This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAES.2019.2895585, IEEE
Transactions on Aerospace and Electronic Systems
TAES-201800774 1

Improving the Doppler Resolution of Ground-

Based Surveillance Radar for Drone Detection
Hongbo Sun, Senior Member, IEEE, Beom-Seok Oh, Member, IEEE,
Xin Guo, and Zhiping Lin, Senior Member, IEEE

classification techniques were reported for classifying different

Abstract—In this Correspondence we propose to use an sizes of drones [7], different types of drones [8][9], drones and
Iterative Adaptive Approach (IAA) to perform the Doppler birds [9], drones and persons/dogs [10], and more interestingly,
processing in ground-based surveillance radar for drone loaded and unloaded drones [11]. It should be noted that almost
detection. It overcomes the limited dwell time and improves the
Doppler resolution, and therefore significantly enhances the
all those results were obtained using the experimental radars
discrimination of micro-Doppler signature and the correct with continuous illumination to the drone targets, and the
classification of drones. In addition, the target detectability is also Doppler resolution is fine enough to discriminate micro-
improved. This approach is validated in the field experiments Doppler signatures generated by the drones’ rotating blades.
conducted with a commercial portable ground-based surveillance However, in practical ground-based surveillance radar systems,
radar. the radar antennas need to scan quickly to cover a wide spatial
sector up to 360º. That means the dwell time of radar beam to
Index Terms— Ground-based surveillance radar, Doppler
resolution, iterative adaptive approach, drone detection, micro- any particular target is quite limited (which is, usually, a few
Doppler tens of milliseconds). Thus, when adopting the conventional
Fast Fourier Transform (FFT) for Doppler processing, the radar
Doppler resolution is very poor and the accurate micro-Doppler
I. INTRODUCTION signatures of drones is difficult to discriminate. To the authors’
knowledge, no literature was found to address this micro-
N OWADAYS the increasing use of remote-controlled mini
drones has become a critical issue in the air traffic
management. They may be misused for criminal acts or even
Doppler discrimination problem for drone detection using
ground-based surveillance radar.
terrorist attacks, and pose serious threats for the public security In principle the super-resolution algorithms that are often
and the protection of restricted zones. Thus, accurate detection used in array processing, such as Minimum Variance
and classification of mini drones are highly essential and Distortionless Response (MVDR) [12] and MUltiple SIgnal
critical. Radar is superior to other Electro-Optical/Infrared Classification (MUSIC) [13], can also be applied to the
(EO/IR) sensors for its wide, long-range and rapid surveillance temporal signal for improving the Doppler resolution. These
capability in all weather conditions, which is therefore widely algorithms usually require multiple signal snapshots to estimate
adopted as an effective sensor to detect the drones. the covariance matrix or perform eigen-analysis. Some
If the target or any structure on the target has mechanical algorithms, e.g., MUSIC, also need to know the number of
vibration or rotation in addition to its bulk translation, it might sources a priori. However, the Doppler processing in
induce a frequency modulation on the returned radar signal that surveillance radar is performed over the slow-time samples at
generates sidebands about the target’s Doppler shift. This is each range bin. That means only one temporal snapshot is
called the micro-Doppler effect, which offers a new approach available. The number of target’s Doppler and micro-Doppler
for target property analysis and target classification [1]. In sources is also unknown. Thus, the classical super-resolution
recent years, micro-Doppler generated by drones’ rotating algorithms cannot be applied. In this Correspondence, we
blades is popularly used for drone detection and classification. propose to use an Iterative Adaptive Approach (IAA) [14] for
Micro-Doppler signatures of drones were extensively analyzed the Doppler processing in surveillance radar. It is demonstrated
using time-frequency representation techniques such as that very high Doppler resolution can be achieved and micro-
spectrogram [2][3], and smoothed-pseudo Wigner-Ville Doppler signatures of low-flying drone can be clearly
distribution [4]. Micro-Doppler was also directly used to discriminated even with single temporal snapshot.
improve the radar detection performance for drone targets The rest of this Correspondence is organized as follows.
[5][6]. Moreover, based on micro-Doppler signatures, various Section II briefly introduces the IAA for Doppler estimation.
Section III presents some simulation results to compare the

Manuscript received July 17, 2018; revised November 9 2018. Beom-Seok Oh and Zhiping Lin are with School of Electrical and Electronic
(Corresponding author: Hongbo Sun.) Engineering, Nanyang Technological University, Singapore.
Hongbo Sun and Xin Guo are with Temasek Laboratories, Nanyang
Technological University, Singapore (e-mail: ehbsun@ntu.edu.sg). Xin Guo is
currently with Thales Solutions Asia Pte. Ltd., Singapore.

0018-9251 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAES.2019.2895585, IEEE
Transactions on Aerospace and Electronic Systems
TAES-201800774 2

Doppler estimation performance of IAA and conventional FFT. microwave ground-based surveillance radar, usually only
Section IV presents the field experiments conducted with a several tens of temporal samples (i.e., radar pulses) can be used
commercial portable ground-based surveillance radar and for the Doppler processing within a limited integration time.
various mini drones. Finally, Section V concludes this For this scenario, less than 10 iterations are needed to achieve
Correspondence. the convergence of IAA. The computational complexity of each
iteration is on the order of O( M 2 K ) , which is comparable to
II. ITERATIVE ADAPTIVE APPROACH (IAA) the conventional MVDR and MUSIC algorithms. If necessary,
IAA is a data-dependent, nonparametric, iterative adaptive this computational complexity can be reduced by exploiting the
algorithm based on weighted least square (WLS) approach. It Toeplitz-block-Toeplitz structure of R [14].
was first introduced in [14] for source localization in array
signal processing. Unlike the conventional MVDR and MUSIC III. SIMULATION RESULTS
algorithms in which many snapshots are required to estimate We consider a Doppler estimation scenario based on a radar
the covariance matrix, IAA can work well with only a few or pulse train consisting of 64 pulses. Two target signals with
even one snapshot to achieve super-resolution. This technique equal power are simulated at the normalized Doppler frequency
has been applied in some applications such as spectral analysis 0.1 and 0.2 respectively. Fig. 1(a) depicts the Doppler spectra
for nonuniformly sampled data [15], MIMO radar imaging [16], estimated by FFT and IAA for the noise-free scenario. It can be
space-time processing for airborne radar ground moving target seen that the two Doppler signals estimated by IAA are
detection [17][18], and direction-of-arrival estimation for approximately the Dirac delta functions with ideal resolution
nonuniform sparse array [19]. In this Correspondence the IAA and zero sidelobe when the Signal-to-Noise Ratio (SNR) is
is slightly adapted for Doppler estimation. The processing steps infinite. Fig. 1(b) and (c) depict the results when additive
are summarized as below. Gaussian noise is injected and the SNRs are 10dB and -5dB
Consider the signal generated by K Doppler sources respectively. We can see that for both high and low SNR
f ∈ [ f1 , f 2 , ... , f K ] and f k is the kth Doppler, k = 1, ... , K . scenarios, IAA gets good estimation for the two Doppler signals
The received signal vector of M temporal samples in the with much higher resolution than the conventional FFT. The
presence of additive noise can be represented as noise levels in the IAA results are almost the same as that in the
= y A(f )s + e , (1) FFT results and no noticeable SNR loss is observed. Note that
no tapering is used in the FFT results shown in Fig. 1. Applying
where A(f ) = [a( f1 ), a( f 2 ), ... , a( f K )] is the M × K Doppler
proper taper window can reduce the FFT sidelobes, but at the
steering matrix for K Doppler sources, a( f k ) is the Doppler cost of suffering a few dB SNR loss. To quantitatively compare
steering vector for Doppler f k , s = [ s1 , s2 , ... , sK ] represents the Doppler estimation accuracy of FFT and IAA under various
the amplitudes of K Doppler sources, and e is the noise vector. SNR scenarios, hundreds of Monte Carlo simulations are
In practice the number of Doppler sources, K, is unknown. conducted for the Doppler signal with 0.1 normalized Doppler
Hence, K is considered to be the number of scanning grids in all frequency and -5dB, 0dB, 5dB, 10dB, and 15dB SNRs
possible Doppler regions. respectively. The Root-Mean-Square Errors (RMSE) of the
Let P be a K × K diagonal matrix, whose diagonals normalized Doppler frequencies estimated by FFT and IAA
with 64 pulses are plotted in Fig. 2. It is seen that the RMSEs
Pk = sk , k = 1, ... , K contain the powers at each Doppler
2

of Doppler frequency estimated by IAA are very close to that

frequency on the scanning grid. Furthermore, define the estimated by FFT and the differences are very trivial. Similar
interference (signals at Dopplers other than the Doppler of result can also be obtained for the 0.2 normalized Doppler
current interest f k ) and noise covariance matrix Q( f k ) as frequency. Note that further evaluation for the input SNR below
Q( f k )=R − Pk a( f k )a H ( f k ) , (2) -5dB is not meaningful. With -5dB input SNR and after the
10 log10 (64) = 18dB coherent integration gain of 64 pulses, the
where R = A(f )PA (f ) . Then the WLS cost function is given
H

output SNR is merely 13dB, which is the minimum threshold

by
that is usually used for target detection.
y − sk a ( f k ) Q ( f ) ,
2
−1 (3)
k

−1
= x Q ( f k )x . Minimizing (3) with respect to
2 H
where x Q −1 ( f k )

sk and using (2) yield [14]

a H ( f k )R −1y
sˆk = . (4)
a ( f k )R −1a( f k )
H

Since the IAA requires R, which depends on the unknown

powers, it must be implemented as an iterative approach. The
initialization can be done by letting R equal to the identity
matrix IM. The number of iterations needed for the convergence
of IAA is not constant for different system parameters. In the (a)

0018-9251 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAES.2019.2895585, IEEE
Transactions on Aerospace and Electronic Systems
TAES-201800774 3

components, but is with higher resolution than the FFT

spectrum.
The previous simulation results are based on the stationary
Doppler signals. Next we evaluate the performance of IAA for
nonstationary Doppler signal. Assuming a nonstationary
Doppler signal whose instantaneous normalized frequency is
0.1 + 0.001× [0 : M − 1] , where M is the number of pulses. For a
pulse train consisting of 64 pulses, the instantaneous
normalized frequency varies from 0.1 to 0.164, while the
normalized Doppler resolution (for FFT-based estimator) is
(b) 1/64=0.0156. That means this nonstationary Doppler signal
occupies 4 Doppler bins in the Doppler spectrum. Fig. 4 depicts
the Doppler spectra estimated by FFT and IAA for this
nonstationary Doppler signal, in which additive Gaussian noise
is also injected with 10dB SNR. We can see that the IAA still
works very well. The spectrum of this nonstationary Doppler is
correctly estimated and no noticeable SNR loss is observed.

(c)

Fig. 1. Simulations of Doppler estimation by FFT and IAA, two Doppler

signals. (a) No noise; (b) With noise, 10dB SNR; (c) With noise, -5dB SNR.

Fig. 3. Simulations of Doppler estimation by FFT and IAA, noise only.

Fig. 2. RMSEs of the 0.1 normalized Doppler frequency estimated by FFT

and IAA under various SNRs.

Based on many other extensive simulations conducted under

various scenarios (e.g., different SNRs, amplitudes, and number
of sources), we have the following additional observations: (i)
The IAA can always converge rapidly within 10 iterations. No
divergence and indefinite estimation had ever occurred. (ii) The
Fig. 4. Simulations of Doppler estimation by FFT and IAA, one nonstationary
estimation performance of IAA is very robust. Each peak in the Doppler signal, 10dB SNR.
estimated Doppler spectrum represents a real Doppler signal
component and no false peak is generated. (iii) The peak
amplitude generated by IAA exactly represents the strength of
IV. EXPERIMENTAL VALIDATIONS
the respective Doppler signal component. Moreover, unlike
some other super-resolution algorithms, IAA doesn’t have any To evaluate the detection performance of ground-based
assumption or restriction on the model of signal to be estimated. surveillance radar for mini drone detection, a commercial
It can be directly applied to any signal with unknown portable ground-based surveillance radar was acquired by
parameters. As an example, a Gaussian noise (64 samples) is Nanyang Technological University. This radar is a
generated to simulate a signal consisting of very rich Doppler mechanically scanned radar, which operates at X-band and
components with various amplitudes. Fig. 3 shows the Doppler transmits Linear Frequency Modulated Continuous Wave
spectra estimated by FFT and IAA. We can see that the IAA (LFMCW) waveform. Some key parameters of this radar are
spectrum exactly emulates the characteristics of all Doppler listed in Table I. Fig. 5(a) shows the radar setup in an

0018-9251 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAES.2019.2895585, IEEE
Transactions on Aerospace and Electronic Systems
TAES-201800774 4

experimental trial conducted in an open space near the

industrial area of Singapore. Fig. 5(b) depicts the geometrical
configuration and a typical drone flight path detected and
tracked by the radar. During the trial, the radar worked in
surveillance mode and the planar antenna scanned repeatedly
for a sector of 120º. The dwell time of radar beam on a
particular target is about 40ms, which restricts the maximum
coherent integration of 64 pulses. Two types of drone targets (a) (b)
were used in the trial. The first one was a DJI Phantom 3 Fig. 6. Drone targets used in the experimental trial.
quadcopter (shown in Fig. 6(a)), whose weight is 1.28kg and (a) DJI Phantom 3 quadcopter; (b) Self-made fixed-wing drone.
the diagonal size is 35cm (excluding propellers). The second
one was a self-made fixed-wing drone with a single propeller Fig. 7(a) and (b) show the range-Doppler plots obtained by
on the tail (shown in Fig. 6(b)), whose wingspan is about 70cm. FFT and IAA-based Doppler processing when the DJI Phantom
3 quadcopter flew slowly away from the radar at the height of
TABLE I 20-25m. The two plots are normalized to their respective noise
KEY PARAMETERS OF THE GROUND-BASED SURVEILLANCE RADAR
floor and with the same dynamic range. Fig. 8 shows the
Operating Frequency X-band
Waveform Linear Frequency Modulated Continuous Wave
Doppler cuts of Fig. 7(a) and (b) at the range bin where the
Transmit Power 1 Watt (max.) quadcopter is detected. Noting that in the FFT result, the
Typical Detection Pedestrian (0.5 m²) 9 km Hanning taper is used for Doppler sidelobe suppression. It can
Range Vehicle (10 m²) 19 km be seen that the drone body is clearly detected at 152m and the
Azimuth ≤ 5 mils
Accuracy
Range (at 3 km range setting) ≤ 5 m micro-Doppler generated by the propeller blades is also clearly
Radar Sensor 18 kg visible in both figures. In addition to the drone, a bird (which
Weight
Operator Unit 4 kg was visually identified during the measurement) is also detected
at 98m. The strong ground clutter reflections after 300m are
from the industrial buildings in the far end of trial scene.
Comparing the results shown in Fig. 7(a)(b) and Fig. 8, we can
see that the Doppler resolution obtained by FFT is very poor
and the accurate micro-Doppler signature of the drone cannot
be discriminated due to the limited dwell time, while the IAA
can significantly improve the Doppler resolution and the
regularly distributed micro-Doppler signature can be clearly
discriminated at both sides of the drone body in Doppler. In
addition, seeing the detection results of the drone body and the
bird in Fig. 7(a) and (b), it is obvious that the higher Doppler
resolution is also helpful to separate the slow-moving targets
(a)
and the ground clutter at 0Hz, and therefore improve the
detectability for the targets with very low velocity.
Fig. 9(a) and (b) show the range-Doppler plots when the
fixed-wing drone flew towards the radar from a different
direction. The drone is detected at 125m, which has a much
larger Doppler for its faster velocity than the quadcopter.
Another bird is detected at 330m. Fig. 10 shows the Doppler
cuts of Fig. 9(a) and (b) at the range bin where the fixed-wing
drone is detected. Similarly, the IAA achieved much higher
Doppler resolution than the FFT, and the detailed micro-
Doppler signature can be easily identified. Moreover,
comparing Fig. 7(b) and Fig. 9(b), we can also see that the
micro-Doppler peaks of the fixed-wing drone are sparser than
those of quadcopter, which reflects the fact of faster rotation
speed of fixed-wing drone’s blades. This result implies that the
(b) higher Doppler resolution achieved by IAA can greatly benefit
the subsequent classification of drone types, if necessary.
Fig. 5. Experimental trial using a commercial X-band portable ground-based
surveillance radar. We also compared the target detection performance achieved
(a) Radar setup; (b) A typical drone flight path detected and tracked by radar. by FFT and IAA in the two measurements. From the results
shown in Figs. 7 ~ 10 we can already see that the IAA obtains
higher SNR than the FFT for all targets’ Doppler and micro-
Doppler components. Some detailed quantitative analyses were

0018-9251 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAES.2019.2895585, IEEE
Transactions on Aerospace and Electronic Systems
TAES-201800774 5

performed and the exact SNR values of drone bodies and birds
obtained in Fig. 7 and Fig. 9 are listed in Table II. In summary,
the IAA achieves 2.1 ~ 7.1 dB SNR improvements comparing
to the FFT. The lower SNR obtained by FFT should be mainly
due to the loss caused by Hanning taper for Doppler sidelobe
suppression. However, it should be noted that in practical
applications, applying taper window in FFT processing is
mandatory, otherwise the Doppler sidelobes of strong
targets/clutters will be too high and mask other weak signals.
These results prove that the IAA can achieve higher Doppler
resolution with even higher SNR than the conventional FFT in
practical applications. We also would like to highlight that
many other extensive experimental measurements and data
analyses were conducted, and the IAA demonstrated robust (a)
performance and outperformed the conventional FFT in all
scenarios.

(b)

Fig. 9. Measured range-Doppler plots for fixed-wing drone.

(a) (a) Doppler estimated by FFT; (b) Doppler estimated by IAA.

(b) Fig. 10. Comparison of Doppler profiles obtained by FFT and IAA for fixed-
wing drone.
Fig. 7. Measured range-Doppler plots for quadcopter.
(a) Doppler estimated by FFT; (b) Doppler estimated by IAA.
TABLE II
SNR COMPARISONS FOR THE DETECTED TARGETS
SNRs obtained by FFT SNRs obtained
(with Hanning taper) by IAA
Bird (in Fig. 7) 31.1 dB 33.2 dB
DJI Phantom 3 quadcopter 53.8 dB 57.8 dB
body (in Fig. 7)
Bird (in Fig. 9) 22.5 dB 29.6 dB
Fixed-wing drone body (in 48.3 dB 51.1 dB
Fig. 9)

With the higher Doppler resolution obtained by IAA, it will

be easier to differentiate micro-Doppler signature of drones
Fig. 8. Comparison of Doppler profiles obtained by FFT and IAA for from other signals, and thus better classification accuracy for
quadcopter.

0018-9251 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAES.2019.2895585, IEEE
Transactions on Aerospace and Electronic Systems
TAES-201800774 6

drones can be achieved. Quite a number of experimental data V. CONCLUSION

were used to evaluate and compare the results of FFT and IAA In this Correspondence we propose to use the IAA technique
for automatic drone classification. We considered the 2- to overcome the limited dwell time and improve the Doppler
category classification, i.e., drone vs. non-drone. The number resolution of ground-based surveillance radar for drone
of drone/non-drone snapshots in training data set and test data detection. The experimental trial results demonstrate that much
set utilized in the classification experiment are listed in Table higher Doppler resolution can be achieved and micro-Doppler
III. The drone data consist of both fixed-wing drone and signatures of drones can be clearly discriminated, which greatly
quadcopter drone. The non-drone data consist of other types of benefit the subsequent classification of drones. The higher
moving targets such as cars, lorries, and bicycles. Note that the Doppler resolution can also help to separate the slow-moving
test data set was acquired on a different day from the training targets and the ground clutter at zero Doppler, and therefore
data set, and there is no overlapping between the two sets. improve the detectability of multirotor drone which usually
moves very slowly. In addition, using the IAA can avoid the
TABLE III
NUMBER OF EXPERIMENTAL DATA USED IN THE DRONE CLASSIFICATION taper loss in FFT-based Doppler processing and the overall
Training data set Test data set radar detection performance for all targets is also improved.
Non-drone 3,056 291 Most importantly, the IAA demonstrated very robust and
3,056 916 superior performance in our extensive experimental trials
Drone (Fixed-wing: 1,271, (Fixed-wing: 631,
Quadcopter: 1,785) Quadcopter: 285) conducted with the commercial ground-based surveillance
radar system. We believe it is an ideal technique which can
We adopted the classification technique proposed by the replace the conventional FFT for Doppler processing.
authors in [8]. First the processed radar signals (of the training
data set) by both FFT and IAA are respectively decomposed
into a set of Intrinsic Mode Functions (IMFs) using the REFERENCES
Empirical Mode Decomposition (EMD) [20]. The 8 features
proposed in [8] are then extracted from the first 4 IMFs [1] Chen, V. C., The Micro-Doppler Effect in Radar, Artech House,
Norwood, MA, USA, 2011.
followed by a feature normalization into a range of [0,1]. The [2] Wit, J. J. M., Harmanny, R. I. A., and Premel-Cabic, G., “Micro-Doppler
normalized features are subsequently used to train a Support analysis of small UAVs”, 2012 European Radar Conf., Amsterdam,
Vector Machine (SVM) classifier [21] with a nonlinear Radial Netherlands, pp. 210-213, Nov. 2012.
Basis Function (RBF) kernel (kernel width ∈ {5 : 2 :19,50,100} [3] Jian, M., Lu, Z., and Chen, V. C., “Experimental study on radar micro-
Doppler signatures of UAV”, 2017 IEEE Radar Conf., Seattle, USA, pp.
). The 8 test features extracted from the test data set are 854-857, May 2017.
normalized with respect to the minimum and maximum values [4] Tan, R., Lim, H. S., Smits, A. B., Harmanny, R. I. A., and Cifola, L.,
“Improved micro-Doppler features extraction using smoothed-pseudo
of the training features. To stay focused on the main objective Wigner-Ville distribution”, 2016 IEEE Region 10 Conf., Singapore, pp.
of this correspondence, we omit the intermediate results and 730-733, Nov. 2016.
directly present the final Receiver Operating Characteristic [5] Hoffmann, F., Ritchie, M., Fioranelli, F., Charlish, A., and Griffiths, H.,
(ROC) curves for the best “IAA+SVM” drone classifier (at “Micro-Doppler based detection and tracking of UAVs with multistatic
radar”, 2016 IEEE Radar Conf., Philadelphia, USA, May 2016.
RBF kernel width = 7) and “FFT+SVM” drone classifier (at [6] Kim, B. K., Kang, H.-S., and Park, S.-O., “Experimental analysis of small
RBF kernel width=11) in Fig. 11. It is obvious that the drone polarimetry based on micro-Doppler signature”, IEEE Geosci.
“IAA+SVM” classifier outperforms the “FFT+SVM” classifier Remote Sens. Lett., vol. 14, no. 10, pp. 1670-1674, Aug. 2017.
with a better true positive rate (i.e., correct drone classification [7] Fuhrmann, L., Biallawons, O., Klare, J., Panhuber, R., Klenke, R., and
Ender, J., “Micro-Doppler analysis and classification of UAVs at Ka
rate) for a given false positive rate (i.e., false drone band”, 2017 Inter. Radar Symp., Prague, Czech, June 2017.
classification rate). This result validates that the better Doppler [8] Oh, B.-S., Guo, X., Wan, F., Toh, K. A., and Lin, Z., “Micro-Doppler
resolution achieved by IAA can also greatly improve the mini-UAV classification using empirical-mode decomposition features”,
automatic classification for drones. IEEE Geosci. Remote Sens. Lett., vol. 15, no. 2, pp. 227-231, Feb. 2018.
[9] Molchanov, P., Egiazarian, K., Astola, J., Harmanny, R. I. A., and Wit, J.
J. M., “Classification of small UAVs and birds by micro-Doppler
signatures”, 2013 European Radar Conf., Nuremberg, Germany, pp. 172-
175, Oct. 2013.
[10] Lam, O. H. Y., Kulke, R., Hagelen, M., and Mollenbeck, G.,
“Classification of moving targets using mirco-Doppler radar”, 2016 Int.
Radar Symp., Krakow, Poland, May 2016.
[11] Fioranelli, F., Ritchie, M., Griffiths, H., and Borrion, H., “Classification
of loaded/unloaded micro-drones using multistatic radar”, Electro. Lett.,
vol. 51, no. 22, pp. 1813-1815, Oct. 2015.
[12] Capon, J., “High resolution frequency-wavenumber spectrum analysis”,
Proc. of IEEE, vol. 57, no. 8, pp. 1408-1418, Aug. 1969.
[13] Schmidt, R. O. , “Multiple emitter location and signal parameter
estimation”, IEEE Trans. Antennas Propag., vol. 34, no. 3, pp.276-280,
Mar. 1986.
[14] Yardibi, T., Li, J., Stoica, P., Xue, M., and Baggeroer, A. B., “Source
Fig. 11. Receiver Operating Characteristic (ROC) curves for the best localization and sensing: a nonparametric iterative adaptive approach
“IAA+SVM” classifier and “FFT+SVM” classifier.

0018-9251 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAES.2019.2895585, IEEE
Transactions on Aerospace and Electronic Systems
TAES-201800774 7

based on weighted least squares”, IEEE Trans. Aerosp. Electron. Syst., research associate, and currently as a
vol. 46, no. 1, pp. 425-443, Jan. 2010.
research fellow in the School of
[15] Stoica, P., Li, J., and He, H., “Spectral analysis of non-uniformly sampled
data: A new approach versus the periodogram”, IEEE Trans. Signal Electrical and Electronic
Process., vol. 57, no. 3, pp. 843-858, Mar. 2009. Engineering, Nanyang Technological
[16] Roberts, W., Stoica, P., Li, J., Yardibi, T., and Sadjadi, F. A., “Iterative University, Singapore. His research
adaptive approaches to MIMO radar imaging”, IEEE J. Sel. Topics Signal
interests include pattern analysis and
Process., vol. 4, no.1, pp. 5-20, Jan. 2010.
[17] Xue, M., Zhu, X., Li, J., Vu, D., and Stoica, P., “MIMO radar angle- classification, and machine learning.
Doppler imaging via iterative space-time adaptive processing”, 2009 Int.
Waveform Diversity and Design Conf., Kissimmee, USA, pp. 129-133,
Feb. 2009.
[18] Sun, H., Lu, Y., and Lesturgie, M., “Experimental investigation of
iterative adaptive approach for ground moving target indication”, 2011
Xin Guo received her B.Eng and
IEEE CIE Int. Conf. on Radar, Chengdu, China, pp. 715-718, Oct. 2011. Ph.D, both in Electrical Engineering,
[19] Sun, H., Wan, L., Lan, X., and Xie, L., “Target DOA estimation using from Nanjing University of Science
nonuniform sparse array for low frequency radar”, 2017 IET Int. Conf. on and Technology, China, in 1999 and
Radar Syst., Belfast, UK, Oct. 2017.
[20] Huang, N. E., Shen, Z., Long, S. R., Wu, M. C., Shih, H. H., Zheng, Q.,
2004 respectively. From 2004 to
Yen, N.-C., Tung, C. C., and Liu, H. H., “The empirical mode 2006, she was with the Department of
decomposition and the Hilbert spectrum for nonlinear and non-stationary Electrical and Computer Engineering,
time series analysis”, Proc. Roy. Soc. London A, Math. Phys. Eng. Sci., National University of Singapore, as a
vol. 454, no. 1971, pp.903-995, Mar. 1998.
[21] Duda, R. O., Hart, P. E., and Stork, D. G., Pattern Classification,
Research Fellow. From 2006 to 2017,
Hoboken, NJ, USA: Wiley, 2012. she was with Temasek Laboratories,
Nanyang Technological University as
Hongbo Sun (M’04-SM’13) a Research Scientist and Senior Research Scientist. In 2017, she
received his B.Eng. and Ph.D. joined Thales Solutions Asia Pte. Ltd. as a Signal Processing
degrees, both in Electrical Engineer. Her research interests focus on advanced signal
Engineering, from Nanjing processing techniques for radar and other applications.
University of Science and
Technology, China, in 1997 and Zhiping Lin (SM’00) received the
2002, respectively. He joined the B.Eng. degree in control engineering
Nanyang Technological University from South China Institute of
(NTU), Singapore, in 2002 as a Technology, Canton, China in 1982
Research Fellow. Presently he is a and the Ph.D. degree in information
Senior Research Scientist and engineering from the University of
Principal Investigator in Temasek Laboratories at NTU. He has Cambridge, England in 1987. He was
authored/co-authored one book chapter and more than 80 with the University of Calgary,
technical papers in refereed journals and conference Canada for 1987-1988, with Shantou
proceedings. He is serving as the Associate Editor for University, China for 1988-1993, and
Electronics Letters and Bulletin of Geosciences. He was also with DSO National Laboratories,
the Guest Associate Editor for an IEEE Geoscience and Remote Singapore for 1993-1999. Since 1999, he has been with
Sensing Letter Special Stream from 2015 to 2017, and the past Nanyang Technological University (NTU), Singapore. He is a
Chairman of IEEE AES/GRS Joint Singapore Chapter from Program Director at Centre for Bio Devices and Signal Analysis,
2016 to 2017. Previously he had ever been the Technical NTU. Dr. Lin was the Editor-in-Chief of Multidimensional
Program Committee Co-chair in the 5th Asia-Pacific Systems and Signal Processing for 2011 – 2015, after being in
Conference on Synthetic Aperture Radar (APSAR 2015) and its editorial board since 1993. He was an Associate Editor of
the Technical Program Committee member in many other Circuits, Systems and Signal Processing for 2000-2007 and an
international conferences such as RADAR 2011, ICARES Associate Editor of IEEE Transactions on Circuits and Systems
2014, APSAR 2015, ICARES 2015, OCRA 2016, RADAR - II for 2010-2011. He was a reviewer for Mathematical
2016, PIERS 2017, ICARES 2018, and AGERS 2018, etc. He Reviews for 2011-2013. He is currently a Subject Editor and a
was also the winner of Excellent Paper Award in RADAR 2006 Guest Editor of the Journal of the Franklin Institute. His
conference. Dr. Sun’s research interests mainly focus on the research interests include multidimensional systems and signal
advanced radar concepts and signal processing techniques. processing, statistical and biomedical signal processing, and
machine learning. He is the co-author of the 2007 Young
Beom-Seok Oh (M’15) received the B.S. degree in Computer Author Best Paper Award from the IEEE Signal Processing
Science from KonKuk University, South Korea, in 2008. He Society, Distinguished Lecturer of the IEEE Circuits and
received the M.S. degree in Biometrics and the Ph.D. degree in Systems Society for 2007-2008, and received the Best Paper
Electrical and Electronic Engineering from Yonsei University, Awards at ELM 2015 and ELM 2017.
South Korea, in February 2010 and August 2015, respectively.
From April to November 2015, he has been working as a

0018-9251 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.