In a general continuous-form (functional) formalism, a random temporal-spatial realization of the data field, u, is considered to be created by some continuous distribution of the far-distant radiation/scattering sources e as plane or spherical wavefronts, which sweep across the radar sensor array (moving antenna in the case of SAR). These fields satisfy an operator-form linear stochastic equation (the so-called equation of observation (EO) [10,21]): (1) u ( p ) = ( 𝒮 e ( r ) ) ( p ) + n ( p ) ;                         e ( r )   ∈   𝔼 ( R ) ;                       u ( p ) ,     n ( p )   ∈   𝕌 ( P )where p = (t, ρ) defines the time (t)—space (ρ) points in the temporal-spatial observation domain p ∈ P = T× P (t ∈ T, ρ ∈ P) (in the SAR case, ρ = ρ(t) specifies the carrier trajectory, i.e., the array is composed of the moving antenna); e(r) represents the random scene reflectivity over the probing surface; r is a vector of the scan parameters, usually the polar, cylindrical or Cartesian coordinates of the probing surface R; n corresponds to the additive noise field, and the linear kernel SFO 𝒮: 𝔼(R) → 𝕌(P) defines a mapping of the source signal space 𝔼(R) onto the observation signal space 𝕌(P). The metrics structures in the corresponding Hilbert signal spaces 𝕌(P) and 𝔼(R) are imposed by scalar (inner) products: (2) [ u , u ′ ] 𝕌 = ∫ P u ( p ) u ′ * ( p ) d p ,                   [ e , e ′ ] 𝔼 = ∫ R e ( r ) e ′ * ( r ) d rrespectively, where asterisk stand for complex conjugate. In the conventional integral form, EO (1) may be rewritten as: (3) u ( p ) = ∫ R S ( p , r ) e ( r ) d r + n ( p ) ,where S(p, r) = S(t, ρ; r) represents the functional kernel of the SFO referred to as the unit signal [10,34] determined by the time-space modulation employed in a particular RS system, and the scene domain R specifies the bounded SFO support. For explicit definitions of the unit signals for array radar systems we refer readers to to [7], (Sec. 6, [21]) and for imaging SAR systems to [4,22], (Sec. 5, [11]) and (Sec. 6, [26]).

It is convenient in the RS applications to assume that due to the integral signal formation model (3), the central limit theorem conditions hold [3,23,34,35], hence the fields e, n, u in (1), (3) are considered to be the zero-mean complex-valued random Gaussian fields. Next, since in all RS applications the regions of high correlation of e(r) are always small in comparison with the resolution element on the probing scene [3,10,11,34], the signals e(r) scattered from different directions r, r′ ∈ R are assumed to be uncorrelated, i.e., characterized by the correlation function: (4) R e ( r ,   r ′ ) = 〈 e ( r ) e * ( r ′ ) 〉 = b ( r ) δ ( r − r ′ ) ;         r , r ′ ∈ Rwhere δ(·) defines the delta function and 〈·〉 is the averaging (expectation) operator. The average (5) b ( r ) = 〈 e ( r ) e * ( r ) 〉 = 〈 | e ( r ) | 2 〉 ;       r ∈ Rof the square modules of the random scattering field e(r) as a function over the analysis domain (scene frame) R∋r has a statistical meaning of the average power scattering function and is traditionally referred to (in the RS and radar imaging literature, e.g., [1,4,31,32,34], etc.) as the SSP of the scattered field. Representing the spatial distribution of the average power of the random scatterers, the SSP b(r) characterizes in an explicit statistical sense the brightness reflectivity of the scene being mapped (for this reason, b is adopted in the notations as an abbreviation from brightness reflectivity). The estimate b̂ (r) of the SSP formed using some statistically grounded method is associated with the scene image to be formed via processing the recorded data observations.

The formulation of the data discretization and sampling in this paper follows the experiment design formalism given in [10,23,30,34] that enables one to generalize the finite-dimensional approximations of Equations (1,3) independent of the particular system configuration and the method of data measurements and recordings employed. Following [10,34], consider the sensor array (synthesized array) specified by a set of distanced in space (i.e., orthogonal) tapering functions { κ l * ( ρ ) ; l = 1 , … , L } (in the SAR case, the { κ l * ( ρ ) } are synthesized by the moving antenna over L spatial recordings [4,34]). Consider next, that the output signals in such spatially distributed measurement channel are then converted to I samples at the outputs of identical temporal sampling filters defined by their impulse response functions { v i * ( t ) ; i = 1 , … , I } where complex conjugate is taken for notational convenience. Without loss of generality [2–4,10,21,34], the sets {κ_l} and {v_i} are assumed to be orthonormal (e.g., via proper filter design and sensor antenna calibration [4,10]). The composition {h_m(p) = κ_l(ρ)v_i(t); m = (l, i) =1,..., L × I = M} ordered by multi-index m = (l, i) composes a set of the orthonormal spatial-temporal decomposition functions (base functions) that explicitly determine the vector of outcomes: (6) u = v e c m { u m = [ u ,   h m ] 𝕌 = ∫ P u ( p ) h m * ( p ) d p ;         m = 1 , … , M }of such an M-dimensional (M-D in our notation) data recording channel, in which the employed base functions {h_m(p)} span the relevant M-D data representation subspace 𝕌_(M) = 𝒫_𝕌(M) 𝕌 = Span_(M){h_m(p)} specifying the corresponding projection operator 𝒫_𝕌(M) defined by Equation (6).

In analogy to Equation (6), one can define now the K-D vector-form approximation of the scene random scattering field: (7) e = v e c k { e k = [ e ,   g k ] 𝔼 = ∫ R e ( r ) g k * ( r ) d r ;           k = 1 , … , K }with the elements composed of the decomposition coefficients {e_k} with respect to some chosen normalized set of expansion functions {g_k(r)} that span such K-D source signal subspace 𝔼_(K) = 𝒫_𝔼(K) 𝔼 = Span_(K){g_k(r)} specifying the corresponding projector 𝒫_𝔼(K). Note, that to satisfy the observability requirement [2,21], it is desirable (but not mandatory in the super-resolution case [22,24]) that these basis functions be properly adjusted in such a way that 𝒫_𝔼(K) defines a projector onto the M-D orthogonal complement to the null-space [2] of the degenerate (projection-dependent) SFO 𝒫_𝕌(M)𝒮 for any given/chosen 𝒫_𝕌(M) so that the observations (6) contain information of the observable M-D signal projection specified by Equation (7), i.e., K ≤ M. The reader is referred to [2,10,21,22,24,25,34] for the mathematical and signal processing details related to the construction of such feasible projectors 𝒫_𝕌(M) and 𝒫_𝔼(K). In this paper we adopt the technically inspired fine representation basis formed by a K_x×K_y regular pixel-formatted lattice with unitary pixel amplitudes and the spacing between lattice points normalized to one pixel width [2,25,34] where K_x defines the dimension of the rectangular pixel grid over the horizontal (azimuth) coordinate x and K_y defines its dimension over the orthogonal (range) coordinate y in the scene frame. Note that such rectangular pixel frame {pix_k(r)} specified by the ordered multi-index k = (k_x, k_y); k_x= 1,.., K_x; k_y= 1,.., K_y; k = 1,…, K = K_x × K_y is practically motivated in a majority of RS imaging applications [1,3–11,18,26,34,37], etc, because the ordinary pixels coincide with their squares { pix k ( r ) = pix k 2 ( r ) } which makes the same pixel grid {g_k(r) = pix_k(r)} applicable for fine discrete-form representation of both the complex scattering function e(r) and the SSP b(r) approximating for large K_x, K_y the continuous scene framing [2].

With the specified decompositions (6), (7), the discrete (vector-form) approximation of the continuous-form EO (1), (3) is given by: (8) u = Se + n ,where the u, n and e define the vectors composed of the decomposition coefficients u_m, n_m and e_k of the finite-dimensional (truncated) approximations of the fields u, n and e defined by the scalar products (6), (7), and S represents the matrix-form approximation of the SFO 𝒮 with elements [21]: (9) { S m k = [ 𝒮 g k ,   h m ] 𝕌 = ∬ R × P S ( p ,   r ) g k ( r ) h m * ( p ) d r d p ;                                 k = 1 , … , K ;       m = 1 , … , M } .Zero-mean Gaussian vectors e, n, and u in Equation (8) are characterized by the correlation matrices R_e, R_n, and: (10) R u = S R e S + + R n = SD ( b ) S + + R nrespectively, with the diagonal-form R_e = D(b) = diag(b), in which the K × 1 vector b of the principal diagonal is composed of elements b k = 〈 e k e k * 〉 and the superscript ⁺ defines the Hermitian conjugate when stands with a matrix (or a vector). The vector b is referred to as a vector-form representation of the SSP, i.e., the SSP vector.

The nonlinear inverse problem of radar/SAR imaging with the discrete-form measurement data (8) is formulated now as follows: to derive an estimator for the SSP vector b and use it to reconstruct the SSP distribution: (11) b ^ ( k ) = est { 〈 | e ( k ) ( r ) | 2 〉 } = ∑ k = 1 K b ^ k g k ( r )over the pixel-formatted observation frame R ∋ r (referred to as the scene image) by processing the recorded data u (in the operational scenario with the single processed data realization, e.g., SAR system) or J > 1 whatever available independent realizations {u_(j); j = 1,.., J} (e.g., in a multiple snapshot scenario [5,15]) collected with a particular imaging radar/SAR system. Recall that in this paper we intend to develop and follow a new DYED framework to derive such the estimator est{·} (11) that unifies the previously proposed DEED regularization with the VA dynamic image enhancement approach.

We, first, recall the conventional continuous-form kernel estimator [2,21] that is an MSF-based extension of the periodogram smoothing spectral analysis technique [36,37]. With such the method, the SSP estimate b̂(r) is derived from only one random observation (realization) of the data field u(p) as the generalized periodogram (the so-called sufficient statistics [21]) formed as a squared modulus of the MSF output | (𝒮⁺u)(r)|² smoothed by the kernel window operator (WO) 𝒲 (i.e., pseudo averaged): (12) b ^ ( r ) = 𝒲 | ( 𝒮 + u ) ( r ) | 2 = ∫ R W ( r − r ′ ) | ∫ P S * ( r ′ ,   p ) u ( p ) d p | 2 d r ′ ,where W(r) is the functional kernel of the WO 𝒲, and S^* (r′, p) represents the functional kernel of the MSF operator 𝒮⁺, that is, the adjoint SFO. Here, superscript + stands for the adjoint operator in the relevant signal spaces. In the Hilbert signal spaces introduced above, the MSF operator 𝒮⁺ adjoint to the SFO 𝒮 is defined via corresponding inner products (2) as follows, [e,𝒮⁺u]_𝔼 = [𝒮e,u]_𝕌. For a detailed analysis of this method and the corresponding synthesis of different windows with special scaling and smoothing properties we refer to [36]. Note that in a classical case of an isotropic kernel WO, the Equation (12) relates to a category of low-resolution kernel-type SSP estimators.

In this Section, we extend the recently proposed high-resolution maximum likelihood-based amplitude phase estimator (ML-APES) [24] to the SSP estimation problem at hand via its modification adapted to the distributed RS scene (not composed of sparse multiple point-type targets as originated in [24]). In the considered low snapshot sample case (e.g., one recorded SAR trajectory data signal), the sample data covariance matrix Y = ( 1 / J ) ∑ j = 1 J u ( j ) u ( j ) + is rank deficient (rank-1 in the single radar snapshot and single look SAR cases, J = 1). As it is shown in [24], minimization of the negative likelihood function lndet { R u } + tr { R u − 1 Y } with respect to the SSP vector b related to R_u = R_u(b) via Equation (10) is equivalent to minimizing the covariance fitting Stein’s loss, − lndet { R u − 1 Y } + tr { R u − 1 Y }. The solution to such minimization problem found in [24] results in the solution-dependent ML-APES estimator ([24], Equation (32)): (13) b ^ k = s k + R u − 1 Y R u − 1 s k ( s k + R u − 1 s k ) 2 ;       k = 1 , … , K .

In the APES terminology (as well as in the minimum variance distortionless response (MVDR) [9,33] and other ML-related approaches [23,35], etc.), s_k represents the so-called steering vector in the kth look direction, which in our notational conventions is essentially the kth column vector of the SFO matrix S. By the authors’ design [24], the numerical implementation of the ML-APES algorithm (13) assumes application of an iterative fixed-point technique by building the model-based estimate R̂_u = R_u (b̂_[i]) of the unknown covariance R_u modeled by Equation (10) from the latest (ith) iterative SSP estimate [b̂_[i]] with the zero step initialization b̂_[0] = b̂_MSF computed applying the conventional MSF estimator.

Let us adapt the algorithm (13) to the considered here single snapshot/single look case (J = 1) substituting Y by uu⁺, taking into account the properties of the convergent MVDR estimates of the SSP, which in a coordinate/pixel form are given by b ^ k ≈ 1 / s k + R u − 1 s k (also referred to as properties of a conventional Capon beamformer [9,24]), and making the use of a fixed-point nature of the algorithm (13) according to which the ML-APES estimates in the vector form b ^ = v e c k { b ^ k ; k = 1 , … , K } are to be found as a numerical solution to the nonlinear matrix-vector equation: (14) b ^ ML-APES = { D ^ S + R u − 1 ( b ^ ) u u + R u − 1 ( b ^ ) S D ^ } diagwith the solution-dependent: (15) D ^ = D ( b ^ ) = diag ( b ^ )                           and                           R u − 1 ( b ^ ) = ( S D ^ S + + R n ) − 1where operator {·}_diag returns the vector of a principal diagonal of the embraced matrix. Specifying the ML-APES matrix-form solution operator (SO): (16) F APES = F ( 1 ) = D ^ S + R u − 1 ( b ^ ) ,we next represent the estimate (14) in a more compact format: (17) b ^ ML-APES = b ^ ( 1 ) = { F ( 1 ) u u + F ( 1 ) + } diag = ( F ( 1 ) u ) ⊙ ( F ( 1 ) u ) * ,where ⊙ defines the Shur-Hadamar (element vise) vector/matrix product. The algorithmic structure of the nonlinear (i.e., solution-dependent) ML-APES estimator (17) guarantees positivity, but does not guarantee the consistency. Next, convergence enforcing regularization via performing projections onto convex solution sets (POCS) at each iteration step should be incorporated into the overall fixed-point iterative scheme for solving Equations (14,17) to guarantee the convergence in the considered here deficient-rank case. We defer the analysis of the consistency and convergence issues, as well as the POCS regularization with the relevant modifications of Equations (14,17) to the next Section proceeding now with the analysis of an alternative high-resolution SSP estimation approach based on the DEED regularization.

The DEED regularization framework proposed and developed in [25,26,30,34] can be viewed as a problem-oriented formalization of the requirements to the signal/image processing/post-processing co-design aimed at satisfying the desirable properties of the reconstructed RS images, namely: (i) maximization of spatial resolution balanced with noise suppression, (ii) consistency, (iii) positivity, (iv) continuity and agreement with the data [2,25,34]. Within the general DEED framework [25,34], all these aspects are formalized via the design of balanced resolution-enhancement-over-noise-suppression SSP estimation techniques unified with the POCS regularization. Such unification-balancing is descriptive in the sense that the user/observer can induce the desirable metrics (geometrical) structures in the image/solution space, and next, specify the type, the order and the amount of the employed regularization via constructing the related resolution-enhancement-over-noise-suppression performance measures with adjustable balancing factors (regularization parameters). Different feasible assignments of such user-controllable “degrees of freedom” specify a family of the DEED-related techniques. For a detailed formalism of the DEED method we refer to [25,30] and for its implementation in a family of fixed-point iterative techniques to [26,34]. Here we provide a modification of the original DEED method [25,26] (in terms related to the presented above ML-APES strategy) to adapt it for the dynamic experiment design (DYED) regularization framework that we next develop in Section 5.

The DEED-optimal SSP estimate b̂ is to be found as the POCS-regularized solution to the nonlinear equation [26]: (18) b ^ DEED = 𝒫 { F DEED u u + F DEED + } diagwhere F_DEED represents the adaptive (i.e., dependent on the SSP estimate b̂) matrix-form DEED solution operator and 𝒫 is the POCS regularization operator. Two fundamental issues constitute the benchmarks of the modified DEED estimator (18) that distinguish it from both the kernel algorithm (12) and the ML-APES method (17). First, the strategy for determining the DEED solution operator F_DEED in Equation (18) is reformulated in the minimum risk (MR)-inspired worst case statistical performance (WCSP) optimization setting [9,25] to provide robustness of the SSP vector estimates against possible model uncertainties, in particular, possible random distortions Δ_S in the perturbed SFO matrix S̃ = S + Δ_S that result in multiplicative noise occurring in practical RS scenarios due to calibration errors and speckle noising effect [25,31,34]. The second issue relates to the problem-oriented co-design of the POCS regularization operator 𝒫 in Equation (18). Such co-design (that we perform in the next Section) is aimed at satisfying some intrinsic and desirable properties of the solution such as positivity, consistency, model agreement (e.g., despeckling with edge preservation), and convergence [2,37].

Following the DEED framework [25], the strategy for deriving the optimal SO F_DEED is formalized by the MR-WCSP optimization problem: (19) F DEED = arg   min F   max 〈 ‖ Δ S ‖ 2 〉 ≤ η { 𝔕 ( F ) }in which: (20) 𝔕 ( F ) = tr { 〈 ( F S ˜ − I ) A ( F S ˜ − I ) + 〉 } + α tr { 〈 F R n F + 〉 }represents the DEED objective function where the averaging in the first term (systematic risk component) is performed over the randomness of the distorted SFO S̃ = S + Δ_S with the uncertainty conditioned by the statistical bound 〈 ‖ Δ S ‖ 2 〉 = 〈 tr { Δ S Δ S + } 〉 ≤ η. The regularization parameter α and the invertible weight matrix A constitute the user controllable/adjustable “degrees of freedom”: α is viewed as a tolerance factor that balances the systematic risk component (specified by the first term in ℜ (F)) and the fluctuation risk component (specified by the second term in ℜ (F)) in the composite risk objective function (20), while A induces the weighted metrics structure in the systematic risk tr{〈(FS̃ – I)A(FS̃ – I)⁺〉} that measures “how far” is the DEED-optimal SO from the pseudo inverse to the uncertain SFO S̃ in the averaged operator metrics induced by the employed weight matrix A. The solution to the conditioned optimization problem Equation (19) derived in the previous study [25,34] yields the DEED-optimal SO: (21) F DEED = F ( 2 ) = K S + R ∑ − 1 ,where K = ( S + R ∑ − 1 S + α A − 1 ) − 1 defines the so-called reconstruction operator (with the regularization parameter α and stabilizer A⁻¹), and R ∑ − 1 is the inverse of the diagonal loaded noise correlation matrix [34], R_Σ = R_Σ (β) = R_n + βI. In the practical RS scenarios (and specifically, in all SAR imaging applications [1,3–7,25,32,34], etc., it is a common practice to adopt the robust white additive noise model, i.e., R_n = N₀I, attributing the unknown correlated noise component as well as the speckle to the composite uncertain noise term, Δ_Se + n, in which case [25,34]: (22) R ∑ = N ∑ I ;               N ∑ = N 0 + βwith the composite noise power N_Σ = N₀ + β, the additive observation noise power N₀ augmented by the loading factor β = γη/α ≥ 0 adjusted to the regularization parameter α, the Loewner SFO ordering factor γ > 0 of the SFO S and the uncertainty bound η imposed by the conditional maximization in Equation (19) (see [25,34] for details). For these assumptions, the robust DEED-related SO becomes: (23) F DEED = F ( 3 ) = ( Ψ + α N ∑ A − 1 ) − 1 S + = K S + ,i.e., a composition of the MSF operator S⁺ and the self adjoint reconstruction operator K = (Ψ + αN_ΣA⁻¹)⁻¹ recognized to be the regularized inverse of the discrete-form ambiguity function (AF) matrix operator: (24) Ψ = S + S ,with the type and amount of regularization specified by the DEED degrees of freedom, A and α, respectively.

Putting F⁽²⁾, F⁽³⁾ in Equation (18) results in two POCS-regularized DEED-related SSP estimators that produce the SSP estimates defined as b̂⁽²⁾ and b̂⁽³⁾, respectively. Note that other feasible adjustments of the processing level degrees of freedom {α, N_Σ, A} summarized in [26,34] for the robust RS adopted model (22) of the correlation matrix R_Σ specify the family of relevant POCS-regularized DEED-related (DEED-POCS) techniques (18) represented in the general form as follows: (25) b ^ ( p ) = 𝒫 { D ( b ^ ( p ) ) } diag = 𝒫 { F ( p ) u u + F ( p ) + } diag = 𝒫 { K ( p ) Q K ( p ) + } diag ;       p = 1 , … , Pwhere: (26) Q = S + u u + Sdefines the measurement statistics matrix independent on the solution b̂, and different reconstruction operators {K^(p); p = 1,…,P} specified for different feasible assignments to {α, N_Σ, A} define the corresponding DEED-POCS estimators (25) with the relevant SOs’ {F^(p) = K^(p)S⁺; p = 1,…,P}.

The relationship between two high-resolution SSP estimators, the ML-APES (17) and the DEED-POCS (18), can now be established using the second equivalent form for representing the SO (16) given by (Appendix B, [21]): (27) F APES = F ( 1 ) = D ^ S + R u − 1 ( b ^ ) = K S + R n − 1 ,which coincides with F_DEED = F⁽²⁾ specified by Equation (21) for the simplified scenario of completely certain SFO (Δ_S = 0), thus, R_Σ = R_n = N₀I in (22). Due to such structural algorithmic similarity, the DEED method (18) can be addressed as a regularized robust version of the APES approach (17) adapted to the uncertain RS scenarios.

Following the DEED-POCS regularization formalism [26,34], the POCS-level regularization operator 𝒫 in Equation (18) could have a composite structure, i.e., could be constructed as a composition of operators/projectors conditioned by non-trivial prior information that formalizes some desirable properties of the solution, e.g., positivity, consistency, etc. We specify such RS-adapted composite regularization in Section 4.4. In this Section, we are going to establish that to guarantee the consistency of the DEED-related estimator(s) (18), (25) the 𝒫 should incorporate a kernel-type WO 𝒲 (not necessarily isotropic) as a necessary requirement. To verify this consistency guarantee, we limit ourselves here with the relevant simplest regularization operator model, 𝒫 = 𝒲. Analysis of the consistency requires the hypothetical continuous asymptotes, lim K → ∞ 𝒫 𝔼 ( K ) = 𝒤, lim M → ∞ 𝒫 𝕌 ( M ) = 𝒤, the identity operators. Adopting these assumptions and white observation noise model, the DEED estimator given by Equation (25) can be expressed in the following generalized continuous functional form: (28) b ^ ( r ) = 𝒲 | ( 𝒡 u ) ( r ) | 2 = ∫ R W ( r − r ′ ) | ∫ P F ( r ′ ; p ) u ( p ) d p | 2 d r ′ ,where W(r) is the functional kernel of the WO 𝒲, and F(r;p) represents the functional kernel of the continuous-form DEED-optimal SO 𝒡 = 𝒦𝒮⁺, a composition of the MSF operator 𝒮⁺ and the DEED-optimal reconstruction operator 𝒦 given by the continuous-form assymptotic to (23) (subscript _DEED is omitted to simplify the notations). To analyze the consistency of the estimator (28), one should consider the large measure [2] of the observation domain, 𝔐 = mesP. For the hypothetical asymptotes, 𝔐 → ∞, N_Σ → 0 the operator composition 𝒡𝒮 = 𝒦𝒮⁺𝒮 tends to the identity operator [2], in which case, the estimator (28) produces the degraded (smoothed by the WO) estimate of the SSP with the asymptotic bias ∥ b − lim 𝔐 → ∞ < b ^ 𝔐 > ∥ 2 = ∥ ( 𝒤 − 𝒲 ) b ∥ 2 where ∥ b ∥ 2 = [ b , b ] 𝕃 2 = ∫ P b ( r ) b * ( r ) d r defines the conventional (Lebesgue) squared norm in the 𝕃₂ Hilbert space, and subscript 𝔐 indicates the measure of the observation domain for which the relevant estimate has been obtained. The ratio of the average fluctuation noise energy in the estimate (28) to the average fluctuation noise energy in the high-resolution sufficient statistics (SS), v 𝔐 = v 𝔐 ( r ) = | ∫ P F 𝔐 ( r ; p ) u ( p ) d p | 2, is evaluated as follows: (29) ξ ( 𝔐 ) = 〈 ‖ 𝒲 v 𝔐 − 𝒲 〈 v 𝔐 〉 ‖ 2 〉 〈 ‖ v 𝔐 − 〈 v 𝔐 〉 ‖ 2 〉 = tr { 𝒲 𝒞 𝔐 𝒲 + } tr { 𝒞 𝔐 } ,where 𝒞_𝔐 defines the covariance operator of the SS, v_𝔐(r), i.e., the linear integral operator with the functional kernel cov{v_𝔐(r)v_𝔐(r′) = 〈v_𝔐(r)v_𝔐(r′)〉 – 〈v_𝔐(r)〉 〈v_𝔐(r′)〉. In the limiting case, 𝔐 → ∞, statistics v_𝔐(r) becomes δ - correlated with the variance b²(r) [21]. Thus, one can evaluate the boundary value of the fluctuation noise ratio (29) as lim 𝔐 → ∞ ξ ( 𝔐 ) ∼ lim 𝔐 → ∞ ( 𝔐 − 1 tr { 𝒲 𝒲 + } ). Hence, the consistency requirement [36], lim 𝔐 → ∞ ξ ( 𝔐 ) = 0, is satisfied for any WO 𝒲 with the bounded operator norm, tr{𝒲𝒲⁺} < ∞ that provides lim 𝔐 → ∞ ( 𝔐 − 1 tr { 𝒲 𝒲 + } ) = 0, that is for a kernel operator [2]. This restricts the class of admissible windows 𝒲 in Equation (28) by the kernel operators that in the engineering interpretation simply means that feasible windows must be restricted by a class of kernel-type filters.

Remark 1: The APES approach, as well as other ML-based high-resolution SSP estimators [21,24], etc., do not imply regularizing windowing at all, i.e., 𝒲 = const·𝒤 while the identity operator is not a kernel operator. Hence, in the uncertain scenarios with rank-deficient data covariance any ML-based approach inevitably produces an inconsistent estimate of the SSP. Technically it means that any ML-based resolution enhancement attempt should be combined with the relevant kernel windowing (not necessarily isotropic linear spatial smoothing) to guarantee the consistency.

The next crucial performance issue relates to construction of convergent iterative scheme for efficient computational implementation of the POCS regularized DEED-related estimators. To convert such the technique to an iterative procedure we, first, transform the Equation (18) into the equivalent equation: (30) 𝒫 { Φ D b ^ } = 𝒫 { q }a numerical solution to which produces the desired SSP estimate b̂, where 𝒫{·} defines application of operator 𝒫 to the embraced vector quantity, Φ_D is the solution-depended diagonal loaded point spread function (PSF) matrix operator: (31) Φ D = Φ D ( b ^ ) = ( Ψ + N ∑ D − 1 ( b ^ ) ) ⊙ ( Ψ + N ∑ D − 1 ( b ^ ) ) * ,constructed from the diagonal loaded AF matrix (24) via the Shur-Hadamar (element vise) product ⊙, and vector q represents the measurement statics vector (32) q = { Q } diag = { S + Y S } diag ,formed from the sampled data matrix Y (Y = uu⁺ in the considered above rank-1 data covariance matrix case) applying the MSF SO S⁺; i.e., the q given by Equation (32) defines the low-resolution image formed by an RS radar or a fractional SAR imaging system that employs the conventional matched spatial processing algorithm. Fixing the Equation (30) at iteration [i], i = 1, 2, …, and inducing the “contractive mapping” term ( b ^ [ i + 1 ] − b ^ [ i ] ) → i → ∞ 0, yields the progressive contractive mapping iterative scheme: (33) 1 τ ( b ^ [ i + 1 ] − b ^ [ i ] ) + 𝒫 { Φ D [ i ] b ^ [ i ] } = 𝒫 { q } ,with the zero-step iteration b̂_[0] = q defined by Equation (32), in which the relaxation parameter τ and the regularization operator 𝒫 constitute the POCS regularization-level degrees of freedom that should be specified to guarantee the contractive mapping of Equation (33), hence, the convergence.

Following the POCS regularization formalism [2,26], the regularization operator 𝒫 could be constructed as a composition of projectors 𝒫_n onto convex sets 𝔺_n; n = 1,…, N with not empty intersection. Then for the composition of the relaxed projectors: (34) 𝒫 n λ = 𝒤 − λ n ( 𝒫 n − 𝒤 ) ,with the “speeding-up” regularization parameters {λ_j}, the general-form POCS-regularized fixed-point iteration rule becomes: (35) b ^ [ i + 1 ] = 𝒫 N λ 𝒫 N − 1 λ ⋯ 𝒫 1 λ b ^ [ i ]and it is guaranteed to converge to a point in the intersection of the sets {𝔺_n} provided 0 < λ_n < 2 for all n in any order regardless of the initialization b̂_[0] that is a direct sequence of the fundamental theorem of POCS (Sec. 15.4.5, [20]) (see also (Sec. 6, [25]) and (Appendix B, [26]). In this study, for simplicity, we fix λ_n = 1 for all n = 1,…, N. Also, any operator that acts in the same convex set (e.g., kernel-type WO) can be incorporated into such composite regularization operator 𝒫 [2]. Our next task is to make the use of the presented convergence enforcing POCS regularization paradigm employing some practical imaging radar/SAR-motivated considerations.

The DEED-POCS framework offers a possibility to design the POCS regularization operator 𝒫 in such a way that to preserve high spatial resolution performances of the resulting DEED-related consistent SSP estimates. Following the VA-based image enhancement approach [16,20], this task could be performed via anisotropic image post-processing that in our statement implies anisotropic regularizing windowing over the properly constructed convex solution set in the image space. To proceed with the derivation of such a WO, in this paper, we incorporate the prior information on the desirable smoothness and geometrical properties of an image and its estimate by constructing the vector image/solution space 𝔹_(K) ∋ b as a K-D discrete-form approximation to the corresponding function image space 𝔹(R), in which the initial continuous SSP functions b (r), r ∈ R reside. To formalize the geometrical information on the image changes and simultaneously on the image edge changes over the scene frame, the metrics structure in 𝔹(R) must incorporate the image norm as well as the image gradient norm [2,3]. This is naturally to perform by adopting the so-called Sobolev metrics [21]: (36) ‖ b ‖ 𝔼 ( R ) 2 = [ b ,   𝒨 b ] 𝕃 2 ( R ) = ∫ R b ( r ) ( ∫ R M ( r ,   r ′ ) b ( r ′ ) d r ′ ) * d r ,where M (r, r′) is the functional kernel of the metrics inducing operator 𝒨 constructed as a composition: (37) 𝒨 = m ( 0 ) 𝒤 + m ( 1 ) ∇ r 2 = m ( 0 ) 𝒤 + m ( 1 ) { D x ( 1 ) + D x ( 1 ) + D y ( 1 ) + D y ( 1 ) }in which D x ( 1 ) = ∂ / ∂ x and D y ( 1 ) = ∂ / ∂ y represent the first order differential operators with respect to the spatial variables x and y, respectively, ∇ r 2 = D x ( 1 ) + D x ( 1 ) + D y ( 1 ) + D y ( 1 ) defines the Laplacian with respect to the 2-D space variable r = (x, y), and m⁽⁰⁾ and m⁽¹⁾ are the nonnegative real-valued scalars that control the balance between two metrics measures in Equation (36). If m⁽⁰⁾ = 1, m⁽¹⁾ = 0, then (36) reduces to the conventional Lebesgue 𝕃₂ metrics in the Hilbert space 𝔹(R) that does not incorporate information on the image derivatives. In the opposite case m⁽⁰⁾ = 0, m⁽¹⁾ = 1, the metrics Equation (36) transforms into the so-called Dirichlet variational functional [37] that controls only the gradient flow [2,37]. In the equibalanced case, m⁽⁰⁾ = m⁽¹⁾ = 1, the same importance is assigned to the both metrics measures specified by the kernel metrics inducing operator 𝒨. Incorporation of such metrics inducing operator as the WO into the general DEED-optimal technique (28), i.e., specifying 𝒲 = 𝒨, results in the required anisotropic kernel-type windowing because it controls not only the SSP (image) discrepancy measure but also its gradient flow over the scene in the Sobolev-type image/solution space 𝔹(R) ∋ b(r). In the next Section we show that due to the gradient-dependent anisotropy, such regularizing post-processing is aimed at edge preservation in the scene regions with high gradient contrast while performing smoothed windowing over the homogeneous image zones corrupted by speckle.

To proceed with designing the related WO adapted to the discrete problem model, the relevant 𝔹_(K) ∋ b as a K-D discrete-form approximation to 𝔹(R) ∋ b(r) has to be defined via specifying the corresponding metrics ‖ b ‖ 𝔹 ( K ) 2 = [ b ,   M b ] with the metrics inducing matrix M constructed as a matrix-form approximation of 𝒨 given by Equation (37). For the adopted pixel-framed discrete image representation format {b_k = b(k_x, k_y)}; k = (k_x, k_y); k_x= 1,.., K_x; k_y= 1,.., K_y; k = 1,…, K = K_x×K_y} this yields the desired metrics: (38) ‖ b ‖ B ( K ) 2 = [ b ,   M b ] = m ( 0 ) ∑ k x , k y = 1 K x ,   K y ( b ( k x ,   k y ) ) 2 + m ( 1 ) ∑ k x , k y = 1 K x , K y ( b ( k x ,   k y ) − 1 4 ( b ( k x − 1 ,   k y ) + b ( k x + 1 ,   k y ) + b ( k x , k y − 1 ) + b ( k ,   k y + 1 ) ) ) 2 .

The second sum on the right hand side of Equation (38) is recognized to be a 4-nearest-neighbors difference-form approximation of the Laplacian operator ∇ r 2 in Equation (37) [2,16]; hence, it represents the metrics measure of the high frequency spatial components in the discretized SSP that corresponds to is gradient variations. From Equation (38) we easily derive the corresponding metrics inducing matrix-form operator: (39) M = m ( 0 ) I + m ( 1 ) ∇ 2 ,where ∇² is the numerically approximated Laplacian operator ∇ r 2. Application of such ∇² to a vector b returns the vector ∇²b with elements defined by the terms in the second sum at the right hand side of Equation (38) ordered by multi index {k = (k_x, k_y) = 1,…, K = K_x × K_y} over the pixel-formatted 2-D frame. Also, in all applications below, we adopt the equibalanced metrics structure specifying m⁽⁰⁾ = m⁽¹⁾ = 1.

Last, we restrict the solution subspace (the so-called active solution set or correctness set in the DEED terminology [25,34]) to the K-D convex set 𝔹₊ ⊂ 𝔹_(K) of SSP vectors with nonnegative elements (as power is always nonnegative). This is formalized by specifying the projector 𝒫₊ onto such convex set 𝔹₊ i.e., the POCS operator (as the positivity operator specifies POCS [2]) that has the effect of clipping off all the negative values. The composition: (40) 𝒫 = 𝒫 2 𝒫 1 ,defines the required composite POCS operator with the regularizing WO 𝒲 = 𝒫₁ = M and 𝒫₂ = 𝒫₊ (in the function image space, 𝔹(R), the continuous-form generalization for the WO is given by 𝒲 = 𝒫₁ = 𝒨).

With the model (40), the discrete-form contractive progressive mapping iterative process (33) transforms into: (41) b ^ [ i + 1 ] = b ^ [ i ] + τ 𝒫 + { M q − M Φ D [ i ] b ^ [ i ] } ;                                     i = 0 ,   1 ,   2 ,   …initialized by the conventional low-resolution MSF image b̂_[0] = q. Some theoretical generalizations of Equation (41) for a hypothetical continuous STAP over the 2-D scene frame R ∋ r = (x, y) in “evolution time” are useful at this point for establishing the asymptotic dynamic adaptive reconstructive processing properties of the DEDR-VA approach.

Associating the iterations i, i+1,… with discrete “evolution time”, i.e., i + 1 → t + Δt;i → t;τ → Δt, the Equation (41) can be rewritten in the “evolution” form: (42) ( 1 / Δ t ) [ b ^ ( t + Δ t ) − b ( t ) ] = 𝒫 + { M q − M Φ D ( t ) b ^ ( t ) }with the corresponding dynamic scheme in hypothetical continuous evolution time (Δt → dt; t +Δt → t + dt) being: (43) ∂ b ^ ( t ) ∂ t = 𝒫 + { M q − M Φ D ( t ) b ^ ( t ) } ,where ∂b̂(t)/∂t represents the derivate with respect to the evolution time. Considering the continuous 2-D rectangular scene frame R ∋ r = (x, y) with the corresponding initial MSF scene image q(r) = b̂(r;0) and the “evolutionary” enhanced SSP estimate b̂(r; t), respectively, we proceed from (43) to the equivalent asymptotic dynamic scheme: (44) ∂ b ^ ( r ; t ) ∂ t = 𝒫 + { 𝒨 { ( q ( r ) ) } − 𝒨 { ∫ R Φ b ^ ( r ,   r ′ ; t ) b ^ ( r ′ ,   t ) d r ′ } } ,where Φ_b̂ (r, r′; t) represents the kernel PSF in evolution time t corresponding to the continuous-form dynamic generalization of the PSF matrix Φ_D[i] in Equation (31), and 𝒨 is the metrics inducing operator defined by Equation (37).

Three practically inspired versions of Equation (37) relate to three feasible assignments to the operator 𝒨. These are as follows:

𝒨 = 𝒤 specifies the conventional Lebesgue metrics, in which case the evolution process (44) does not involve control of the image gradient flow over the scene.

For the purpose of generality, instead of two metrics balancing coefficients m⁽⁰⁾ and m⁽¹⁾ we incorporated into the PDE (45) three regularizing factors c₀, c₁ and c₂, respectively, viewed as VA-level user-controllable degrees of freedom to compete between smoothing and edge enhancement. Although due to the solution-depended nature the dynamic DEED-VA scheme in its continuous PDE form (45) cannot be addressed as a practically realizable procedure, the undertaken theoretical developments are useful for establishing the relationship between the general-form VA scheme (45) and the already existing dynamic image enhancement approaches [16–20].

Different feasible assignments to the processing level degrees of freedom in the PDE (45) specify different VA-related procedures. Here beneath we consider the following ones:

The simplest case relates to the specifications: c₀ = 0, c₁ = 0, c₂ = const = – c, c > 0, and Φ(r, r′; t) = δ(r – r′) with excluded projector 𝒫₊. In this case, the PDE (45) reduces to the isotropic diffusion (so-called heat diffusion [16]) equation ∂ b ^ ( r ; t ) / ∂ t = c ∇ r 2 b ^ ( r ; t ) with constant (isotropic) conduction factor c. We reject the isotropic diffusion for the purposes of radar/SAR image processing because of its resolution deteriorating nature.

The discrete-form approximation of the PDE (45) in “iterative time” {i = 0, 1, 2, …} yields the iterative numerical procedure: (46) b ^ [ i + 1 ] = b ^ [ i ] + 𝒫 + { c 0 ( q − Φ D [ i ] b ^ [ i ] ) + c 1 ∇ 2 { q } − c 2 ∇ 2 { Φ D [ i ] b ^ [ i ] } } ;                 i = 0 , 1 , 2 … ,with the same MSF initialization b̂_[0] = q, where we have attributed the relaxation parameter τ to the corresponding VA regularization factors, for simplicity. The numerical Laplacian ∇²{·} applied to the embraces quantity is defined by the 4-nearest-neighbors difference-form approximation of the continuous Laplacian operator ∇ r 2 specified by the terms in the second sum on the right hand side of Equation (38). Different feasible assignments to these degrees of freedom specify different related reconstruction techniques exemplified in the previous section, namely: isotropic diffusion, anisotropic diffusion, locally selective DEED-based reconstruction, and selective anisotropic reconstruction techniques. In particular, our DEED-VA-fused DYED method coincides with the previously developed conventional DEED regularization technique [26] in the case when no VA inspired discrepancy terms are adopted, i.e., patting c₁ = c₂ = 0 in Equation (46). In contrary, the extended DEED-VA approach combines the VA-based isotropic diffusion with the anisotropic DEED reconstruction in a balanced fashion incorporating also the convergence enforcing POCS regularization. This not only speeds up the iterative process but provides perceptually enhanced imaging results as we illustrate in the comparative simulations presented in the next Section.

Remark 2: With the performed extension of the DEED regularization method into the unified DEED-VA framework, the warnings about the dynamic process (46) being ill conditioned do not apply, since the purpose of the two-level regularization (the DEED level and the VA level) is aimed at curing that same ill conditioning providing the POCS-regularized iterative DYED technique (46) converges to a point in the specified convex solution set 𝔹₊. Nevertheless (as it is frequently observed with nonlinear iterative processes [2,36,37]) such nonlinear iterative procedure (46) may suffer from some numerical instabilities demonstrating only local convergence.

In the simulation experiment, we considered a fractional SAR as a sensor system, analogous to a single look fraction of a multi look focused SAR [4,5,26,31]. The resolution properties of such the RS imaging system that employs the conventional MSF processing are explicitly characterized by the AF of the unit signal S(t,ρ;r) given by the composition [4,5,26]: (47) Ψ ( Δ τ ,   Δ θ ) = C Ψ Ψ r ( Δ τ ) Ψ a ( Δ θ ) .Here Ψ_a(Δ_θ) represents the azimuth AF over the azimuth angular spacing coordinate Δ_θ = arctan(Δx/r_s) related to the cross-range spacing Δx = x − x′ between two scatterers at the particular slant range r_s ; Ψ_r(Δ_τ) represents the range AF of the probe pulse signals as a function of the time delay variable Δ_τ = 2Δr_s/c related to the corresponding displacement Δr_s = r_s – r′_s of the scatterers along the slant range directions explicitly specified by the pulse modulation employed [4,26], c is the speed of light, and C_Ψ is a normalizing constant, not essential in the simulations [26]. To benefit from the range-angular AF factorization (47), the MSF images can be originally formed in the slant range planes (r_s, θ) and then projected to the ground scene (x, y) = r∈R with the corresponding pixel spacing. For the Gaussian antenna tapering function [26], the related azimuth PSF Φ_a(Δx) expressed over the Cartesian coordinate x in the ground scene plane R is given by [26]: (48) Φ a ( Δ x ) = | Ψ a ( Δ x ) | 2 ≈ ( e − π ( Δ x ) 2 / r s 2 ( λ 0 / 2 L a ) 2 ) 2 ,where λ₀ is the wavelength of the radar signal transmitted, and L_a is the effective (fractional) synthesized aperture max L_a = ψ_aL_max, a fraction ψ_a of the maximum focused synthesized antenna length L_max ≈ λ₀r_s/L_A corresponding to the physical antenna with horizontal aperture L_A. To be specific, the effective width κ_a of the azimuth PSF Φ_a(Δx) is measured in pixels at the user selected threshold (e.g., κ_a/2 at 0.5 from the maximum value Φ_a(0) at the midrange 0 r_s0 [6,26]). In the same manner, κ_r specifies the effective pixel width of the range PSF Φ_r(Δy) related to the discretized |sinc[2πψ_rB((2r_s / c) – (2r_s′ / c))]|²-type ψ_r-fractional range AF [6,22,26] in the slant range gates {r_s, r_s′} measured at the same user selected threshold (κ_r /2 at 0.5 from the maximum value Φ_r(0) [22,26]) where B represents the employed modulation signal bandwidth [4,6]. Because the DEED-VA-related fractional SAR image enhancement algorithms belong to the category of the MSF image post-processing techniques, the simulations were performed at the image processing level, i.e., avoiding the SAR raw signal simulations [31]. We tested the (1024 × 1024)-pixel (i.e., a large scale) scene image shown in Figure 1 in two hypothetical operational scenarios. The original scene was borrowed from the real-world high-resolution RS imagery [38]. Following [4,5,26,31] the degradations in the spatial resolution due to the fractional aperture synthesis mode were simulated via blurring the original image of Figure 1 with the range PSF Φ_r(Δy) along the y-axis and with the azimuth PSF Φ_a(Δx) along the x-axis, respectively. The degradations at the image formation level due to the propagation uncertainties were simulated using the statistical model of a SAR image defocusing [4,22,32]. The fractional resolution along the x and y scene coordinates were controlled by assigning different effective pixel widths κ_r and κ_a of the range and the azimuth PSFs and their varying over the scene that account to the range variation effect and uncompensated carrier trajectory deviations [4,22,26,31].

Next, to comply with the technically motivated MSF fractional image formation mode, the blurred scene image was degraded with the composite (signal-dependent) noise simulated as a realization of χ 2 2 -distributed random variables with the pixel mean value assigned to the actual degraded scene image pixel. The simulation experiment compares three most prominent SAR-adapted enhanced imaging techniques, namely: the celebrated VA-based anisotropic diffusion method [16,18] specified in Section 5.2.(ii); the ML-APES method [24] detailed in Section 3.1, and the developed fused anisotropic DYED reconstruction method aggregated with the POCS regularization performed via Equation (46). The simulations were run for two hypothetical operational scenarios. The first one corresponds to the partially compensated defocusing errors [4,22]. In the second scenario, no autofocusing was assumed, thus the degradations encompass both uncontrolled SFO distortions and MSF mismatches attributed to “heavy” propagation medium perturbations [22,31], range migration effect [4] and uncompensated carrier trajectory deviations that may occur in much more severe operational scenarios [26,29,31]. For both scenarios, the simulations were run for different values of the composite signal-to-noise ratio (SNR) μ_SAR defined as the ratio of the average signal component in the degraded image b̂⁽¹⁾ formed using the MSF algorithm (32) to the relevant composite noise component in that same speckle corrupted MSF image.

For objective evaluating of the reconstructive imaging quality, we have adopted two quality metrics traditionally used in image restoration/enhancement [2,3,11,37]. The first one evaluates the mean absolute error (MAE): (49) M A E ( p ) = 10 log 10 { 1 K x K y ∑ k x = 1 K x ∑ k y = 1 K y | b ^ ( p ) ( k x , k y ) − b ( k x , k y ) | } ,where {b(k_x, k_y)} represents the pixel values of the initial SSP and {b̂^(p) (k_x, k_y)} represents the pixel values of the SSP reconstructed applying the pth tested technique. In the performed simulations, p = 1 corresponds to the MSF algorithm (32), p = 2 corresponds to the VA-based anisotropic diffusion enhancement technique [16], p = 3 specifies the ML-APES method [24] resulting in the estimator (17), and p = 4 relates to the developed DYED-optimal fused DEED-VA algorithm (46). This metrics is well suitable for quantification of fine image reconstruction details, such as edge preservation (sharpening) and resolution of small targets on the extended scene [37].

The second employed metrics is the so-called improvement in the output signal-to-noise ratio (IOSNR) [2,3,29] measured via the ratio of the corresponding squared l₂ error norms: (50) IOSNR ( p ) = 10 log 10 ( ‖ q − b ‖ 2 / ‖ b ^ ( p ) − b ‖ 2 ) ,where b represents the original SSP frame, q = b̂⁽¹⁾ is the low-resolution speckle-corrupted image formed by a fractional SAR system that employs the conventional MSF method (32), and {b̂^(p)} represents the SSP reconstructed from the corrupted MSF image q applying the pth enhanced imaging method from the simulated family, the same as in (49). The lower is the MAE and the higher is the IOSNR, the better is the image enhancement/reconstruction performed with the particular employed method.

Figure 1 shows the original scene image (borrowed from the high-resolution RS imagery [38]) not observable with the simulated fractional SAR imaging systems. The images in Figures 2 and 3 present the results of image formation/enhancement applying different tested DEED-VA-related techniques in two operational scenarios as specified in the figure captions. Figures 2(a) and 3(a) demonstrate the images formed applying the conventional MSF algorithm. From these figures, one may easily observe that the MSF images suffer from imperfect spatial resolution due to the fractional aperture synthesis mode and composite observation/focusing mismatches and are corrupted by multiplicative signal-dependent noise. In the first scenario, the simulated degradations in the resolution are moderate over the range direction (κ_r = 10) and significantly larger over the azimuth direction (κ_a = 20). In the second scenario, the fractional SAR system suffers from much more severe degradations due to additional defocusing in both directions (κ_r = 20; κ_a = 40) and lower SNR.

Next, Figures 2(b) and 3(b) show the images enhanced applying the anisotropic diffusion method [16,18]. The images reconstructed using the ML-APES method [24] are shown in Figures 2(c) and 3(c), and the corresponding images optimally reconstructed applying the DEED-VA-optimal technique (46) with the same equibalanced regularization factors c₀ = c₁ = c₂ = 1, after the same 30 iterations are presented in Figures 2(d) and 3(d), respectively. Figures 4 and 5 present comparative results of reconstructive imaging in a 1-D format with the operational parameters and scenarios specified in the figure captions. Figures 6 and 7 report the quantitative performances evaluated via the two quality metrics (49) and (50) gained with three tested SSP estimation methods, namely: the VA-related anisotropic diffusion (VA-AD); the ML-APES and the DYED-optimal DEED-VA.

From the reported simulation results, the advantage of the well-designed imaging experiments (cases of the ML-APES and the optimal DEED-VA techniques) over the poorer design enhancement experiments (MSF and anisotropic diffusion (VA-AD) without ML-APES reconstruction) is evident for both scenarios. Due to the performed regularized inversions, the resolution was substantially improved. Quantitative performance improvement measures are reported in Figures 6 and 7 for the same 30 performed iterations.

The highest values of the IOSNR, as well as the lowest values of the MAE, were obtained with the DYED-optimal SSP estimator, i.e., with the fused DEED-VA technique (46) adapted to the particular operational scenarios. Note that IOSNR (50) is basically a square-type error metric; thus, it does not qualify quantitatively the “delicate” visual features in the reconstructed RS images; hence, small differences in the corresponding IOSNRs reported in Figures 6 and 7. Furthermore, all the DYED-related estimators manifest the higher IOSNRs and lower MAEs in the case of higher SNR. Both the 2-D and 1-D RS imaging results are indicative of the superior qualitative reconstructive performances achieved with the high-resolution DYED-related estimators (17) and (46), while the DYED-optimal DEED-VA approach outperforms the ML-APES method. Last, in Figure 8, we report the convergence rates (specified via the dynamics of the corresponding IOSNR and MAE metrics versus the number of iterations) evaluated for the first test scenario for the same three dynamic enhancement/reconstruction techniques: VA-AD, ML-APES, and the developed DYED-optimal unified DEED-VA technique. The reported convergence rates are indicative of the considerably speeded-up performances manifested by the DEED-VA algorithm (46) that outperforms the most prominent existing ML-APES and anisotropic diffusion methods in the both quality metrics requiring 5–6 times less number of iterations to approach the same asymptotic convergence.