IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 57, NO.

8, AUGUST 2009 2279

An Innovative Multiresolution Approach for DOA

Estimation Based on a Support Vector Classification
Massimo Donelli, Member, IEEE, Federico Viani, Paolo Rocca, and Andrea Massa, Member, IEEE

Abstract—The knowledge of the directions of arrival (DOAs) of on an array of narrow band sensors. Among them, the most
the signals impinging on an antenna receiver enables the use of widely known and used are estimation of signal parameters via
adaptive control algorithm suitable for limiting the effects of in- rotational invariance technique (ESPRIT) [5]–[7] and multiple
terferences and increasing the gain towards the desired signals in
order to improve the performances of wireless communication sys- signal classification (MUSIC) [8], [9]. Other approaches based
tems. In this paper, an innovative multi-resolution approach for the on the maximum likelihood (ML) DOA estimation have been
real-time DOA estimation of multiple signals impinging on a planar proposed [10], [11], as well.
array is presented. The method is based on a support vector clas- In the last years, great attention has been also paid to the
sifier and it exploits a multi-scaling procedure to enhance the an- use of learning-by-examples (LBE) techniques. LBE-based
gular resolution of the detection process in the regions of incidence
of the incoming waves. The data acquired from the array sensors approaches are able to provide a good trade-off between ac-
are iteratively processed with a support vector machine (SVM) cus- curacy and convergence, which is mandatory for real time
tomized to the problem at hand. The final result is the definition of a systems where fast reactions are required. Furthermore, they
map of the probability that a signal impinges on the antenna from a satisfactory deal with unknown configurations (i.e., different
fixed angular direction. Selected numerical results, concerned with from those “learned” during the training process) thanks to their
both single and multiple signals, are provided to assess potentiali-
ties and current limitations of the proposed approach. generalization capability. Within this framework, the benefits
of using radial basis function neural networks (RBFNN) have
Index Terms—Classification, direction of arrival (DOA) estima- been carefully analyzed in [12]. As a matter of fact, neural net-
tion, multiresolution, planar arrays, support vector machine.
works (NNS) are suitable in approximating nonlinear functions
as those in DOAs estimation. Moreover, they can be easily
implemented in analog circuits. An improved RBFNN-based
I. INTRODUCTION
approach has been presented by the same authors of [12] in
[13] to address the problem of tracking an unknown number of
multiple sources when no a-priori information on the number

I N THE LAST decades, the technology of adaptive antenna

arrays has been greatly advanced and applied to many mo-
bile and wireless communication systems [1], [2]. Within this
of impinging signals is available. More specifically, the region
above the antenna has been partitioned into angular sectors
and each sector “assigned” to a simpler NN, thus reducing
framework, the antenna beam-forming plays an important role with respect to [12] the problem complexity as well as the
and the estimation of the directions of arrival (DOAs) of sig- computational burden of the learning phase. Towards this end,
nals impinging on the array is a crucial task in order to enhance each network has been trained to detect the subset of incoming
the spatial diversity and consequently the spectral efficiency. As signals that impinge on the corresponding angular sector. Ac-
a matter of fact, such an information enables the generation or cordingly, only those NNs of the regions where the signals have
steering of the radiation pattern with a maximum towards the de- been detected in the first stage of the process are activated in
sired signals and nulls along the directions of interfering signals the second one to estimate the DOAs of the incoming signals.
[3], [4]. The effects of interferences are mitigated and both the More recently, some techniques based on support vector
gain and the performance of the whole communication system machines (SVMs) [14] have been analyzed to profitably ex-
are enhanced. For such reasons, the estimation of the DOAs of ploit their solid mathematical foundation in statistical learning
unknown interfering and desired signals is of great interest and theory [15]. The main advantages of those approaches lie in
it is still an open problem as confirmed by the number of papers their ability to deal with various and complex electromagnetic
published on this topic. problems [16], [17], and, analogously to NNs, in an easy
In the scientific literature, several methods have been pro- hardware implementation [18]. As far as the DOA estimation
posed for the direction finding of multiple signals impinging is concerned, a support vector regression (SVR) procedure has
been presented in [19] when dealing with linear arrays. In such
Manuscript received August 06, 2008; revised January 20, 2009. First pub- a case, a SVM has been used to estimate the DOA of each
lished June 05, 2009; current version published August 05, 2009.
impinging electromagnetic wave starting from a set of known
The authors are with the Department of Information and Communica-
tion Technology, University of Trento, 38050 Trento, Italy (e-mail: andrea. input-output examples where the DOAs of the signals were
[email protected]; [email protected]; [email protected]; uniformly distributed in the whole angular region above the re-
[email protected]). ceiver. Despite the generalization capability of the SVR-based
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. method, an a-priori information on the number of sources and
Digital Object Identifier 10.1109/TAP.2009.2024485 pre-fixed angular separations between the DOAs (as in [12])
0018-926X/$26.00 © 2009 IEEE
Authorized licensed use limited to: International Institute of Information Technology Bangalore. Downloaded on January 27,2025 at 07:02:46 UTC from IEEE Xplore. Restrictions apply.
 2280 IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 57, NO. 8, AUGUST 2009

have been considered to increase the reliability of the estimation

procedure. An extension of such a model has been presented in
[20] and experimentally validated in [21] successively.
In this paper, an innovative procedure for real-time direction
finding of signals impinging on a planar array of electromag-
netic sensors is presented. The problem of the DOAs estimation
is formulated as a two step procedure, where the first step is
aimed at determining the decision function that correctly classi-
fies whatever input pattern by means of a SVM-based approach.
In the second step, the output of the decision function is mapped
into the a-posteriori probability that a signal impinges on the
antenna from a fixed direction. In order to increase the accu-
racy of the estimation process and to reduce the computational Fig. 1. Planar array geometry.
burden affecting other DOAs procedures, the proposed two-step
strategy is nested into an iterative multi-scaling process [22].
Accordingly, the resolution accuracy is improved only in those are generated by a set of electromagnetic sources placed in the
angular regions where the unknown sources are supposed to be far-field of the receiving antenna. The open-circuit voltage at
located at the previous iteration. More specifically, the algorithm the output of the th sensor can be expressed as [20]
first determines a coarse probability map of the DOAs starting
from a training set where the incoming signals are nonuniformly
distributed along the elevation direction, , and the azimuthal
one, . Then, the SVM is used to classify the input test dataset
at successive resolution levels by performing a kind of synthetic
(1)
zoom in the angular regions of interest (ARoIs) where a higher
probability is detected and considering the same training set, where ,
thus performed only once and off-line. Concerning the antenna being the location of the th sensor expressed in wavelength,
architecture and unlike [13] and [20], planar arrays of sensors and is the background random noise at the th locations.
are considered since linear arrays lack the ability to scan in The noise samples are supposed to be statistically independent
3D-space and the estimation of both the elevation and the and characterized by a random Gaussian distribution with zero
azimuth angles is crucial and has many applications in var- mean value. Moreover, and are the electric field associ-
ious fields of engineering. For instance, a complete DOA infor- ated to the th impinging wave and the effective length of the
mation it is possible to improve the coverage of transmission th array element.
in wireless communications by avoiding interferences and en- According to the guidelines described in [3] and [4] about the
hancing the system capacity [23]. More specifically, planar ar- control of adaptive/smart antennas, the solution of the DOAs es-
rangements are very attractive in mobile communications with timation problem is based also in this work on the measurement
portable devices where the main beam must be scanned in any of the total correlation matrix, defined as
direction [24]. Moreover, the number of impinging signals is un-
known as well as their directions belonging to the whole angular
range above the planar antenna system (i.e., and (2)
).
The paper is organized as follows. The formulation of the where and the superscript stands
iterative two-step multi-resolution DOA approach (in the fol- for complex conjugation, at the output of the planar array since
lowing denoted by the acronym IMSA-SVM) is described in it contains sufficient information on the received signals [13].
Section II. In order to show the innovative features of the ap- From a statistical point of view, the problem at hand can be for-
proach and to assess its effectiveness, a selected set of numer- mulated as the definition of the probability map of the angular
ical results concerned with both single and multiple signals is incidence of the incoming waves starting from the knowledge of
reported and discussed (Section III). Moreover, some compar- the total correlation matrix . Towards this end, let us partition
isons with state-of-the-art techniques are also reported. Finally, the angular region above the array into a two-dimensional lattice
some conclusions are drawn in Section IV. of cells, each one corresponding to an angular
sector of sides and [Fig. 2(a)]. The status of each cell
II. MATHEMATICAL FORMULATION can be empty [ ], if any signal impinges
Let us consider a planar array of isotropic elements on the array from the angular region identified by the same cell,
displaced on a regular and rectangular grid with inter-element or occupied [ ], otherwise. Accordingly, the
spacing on the plane. A set of electromagnetic original problem can be stated as follows: “find the a-posteriori
waves impinge on the array from unknown angular directions probability function given a measured value of the total
, as sketched in Fig. 1. The signals, correlation matrix at the receiver.” Mathematically,
supposed to be narrow-band and centered at the carrier fre- can be also expressed as the linear combination of the nonover-
quency ( being the corresponding free-space wavelength), lapping basis functions , defined over
Authorized licensed use limited to: International Institute of Information Technology Bangalore. Downloaded on January 27,2025 at 07:02:46 UTC from IEEE Xplore. Restrictions apply.
 DONELLI et al.: AN INNOVATIVE MULTIRESOLUTION APPROACH FOR DOA ESTIMATION 2281

being the resolution index and . The summa-

tion over ranges from 0 [Fig. 2(a)], which corresponds to the
largest characteristic length scale, to [Fig. 2(b)], which
corresponds to the smallest angular basis-function support at the
th scaling step. For a given value of ,
is the number of nonoverlapped basis functions centered in the
angular sub-domain represented at the th resolution. Accord-
ingly, the iterative DOA detection procedure is aimed at locating
the terms of small length scale at those ARoIs [e.g., the shaded
cells in Fig. 2(a)–(b)] where the signals are supposed to impinge
with higher probability.
In order to profitably exploit the multiresolution represen-
tation of the a-posteriori probability function (4) and solving
the arising DOA problem, the following multistep classifica-
tion process is performed by means of a SVM-based technique.
More in detail
• Step 0— SVM Training Phase. The SVM is trained once
and off-line starting from the knowledge of a set of known
examples (i.e., input/output relationships)

(5)

called training set, where is the number of training data.

The samples of each training data are composed by
examples concerned with angular positions ,
, where a signal impinges on
the array [i.e., occupied directions—
], while the remaining are
related to empty directions [i.e.,
].
Starting from the knowledge of the training set, the
problem turns out to be the definition of a suitable dis-
Fig. 2. IMSA-DOA Procedure—Angular region partitioning and ARoIs iden- criminant function
=1
tification at the steps s (a) and s =2
(b).

(6)
the angular lattice
that separates the two classes and
on the basis of the total correlation matrix measured
(3) at the output of the planar array. In order to approach the
problem with a single classifier, the problem at hand is
reformulated as that of building the following single output
where the weighting coefficient is the probability function
value that a wave impinges on the array from the th angular
sector [i.e., ] and
if belongs to the th cell and otherwise.
In order to improve the achievable angular resolution, a multi- (7)
resolution representation of the unknown function is
looked for [Fig. 2(b)— ] by exploiting an iterative process Towards this purpose and according to the SVM theory
analogously to [22]. More specifically, the probability function [15], the following linear decision function is adopted
is expressed at the th step of the iterative procedure as a twofold
summation of shifted and dilated spatial basis functions

(8)

is determined in a space (called “feature space”) with

(4) a higher dimensionality than the original input data space
Authorized licensed use limited to: International Institute of Information Technology Bangalore. Downloaded on January 27,2025 at 07:02:46 UTC from IEEE Xplore. Restrictions apply.
 2282 IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 57, NO. 8, AUGUST 2009

and obtained through the nonlinear operator [15]. The where if

unknown terms and , which unequivocally define the and
decision hyperplane , are the normal vector and a bias, otherwise, allows one to identify the ARoIs, ,
respectively. They are computed during the Training Phase defined as those angular sub-domains
according to the guidelines described in [17]; where ;
• Step 1—Low-Order DOA Estimation . At the first — Step 2.b—Multiresolution DOA Estimation. A syn-
step, a coarse probability map [(4)— ] is determined thetic zoom is performed by refining the representation
by means of the SVM classifier mapping the decision func- of the unknown function and increasing
tion into the a-posteriori probability function. the angular resolution only in the ARoIs
The unknown probability coefficients , identified at (Step 2.a). Therefore, the multiresolution
are approximated with a sigmoid function a-posteriori probability function (4) is updated 1 by
[15] as follows setting and computing the
new highest resolution coefficients, ,
when , as in (9);
(9) • Step 3—Termination Criterion . The sequence
of operations of Step 2 is repeated until both the dimen-
sions and the number of ARoIs between two consecutive
where and are two parameters computed according to a
cycles are stationary [i.e., and the vari-
fitting process [17] starting from a subset of the training
ations of the dimensions of the ARoIs are not greater than
data of the Training Set;
• Step 2—IMSA-SVM Process ( ). the highest angular resolution at the th step,
— Step 2.a—Angular Regions of Interest Identi- ].
fication . Starting from the probability map
previously (i.e., at the th iteration) determined,
such a step is aimed at identifying the angular sectors III. NUMERICAL SIMULATIONS AND RESULTS
, where the signals are supposed
to impinge in order to improve the resolution only in In order to assess the effectiveness and reliability of the
those regions and enhance the accuracy of the DOA es- proposed approach, an exhaustive set of numerical experiments
timation. Towards this end, first the values of the func- has been performed and some selected results will be reported
tion are scaled, thus defining the following in the following for illustrative purposes. The remaining of
new set of normalized probability coefficients this section will firstly (Section III-A) illustrate the behavior
of the multi-scaling procedure also in comparison with other
state-of-the-art approaches for DOA estimation. The second
part (Section III-B) will be devoted to analyze the poten-
tialities and current limitations of the IMSA-SVM approach
when dealing with various and challenging electromagnetic
scenarios. In such a framework, some configurations in which
(10)
conventional state-of-the-art signal subspace-based array pro-
cessing techniques cannot be applied are also dealt with in
where
order to point out the enhanced range of applicability of SVM
approaches. Finally, a uniform array of -dipoles is consid-
and ered (Section III-C) to verify the suitability and reliability of
. the proposed method in correspondence with a realistic array
Successively, the new probability function modelling With reference to the geometry shown in Fig. 1, a
square planar array of isotropic radiators spaced by
is considered. The power of the impinging signals has
been set to , above the level of the
background noise.
Concerning the training set, the following setup and
is thresholded by nulling the scaled coefficients greater has been assumed and the SVM classifier has been
than a user-defined threshold . Finally, the thresholded trained once and off-line on the same data set whatever the test
function experiment. As regards to the training examples,
different scenarios have been considered, being the
number of configurations with signals. Moreover, the actual
DOAs of the signals of the training data have been randomly
1It is worth noting that at the sth step of the multi-scaling procedure only the
angular ranges belonging to the ARoIs are processed by the SVM classifier with
(11) a non-negligible saving of computational resources.
Authorized licensed use limited to: International Institute of Information Technology Bangalore. Downloaded on January 27,2025 at 07:02:46 UTC from IEEE Xplore. Restrictions apply.
 DONELLI et al.: AN INNOVATIVE MULTIRESOLUTION APPROACH FOR DOA ESTIMATION 2283

step ( ), the planar angular region is partitioned

into cells (being and ,
, the angular steps along the elevation direction, , and
the azimuthal one, , respectively) and a coarse DOA prob-
ability map is determined following the procedure described
in Section II (Step 1). Then, the multi-scaling procedure takes
place ( ). The ARoIs are identified and partitioned into
cells with an angular resolution of
and . For the sake of space, only the
DOA probability map obtained at the end of the second step
is shown in Fig. 3(a). The procedure is then iterated
until with the final result reported
in Fig. 3(b) characterized by an angular resolution in
equal to and . As it can be observed
(Fig. 3), the region with higher probability of incidence turns
out to be closer and closer to the actual angular location of the
signal when increasing the step number. Quantitatively such an
event can be analytically quantified by computing the values of
the location index (Fig. 1) and of the incidence area
defined as follows

(13)

where, see equation at the bottom of the page, and

(14)

Fig. 3. Single signal scenario, I = 1—Probability map determined by the

IMSA-DOA procedure at: (a) s = 2, (b) s = S = 4.
being
chosen in a discrete grid of locations ,
belonging to the the angular region above the antenna
, are the actual coordinates of the signal
incidence point, whereas
(12)

and being the floor function and the ceiling function, re-
spectively. Moreover, in order to fully assess the generalization
properties of the SVM-based approach, the DOAs of the test ex-
amples are different from those of the training dataset.

A. Single Signal Scenario—Comparative Assessment

The first experiment deals with the DOA detection of a single
(15)
signal and a test set of examples related to the
single-signal scenario has been considered. An illustrative de-
scription of the behavior of the proposed IMSA-SVM approach
is shown in Fig. 3 dealing with the “representative” (of the identify the center of the th ARoI where the signal/signals
method performance on the whole test dataset) configuration is/are supposed to impinge. As a matter of fact, the value of the
of a signal coming from . At the first location index reduces from down to

Authorized licensed use limited to: International Institute of Information Technology Bangalore. Downloaded on January 27,2025 at 07:02:46 UTC from IEEE Xplore. Restrictions apply.
 2284 IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 57, NO. 8, AUGUST 2009

TABLE I
STATISTICS OF THE AVERAGED PERFORMANCE INDEXES (&^ = & AND
^= ) FOR DIFFERENT SIGNAL CONFIGURATIONS (I = 1; 2; 3; 4)

(being and ). Analogously,

, , , until . As re-
gards to the whole set of test examples, the statistics of the “con-
vergence” values of the indexes (13) and (14) are given in the
first block of Table I.
In order to get an insight into the advantages of the pro-
posed multi-resolution approach over the classification single-
step techniques, a bare DOA SVM-based method has been con-
sidered and applied to the same test example. To fairly com-
pare the two methods, the same training dataset has been used.
Moreover, the same angular resolution has been adopted in both
cases. Towards this purpose, an angular lattice characterized
by a uniform grid whose cell side was equal to the finest dis-
cretization of the multi-resolution procedure (i.e.,
and ), has been defined over the whole angular
investigation domain of the single step SVM approach. As it
can be observed [Fig. 4(a)], although the value of is quite
close to that of the IMSA strategy (i.e.,
versus ), the extension of the incidence area
turns out to be significantly wider ( versus
). On the other hand, it cannot be neglected that
the CPU-time of the test phase of the bare procedure is approx- Fig. 4. Single signal scenario, I = 1—Probability maps obtained with dif-
imately fifty times the one of the IMSA-SVM because of the ferent classification approaches: (a) single-step SVM, (b) MLP neural network,
need to obtain a detailed map in the whole investigation area and (c) RBF neural network [1 = 1:25 and 1 = 5 ].
instead of in a limited ARoI, , only. As a matter of
fact, the number of test points used by the IMSA approach turns
out to be widely reduced. confirmed by the values of the location index:
For completeness, the results from other standard nonlinear and .
classification methods, such as the multilayer perceptron The last analysis is concerned with the comparison between
(MLP) and the radial basis functions (RBF) neural network, the IMSA-SVM and those state-of-the-art methods for DOA es-
have been analyzed, as well. More specifically, the DOA prob- timation aimed at determining the angular incidence of the sig-
ability maps obtained with the MLP-based and RBFNN-based nals, namely MUSIC, ESPRIT (i.e., two one-dimensional ES-
classifiers are reported in Figs. 4(b) and 4(c), respectively. PRITs independently-applied to the arrays followed by an align-
Whatever the method, the achieved estimate does not appear ment procedure to associate the estimated azimuth and elevation
to be adequate and certainly not comparable neither with that angle), -unitary ESPRIT [7], and a support vector regres-
of the IMSA-SVM [Fig. 4(b)–(c) versus Fig. 3(b)] nor with sion-based (SVR) approach. Towards this end, the azimuthal di-
that of the bare SVM [Fig. 4(b)–(c) versus Fig. 4(a)] as also rection of the actual signal has been fixed to , while
Authorized licensed use limited to: International Institute of Information Technology Bangalore. Downloaded on January 27,2025 at 07:02:46 UTC from IEEE Xplore. Restrictions apply.
 DONELLI et al.: AN INNOVATIVE MULTIRESOLUTION APPROACH FOR DOA ESTIMATION 2285

TABLE II
SINGLE SIGNAL SCENARIO, I = 1—COMPARATIVE ASSESSMENT. VALUES OF THE LOCATION INDEX & WHEN APPLYING IMSA-DOA, SVR, MUSIC, AND ESPRIT

Fig. 6. Single signal scenario, I =1—Behavior of the location index versus

Fig. 5. Single signal scenario, I = 1—Uniform (points) and nonuniform (tri- the number of failed array elements.
angles) angular training sets.

and the standard ESPRIT implementation. Towards this end,

the elevation angle has been varied in the range . an increasing number of array elements has been switched off.
Moreover, the SVR algorithm has been previously trained with Moreover, the a-priori information on the failure of some array
a dataset composed by examples concerned elements has not been exploited through the definition of an
with only one signal ( ). The methods are then compared ad-hoc training set, but the same nonuniform set of input-output
by means of the resulting signal location error, . examples concerned with the unperturbed array structure has
Because of the planar array of isotropic elements and as ex- been used. The results of the comparative assessment when
pected [25], the performances of the DOA techniques in ele- are reported in Fig. 6.
vation-estimation turn out to better at high elevations
[Table II], while the azimuth-estimation is greatest at low el- B. Complex Scenarios—Performance Analysis
evations . Moreover, the values of the estimation in- The following experiments are aimed at assessing the effec-
dexes point out that the IMSA-SVM (last column—Table II) is tiveness of the IMSA-SVM in detecting the DOAs of multiple
able to obtain similar results, in terms of angular resolution, than signals.
those provided by the SVR and of the same order in magnitude Dealing with the detection of two different incidence points,
of MUSIC and ESPRITs except for wider angles , the first example is concerned with test signals coming from
even though these latter need more CPU-time (i.e., an optimized and , respec-
IMSA-SVM implementation just needs few milliseconds on a tively. The probability maps estimated by the IMSA-SVM at
PC equipped with a processor and of RAM). different steps are shown in Fig. 7 together with those obtained
As regards to the growing of the location index around 60 , its with the single-step SVM classification procedure [Fig. 7(d)],
mainly depends on the training set. As a matter of fact, it can be the MLP-based approach [Fig. 7(e)], and the RBF technique
avoided by modifying the off-line training phase. For instance, [Fig. 7(f)]. As expected and confirming the outcomes from the
the choice of a uniform angular distribution of the training sam- study of the single-signal detection, the multiscaling process
ples (Fig. 5), instead of a nonuniform arrangement, allows one allows one to significantly enhance the performances of the
to obtain a behavior of almost invariant to for medium-high single-step classification approaches as pictorially shown in
elevations. Fig. 7 and quantitatively confirmed by the indexes in Table III.
In order to point out the generalization capabilities of the pro- Moreover, it is worth noting that this conclusion is not limited
posed approach as well as its robustness to the model tolerances to a particular configuration of incidence angles, but it holds
[14], [26], the effect of the array failure has been evaluated true whatever the two-signals scenario under test.
and the arising results compared to those with 2D-unitary In order to assess the stability of the proposed approach, a
ESPRIT which demonstrated several advantages over MUSIC test set composed by examples has been con-
Authorized licensed use limited to: International Institute of Information Technology Bangalore. Downloaded on January 27,2025 at 07:02:46 UTC from IEEE Xplore. Restrictions apply.
 2286 IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 57, NO. 8, AUGUST 2009

Fig. 7. Multiple signals scenario, I =2—Probability maps obtained with Fig. 7. (continued) Multiple signals scenario, I =2 —Probability maps ob-
different classification approaches: IMSA-SVM [(a) s =1
, (b) s =2
, (c) tained with different classification approaches: [(d) single-step SVM, (e) mul-
s = S =3 ]. tilayer perceptron (MLP) neural network, and (f) radial basis function (RBF)
1 =1
neural network [   =25 1 =1
: and   = 10 ].

sidered. The results obtained with the IMSA-SVM are summa-

rized in Table I (second block). As expected, the mean values of nals scenarios [ , versus
the averaged performance indexes ( and
, and , ].
) turn out to be very close to those of the previous
As an illustrative example, let us consider the case of a set
test example [i.e., , versus
of signals impinging on the array from ,
, and , , . Starting from
]. the coarse map determined, three different ARoIs are succes-
The second numerical experiment, concerned with multiple sively identified [Fig. 8(a)] and better resolved thus iteratively
incidences, considers three-signals configurations. As regards improving the DOA resolution accuracy as pointed out by the
to the results for a test set of three-signals ex- indexes in Table IV where the values estimated by the other clas-
amples, the values in the third block of Table I indicate that sification approaches are reported [Fig. 8(b)], as well. By com-
the resolution accuracy of the proposed approach does not sig- paring the distribution at the th step of the IMSA and the
nificantly reduce with respect to the single-signal or two-sig- one from the bare SVM, it is evident the improvement guaran-
Authorized licensed use limited to: International Institute of Information Technology Bangalore. Downloaded on January 27,2025 at 07:02:46 UTC from IEEE Xplore. Restrictions apply.
 DONELLI et al.: AN INNOVATIVE MULTIRESOLUTION APPROACH FOR DOA ESTIMATION 2287

TABLE III TABLE IV

MULTIPLE SIGNALS SCENARIO, I =2
. PERFORMANCE INDEXES WHEN MULTIPLE SIGNALS SCENARIO, I =3
(CONFIGURATION = = ). 111
APPLYING IMSA-DOA, SINGLE-STEP SVM, MLP NEURAL NETWORK, AND PERFORMANCE INDEXES WHEN APPLYING IMSA-DOA, SINGLE-STEP SVM,
RBF NEURAL NETWORK MLP NEURAL NETWORK, and RBF NEURAL NETWORK

configuration where four separated signals can be rec-

ognized [ , ,
, — Fig. 9(a)
and (c) ]. The second example [Fig. 9(d)] deals with a two-clus-
ters setup [Configuration ,
, ,
], while a single signal and a cluster
of three-signals are present in the last example [Configura-
tion , ,
, ]. Whatever
the example, the multiscaling process is able to identify with
an ever increasing resolution from [Fig. 9(a)] up to
[Fig. 9(c)] the ARoIs to which the incidence
directions of the actual signals belong as pointed out by the
numerical indexes , in Table V. On the other
hand, it should be noticed that the DOA estimation process
tends to cluster multiple regions-of-incidence in a single ARoI
when the angular separations among the signals reduce. Such
an event takes place also in correspondence with the “Config-
uration ” [Fig. 9(d) and Table VI] where two ARoIs are
identified. It is even more evident in Fig. 9(e) (Table VII) where
the angular incidences of three signals are detected in only one
ARoI. The “clustering” effect is quantitatively pointed out by
the behavior of the averaged localization index (Table I—fourth
block) when dealing with the complete test set
to which previous examples belong. As a matter of fact, there is
a significant increase of the compared to the values of
the same quantity when [ versus
, , ], even
Fig. 8. Multiple signals scenario, I =3 111
(Configuration = = )—Probability
though the value of remains close to those of other
maps obtained with different classification approaches: (a) IMSA-SVM [s = multiple-signals configurations since the estimated ARoIs still
S =3 1 =1
] and (b) single-step SVM [   and 1 =1  ]. carefully identify the actual incidence areas.
The fourth and fifth experiments deal with more critical test
scenarios since the examples under test are concerned with a
teed by the multiscaling process both in resolving and properly number of signals different from that in the training set (i.e.,
locating a number of ARoIs equal to the number of signals ( ). ). More specifically, let us consider the Clus-
In the third experiment, sig- tered Distribution of signals with incidence directions
nals impinge on the planar array. Fig. 9 shows the re- indicated by the white points in Fig. 10. It is worth noticing that
sults provided by the IMSA-SVM and in correspondence such a configuration turns out to be not admissible (i.e.,
with a set of representative examples. More in detail, estimates cannot be obtained) for signal subspace-based array
the first example (Configuration ) refers to a processing techniques as 2D-unitary ESPRIT when the planar
Authorized licensed use limited to: International Institute of Information Technology Bangalore. Downloaded on January 27,2025 at 07:02:46 UTC from IEEE Xplore. Restrictions apply.
 2288 IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 57, NO. 8, AUGUST 2009

=4
Fig. 9. (continued) Multiple signals scenario I —Probabilitymaps ob-
1111
tained with the IMSA-SVM. Configuration = = = :Configuration = : step22
s = S =3 13 =
(d); Configuration = s S =3 (e).

TABLE V
MULTIPLE SIGNALS SCENARIO, I =4
(CONFIGURATION = = = ). 1111
PERFORMANCE INDEXES WHEN APPLYING IMSA-DOA, SINGLE-STEP SVM,
MLP NEURAL NETWORK, AND RBF NEURAL NETWORK

=4
Fig. 9. Multiple signals scenario I —Probability maps obtained with the
1111
IMSA-SVM. Configuration = = = : step (a) s =1
, (b) s =2
, and (c) s =
S =3.

array structure at hand is used. As a matter of fact, the maximum

number of sources 2D-unitary ESPRIT can handle is equal to [7]
still able to carefully estimate the ARoI to which the actual sig-
nals belong with a degree of accuracy higher than that from the
(16) other techniques both in terms of localization and area extension
(Table VIII). Similar conclusions hold true when dealing with
being . On the other hand, it should be considered the detection of the signals distribution displayed in Fig. 11, al-
that an high dimensional array processing is enabled widening though the detection of the single signal on the bottom of the
the size of the planar array (i.e., the number of array sensors) at region of analysis appears to be more critical probably because
the expense of the computational complexity that, unlike SVM- of the absence of similar spatial configurations in the training
based methods, exponentially grows. set.
Fig. 10 compares the “convergence” map Finally, the last experiment is concerned with a scenario
provided by the IMSA-SVM and the ones from other single-step where there are not signals that impinge on the array and the
classifiers. As it can be observed, the multiscaling process is noise level has been varied from the reference value used for
Authorized licensed use limited to: International Institute of Information Technology Bangalore. Downloaded on January 27,2025 at 07:02:46 UTC from IEEE Xplore. Restrictions apply.
 DONELLI et al.: AN INNOVATIVE MULTIRESOLUTION APPROACH FOR DOA ESTIMATION 2289

TABLE VI
MULTIPLE SIGNALS SCENARIO, I =4 22
(CONFIGURATION = ). PERFORMANCE
INDEXES WHEN APPLYING IMSA-DOA, SINGLE-STEP SVM, MLP NEURAL
NETWORK, AND RBF NEURAL NETWORK

TABLE VII
MULTIPLE SIGNALS SCENARIO, I =4 13
(CONFIGURATION = ). PERFORMANCE
INDEXES WHEN APPLYING IMSA-DOA, SINGLE-STEP SVM, MLP NEURAL
NETWORK, AND RBF NEURAL NETWORK.

the SVM training [ (Test Set) versus

(Training Set)] thus further complicating the test case. As
a matter of fact, neither the free-case example is present in
the training set nor the same noise level has been “learned.”
Nonetheless, the SVM-based classifier did not detected the
presence of any signal thus defining a uniform distribution of
probability [Fig. 12(a)]. Otherwise, the other methods give
color-maps with some “artifacts” [see Fig. 12(b) and (c)]
although characterized by very small values of the probability
of signal incidence.

C. Dipole Array Antenna

In the last experiment, a uniform array of -dipoles is taken
into account with dipoles oriented along the axis. Therefore,
the effective length [24] of the array element turns out to be

(17) Fig. 10. Multiple signals scenario (I = 18 —Clustered Distribution) —Prob-

Moreover, the inter-element distance has been chosen equal to ability maps obtained with different classification approaches: (a) IMSA-SVM
( =
s S = 3) , (b) single-step SVM, (c) MLP neural network, and (d) RBF
and [27]. Then, a subset of the exper-
1 =1
neural network [   1 =1
and   ].
iments of the previous sections, but with the dipole array, has
been dealt with to evaluate the applicability of the IMSA-SVM
approach to nonideal electromagnetic scenarios, as well. in Fig. 13. Likewise the case with point-like sources, the esti-
In the first example , the multiscaling procedure mations of both the location and the incidence area improve at
stops after iterations and the final result is shown each iteration starting from and down
Authorized licensed use limited to: International Institute of Information Technology Bangalore. Downloaded on January 27,2025 at 07:02:46 UTC from IEEE Xplore. Restrictions apply.
 2290 IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 57, NO. 8, AUGUST 2009

Fig. 11. Multiple signals scenario (I = 18

— Sparse Distribution) — Prob-
ability maps determined by the IMSA-SVM at the convergence (s S = =
201  =5 and 1 = 20 ).

TABLE VIII
MULTIPLE SIGNALS SCENARIO, I = 18
(CLUSTERED DISTRIBUTION).
PERFORMANCE INDEXES WHEN APPLYING IMSA-DOA, SINGLE-STEP SVM,
MLP NEURAL NETWORK, AND RBF NEURAL NETWORK.

to and , where ,
and , . In this case, the
performance are comparable to that in Section III-A. Different
conclusions arise when processing the data of the two-signal
scenario [Fig. 14(a)]. In such a case, only the (i.e., the signal
with the lowest elevation ) is detected [Fig. 14(a)]. Such an
event does not depend on the DOA detection method, but from
the antenna array at hand. As a matter of fact, the radiation pat-
tern of the array element is omnidirectional in the plane
(i.e., and ) with a angle of almost 80
degrees [24]. Therefore, the gain of the dipole is lower along
the direction with higher , being . Oth-
erwise, when the actual configuration is described by a set of Fig. 12. No-signals scenario [I =0;P = 20 dB (Test Set) — P = 0 dB
(Training Set)] — Probability maps obtained with different classification ap-
signals coming from the directions and proaches: (a) IMSA-SVM s ( = S = 1)
, (b) MLP neural network, and (c)
, the IMSA-SVM method still gives ac- RBF neural network.
curate estimates [Fig. 14(b)] although with non-ideal isotropic
receiving sensors.
The effectiveness of the proposed approach has been assessed
dealing with different scenarios and working conditions. More-
IV. CONCLUSION
over, a comparative analysis has been carried out by considering
In this paper, a multiresolution approach for the DOA esti- state-of-the-art DOA methods. The obtained results have shown
mation of multiple signals based on a support vector classifier that:
has been presented. The procedure is aimed at defining a prob- • the use of a classifier based on SVM allows one to estimate
ability map of the incidence of an electromagnetic signal on a the DOA probability map in real time;
planar array of sensors. Starting from a coarse map, a synthetic • thanks to the SVM generalization capability, the
zoom is iteratively performed in the angular sector where the IMSA-SVM behaves properly when dealing with complex
incidence of a signal has been detected with higher probability electromagnetic scenarios non-necessarily belonging to
at the previous step of the multiscaling procedure. the set of training examples;
Authorized licensed use limited to: International Institute of Information Technology Bangalore. Downloaded on January 27,2025 at 07:02:46 UTC from IEEE Xplore. Restrictions apply.
 DONELLI et al.: AN INNOVATIVE MULTIRESOLUTION APPROACH FOR DOA ESTIMATION 2291

• unlike subspace-based algorithms, the computational

complexity does not increase with the size of the rectan-
gular array;
• the proposed LBE technique adapts to element failure
or other source of errors coming from the tolerances in
the array structure that cause non-negligible performance
degradation in conventional estimation techniques which
require highly calibrated antennas with identical radiation
properties;
• the a-priori knowledge (deterministic or statistical) on the
array configuration and radiation pattern characteristics
can be easily and usefully exploited by defining suitable
IMSA-SVM training sets;
• the multiscaling procedure (IMSA) provides good results
Fig. 13. Dipole Array, I = 1—Probability map determined by the dealing with both single-signal and multiple-signals con-
IMSA-DOA [s =
S =4
].
figurations with an angular resolution comparable to that
of other state-of-the-art DOA algorithms;
• system complexity, classifier architecture, and computa-
tional costs significantly reduce with respect to the “bare”
classification.

ACKNOWLEDGMENT
The authors wish to thank the reviewers for many valuable
suggestions and comments that helped improve this paper. In
addition, discussions with Dr. L. Lizzi and Dr. Oliveri are much
appreciated.

REFERENCES
[1] M. Chryssomallis, “Smart antennas,” IEEE Antennas Propag. Mag.,
vol. 42, no. 3, pp. 129–136, Jun. 2000.
[2] E. L. Hines, M. S. Leeson, M. M. Ramon, M. Pardo, E. Llobet, D.
D. Iliescu, and J. Yang, Intelligent Systems: Techniques and Applica-
tions. Maastricht, Germany: Shaker, 2008.
[3] S. P. Applebaum, “Adaptive arrays,” IEEE Trans. Antennas Propag.,
vol. 24, no. 5, pp. 585–598, May 1976.
[4] D. S. Weile and E. Michielssen, “The control of adaptive antenna arrays
with genetic algorithms using dominances and diploidy,” IEEE Trans.
Antennas Propag., vol. 49, no. 10, pp. 1424–1433, Oct. 2003.
[5] R. Roy and T. Kailath, “ESPRIT-Estimation of signal parameters via
rotational invariance techniques,” IEEE Trans. Acoust., Speech, Signal
Process., vol. 37, no. 7, pp. 984–995, Jul. 1989.
[6] F. Gao and B. Gershman, “A generalized ESPRIT approach to direc-
tion-of-arrival estimator,” IEEE Signal Process. Lett., vol. 12, no. 3, pp.
254–257, Mar. 2005.
[7] M. D. Zoltowski, M. Haardt, and C. P. Mathews, “Closed-form 2-D
angle estimation with rectangular arrays in element space or beamspace
via unitary ESPRIT,” IEEE Trans. Signal Process., vol. 44, no. 2, pp.
316–328, Feb. 1996.
[8] R. O. Schmidt, “Multiple emitter location and signal parameter estima-
tion,” IEEE Trans. Antennas Propag., vol. 34, no. 3, pp. 276–280, Mar.
1986.
[9] A. Swindlehurst and T. Kailath, “A performance analysis of sub space-
based methods in the presence of model errors. I. The MUSIC algo-
rithm,” IEEE Trans. Signal Process., vol. 40, no. 7, pp. 1578–1774,

= 2—Probability map determined

Jul. 1992.
Fig. 14. Dipole Array, I
= ( = 12 ;  = 165 ),
[10] I. Ziskind and M. Wax, “Maximum likelihood localization of multiple
by the IMSA-DOA when (a) I
= ( = 82 ;  = 165 ) = S = 3] and (b)
sources by alternating projection,” IEEE Trans. Acoust., Speech, Signal
I [s
= ( = 30 ;  = 60 ), = ( = 30 ;  = 300 )
Process., vol. 36, no. 10, pp. 1553–1560, Oct. 1988.
I I
[s = S = 3].
[11] P. Stoica and A. B. Gershman, “Maximum-likelihood DOA estimation
by data-supported grid search,” IEEE Signal Process. Lett., vol. 6, no.
10, pp. 273–275, Oct. 1999.
[12] A. H. El Zooghby, C. G. Christodoulou, and M. Georgiopulos, “Per-
formance of radial-basis function networks for direction of arrival es-
• the SVM-based approach is able to estimate the DOAs of timation with antenna arrays,” IEEE Trans. Antennas Propag., vol. 45,
a number of sources greater than the maximum allowed no. 11, pp. 1611–1617, Nov. 1997.
[13] A. H. El Zooghby, C. G. Christodoulou, and M. Georgiopulos, “A
by conventional eigenvalue decomposition methods for a neural network-based smart antenna for multiple source tracking,”
fixed planar array geometry; IEEE Trans. Antennas Propag., vol. 48, pp. 768–776, May 2000.
Authorized licensed use limited to: International Institute of Information Technology Bangalore. Downloaded on January 27,2025 at 07:02:46 UTC from IEEE Xplore. Restrictions apply.
 2292 IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 57, NO. 8, AUGUST 2009

[14] M. Martinez-Ramon and C. G. Christodoulou, Support Vector Ma- Federico Viani was born in Trento, Italy, in 1981.
chines for Antenna Array Processing and Electromagnetics Synthesis He received the M.S. degree in telecommunications
Lectures on Computational Electromagnetics Lecture # 5. San engineering from the University of Trento, Italy, in
Rafael, CA: Morgan & Claypool, 2006. 2007.
[15] V. Vapnik, Statistical Learning Theory. New York: Wiley, 1998. He is a member of the ELEDIA Research Group,
[16] E. Bermani, A. Boni, S. Caorsi, M. Donelli, and A. Massa, “A University of Trento, and his main interests are in an-
multi-source strategy based on a learning-by-examples technique for tennas synthesis and design, ultrawideband commu-
buried object detection,” in Progr. Electromagn. Res.. Philadelphia, nications and wireless sensor networks.
PA: PIER, 2004, vol. 48, pp. 185–200.
[17] A. Massa, A. Boni, and M. Donelli, “A classification approach based
on SVM for electromagnetic subsurface sensing,” IEEE Trans. Geosci.
Remote Sens., vol. 43, no. 9, pp. 2084–2093, Sep. 2005.
[18] D. Anguita, A. Boni, and S. Ridella, “A digital architecture for support
vector machines: Theory, algorithm and FPGA implementation,” IEEE
Trans. Neural Netw., vol. 14, no. 5, pp. 993–1009, Sep. 2003. Paolo Rocca received the B.S. and M.S. degrees in
[19] M. Pastorino and A. Randazzo, “A smart antenna system for direction telecommunications engineering both from the Uni-
of arrival estimation based on a support vector regression,” IEEE Trans. versity of Trento, Italy, in 2004 and 2005, respec-
Antennas Propag., vol. 53, no. 7, pp. 2161–2168, Jul. 2005. tively, and the Ph.D. degree from the International
[20] M. Pastorino and A. Randazzo, “The SVM-based smart antenna for Graduate School in Information and Communication
estimation of the directions of arrival of electromagnetic waves,” IEEE Technologies, University of Trento, in 2008.
Trans. Instrum. Meas., vol. 55, pp. 1918–1925, Dec. 2006. He is a member of the ELEDIA Research Group,
[21] A. Randazzo, M. A. Abou-Khousa, M. Pastorino, and R. Zoughi, “Di- University of Trento. His main interests are in the
rection of arrival estimation based on support vector regression: Ex- framework of array antenna synthesis and design, op-
perimental validation and comparison with MUSIC,” IEEE Trans. An- timization techniques for electromagnetics, electro-
tennas Wireless Propag. Lett., vol. 6, pp. 379–382, 2007. magnetic inverse scattering.
[22] S. Caorsi, M. Donelli, D. Franceschini, and A. Massa, “A new method-
ology based on an iterative multiscaling for microwave imaging,” IEEE
Trans. Microw. Theory Tech., vol. 51, no. 4, pp. 1162–1173, Apr. 2003.
[23] L. C. Godara, “Applications of antenna arrays to mobile communica- Andrea Massa (M’03) received the “Laurea” degree
tions—Part I: Performance improvement, feasibility, and system con- in electronic engineering and the Ph.D. degree in
siderations,” Proc. IEEE, vol. 85, pp. 1031–1060, Jul. 1997. electronics and computer science from the Uni-
[24] C. A. Balanis, Antenna Theory: Analysis and Design. New York: versity of Genoa, Genoa, Italy, in 1992 and 1996,
Wiley, 2005. respectively.
[25] A. Manikas, A. Alexiou, and H. R. Karimi, “Comparison of the ulti- From 1997 to 1999, he was an Assistant Professor
mate direction-finding capabilities of a number of planar array geome- of Electromagnetic Fields at the Department of Bio-
tries,” Proc. Inst. Elect. Eng. Radar, Sonar Navig., vol. 144, no. 6, pp. physical and Electronic Engineering, University of
321–329, Dec. 1997. Genoa, teaching the university course of Electromag-
[26] C. G. Christodoulou and M. Georgiopoulos, Applications of Neural netic Fields 1. From 2001 to 2004, he was an Asso-
Networks in Electromagnetics. Boston, MA: Artech House, 2001. ciate Professor at the University of Trento, Trento,
[27] W. Chen, J. Jen, and S. M. Zhang, “Planar dipole arrays with equal Italy. Since 2005, he has been a Full Professor of Electromagnetic Fields at the
element input impedances,” Electron. Lett., vol. 31, no. 24, pp. University of Trento, where he currently teaches electromagnetic fields, inverse
2061–2062, 1995. scattering techniques, antennas and wireless communications, and optimization
techniques. At present, he is the Director of the ELEDIALab at the University
Massimo Donelli (M’04) received the Electronic En- of Trento and Deputy Dean of the Faculty of Engineering. His research work
gineering degree and the Ph.D. degree in space sci- since 1992 has been principally on electromagnetic direct and inverse scattering,
ence and engineering from the University of Genoa, microwave imaging, optimization techniques, wave propagation in presence of
Italy, in 1998 and 2003, respectively. nonlinear media, wireless communications and applications of electromagnetic
He is with the Department of Information and fields to telecommunications, medicine and biology.
Communication Technology, University of Trento, Dr. Massa is a member of the PIERS Technical Committee, the European
Italy. His main interests are electromagnetic inverse Microwave Association (EuMA), and of the Inter-University Research Center
scattering, adaptive antennas synthesis, optimization for Interactions Between Electromagnetic Fields and Biological Systems
techniques for microwave imaging, wave prop- (ICEmB).
agation in superconducting materials and urban
environment.

Authorized licensed use limited to: International Institute of Information Technology Bangalore. Downloaded on January 27,2025 at 07:02:46 UTC from IEEE Xplore. Restrictions apply.