You are on page 1of 6

A Machine Learning Approach to Radar Sea Clutter

Suppression
D. Callaghan, J. Burger, Amit K Mishra
University of Cape Town

Abstract—The radar detection of small maritime targets This paper explores a novel approach to sea clutter suppres-
requires special attention to the suppression of sea clutter returns sion by treating it as a machine learning problem. In this we
that, in certain circumstances, may be difficult to discriminate compare the use of two machine learning techniques to model
from target returns. In this paper, the novel use of machine the classification boundary between sea clutter and small
learning techniques is explored to address this familiar problem. maritime targets. A comparison of the success of different
A comparison of the results obtained using two machine learning
techniques for the suppression of sea clutter is presented. Data
machine learning algorithms has been conducted by [4] for
for this experiment was gathered using an experimental S-band atmospheric radar. The clutter types requiring suppression from
radar called NetRAD. This radar system was observing a coastal [4] were ground clutter, sea clutter and anomalous propagation.
scene and the data collected was classified as either target or In this study it was found that SVM and other machine
clutter using k-Nearest-Neighbour (kNN) and Support Vector learning algorithms outperformed kNN by a significant margin.
Machine (SVM) algorithms. The results are presented as averaged It was also noted however that kNN was not suited to the
probability of detection and probability of false alarm. classification problem due to a low number of features. In this
paper, however the machine learning algorithms made the class
I. I NTRODUCTION distinction between clutter and target based on a high number
of features. Other research papers have explored the alternative
Detecting small targets in the presence of sea clutter is use of machine learning algorithms for clutter suppression
a typical maritime radar problem. The nature of sea clutter [5], [6], [7], but not in the context of small maritime target
returns can be very close to that of returns from a desired detection.
target. Without further processing, this would result in either The machine learning approach leverages a vast body of
a low probability of target detection (PD ) or unacceptable knowledge, analysis and techniques to provide a more abstract,
probability of false alarm (Pf a ). Typical performance spec- scalable and generalised treatment of the number and type of
ifications for a radar system will require Pd ≥ 0.8 and inputs as well as the discrimination algorithms to be employed.
Pf a ≤ 10−4 . This problem is commonly solved by applying
CFAR (Constant False Alarm Rate) processing [1] which is In this paper the performance of two powerful but distinctly
an adaptive process that determines a local optimum threshold different machine learning algorithms, viz. SVM and kNN,
level which improves the probability of target detection (Pd ) are compared by applying them to the discrimination of small
and reduces the probability of false alarm Pf a . maritime targets and sea clutter. SVM is based on linear
modelling of the classification boundary using a least squares
A great deal of investigation and modelling of sea clutter approach. kNN on the other hand is a non-linear modelling
statistics have been done by Ward et al. [2], amongst others. process. The success of each algorithm is dependent on the
Ward et al. have proposed a compound statistical model for underlying statistical structure of the data. In general, SVM is
sea clutter that characterises the statistics of different scattering better suited for a single Gaussian distributed data whereas
mechanisms. In [2], two distinct groups of scattering mecha- kNN is better suited for data consisting of a mixture of
nisms are identified. First the large scale sea wave structure Gaussian distributions [8]. Sea clutter in general is not readily
involving long wavelengths relative to the resolution of the described by simple statistical features. Multifratal features
radar system. This resolved sea structure results in spatial have been proposed to characterise the complicated statistical
variations in mean sea clutter return. The second group consists nature of sea clutter [9].
of the small scale structures, white caps (breaking waves),
bragg scattering from small scale capillary waves and spikey The radar data was obtained from the experimental radar
specular returns from cresting waves shortly before they break. called NetRAD which was set up to collect maritime radar
The second group of mechanisms has a Gaussian distribution data over several days from a littoral zone near Simons Town
while the first group is modelled by a Gamma distribution. on the Cape Peninsula.
The Gamma distributed first group clutter modulates the mean
power level of the second group clutter return. II. BACKGROUND
NetRAD is a multistatic radar system consisting of a
Currently, CFAR processing is commonly used to mitigate
transceiver node and several receiver nodes. The radar operates
the problem of variation in the sea clutter mean power return
at 2.4 GHz with a bandwidth of 50 MHz and is used to detect
[3] [2]. Cell averaging CFAR processing achieves constant
maritime targets. An illustration of the position of NetRAD for
probability of false alarm using an average of local radar cell
the datasets used is presented in Fig. 1. For this paper, only
power returns to adjust an amplitude threshold to discriminate
data received at the transceiver node was used and therefore
between a target and sea clutter. Other CFAR processing
the configuration is effectively monostatic.
may employ Doppler frequency or polarization returns or a
combination of all of the above to arrive at a suitable set of The transmitted chirp is selected from a set of reference
local detection thresholds. chirps ranging from 0.1 µs to 10 µs. For all the datasets used

978-1-4673-8823-8/17/$31.00 ©2017 IEEE 1222


Kalk Bay
that are used to calculate the least squares optimisation. This
Distant Target set of objects is called the support vectors and is shown in Fig.
Land 3. The method employed is linear regression [8] using a vector
Sea
Fish Hoek function for the equation of the boundary. In linear regression,
the mean square of the difference between the support vectors
and the boundary is minimised. The equation for the boundary
Receiver has the form:
Node f (x) = wT x + b (1)
2
km Roman Rock
Transceiver
where x is a vector of input values and w is a vector of
Node constant weights. The training process in SVM finds the
weights vector w that minimises the sum of squares between
Receiver the support vectors and the boundary.
Node
x x Support
Cape Peninsula
x
x Vector
Fig. 1: Illustrative position of NetRAD. x Margin
x y
s s b o undar
Cla o
the sea state was 1 and the look angle was between 78◦ and o
94◦ . The antenna was set up on the beach at shoulder height o o
with 0◦ elevation resulting in a near 0◦ grazing angle. Each
node records at 100 Mega-samples per second. An example of
o
NetRAD data after range pulse compression is shown in Fig. Fig. 3: Example of a hyperplane (boundary line) boundary for
2 where there are clearly two targets, one stationary target and data with two features. The hyperplane in this example is a 1
one moving target. There is also typical substantial close-range dimensional line in a two dimensional space.
clutter on the far left of Fig. 2.

B. K Nearest Neighbour (kNN)


This algorithm employs a simple non-linear technique
for defining the decision boundary. Each object is classified
according to the classification of its k nearest neighbours [8].

f1
x x
x
o x
x x x
o x o x
o x
x
o
o o
x o
oo o

Fig. 2: Pulse compressed data from NetRAD operating in


Simon’s Town on the 21st of April 2011 at 12:35 SAST. The f2
x axis is range (fast time) and the y axis is slow time. Note Fig. 4: Example of the kNN classification boundary.
the solid red vertical line in the middle of the figure is Roman
Rock and the weaker diagonal red line is a moving vessel. The
red lines on the far left are close-range clutter returns. Notice in the example of Fig. 4 that objects may be misclas-
sified because their k nearest neighbours are of the opposite
class. This allows a more flexible and generalised boundary
A brief description of SVM and kNN will be provided in condition that might be more appropriate for some classifica-
the next two subsections followed by an explanation of how tion problems. The higher the value of k, the “smoother” the
the NetRAD data was preprocessed before applying the data boundary condition becomes often resulting in more misclas-
to the machine learning algorithms. sified objects but resulting in a boundary definition that better
generalises the class descrimination.
A. Support Vector Machines (SVM)
III. DATA P REPROCESSING
The SVM algorithm uses a least squares minimisation
technique to determine the optimum boundary between classes. NetRAD data consists of raw ADC datacubes along with
The algorithm identifies a set of objects close to the boundary conditions of each experiment including elevation, azimuth,

978-1-4673-8823-8/17/$31.00 ©2017 IEEE 1223


wind direction, wind speed, temperature and time of each
experiment. This includes any identified target and the transmit
pulse used. The axes of the raw datacube can be labelled fast
and slow time where fast time is used to calculate range.
Pulse compression must first be applied to the data by a

Range [m]
matched filter operation using the reference pulse over fast time
(labelled as (1) shown in pink in Fig. 5). Doppler processing
is then applied over 256 pulses by applying the FFT of N
samples over slow time at each range (labelled as (2) shown in
green in Fig. 5). The resultant image is a Range-Doppler map
of the scene and an example is shown in Figure 6. Multiple
frames can be extracted from the original datacube resulting in
a sequence of Range-Doppler frames. The data shown in Fig.
2 is transformed by the preprocessing (illustrated in Fig. 5) to
the form shown in Fig. 6. The form in Fig. 6 is range-Doppler Doppler Bin
data showing close-range stationary clutter at around Doppler
bin 128, a stationary rock called Roman Rock at range 1400 Fig. 6: A pulse Doppler frame after applying the transform
samples and one moving target at range 950 samples having a illustrated in Fig. 5 to the data shown in Fig. 2.
positive Doppler indicating that the target is moving towards
the radar. The figure shows typical high levels of close-range
clutter. There is a generally low level of clutter return due to
IV. F EATURE S ELECTION
a sea state of 1 and a near perpendicular look angle. The high
levels of return at 2000 m seems to be a system error. There are two processes that are applied. The first process
is a set of training exercises in which typical datasets are
used where the two classes are pre-identified to the process.
Fast Time (Range)
The feature values at the boundary are the decision thresholds
for discriminating between the classes. A number of training
(1) Pulse compression exercises may be applied in order to reach a final boundary
over range line
definition.
Slow Time

The second process is to apply the boundary to a dataset


in which the classes have not been identified to the process.
(2) Doppler processing The class identification of the feature blocks may still however
over 256 pulses at
every range be known and the result is used as a validation test. Cross
Range
validation technique called leave-p-out was used to divide up
the available dataset into training and validation tests.
Doppler

The diagram in Fig. 6 was broken up into horizontal


rectangular sections consisting of the full Doppler width and
10 range rows each. The magnitude of each of the 256 Doppler
Fig. 5: Illustration of how the raw NetRAD datacube is by 10 range pixels constitute the features that are processed by
preprocessed. The top datacube represents the raw data in fast the machine learning algorithm. Each of these feature blocks
time and slow time. (1) First all the range lines are pulse were preclassified as representing a target or clutter which was
compressed with the reference pulse. (2) then an FFT operation given class identifiers 1 and 0 respectively. An example of a
is applied to 256 pulses at each range. The result is a range feature block containing a target is shown in Fig. 7.
Doppler datacube consisting of 256 Doppler bins. A potential misclassification results when a target is tran-
sitioning between adjacent feature blocks. This is shown in
Fig. 8. The solution to this was to remove the frames that
exhibited a transitioning target from the training and validation
The processing chain presented in Figure 5 was imple- data. A more optimal solution would be to implement a sliding
mented in Julia [10] and was used to process all the data window.
for this project. Since data diversity is important for machine
learning, only several pulse-Doppler images were considered
from each dataset (5 to 15 images per experiment) which V. R ESULTS FROM SVM C LASSIFICATION
significantly reduced the total data size. No kernel function was used resulting in a linear model.
The alternative is a non-linear model by applying a kernel
The measurements that a machine learning algorithm uses function such as polynomial or radial basis function. These
for classification are called features. Careful consideration kernel functions were tried but did not give meaningful results.
must be given when deciding what features to use since well There is a data imbalance favouring the 0 class (sea clutter)
selected features can significantly improve the results. The since all the frames have 1 target feature block out of the
features considered in this paper are discussed in the next 204 feature blocks. To reduce the amount of data and balance
section. the classes out, random feature blocks of no-target data were

978-1-4673-8823-8/17/$31.00 ©2017 IEEE 1224


Predicted Class
Target
Target
Clutter

Range [m]
Class
Identifier

Clutter

Feature Block Number


True Class
Target

Doppler Bin
Class
Identifier
Fig. 7: Feature block containing a target. This block would be
classified as class 1.
Clutter
Feature Block Number

Fig. 9: Example plot of SVM validation test results comparing


the predicted class to the true class for each feature block in
the dataset. This example shows a partially successful test and
Range [m]

is the results of iteration 7 in Fig 12.

Predicted Class
Target
Target
Clutter

Class
Identifier

Clutter

Feature Block Number


True Class
Range [m]

Target

Class
Identifier

Clutter
Feature Block Number

Doppler Bin
Fig. 10: Example plot of SVM validation test results comparing
the predicted class to the true class for each feature block in
Fig. 8: The transitioning of a target between adjacent feature
the dataset.. This is the results of iteration 2 in Fig. 12.
blocks.

was repeated 10 times. The averaged results for these tests are
removed from the dataset. Comparative tests were performed
shown in TABLE II. For the observed optimal value of k (3),
removing 95%, 75%, 50% and 0% of no-target data. Pd is
tests were performed removing 95%, 75%, 50% and 0% of
calculated as the number of correctly predicted target present
no-target data. These results are shown in TABLE III.
feature blocks divided by the total number of true target present
feature blocks. Pf a is calculated as the number of incorrectly The success of the results from the kNN test prompted the
predicted target present feature blocks divided by the total application of this algorithm to a dataset containing several
number of true clutter only feature blocks. Each of these tests targets from a boat race. The results of this are shown in Fig.
were repeated 20 times and the average Pd and Pf a are shown 15 where three successful target detections can be clearly seen.
in TABLE I. Fig. 12 is an example of the results from one of
the 20 tests conducted with 75% of the no-target data removed.
Examples of tests comparing the predicted class with the actual
class are shown in Figs. 9, 10 and 11 where Fig. 9 is iteration % data removed Pd Pf a
7 of Fig. 12, Fig. 10 is iteration 2 and Fig. 11 is iteration 3. 95% 0.850 0.507
75% 0.777 0.465
50% 0.740 0.505
VI. R ESULTS FROM K NN C LASSIFICATION 0% 0.743 0.502
An Example of a test comparing the predicted class with
the actual class is shown in Fig. 13. Tests were performed TABLE I: Summary of results from SVM classifier tests.
with different values of k ranging from 2 to 5 where each test

978-1-4673-8823-8/17/$31.00 ©2017 IEEE 1225


Predicted Class % data removed Pd Pf a
Target 95% 0.920 <1×10−5
Target
Clutter 75% 0.910 2.16×10−5
50% 0.901 1.07×10−5
Class
Identifier 0% 0.900 1.07×10−5

Clutter
TABLE III: Summary of results from kNN classifier tests for
Feature Block Number different amounts of data removed.
True Class
Target

Predicted Class
Class Target
Target
Identifier
Clutter

Class
Clutter Identifier
Feature Block Number

Clutter
Fig. 11: Example plot of SVM validation test results comparing Feature Block Number
the predicted class to the true class for each feature block in True Class
Target
the dataset. This is the results of iteration 3 in Fig 12 and
represents a completely failed test.
Class
Identifier

Clutter

Feature Block Number

Fig. 13: Example plot of kNN validation test results comparing


the predicted class to the true class for each feature block in
the dataset. This is the results of iteration 4 in Fig 14.

Fig. 12: Example of SVM Pd and Pf a per leave-p-out iteration


for one of the 20 tests. 75% of no-target data was removed.

VII. C ONCLUSION AND FUTURE WORK


In this work we present a machine learning approach to
target detection in the presence of sea clutter. This approach
has its own limitations in terms of potential lack of enough Fig. 14: Example of kNN Pd and Pf a per leave-p-out iteration
training data. However with the newer systems under devel- for one of the 20 tests. 75% of no-target data was removed.
opment like NeXtRad [11] we expect to get sufficient amount
of training data to make this approach successful. Secondly
the new generation of deep learning algorithms [12] are very
robust to this problem of insufficient trianing data. This investigation has identified kNN as more successful
than SVM in this application because of consistently low Pf a
for all iterations of the tests as seen in Fig. 14 compared to
k Pd Pf a erratic Pf a performance of the SVM algorithm as shown in
2 0.883 <0.001 Fig. 12.
3 0.909 <0.001
4 0.866 <0.001
5 0.886 <0.001
This finding should guide future work on the improvement
of machine learning algorithm performance. Finally a com-
parison of the algorithm with conventional methods such as
TABLE II: Summary of results from kNN classifier tests for CFAR should be conducted in marginal detection conditions.
different values of k.
Overall we are confident that this will open up a new

978-1-4673-8823-8/17/$31.00 ©2017 IEEE 1226


Fig. 15: Results of applying the kNN trained classifier on a
dataset from a boat race. The class identifier is colour coded
with blue indicating no target and yellow indicating a target
detection. The figure on the right is the range-Doppler data
that was passed through the algorithm.

chapter in the domain of small target detection in sea clutter.

R EFERENCES
[1] S. Watts, C. Baker, and K. Ward, “Maritime surveillance radar. ii.
detection performance prediction in sea clutter,” in IEE Proceedings
F-Radar and Signal Processing, vol. 137, no. 2. IET, 1990, pp. 63–
72.
[2] K. D. Ward, S. Watts, and R. J. Tough, Sea clutter: scattering, the K
distribution and radar performance. IET, 2006, vol. 20.
[3] M. A. Richards, J. A. Scheer, W. A. Holm et al., Principles of modern
radar. Citeseer, 2010.
[4] T. Islam, M. A. Rico-Ramirez, D. Han, and P. K. Srivastava, “Artificial
intelligence techniques for clutter identification with polarimetric radar
signatures,” Atmospheric Research, vol. 109, pp. 95–113, 2012.
[5] R. Vicen-Bueno, R. Carrasco-Álvarez, M. Rosa-Zurera, J. C. Nieto-
Borge, and M.-P. Jarabo-Amores, “Artificial neural network-based clut-
ter reduction systems for ship size estimation in maritime radars,”
EURASIP Journal on Advances in Signal Processing, vol. 2010, no. 1,
p. 1, 2010.
[6] G. Mountrakis, J. Im, and C. Ogole, “Support vector machines in remote
sensing: A review,” ISPRS Journal of Photogrammetry and Remote
Sensing, vol. 66, no. 3, pp. 247–259, 2011.
[7] P.-L. Shui, D.-C. Li, and S.-W. Xu, “Tri-feature-based detection of
floating small targets in sea clutter,” IEEE Transactions on Aerospace
and Electronic Systems, vol. 50, no. 2, pp. 1416–1430, 2014.
[8] J. Friedman, T. Hastie, and R. Tibshirani, The elements of statistical
learning. Springer series in statistics Springer, Berlin, 2001, vol. 1.
[9] J. Gao and K. Yao, “Multifractal features of sea clutter,” in Radar
Conference, 2002. Proceedings of the IEEE, 2002, pp. 500–505.
[10] S. Jonkers, “Software Infrastructure for NeXtRAD Development in Julia
Programming Language,” Master’s thesis, University of Cape Town,
2016.
[11] M. Inggs, H. Griffiths, F. Fioranelli, M. Ritchie, and K. Woodbridge,
“Multistatic radar: System requirements and experimental validation,”
in 2014 International Radar Conference. IEEE, 2014, pp. 1–6.
[12] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521,
no. 7553, pp. 436–444, 2015.

978-1-4673-8823-8/17/$31.00 ©2017 IEEE 1227

You might also like