[Joseph R. Guerci and Edward J. Baranoski] © EYEWIRE Knowledge-Aided Adaptive Radar at DARPA [An overview] F or the past several years, the Defense Advanced Research Projects Agency (DARPA) has been pioneering the development of the first ever real-time knowledge-aided (KA) adaptive radar architecture. The impetus for this program is the ever increasingly complex missions and operational environments encountered by modern radars and the inability of traditional adaptation methods to address rapidly varying interference environments. The DARPA KA sensor signal processing and expert reasoning (KASSPER) program has as its goal the demonstration of a high performance embedded computing (HPEC) architecture capable 1053-5888/06/$20.00©2006IEEE of integrating high-fidelity environmental knowledge (i.e., priors) into the most computationally demanding subsystem of a modern radar: the adaptive space-time beamformer. This is no mean feat as environmental knowledge is a memory quantity that is inherently difficult (if not impossible) to access at the rates required to meet radar front-end throughput requirements. In this article, we will provide an overview of the KASSPER program highlighting both the benefits of KA adaptive radar, key algorithmic concepts, and the breakthrough lookahead radar scheduling approach that is the keystone to the KASSPER HPEC architecture. IEEE SIGNAL PROCESSING MAGAZINE [41] JANUARY 2006 BACKGROUND ON SPACE-TIME ADAPTIVE PROCESSOR Even for a nonspecialist, it is not hard to qualitatively imagine the enormous challenges presented to modern airborne radar systems when attempting to separate ephemeral ground target echoes (associated with moving vehicles, for example) from overwhelming land clutter returns, which can be eight or more orders-of-magnitude stronger than the desired target, especially in highly complex and nonstationary clutter environments (e.g., urban clutter) [1]–[4]. The lion’s share of this signal separation/detection task falls to the real-time, space-time adaptive processor (STAP) [5]–[7]. STAP attempts to filter out the clutter and noise returns by an adaptive multidimensional finite impulse response (FIR) filter structure, as depicted in Figure 1, consisting of M time taps to support Doppler filtering, and N spatial taps to support angle processing. By judicious setting of the complex weights, {wi, j }, a two-dimensional (2-D) angle-Doppler adaptive antenna pattern can be obtained that maximizes the return from a desired target resolution cell, while simultaneously minimizing the returns from clutter and possibly N Antenna Elements jamming. A notional example of this es T Antenna process is depicted in Figure 2(a) s T T l Pu N M 2 1 and (b), where the clutter plus noise xNM . × × × T T T wNM power spectrum of Figure 2(a) is filxN+1 x2N xN+2 . PRI tered by the pattern in Figure 2(b). . . × w2N wN+2 × Delay × xN wN+1 Note that an airborne radar is x2 x . . × × 1 w1. + × wN + w2 + assumed as evidenced by the angleDoppler coupled clutter ridge which Σ + Σ + Σ + + + + is the result of aircraft motion [7]. Σ + Σ + Σ + The multidimensional adaptive FIR filter thus consists of a total of + NM taps [or spatio-temporal degrees-of-freedom (DOFs)]. The + Σ + spatial DOFs could be the outputs of individual antenna array elements (as part of an electronically y = w'x scanned array for example), subarrays, or outputs of a beamformer [FIG1] Space-time (angle-Doppler) beamformer consisting of N independent antenna channels and M time taps (pulse returns) comprising the coherent processing interval (CPI). A [7]; while the temporal DOFs could specific angle-Doppler pattern is obtained by judicious selection of the complex linear be the sampled pulse returns or combiner weights. For other space-time variations of the above architecture (e.g., postDoppler bins [7]. In any case, the Doppler beamspace), see [6] and [7]. −10 D D Di opp str le ibu r C tio lut ν= 0 n ter (Impulse −20 −30 ν≠ −40 0.5 No rm 0 ν=0 Target e 0 ize 0 idelob ν≠ eting S p d m o C r Do Clutte 0 pp le −0.5 Angle r alized −0.5 Norm al 0.5 θJ = –60° 1 θJ = –30° θJ = +45° 2 3 Jammer Nulls 0 −10 −20 −30 ch −40 ot 0 t ut Cl N er −50 −60 Relative Amplitude (dB) 0 1- Normalized Doppler Relative Amplitude (dB) n tributio tter Dis tial Clu a p S 1-D −70 0.5 −0.5 −0.5 −80 0 Normalized Angle (a) 0.5 (b) [FIG2] (a) Notional clutter spectrum for an airborne MTI radar and (b) the optimal 2-D space-time filter response for a boresight aligned target with a normalized Doppler of 0.25 [7]. IEEE SIGNAL PROCESSING MAGAZINE [42] JANUARY 2006 optimum STAP weights vector, w ∈ C NM , that maximizes the signal-to-interference-plus-noise ratio (SINR), satisfies the matrix Weiner-Hopf equation w = R−1 s, (1) ed by the STAP algorithm, and L data samples are used in the formation of (2). Ideally, if the snapshot data samples were drawn from a Gaussian independent and identically distributed (i.i.d.) stochastic process, then (2) corresponds to the maximum likelihood (ML) covariance estimate, i.e., R̂ = arg max f (x i : i = 1, . . . , L | R), (3) where s ∈ C NM is the desired target signal of interest (i.e., {R } steering vector [7]), and R ∈ C NM×NM is the clutter-plusjamming-plus-noise covariance matrix [7], which is guaranwhere f (x i : i = 1, . . . , L | R) denotes the probability density teed to be positive definite in function (pdf) of L snapshots conpractice due to the omnipresent ditioned on R. Reed, Mallet, and receiver noise [7]. Brennan (RMB) have shown that THE IMPETUS FOR THIS PROGRAM A meditation of (1), in the in this case, the expected SINR IS THE EVER INCREASINGLY COMPLEX context of practical real-world loss ratio (adaptive/ideal), ρ, due MISSIONS AND OPERATIONAL considerations, leads one to to finite sample estimation is [10] ENVIRONMENTS immediately recognize a set of serious fundamental issues [7]. L − NM + 2 L ≥ NM, (4) ρ= ■ R is potentially very large (due to the product NM). L+1 ■ R cannot be known a priori since it depends on the clutter where it is assumed that there are at least NM samples. Note and jamming environment. It must therefore be estimated, at that ρ → 1 as NM → ∞ as expected. A useful rule-of-thumb least in part, on the fly. derived from (4) is that, to achieve a loss no greater than 3 dB, ■ Due to the unknown statistics of the radar interference, at least on order of 2NM i.i.d. samples are required. For an R must be determined adaptively from the data. N = 16, M = 16, case, this corresponds to an assumption that ■ The need for immediate results will require (1) to be impleat least 512 i.i.d. samples are available that share identical statismented on the aircraft in real time. tics with the cell under test. A first-generation solution to these issues essentially consists of In practice, the L training samples are derived from a subset the following elements [5]–[7]: of the total space-time data cube measured during a coherent ■ selection of the STAP domain (e.g., pre-Doppler element processing interval (CPI) [5]. One such commonly discussed space and post-Doppler beamspace [7]) scheme is depicted in Figure 3. Unfortunately, as will be made ■ selection of a rank-reduction method since a full dimensional implementation of (1) is not only very difficult and costly but often not a good idea since adding adaptive DOFs (ADOFs) can actually hurt performance in nonstationary environments − [1]–[3], [7] − ■ selection of a covariance estimation scheme (implicit or − Array explicit) [5]–[7] Snapshot Vector ■ a real-time computing architecture [8]. X i+2 Although many ingenious solutions to the first, second, and fourth items have been devised, they have been predicated on a X i+1 single basic approach to the third: sample covariance estimation derived from local radar measurements [5]–[7]. In practice, this Range Test Cell Xi estimation is generally implicit since the solution to a real-time computing architecture generally consists of a data domain implementation (e.g., a QR-factorization implemented via a parX i −1 Guard Cells allel pipelined computing architecture [8], [9]) in which a covariance matrix is not explicitly formed. It is, nonetheless, X i −2 mathematically equivalent to replacing R in (1) with a sample − estimate R̂ generally of the form − − R̂ = 1 L
x i x′i , (2) i∈{L data samples } where x i ∈ C NM is a space-time snapshot [7] data vector for the ith range bin measured by the radar in the domain select- [FIG3] An example data selection strategy for estimating a sample space-time covariance matrix. L data samples are selected from range bins in proximity to a range cell under test (ith range bin in this example). The actual test cell, and cells adjacent to it (guard cells), may be excluded to avoid contamination from the actual target signal [5]. IEEE SIGNAL PROCESSING MAGAZINE [43] JANUARY 2006 readily apparent in the next section, real-world clutter often violates any such stationarity assumptions. KA-STAP Range Bin BACK TO “BAYES-ICS” EXAMPLES OF REAL-WORLD CLUTTER In this section, we describe in some detail basic ways in which It is obvious that real-world ground clutter is not well modprior knowledge can be incorporated into the most demanding eled by a homogenous stationary stochastic process [1]–[3]. component of a modern moving target indication radar: the Variations in underlying terrain, foliage, land-sea interfaces, space-time adaptive beamformer. Though radar centric, the and urban/manmade strucmethods discussed in the followtures, as well as nonlinear array ing were specifically chosen due responses (e.g., circular and/or to their general applicability in LOOK-AHEAD RADAR SCHEDULING IS tilted arrays [24], [25]), all conmany other adaptive sensor sigTHE KEYSTONE TO THE KASSPER tribute to stationarity violanal processing systems such as HPEC ARCHITECTURE tions. Consequently, significant sonar, lidar, and other multidideviations from predicted ideal mensional sensor arrays, where performance are to be expected—their exact nature of course environmental clutter (as opposed to say random thermal depends intimately on the interference [3]. receiver noise) is a dominant source of interference. We will Figure 4 shows a comparison between actual measured describe two broad categories of KA processing: Case I— radar returns for the DARPA Mountain Top experimental UHF Intelligent Training and Filter Selection (ITFS) and Case II— radar [11] located at the White Sands Missile Range in New Bayesian Filtering and Data Prewhitening. The former Mexico [Figure 4(a)] and what would have been measured if consists of primarily indirect exploitation of prior knowledge the terrain were homogenous [Figure 4(b)]. This form of sources (such as training data selection), while the latter entails inhomogeneity leads to either over or under nulling of the direct filtering of the incoming multidimensional data stream clutter [7], with resulting poor detection or false alarm rate based on prior information. performance, respectively. Figure 5 shows data from an experimental X-band airborne CASE I—ITFS radar [12]. Clearly evident are distinct bright clutter discretes In the ITFS approach, prior knowledge of the interference envithat will not be adequately nulled if an averaging process like (2) ronment is used to optimize two adaptive filtering processes: the is simply applied. filter selection and the filter training strategy. In the case of For details on the many deleterious effects of nonstationary radar clutter, this is accomplished by first conducting an envinon-Gaussian, real-world clutter, the reader is referred to the ronmental segmentation analysis based on whatever prior terproceedings of the KASSPER workshops [13], available via the rain/clutter database is available. Everything from digital terrain internet at www.darpa.mil/spo/programs/kassper.htm. and elevation data, land cover/land use (LCLU) to synthetic aperture radar (SAR) imagery; even hyperspectral imagery can be used [1]. Land clutter tends to be clumpy, that is it tends to be locally similar but with distinct and often abrupt boundaries (see Figure 6 for example). Clearly 600 600 from physical principles, an adaptive filter should not attempt to lump all these regions together and apply a 500 500 single filtering strategy. Instead, a segmentation analysis should be performed and an adaptive filter tailored to that 400 400 region should be applied. Generally speaking, the filter selection stage deter300 300 mines what type of adaptive filter is best suited to a given segmented region. In the case of STAP filtering 200 200 for clutter suppression in radar, a pivotal step is the domain in which the actual filtering is performed (e.g., 100 100 pre- or post-Doppler or element or beamspace [5]–[7]) and the number of ADOFs, which manifests itself ultimately in the size of the adaptive filter. For example, in 0 200 0 200 −200 −200 the case of the principal components (PC) method, the Doppler (Hz) Doppler (Hz) number of ADOFs refers to the number of significant (a) (b) eigenvectors to be included in the adaptive weight cal[FIG4] Comparison between (a) real-world clutter from the DARPA culation. Similarly, for the multistage Weiner filter, the Mountain Top radar and (b) returns assuming homogenous clutter [7]. number of stages is the metric for ADOFs [14]. What is The colormap scale is in decibels and ranges from 0 (noise floor, dark critical is that the number of ADOFs be matched to the blue) to 70 dB (dark red). IEEE SIGNAL PROCESSING MAGAZINE [44] JANUARY 2006 Scatterer Power (dB) 600 20 400 10 20 600 200 0 0 −200 −10 −400 −20 Down Range (m) Down Range (m) Resolution Cell Power (dB) 10 400 200 0 0 −10 −200 −20 −400 −600 −600 −30 −30 −600 −400 −200 0 200 400 600 800 1,000 Cross Range (m) −600−400 −200 0 200 400 600 800 1,000 Cross Range (m) (b) (a) [FIG5] Example real-world X-band radar measurements corresponding to a geographical location with discrete clutter (see [14] for details). (a) Hi-resolution geo-registered reflectivity image. (b) Corresponding discrete map. Range Bin % available training data (and of course the real-time computing this approach is the Bayesian covariance estimation approach of architecture). A general rule of thumb, which has its origins in Anderson [16]. the RMB result (4) but is rigorously proved by Smith [27], is Wishart [17] established that the elements of a sample that there be on order 2k i.i.d. samples available for training covariance matrix [L R̂] i, j formed from an outer product sum of the adaptive weights (e.g., sample covariance estimation), L Gaussian i.i.d. samples, i.e., where k is the number of effective ADOFs. L 1
Once a basic filtering structure has been selected, a training x i x′i (5) R̂ = strategy can be selected and optimized for that choice. Basically, L i=1 all or a subset of the samples from the locally stationary region obey a Wishart distribution (actually complex Wishart [18]) of are utilized in the weight training stage. In the case of PC, all of degree L, i.e., R̂ ∼ W(L R̂, L). the range bins—including the cell under test—might be included since it has been shown that this approach is robust to target cancellation. In contrast, a multibin post-Doppler approach without a PC technique might need to take extra Scan % = 4,796 care and introduce exclusion and guard cells to prevent 1,000 target signal cancellation [5]. 80 Figure 7 illustrates the impact that ITFS can have 900 when applied to real-world data. As described in [15], 800 70 the multichannel airborne radar measurement 700 (MCARM) data set (see Figure 6) included a number of significant highways, i.e., moving clutter. If one simply 60 600 applied traditional sample averaging techniques such 500 as those previously described, one could suffer signifi50 400 cant detection losses at roadway speeds [15]. Using an intelligent training and adaptation scheme that essen300 40 tially took account of the road networks, a significant 200 improvement in detection was achieved. 100 CASE II—BAYESIAN FILTERING AND DATA PREWHITENING In the Bayesian approach to radar STAP, prior knowledge is used directly by the filter to aid in adapting to nonstationary clutter. A convenient pedagogical framework for 30 −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 Fractional Doppler Frequency [FIG6] Measured range-Doppler returns for the MCARM dataset [12] showing the highly segmented nature of radar clutter returns. IEEE SIGNAL PROCESSING MAGAZINE [45] JANUARY 2006 If a prior estimate of the covariance matrix exists, R̂0 , it is not unreasonable to assume it too is Wishart distributed [16]. The rationale for this is simple. If it is based on prior radar observations, then it is also of the form (5). Anderson has shown that in this case, the Bayesian estimate (maximum a posteriori) for the covariance matrix is given by [16] An obvious yet useful generalization of (6) is R̂ = α R̂0 + β R̂1 α + β = 1, which is the familiar colored loading or blending approach of [19] and [20]. The practical advantages of (7) relative to (6) are many. For example, the data used to form the prior covariance might lose its relevance with time—the so-called stale weights problem [21]. In that case, even though R̂0 might have been formed from L0 samples, it effectively has less information and should be commensurately deweighted. A common method for accomplishing this, borrowed from Kalman filtering, is the fading memory approach in which case α, in (7), is R̂ = max f (x i : i = 1, . . . , L | R) f(R) R =
 1 L0 R̂0 + L1 R̂1 , L0 + L1 (6) where f (R) denotes the prior pdf associated with the prior covariance estimate R̂0 based on L0 samples, and thus is W(L0 R̂0 , L0 ); and R̂1 denotes the ML estimate based on L1 samples. Equation (6) has an obvious intuitive appeal: the a posteriori covariance estimate is formed as a weighted sum of the prior and current estimates with weighting factors proportional to the amount of data used in the formation of the respective sample covariances. α = e−γ t L0 , (8) where t is the time elapsed since the prior covariance estimate was formed, and the positive scalar γ is the fading memory constant [22]. Sliding Window: 2×DoF with Four Guard Cells Knowledge-Based Training MCARM 5_575, Doppler 10, Symmetric Window MCARM 5_575, Doppler 10, Data Selective Training 25 25 Target Not Detected 20 15 10 5 0 HWY9 15 10 5 0 −5 −5 −10 15 Target 20 M_MSMI (dB) M_MSMI (dB) (7) 16 17 18 19 20 21 −10 15 16 Range (Miles) (a) 13 71 15 6 300 42 1 17 18 19 20 21 Range (Miles) (b) Cells Excluded Rationale 13.5–14 Miles Rt. 15 16–17.5 Miles Rt. 13 19–20 Miles Rt. 9 in Sidelobe Region, Radial Alignment 21–22 Miles Rt. 9 in Main lobe 9 (c) (d) [FIG7] Illustration of the impact prior knowledge (in this case, prior road network data) can have on improving detection performance for the MCARM dataset [15]. (a) STAP filter residue without KA processing. (b) Target, which was previously undetected, is clearly visible after intelligent training [15]. (c) Local map of region indicating locations of road networks. (d) A training cell exclusion rule based on the map data. IEEE SIGNAL PROCESSING MAGAZINE [46] JANUARY 2006 In a more general setting, the blending parameters (α, β) could be chosen based on the relative confidence in the estimates. For example, R̂0 could be derived from a physical scattering model of the terrain. In which case it is also typically of the form (5) with the distinction that the outer products represent clutter patch steering vectors weighted by the estimated reflectivity [7], [23], i.e., Nc 1
G i v i v′i . R̂0 = Nc i=1 (9) Nc clutter patches have been utilized in the formation of R̂0 (typically corresponding to a particular iso-range ring [7]), where v i ∈ C NM is the space-time (angle-Doppler) steering vector corresponding to the ith clutter patch and G i its corresponding power [7]. Such information could be available a priori from SAR imagery [24] (essentially a high-resolution clutter reflectivity map) or physics-based models [23]. How the confidence metric applies, in the form of the weighting parameter α, is difficult to ascribe in practice since the quality of the a priori estimate is vulnerable to a number of error sources. A straightforward remedy is to choose α adaptively so as to maximally whiten the observed interference data. For example: min ZL(α), (10) {α} where    
  ′ ZL(α) =  yi yi − I   (11) i and  − 1 2 xi. y i = α R̂0 + β R̂1 (12) In (10)–(12), x i is the space-time snapshot vector for the ith range bin; (α R̂0 + β R̂1 )(−1/2) is the whitening matrix corresponding to a particular α; y i is the vector residue with dim(y i ) = dim (x i ); and the summation in (11) is performed over a suitable subset of the radar observations for which R̂0 is believed valid. If an a priori covariance estimate is available for each range bin, then the vector residue (12) can be replaced with  − 1 2 xi y i = α R̂0 (i) + β R̂1 (13) where R̂0 (i ) is the a priori estimate for the ith range bin. The above adaptive α approach is but a special case of an entire class of direct filtering methods incorporating prior information, viz., data prewhitening (or simply data detrending). In a more general setting, the space-time vector residues, {y i }, can be viewed as a detrended vector time series using prior knowledge in the form of a covariance based whitening filter. The major potential advantage of this is to remove (or attenuate) the major quasi-deterministic trends in the data (e.g., clutter discretes and mountains) so that the resulting residue vector time series is less nonstationary or inhomogeneous. An interesting example of this can be found in [12]. In this prewhitening example, a CLEAN algorithm was applied to the APTI data set of Figure 5(a), resulting in the discrete map of Figure 5(b). A deterministic covariance [23] was then formed as in (9), from which a square root whitening filter matrix could be derived. Figure 8 shows a log exceedance plot of the difference between the unwhitened data and the prewhitened data. Note the presence of spiky clutter as evidence by the so-called fat-tails in the unwhitened data. In the next section, we address the seemingly daunting challenge of incorporating prior knowledge—an inherently memory intensive process—into a high performance embedded computer. 100 Fraction Exceeding Value STAP Only STAP w/Prewhitening 10−1 c = r / √|r | 2 + |x | 2 10−2 s = x / √|r | 2 + |x | 2 10−3 r ⇐ √|r | 2 + |x | 2 x´ = –sr + cx Better Behaved "Tail" 10−4 r = c*r + s*x 13 dBI 10−5 −10 x x −5 0 5 10 15 Pixel SINR (dB) 20 25 30 r c,s c,s r c,s x´ [FIG8] Illustration of the effectiveness of the prewhitening approach on real-world data [24]. Prewhitening the data corresponding to Figure 5 resulted in a significant reduction in the tails of the clutter residue. [FIG9] An example of computer array implementing a data domain reformulation of the sample matrix-based Weiner-Hopf equation (from [8]). IEEE SIGNAL PROCESSING MAGAZINE [47] JANUARY 2006 REAL-TIME KA-STAP: THE DARPA KASSPER PROGRAM OBSTACLES TO REAL-TIME KA-STAP As mentioned previously, ingenious real-time computing architecture solutions have been devised to implement the sample matrix based (ML) solutions to STAP. In particular, to achieve the enormous throughput burden of a modern multichannel STAP radar, highly parallel HPEC systems based on so-called data domain reformulations of the Weiner-Hopf equation (1) have been devised [8], [9]. Figure 9 shows one such architecture based on a QR-factorization solution to (1) and (5). The basic parallel processing architecture of Figure 9 solves the adaptive Weiner-Hopf equation R̂ w = s by first performing a QR-factorization of the data matrix consisting of L space-time snapshots [7], then solving two triangular systems of equations (back substitution). More specifically, since R̂ ∝ X X ′ , (15) where X ∈ CNMLx is defined as X
[ x1 xL ] (16) (14) can be solved in two steps involving the R matrix of the QR factorization, i.e., if QX ′ =
 r , ⊘ (17) (14) pr o to xima Ac te ce Nu ss m Me be mo r of ry Cl Co ock nte C nts ycle s then the unknown adaptive weight vector w can be solved in two steps: 1 5–10 r ′ a =s Registers Internal, or L1 Cache L2, L3 25–50 Ap Faster Speed Higher Cost <1 102–103 106– ∞ Memory Disk Larger Size Lower Cost Tape [FIG10] Illustration of the time scales involved in accessing different memory storage media. Real-Time Data Stream Causal Processor KA-STAP Output Noncausal Processor [FIG11] Example of a KA-HPEC architecture exploiting the high degree of radar determinism with look-ahead time-scales on the order of seconds. The noncausal processor, running in parallel with a more conventional HPEC STAP processor, is used to look-ahead and detect regions of the radar field of regard requiring KA processing—and thus modifications to the normal adaptive weights calculations. r w =a. (18) With such architectures, tens to hundreds of giga floating point operations (GFLOPs) of real-time computing power can be achieved in hardware that can fit on an airborne radar aircraft. Though marvels of modern technology, these machines are cyber savants: they can solve (1) and (5) at blinding speeds in a strict pipelined fashion but grind to a snail’s pace if the data flow is disrupted for nonpipelined operations. This is a major fundamental obstacle to implementing KA or general Bayesian approaches which are inherently memory intensive (prior information needs to be stored). Figure 10 shows the order-ofmagnitude time scales for accessing different memory storage devices. Thus, to create a real-time KA HPEC (KA-HPEC, pronounced K-PEC) architecture, a major breakthrough in memory management must be achieved since much of the prior information (e.g., terrain maps, road networks, and discrete maps) will reside on mass storage (and thus slow) media. SOLUTION: LOOK-AHEAD SCHEDULING The key KA-HPEC breakthrough in the DARPA KASSPER project is based on a basic fundamental insight. There is a significant degree of determinism and thus predictability to radar clutter returns, particularly if the prediction horizon is only on the order of seconds. For example, let t0 denote the present time of the airborne radar depicted in Figure 11. Let t0 + δ t denote a time slightly in the future—say δ t = 1 s. Then in practice, the following are true: 1) The location of the aircraft at t0 + δ t can be predicted to a very high degree of accuracy assuming that no radical maneuvering is occurring. 2) The future state of the radar (look-direction, frequency, and PRF) at t0 + δ t is also known to a very high degree of accuracy. IEEE SIGNAL PROCESSING MAGAZINE [48] JANUARY 2006 The justification for the first assertion is simply that given the full kinematic KASSPER Hardware Architecture state vector of the aircraft (position, speed, and heading), Newtonian Holds Receiver Data, Platform mechanics insures fairly deterministic Data, etc to Be Read by the Signal Processor behavior—particularly for just a few KASSPER seconds into the future. Justification for Timing and Control etc. Signal Processor the second assertion arises from the simple fact that modern airborne radar systems typically utilize a radar scheduler. Since the radar is computer conControl trolled, it must have a tasking schedule. I/Q Data, The scheduler is highly deterministic INS Data, GPS Data, when considering a future time horizon Air Data, Display on the order of seconds. Sensor Data etc. Why are the above assertions so critiStorage "Knowledge" cal to solving the memory access probData lem described previously? Simple: they allow for look-ahead scheduling. More specifically, they allow for noncausal proHolds a-priori cessing whose prediction horizon is Knowledge to commensurate with the memory access be Read by the Surrogate for Signal delays! To see how this can be exploited Actual Radar Processor "Knowledge" System by a KA-HPEC architecture, consider Storage Figure 11. In this instantiation, a noncausal look-ahead computer is running in parallel with a more conventional [FIG12] The MIT Lincoln Laboratory 96-node real-time KASSPER HPEC system. causal STAP HPEC processor. The noncausal processor is used to spot trouble before it occurs and perform the necessary memory retrieval and prePlatform Moves to This Site ... computations to ensure that the right weight modificaNew Boundary Tiles Loaded to Cache from ... From This Tile tion scheme is ready to go when the data appear. Mass Storage Note that the KA information arising from the noncausal look-ahead processor can often be integrated into the casual (conventional) STAP processor via a straightOld Boundary Tiles Discarded from Cache forward augmentation of the sample snapshot matrix X defined in (16). For example, if R̂K A ∈ CNM×NM is an a priori estimate of the covariance matrix, then the [FIG13] Illustration of the environmental database manipulation illustrating the sliding window approach to migrating data from mass storage to RAM Bayesian linear combiner of (7) can be implemented by and ultimately to cache. X= (1/2)
√ βX1 √ βXL √ 12  α R̂K A , (19) where R̂K A ∈ CNM×NM is the matrix square-root of R̂K A. This KA instantiation really drives home the point that prior information is mathematically tantamount to having more data. Figure 12 shows the MIT Lincoln Laboratory KASSPER HPEC system, a real-time 96 noe parallel processing architecture implementing the noncausal look-ahead scheduling scheme of Figure 11 [26]. The system has the capability of receiving real-time in-phase and quadrature (I&Q) digitized samples from multiple receive channels over the full range extent of a radar and implementing a variety of KA algorithms throughout the entire radar signal processing chain from STAP to constant false alarm rate. Though when it comes to real-time HPEC, the devil is most certainly in the details; Figure 13 gives the basic gist of how the look-ahead scheduling is implemented. As the aircraft moves, a sliding window of data is migrated from a mass storage medium (e.g., disk drives) to a more readily accessible location (e.g., RAM). Depending on the particulars of the radar tasking, a first-pass decision is made as to what regions require KA processing. For example, if the radar is scheduled to point in a direction where a major road network is known to exist, essential details regarding this road network (orientation and range extent) are extracted and exploited in potentially several stages of the radar signal processing chain. Given the lookahead time buffer, this is all accomplished prior to the actual radar event. The exact extent of the sliding window depicted in IEEE SIGNAL PROCESSING MAGAZINE [49] JANUARY 2006 Figure 13 depends (of course) on the particular radar parameters (min/max range and altitude). THE FUTURE: NEXT GENERATION INTELLIGENT ADAPTIVE SENSORS The heart of the DARPA KASSPER program is an architecture for performing KA/Bayesian adaptive sensor signal processing. It is not a specific set of algorithms. Indeed, this article and the entire special issue are merely scratching the surface of potential instantiations not only for radar but for any sensor interacting with the environment. For example, multichannel sonar systems face an analogous set of problems when attempting to detect small target echoes or emanations in a background of largely environmentally induced noise. If environmental databases were available (bathymetry and sea state), KA processing could be employed. The same is true of automated target detection sensors working in the EO/IR (electro-optical/infrared) regime (e.g., hyperspectral and lidar). Though radar centric, it is hoped that this article will spark interest in others outside of radar and usher in, or should we say, re-usher in real-time Bayesian adaptive sensor signal processing. AUTHORS Joseph R. Guerci is director of the Special Projects Office of DARPA. In this capacity, he is responsible for leading the development of some of the nation’s most advanced technologies and systems aimed at meeting emerging national defense needs. An alumnus of Polytechnic University of New York with a Ph.D. in systems engineering, he has more than 60 peer-reviewed publications, eight U.S. patents, and is the author of Space-Time Adaptive Processing for Radar. He is also a Member of the IEEE Radar Systems Panel and a Fellow of the IEEE. Edward J. Baranoski is a program manager in the Special Projects Office at the DARPA where his focus is on next generation sensors and signal processing. From 1990 through 2004, he worked at MIT Lincoln Laboratory, serving on the technical staff developing space-time adaptive processing algorithms and as group leader of the Embedded Digital Systems Group and ISR Systems Group. He has previously worked at the Johns Hopkins Applied Physics Laboratory. He received his B.S. degree from Drexel University, the M.S. degree from the George Washington University, and the Ph.D. degree from Carnegie Mellon University, all in electrical and computer engineering. He was an Associate Editor for IEEE Transactions on Antennas and Propagation and has served on the IEEE Underwater Acoustics Signal Processing and Sensor Array and Multichannel (SAM) Technical Committees, serving as VP of the SAM Technical Committee from 2000–2003, and was cochair of the first IEEE Sensor Array and Multichannel (SAM 2000) Signal Processing Workshop. REFERENCES [1] J.R. Guerci, “DARPA KASSPER overview,” in Proc. 2004 DARPA Workshop Knowledge-Aided Sensor Signal Processing Expert Reasoning (KASSPER), Clearwater, FL, 5–7 Apr. 2004 [Online]. Available: http://www.darpa.mil/spo/programs/kassper/kassper04/day1/Guerci_KASSPER_04.pdf [2] A.O. Steinhardt and J.R. Guerci, “STAP for RADAR: What works, what doesn’t, and what’s in store,” in Proc. IEEE Radar Conf., 26–29 Apr. 2004, pp. 469–473. [3] W.L. Melvin, “Space-time adaptive radar performance in heterogeneous clutter,” IEEE Trans. Aerosp. Electron. Syst., vol. 36, no. 2, pp. 621–633, Apr. 2000. [4] W.L. Melvin, “Introduction to the special section on STAP,” IEEE Trans. Aerosp. Electron. Syst., vol. 36, no. 2, Apr. 2000, pp. 508–509. [5] J. Ward, “Space time adaptive processing for airborne radar,” MIT Lincoln Laboratory, Tech. Rep. 1015, Dec. 1994. [6] R. Klemm, Principles of Space-Time Adaptive Processing. London, U.K.: IEE Press, 2002. [7] J.R. Guerci, Space-Time Adaptive Processing for Radar. Norwood, MA: Artech House, 2003. [8] A. Farina, Antenna-Based Signal Processing for Radar Systems. Norwood, MA: Artech House, 1992. [9] A.O. Steinhardt and C. Rader, “Householder transforms in signal processing,” IEEE Signal Processing Mag., vol. 5, no. 3, pp. 4–12, July 1988. [10] I.S. Reed, J.D. Mallet, and L.E. Brennan, “Rapid convergence rate in adaptive arrays,” IEEE Trans. Aerosp. Electron. Syst., AES 10.6, pp. 853–863, Nov. 1974. [11] G.W. Titi, “The ARPA/Navy Mountaintop Program: Signal processing for airborne early warning,” in Proc. 1996 IEEE Int. Conf. Acoustics, Sonar Signal Processing (ICASSP), Atlanta, GA, 7–10 May 1996, vol. 2, pp. 1165–1168. [12] D.J. Zywicki, W.L. Melvin, G.A. Showman, and J.R. Guerci, “STAP performance in site-specific clutter environments,” in Proc. 2003 IEEE Aerospace Conf., Big Sky, MT, USA, 8–15 Mar. 2003, 2005–2020. [13] Proc. KASSPER Workshops, 2002–2005 [Online]. Available: http://www.darpa.mil/spo/programs/kassper.htm [14] J.R. Guerci, J.S. Goldstein, and I.S. Reed, “Optimal and adaptive reduced-rank STAP,” IEEE Trans. Aerosp. Electron. Syst. (Special Section on Space-Time Adaptive Processing), vol. 36, no. 2, pp. 647–663, Apr. 2000. [15] W. Melvin, M. Wicks, P. Antonik, Y. Salama, L. Ping, and H. Schuman, “Knowledge-based space-time adaptive processing for airborne early warning radar,” IEEE Aerosp. Electron. Syst. Mag., vol. 13, no. 4, pp. 37–42, Apr. 1998. [16] T.W. Anderson, An Introduction to Multivariate Statistical Analysis, 2nd ed. New York: Wiley, 1984. [17] J. Wishart, “The generalized product moment distribution in samples from a normal multivariate population,” Biometrika, vol. 20A, pp. 32–52, 1928. [18] S.U. Pillai, Array Signal Processing. New York: Springer-Verlag, 1989. [19] J.S. Bergin, C.M. Teixeira, P.M. Techau, and J.R. Guerci, “STAP with knowledge-aided data pre-whitening,” in Proc. 2004 IEEE Radar Conf., Apr. 2004, pp. 289–294. [20] J.D. Hiemstra, “Colored diagonal loading,” in Proc. 2002 IEEE Radar Conf., Apr. 2002, pp. 386–390. [21] J.R. Guerci, “Theory and application of covariance matrix tapering for robust adaptive beamforming,” IEEE Trans. Signal Processing, vol. 47, no. 4, pp. 977–985 Apr. 1999. [22] A. Gelb, Applied Optimal Estimation. Cambridge MA: MIT Press, 1984. [23] P.M. Techau, J.R. Guerci, T.H. Slocumb, and L.J. Griffiths, “Performance bounds for hot and cold clutter mitigation,” IEEE Trans. Aerosp. Electron. Syst., vol. 35, no. 4, pp. 1253–1265, Oct. 1999. [24] M. Zatman, “Circular array STAP,” in Proc. 1999 IEEE Radar Conf., pp. 108–113, 20–22 Apr. 1999. [25] G.K. Borsari, “Mitigating effects on STAP processing caused by an inclined array,” in Proc. 1998 IEEE Radar Conf., 11–14 May 1998, pp. 135–140. [26] Proc. 23rd Systems Technology Symp. (DARPA Tech), Anaheim, CA, Mar. 2004 [Online]. Available: http://www.darpa.mil [27] S.T. Smith, “Covariance, subspace, and intrinsic Cramer-Rao bounds,” IEEE Trans. Signal Processing, vol. 53, no. 5, pp. 1610–1630, May 2005. [SP] IEEE SIGNAL PROCESSING MAGAZINE [50] JANUARY 2006