as follows. The prerequisite knowledge of the multipath-induced imaging degradation in the RTI system are introduced in Section 2; the proposed RTI method for modelling the spatial correlation in the target's shadowing is presented in Section 3; RTI-based DFL experiments in the actual scenes are conducted, with adopting empirical evaluations to verify the RTI performance of the proposed method, comparing with some methods in use without the correlation analysis, in Section 4; the experimental results including the imaging quality, DFL accuracy and time complexity are discussed in Section 5; the conclusion is given in Section 6. 2 RADIO TOMOGRAPHIC IMAGING SYSTEM In the RTI system, M RF nodes are deployed along the perimeter of a two-dimensional monitored area at first. Each node transmits and receives the signal with others, then an RF sensing network is conducted. An example of the RF sensing network with M = 20 nodes and m = M(M − 1)/2 = 190 links are shown in Figure 1a. Besides, the target in the RF network may block the RF-signal transmission and induce the RF-link shadowing, resulting in the variation of RSS measurements. Then, the monitored area is divided into n = n1 × n2 virtual pixels, and an image X ∈ R n 1 × n 2 $X\in {\mathbb{R}}^{{n}_{1}\times {n}_{2}}$ of RF-link shadowing is constructed, as the example (n = 600, n1 = 30, n2 = 20) illustrated in Figure 1b. In actual computation, the reconstructed shadowing image X is vectorised into a shadowing vector x ∈ R n $\boldsymbol{x}\in {\mathbb{R}}^{n}$ . FIGURE 1Open in figure viewerPowerPoint An illustration of the Radio Frequency (RF) network and shadowing image The RSS variation in the RF-link sensing network (in Figure 1a) forms the RSS measurement vector y ∈ R m $\boldsymbol{y}\in {\mathbb{R}}^{m}$ , which is estimated as the weighted sum of the target's shadowing, according to y = Φ x + e , $\boldsymbol{y}={\Phi }\boldsymbol{x}+\boldsymbol{e},$ (1)where e ∈ R m $\boldsymbol{e}\in {\mathbb{R}}^{m}$ is the measurement noise; Φ is the projection model to quantify the RF-link shadowing, in which weight Φij evaluates the shadowing at pixel j to the RSS of the RF link i. The task in the RTI system is to use the RSS measurement y (in Figure 1a) to reconstruct the normalised probability image p x ^ | y $p\left(\hat{\boldsymbol{x}}\vert \boldsymbol{y}\right)$ as the reconstructed shadowing vector x ^ $\hat{\boldsymbol{x}}$ (in Figure 1b) by solving Equation (1), according to x ^ = arg min x ‖ y − Φ x ‖ 2 2 . $\hat{\boldsymbol{x}}=\underset{\boldsymbol{x}}{\mathrm{arg min}\,}\ \Vert \boldsymbol{y}-{\Phi }\boldsymbol{x}{\Vert }_{2}^{2}.$ (2)Then the target's position is estimated as the central coordinate of the pixel, the probability of which is the largest in the shadowing x ^ $\hat{\boldsymbol{x}}$ . However, the RTI reconstruction process is often an ill-posed inverse problem as Equation (1) are underdetermined with m < n, meaning that the multipath noise in the RSS measurements may be amplified and then induce the imaging degradation. Thus, to reduce the multipath-induced imaging degradation in x ^ $\hat{\boldsymbol{x}}$ , some extra regularisation conditions should be added in Equation (1) as follows, x ^ = arg min x ‖ y − Φ x ‖ 2 2 + R ( x ) , $\hat{\boldsymbol{x}}=\underset{\boldsymbol{x}}{\mathrm{arg min}\,}\ \Vert \boldsymbol{y}-{\Phi }\boldsymbol{x}{\Vert }_{2}^{2}+R(\boldsymbol{x}),$ (3)where R(x) is the additive auxiliary regularisation conditions. In fact, the target's shadowing often occupies only a small portion of the spatial region in the RF sensing network. Thus, the target's shadowing in the reconstructed image with a few of sparse pixels contain the target's position information. Then the RTI reconstruction process is transformed into an underdetermined problem based on the sparse recovery theory [23], according to x ^ = arg min x ‖ y − Φ x ‖ 2 2 + λ ‖ x ‖ 1 , $\hat{\boldsymbol{x}}=\underset{\boldsymbol{x}}{\mathrm{arg min}\,}\ \Vert \boldsymbol{y}-{\Phi }\boldsymbol{x}{\Vert }_{2}^{2}+\lambda \Vert \boldsymbol{x}{\Vert }_{1},$ (4)where λ is a regularisation constant to balance the reconstruction error and the prior sparse constraints. To improve the RTI quality by modelling the prior distribution of the target's shadowing and the multipath noise, SBL is introduced into the RTI system [27]. Then Heterogeneous SBL (HSBL) [28] and Enhanced SBL (ESBL) [29] are proposed by adaptively estimating the level of the multipath noise. However, the unpredictable multipath interference in the RF sensing network, including the RF scattering of targets and the RF reflection by obstacles, may accumulate and increase the RSS measurement noise and then induce the multipath-induced imaging degradation, as an example shown in Figure 2. Figure 2a and Figure 2b express the target-shadowed RF links and the reconstructed shadowing image by the original SBL method [27], respectively. Besides, to reflect the target-induced multipath interference such as scattering [18], five multipath RF links with low Signal-to-Noise Ratio (SNR) were added in the RF sensing network, as shown in Figure 2c compared to Figure 2a. However, the multipath-induced isolated imaging degradation may be wrongly reconstructed as the sparse shadowing by the SBL method, as shown in Figure 2d with comparison to Figure 2b. Thus, to decrease the multipath-induced sparse imaging degradation and enhance the RTI reconstruction quality, it is needed to improve the SBL method. FIGURE 2Open in figure viewerPowerPoint Multipath-induced imaging degradation in the Radio Tomographic Imaging (RTI) system In fact, in the SBL methods, both the distribution of the target's shadowing and the multipath noise are estimated. When the target's shadowing is more accurately modelled, the robustness to the multipath interference is able to be also enhanced even with the simple noise models. Thus, an improved SBL method by utilising more spatial information of the target is helpful to enhance the robustness of the imaging degradation and promote the reconstruction quality of the sparse shadowing image. Further, in consideration of that, the pixels adjacent to the target may be also affected by the target's shadowing. Representing the spatial correlation as shown in Figure 1b, this paper decides to explore the spatial correlation in the target's shadowing. 3 PROPOSED METHODS 3.1 Modelling of shadowing x and received signal strength measurements y The reconstructed sparse shadowing vector x is modelled by assuming that x follows the Gaussian distribution as x ∼ N 0 , α − 1 I n $\boldsymbol{x}\sim \mathcal{N}\left(\mathbf{0},{\boldsymbol{\alpha }}^{-1}{\boldsymbol{I}}_{n}\right)$ with α = α 1 , … , α n ⊤ $\boldsymbol{\alpha }={\left[{\alpha }_{1},\text{\ldots },{\alpha }_{n}\right]}^{\top }$ for the sparsity estimation. Then to model the block spatial correlation in the shadowing image X with n = n1 × n2 pixels, if the size of the block region for analysing the spatial correlation is d = d1 × d2, there are Q = Q1 × Q2 = (n1 − d1 + 1) × (n2 − d2 + 1) block regions in X. For example, if the block size is d = d1 × d2 = 2 × 2 = 4, the qth block Xq includes the pixels of q , q + n 1 , q + 1 , q + n 1 + 1 $\left[q,q+{n}_{1},q+1,q+{n}_{1}+1\right]$ in X, as shown in Figure 3. FIGURE 3Open in figure viewerPowerPoint Block spatial correlation of shadowing The qth block vector x q = x q , x q + 1 , x q + n 1 , x q + n 1 + 1 ⊤ ${\boldsymbol{x}}_{q}={\left[{x}_{q},{x}_{q+1},{x}_{q+{n}_{1}},{x}_{q+{n}_{1}+1}\right]}^{\top }$ of the qth block region X q ∈ R d 1 × d 2 ${X}_{q}\in {\mathbb{R}}^{{d}_{1}\times {d}_{2}}$ in the shadowing vector x is assumed to follow the Gaussian distribution, according to x q ∼ N 0 , ζ q B q , q = 1 , … , Q , ${\boldsymbol{x}}_{q}\sim \mathcal{N}\left(\mathbf{0},{\zeta }_{q}{B}_{q}\right),\quad q=1,\text{\ldots },Q,$ (5)as a regularisation condition R(x) in Equation (3). The positive definite matrix B q ∈ R d 1 d 2 × d 1 d 2 ${B}_{q}\in {\mathbb{R}}^{{d}_{1}{d}_{2}\times {d}_{1}{d}_{2}}$ and ζ q ∈ R d 1 d 2 × d 1 d 2 ${\zeta }_{q}\in {\mathbb{R}}^{{d}_{1}{d}_{2}\times {d}_{1}{d}_{2}}$ in Equation (5) reflect the structure parameter and the coefficient of spatial correlation in the qth block vector xq respectively. Besides, the shadowing x ∈ R n $\boldsymbol{x}\in {\mathbb{R}}^{n}$ follows x ∼ N 0 , Σ x $\boldsymbol{x}\sim \mathcal{N}\left(\mathbf{0},{{\Sigma }}_{\boldsymbol{x}}\right)$ , where the diagonal direction of Σx reflects the overlapping covariance sub-matrix ζqBq of the qth block Xq, as the example shown in Figure 4a, with the block size as d1 = d2 = 2. FIGURE 4Open in figure viewerPowerPoint Expansion of block covariance matrix Besides, the measurement noise e is assumed to follow the Gaussian distribution with the variance of σ, according to e ∼ N 0 , σ I m $\boldsymbol{e}\sim \mathcal{N}\left(\mathbf{0},\sigma {\boldsymbol{I}}_{m}\right)$ where Im stands for the identity matrix with the order of m. Then the measurement vector y also follows the Gaussian distribution with the variance of σ, as y | x ∼ N Φ x , σ I m $\boldsymbol{y}\vert \boldsymbol{x}\sim \mathcal{N}\left({\Phi }\boldsymbol{x},\sigma {\boldsymbol{I}}_{m}\right)$ . 3.2 Modelling of the spatial correlation in shadowing x However, in the covariance matrix Σx of the shadowing vector x with the two-dimensional correlation shown in Figure 3, as shown in Figure 4a, there are overlaps between the adjacent sub-matrix in the total covariance matrix Σx, such as ζqBq and ζq+1Bq+1. The overlaps greatly increase the computational time in the matrix inversion of this non-diagonal covariance matrix Σx when the size n of the shadowing vector x ∈ R n $\boldsymbol{x}\in {\mathbb{R}}^{n}$ is large [43, 44]. Besides, the covariance submatrix ζqBq of the two-dimensional block region Xq is actually divided into d = d1 × d2 parts, as the same colour region shown in Figure 4a. Further, there are additional overlaps between the adjacent blocks in same row of total covariance matrix Σx, such as ζqBq and ζ q + n 1 B q + n 1 ${\zeta }_{q+{n}_{1}}{B}_{q+{n}_{1}}$ , which have never been analysed in [43, 44]. To reduce the computational complexity induced by the overlaps in the covariance matrix Σx, a block diagonal covariance matrix without the overlaps are needed [45]. To extend the covariance matrix into the block diagonal, the shadowing vector x ∈ R n $\boldsymbol{x}\in {\mathbb{R}}^{n}$ , with the size n = n1 × n2 and block size of spatial correlation d = d1 × d2, is transformed into a representation vector z ∈ R d 1 Q 1 × d 2 Q 2 ∼ N 0 , Σ z 0 $\boldsymbol{z}\in {\mathbb{R}}^{{d}_{1}{Q}_{1}\times {d}_{2}{Q}_{2}}\sim \mathcal{N}\left(\mathbf{0},{{\Sigma }}_{\boldsymbol{z}0}\right)$ by changing the positions of pixels in x, in which the covariance matrix Σz0 of z is the block diagonal, according to Σ z 0 = diag ζ 1 B 1 , … , ζ Q B Q ∈ R d 1 d 2 Q × d 1 d 2 Q , ${{\Sigma }}_{\boldsymbol{z}0}=\mathrm{diag}\left({\zeta }_{1}{B}_{1},\text{\ldots },{\zeta }_{Q}{B}_{Q}\right)\in {\mathbb{R}}^{{d}_{1}{d}_{2}Q\times {d}_{1}{d}_{2}Q},$ (6)as there are no overlaps between different covariance sub-matrix ζ q B q ∈ R d 1 d 2 , q = 1 , … , Q ${\zeta }_{q}{B}_{q}\in {\mathbb{R}}^{{d}_{1}{d}_{2}},q=1,\text{\ldots },Q$ in the covariance matrix Σz0 of z. Then, the example in Figure 3 with the block size d = d1 × d2 = 4, the covariance matrix Σx, Σz of the shadowing vector x and representation vector z are shown in Figure 4a and Figure 4b, respectively. Besides, the transformation from the shadowing vector x ∈ R n $\boldsymbol{x}\in {\mathbb{R}}^{n}$ to the representation vector z ∈ R d 1 Q 1 × d 2 Q 2 $\boldsymbol{z}\in {\mathbb{R}}^{{d}_{1}{Q}_{1}\times {d}_{2}{Q}_{2}}$ is used to move the pixels with the spatial correlation into the same block, defined as x = ∑ q = 1 Q E q z q . $\boldsymbol{x}={\sum }_{q=1}^{Q}{E}_{q}{\boldsymbol{z}}_{q}.$ (7)The element in the representation of z q ∈ R d 1 d 2 ${\boldsymbol{z}}_{q}\in {\mathbb{R}}^{{d}_{1}{d}_{2}}$ satisfies E z q = 0 , E z q , z r = δ q , r ζ q B q , $E\left\{{\boldsymbol{z}}_{q}\right\}=0,E\left\{{\boldsymbol{z}}_{q},{\boldsymbol{z}}_{r}\right\}={\delta }_{q,r}{\zeta }_{q}{B}_{q},$ (8)where δ q , r = 1 , q = r , 0 , o t h e r w i s e . ${\delta }_{q,r}=\left\{\begin{array}{rl}\hfill 1,\quad & q=r,\hfill \\ \hfill 0,\quad & \mathrm{o}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{r}\mathrm{w}\mathrm{i}\mathrm{s}\mathrm{e}.\hfill \end{array}\right.$ Then z = z 1 ⊤ , … , z Q ⊤ ⊤ $\boldsymbol{z}={\left[{\boldsymbol{z}}_{1}^{\top },\text{\ldots },{\boldsymbol{z}}_{Q}^{\top }\right]}^{\top }$ and the matrix E q ∈ R n d 1 d 2 ${E}_{q}\in {\mathbb{R}}^{n{d}_{1}{d}_{2}}$ is obtained by substituting the element 0 with 1 at the coordinates of (q + k1, q + k2), k1 = 0, …, d1 − 1, k2 = 0, …, d2 − 1 in the zero matrix, so that the pixels with the spatial correlation is moved into the same block. Then the measurement model Equation (1) is transformed into y = Φ x + e = Φ ∑ q = 1 Q E q z q + e = A z + e , $\boldsymbol{y}={\Phi }\boldsymbol{x}+\boldsymbol{e}={\Phi }{\sum }_{q=1}^{Q}{E}_{q}{\boldsymbol{z}}_{q}+\boldsymbol{e}=A\boldsymbol{z}+\boldsymbol{e},$ (9)where A = [A1, …, AQ] with Aq = ΦEq, q = 1, …, Q, and the covariance matrix Σz0 of z is diagonal. Besides, the representation vector z is also sparse and can be estimated by changing Equation (4) into z ^ = arg min z ‖ y − A z ‖ 2 2 + λ ‖ z ‖ 1 . $\hat{\boldsymbol{z}}=\underset{\boldsymbol{z}}{\mathrm{arg min}\,}\ \Vert \boldsymbol{y}-A\boldsymbol{z}{\Vert }_{2}^{2}+\lambda \Vert \boldsymbol{z}{\Vert }_{1}.$ (10) 3.3 Solving the spatial correlation in shadowing x To solve Equation (10) by SSBL [43-45], the vector z is assumed to follow the Gaussian distribution as z ∼ N 0 , Σ z 0 $\boldsymbol{z}\sim \mathcal{N}\left(\mathbf{0},{{\Sigma }}_{\boldsymbol{z}0}\right)$ , with the qth block of z satisfying z q ∼ N 0 , ζ q B q , q = 1 , … , Q ${\boldsymbol{z}}_{q}\sim \mathcal{N}\left(\mathbf{0},{\zeta }_{q}{B}_{q}\right),q=1,\text{\ldots },Q$ . Besides, as the RSS measurement vector y | z ∼ N A z , σ I m $\boldsymbol{y}\vert \boldsymbol{z}\sim \mathcal{N}\left(A\boldsymbol{z},\sigma {\boldsymbol{I}}_{m}\right)$ satisfies the Gaussian distribution, then according to p(z|y) ∝ p(y|z)p(z) based on the Bayesian Theorem, the posterior probability of the vector z also follows the Gaussian distribution as z ^ | y ∼ N μ z , Σ z $\hat{\boldsymbol{z}}\vert \boldsymbol{y}\sim \mathcal{N}\left({\boldsymbol{\mu }}_{\boldsymbol{z}},{{\Sigma }}_{\boldsymbol{z}}\right)$ . By using the Type-II Maximum Likelihood Estimation based on the Expectation-Maximisation, the expectation μz of z is estimated by μ ^ z ← σ Σ z 0 − 1 + A ⊤ A − 1 A ⊤ y = Σ z 0 A ⊤ σ I + A Σ z 0 A ⊤ − 1 y , $\begin{array}{rl}\hfill {\hat{\boldsymbol{\mu }}}_{\boldsymbol{z}}& {\leftarrow}{\left(\sigma {{\Sigma }}_{\boldsymbol{z}0}^{-1}+{A}^{\top }A\right)}^{-1}{A}^{\top }\boldsymbol{y}\hfill \\ \hfill & ={{\Sigma }}_{\boldsymbol{z}0}{A}^{\top }{\left(\sigma I+A{{\Sigma }}_{\boldsymbol{z}0}{A}^{\top }\right)}^{-1}\boldsymbol{y},\hfill \end{array}$ (11)where the covariance matrix Σz of z is represented as follows: Σ ^ z ← Σ z 0 − 1 + 1 σ A ⊤ A − 1 = Σ z 0 − Σ z 0 A ⊤ σ I + A Σ z 0 A ⊤ − 1 A Σ z 0 . $\begin{array}{rl}\hfill {\hat{{\Sigma }}}_{\boldsymbol{z}}& {\leftarrow}{\left({{\Sigma }}_{\boldsymbol{z}0}^{-1}+\frac{1}{\sigma }{A}^{\top }A\right)}^{-1}\hfill \\ \hfill & ={{\Sigma }}_{\boldsymbol{z}0}-{{\Sigma }}_{\boldsymbol{z}0}{A}^{\top }{\left(\sigma I+A{{\Sigma }}_{\boldsymbol{z}0}{A}^{\top }\right)}^{-1}A{{\Sigma }}_{\boldsymbol{z}0}.\hfill \end{array}$ (12) Thereafter, the marginal likelihood of the RSS measurements y with respect to the hyper-parameters ζq, Bq, σ is formulated as follows: L ζ q , B q , σ = − log y ; ζ q , B q , σ = log | Σ y | + y ⊤ Σ y − 1 y . $\mathcal{L}\left({\zeta }_{q},{B}_{q},\sigma \right)=-\mathrm{log}\left(\boldsymbol{y};{\zeta }_{q},{B}_{q},\sigma \right)=\mathrm{log}\,\vert {{\Sigma }}_{\boldsymbol{y}}\vert +{\boldsymbol{y}}^{\top }{{\Sigma }}_{\boldsymbol{y}}^{-1}\boldsymbol{y}.$ (13)By setting ∂ L / ∂ d q = 0 $\partial \mathcal{L}/\partial {d}_{q}=0$ , the correlation coefficient ζq is updated by ζ ^ q ← Tr B − 1 Σ z q + μ z q μ z q ⊤ d 1 d 2 , q = 1 , … , Q , ${\hat{\zeta }}_{q}{\leftarrow}\frac{\mathrm{Tr}\left[{B}^{-1}\left({{\Sigma }}_{\boldsymbol{z}}^{q}+{\boldsymbol{\mu }}_{\boldsymbol{z}}^{q}{\left({\boldsymbol{\mu }}_{\boldsymbol{z}}^{q}\right)}^{\top }\right)\right]}{{d}_{1}{d}_{2}},\quad q=1,\text{\ldots },Q,$ (14)where Tr is the trace operator of the matrix, μ z q ${\boldsymbol{\mu }}_{\boldsymbol{z}}^{q}$ and Σ z q ${{\Sigma }}_{\boldsymbol{z}}^{q}$ are the expectation vector and the covariance matrix of the qth block zq. 3.4 Acceleration of solving the spatial correlation However, when the correlation coefficient ζq of each block vector xq is estimated,