The coherent temporal characteristics of
medium-to-low grazing angle sea clutter and small boat reflectivity
are considered for different radar waveforms under a range
of environmental conditions and geometrical configurations. Accurate
empirical modelling of sea clutter enables the inference
of the local sea conditions from radar returns, pertinent for
port safety and navigation. Understanding the dynamics and
associated reflectivity of small boats, in addition to empirical
sea clutter models, allows the development of advanced detection
and tracking algorithms, which will improve the performance of
surveillance and marine navigation radar against small boats.
Work presented is based on the empirical analysis of data
recorded with two calibrated, coherent, pulsed radar systems at
X-band frequencies. Specifically, target echoes from small boats
are included in the datasets and subsequent analysis.
1. Introduction
There is a growing need for accurate, real-time
instrumentation of the sea surface for safe navigation of vessels in and around
harbours and shipping lanes. A
commercial product, wave and surface current monitoring system (WaMoS)
[1], can be connected
to a conventional X-band marine radar. WaMoS II processes the unfiltered sea
clutter to estimate the wave and surface current parameters in near real-time.
According to the manufacturers the instrumented range is 0.1–3km.
At X-band operating frequencies, the main scattering mechanism is that of Bragg
scattering, associated with the resonant capillary waves [2]. Capillary waves, in turn,
are generated by the local near-surface wind and do not propagate beyond the breeze
area. Only in a fully developed sea can the wave height be directly related to
the present wind speed and can the significant wave height Hs be accurately inferred from the average sea
clutter reflectivity [3]. In transient sea conditions, the best fit for the
Georgia Institute of Technology (GIT) mean sea clutter reflectivity model
[4] is found if the
sea state is related to the local mean wind speed rather than Hs [5, 6]. It can therefore be deduced that it is possible to
also infer local wind conditions from X-band sea clutter.
This article investigates the
temporal characteristics of coherent sea clutter, with specific interest in the
Doppler characteristics of sea clutter and the relationship thereof to the
local wind and wave conditions (e.g., average wind speed, wind
direction, and wind gusts).
The significant increase in sea clutter reflectivity
in rough seas with strong winds, together with relatively small radar cross
section (RCS) of small boats (e.g., yachts, ski boats, and rigid
inflatable boats (RIBs)),
has often been blamed for disasters at sea where
large ships collided with these small boats [7]. In certain cases, the marine radar could not discern
the boat signature from the clutter, while in other cases there were too many
false tracks established leading to the subsequent disabling of the automatic
tracker. For safe navigation it is pertinent that the detection capabilities of
marine radar in adverse conditions are improved. With the introduction of
cheaper, solid-sate, coherent marine radar [8] a whole new class of coherent detection algorithms has
become applicable to marine navigation radar, for example, [9–11]. Theoretical and
first-order empirical analysis suggest subclutter visibility [11]. Little work has been presented on the performance of
this class of detectors (often referred to as asymptotically optimal) on
measured data of small boats [12]. No reference is made to the performance as a
function of the type of manoeuvring of the boat and the influence of the boat
on the surrounding sea surface. This article investigates the performance of
the adaptive linear quadratic (ALQ) detector [11] under different boat
manoeuvres for different types of RIBs, including
the 4.2m pencilduck type that is often used for
watersport racing. This boat has a very small RCS, but since it has a racing
speed of up to 40kts the local disturbance of the sea surface often
exceeds the boat RCS by up to 10dB.
With improved subclutter visibility, the problem arises that first detections
are declared not only for small boats, but also for large birds such as
seagulls, with a typical RCS of 0.01–0.1m2 [3], and angels (flocks of birds flying together).
Effective algorithms to discard tracks established on birds have to be
developed [13].
Typical scanning surveillance systems (including
maritime radar) have to declare a detection using only a limited number of
pulses. Due to the long decorrelation time of sea clutter [14] and the more often than not
spiky amplitude statistics [15], detection is quite difficult due to the short dwell
time. Persistent, ubiquitous surveillance has become a top priority
internationally. Typical entities of interest range from small recreational
watercraft to large tanker ships. One of the characteristics of such systems, for
instance, AwareNet [16], is the ability to employ long dwell times at
specific areas of interest through the utilization of multiple, electronically
steered receiver channels. Improved discernibility of small boats with long
dwell times is therefore investigated in this article.
Two sea clutter and boat reflectivity measurement
trials were conducted in 2006 and 2007 on the south western coast of South
Africa. The aim of these trials was firstly to record datasets of sea clutter
returns at different frequencies, range resolutions, grazing angles, look
angles, and environmental conditions to validate current sea clutter models.
Secondly, the aim was to record boat reflectivity datasets for a number of
small boats to investigate its detectability with open literature detectors
that will hopefully lead to the development of improved detection algorithms
for radar systems employing adaptive dwell times.
The layout of the article is as follows. Section 2
presents an overview of the two measurement trials. The results presented in
this article were obtained from the analysis of the data recorded during these
trials. A description is given of the different radar systems, the experimental
set-up, as well as the system and data integrity verification procedures.
Section 3 investigates a subset of the sea clutter measurements, focusing on
the amplitude statistics and temporal characteristics for fixed frequency and
frequency agile waveforms. The RCS and temporal fluctuation of a variety of
small boats are investigated in Section 4 for different manoeuvres. Of
particular interest is the effect of the boat manoeuvring on the local sea
surface and its subsequent reflectivity. Section 5 presents an analysis of the
detectability of these small boats, seagulls, as well as angels.
2. Overview of Measurement Trials
The first
measurement trial was conducted with the Fynmeet dynamic RCS measurement
facility (Figure 1) at the Overberg Test Range (OTB). The site provided azimuth
coverage of 135∘ predominantly up-swell with well-developed
waves and with a significant variation in wind direction. Sea clutter at
grazing angles 0.3–3∘ were recorded. The second measurement trial
was conducted with an experimental, monopulse, X-band radar (Figure 5) deployed
on top of Signal Hill in Cape Town. This site provided azimuth coverage of 140∘ from up- to cross-swell, but with only two
predominant wind directions for the duration of the trial. The sea was more
representative of open sea conditions. Sea clutter and littoral clutter at
grazing angles 0.3–10∘ were recorded. The experimental radar uses
pulse compression to increase the system gain and subsequently yields extended
range capabilities compared to Fynmeet.
Fynmeet deployed at OTB.
2.1. Overberg Test Range 2006 Trial2.1.1. Radar and Experimental Set-Up
The radar was deployed at OTB at location 34∘36′56.52′′S,20∘17′17.46′′E,67m above mean sea level (AMSL). The shortest
distance to the coastline was 1.2km due south. A plan overview of the deployment
site is depicted in Figure 2. The important specifications of Fynmeet are
listed in Table 1.
Fynmeet system and performance specifications.
Transmitter
Frequency range
6.5–17.5 GHz
Peak power
2 kW
PRF range
0–30 kHz
Waveforms
100 and 300 ns pulsed Continuous Wave (CW), fixed/pulse-to-pulse frequency agile (500 MHz)
Antenna
Type
Dual offset reflector
Gain
≥30 dB
Beam width
≤2° (3 dB beam width)
Side lobes
≤−25 dB
Receiver
Dynamic range
60 dB (instantaneous)/120 dB (total)
Capture range
200 m–15 km
Range gates
1–64; ΔR=15 m or 45 m
Sampler type
I/Q intermediate frequency sampler
Image rejection
≤−41 dBc
Radar deployed on Signal Hill with open view of sea.
Plan overview of Fynmeet deployment site.
Local wind speed and direction (Figure 4(a)) were logged with two
weather stations separated by 1km.
The local wave direction ϕwave,
significant wave height Hs,
maximum wave height Hmax and wave period Twave (Figure 4(b)) were logged with a directional recording wave
buoy. The local wave structure is influenced greatly by significant weather
patterns from the south west and further perturbed by the diffraction patterns
due to the cape, southwest of the deployment site which is located in a small
bay area with a sea bed depth varying between 10–30m at ranges of 3–10km.
The ground truth tracks of the boats were estimated using a differential
processing global positioning system (GPS) receiver.
2.1.2. System and Data Integrity Verification
For absolute RCS calibration the response from a
sphere suspended below a helicopter, tracked in range with a typical α−β tracker and in angle with a video centroid
tracker, was measured and the calibration coefficient was calculated
asCcal=20log10(1N∑n=1N|∑m=M(n)−1M(n)+1x(n,m)|AR4σcs), where N is the number of pulses transmitted, M(n) the range gate with maximum return for the nth pulse, A the receiver attenuation, and σcs the sphere RCS. A standard deviation of 1dB was achieved. Daily stability verification
measurements were done with a corner reflector, exhibiting variations in the
order of about 1dB across the measurement period.
Linearity of the quadrature receiver channels were
ascertained by the analysis of calibrated noise source, receiver noise, and
blue sky measurements taken throughout the trial. This analysis included the
estimation of channel skewness, kurtosis as well as the 2nd to 4th normalized intensity moments, I2−I4.
The amplitude and phase imbalance of the quadrature channels were estimated as 0.03dB and 1∘ resulting in a negative Doppler image of ≤−41dBc.
Additionally, there were also harmonically related spurious responses at a
level of ≤−50dBc.
A 5MHz leak-through signal was identified and removed
from the data. The signal phase was nondeterministic and therefore the amplitude
and phase were estimated from the dataset itself. Sea clutter with a strong
steady state component biases this estimate and the best results were obtained
by using a censored mean level technique in the estimation process. The
percentage of the dataset censored was chosen such that the resultant estimate
yielded the lowest variance. Applying the discrete fourier transform (DFT) to
the corrected data, a 0 Hz frequency bin with a comparable power density to
adjacent frequency bins was obtained without suppressing the steady state
clutter response.
2.1.3. Trial Summary
Sea clutter
datasets were recorded over a period of 11 days, with 112 fixed frequency and 38 stepped frequency datasets centred at four
transmit (Tx) frequencies over an azimuth angle range from 90∘N to 225∘N and a grazing angle range 0.3–3∘.
The local weather pattern may be described as roughly following a 6-day cycle
[17] as cold front
systems pass by from the west to the east. Over the trial period, the average
wind speed ranged between 1–20kts,
with a maximum gust of 40kts.
Wind direction spanned 360∘,
but the high wind speeds were mainly from the south west. Hs varied between 1–3.8m,
with a maximum recorded wave height of 7.31m.
The wave direction was roughly 180∘N,
and slowly changed direction toward the end of the trial to 135∘N.
Even with a low wind speed of 1kt,
the significant wave height was ≥1m.
This is due to the strong incoming south westerly swell combined with the
diffraction patterns of the close-by cape and the reduction in sea depth. The
illuminated sea area can therefore not be defined as a fully developed sea
[3] and the standard
tables relating wind speed and wave height to sea state do not apply. In terms
of the average wind speed, sea states 1 to a low 5 were observed. In terms of the wave height,
sea states from a high 2 to a high 5 were observed.
A 5.7 m RIB, a glass fibre ski boat and a wooden
chokka fishing vessel (Figure 4) were deployed on 4 days with conditions ranging from calm to
rough seas. The boats sailed a number of manoeuvres at different ranges and
azimuth angles. A total of 55 fixed frequency and 43 stepped frequency datasets were recorded
centred at four Tx frequencies.
Environmental conditions during OTB trial: (a) wind, and (b) wave.
Boats deployed during OTB trial: (a) 5.7 m RIB, (b) ski boat, and
(c) chokka fishing vessel.
The recorded datasets have been made available to the
international research community. For information on the available datasets and
how to access these datasets, refer to http://www.csir.co.za/small_boat_detection/.
2.2. Signal Hill 2007 Trial2.2.1. Radar and Experimental Set-Up
The experimental, X-band, monopulse radar was deployed on Signal Hill at location 33∘55′15.62′′S,18∘23′53.76′′E,308m AMSL, as indicated on the plan view in Figure 6.
The shortest distance to the coast line was 1250m at a bearing of 288∘N.
The site provided 140∘ azimuth coverage from 240∘N to 20∘N,
of which a large sector spanned open sea whilst the remainder looked towards
the West Coast coastline from the direction of the open sea. The radar had an
open view of Robben Island at a distance of 11km.
Grazing angles ranging from 10∘ at the coastline to 0.3∘ at the radar instrumented range of 60 km were
obtained.
Plan overview of radar deployment site.
Local wind conditions (Figure 8(a)) were measured at the radar,
Robben Island, Cape Town Harbour as well as Slangkop (south-southwest of the
radar). The local wave conditions (Figure 8(b)) were measured with a seabed-based wave sensor
at Camp's Bay and a directional wave buoy at Cape Point whilst numerically modelled
at eight other locations in Table Bay and around Robben Island. The tracks of
the instrumented boats (Figure 8) were estimated using a differential-processing GPS
receiver.
Boats deployed during Signal Hill
trial: (a) Nadine Gordimer, (b) Rotary Endeavour, (c) pencilduck, and (d)
SANParks RIB.
Environmental conditions during Signal Hill trial: (a) wind, and (b) wave.
2.2.2. System and Data Integrity Verification
Due to the
similarity of the two radar systems, similar system and data integrity
verification process were followed as for the Fynmeet radar described in
Section 2.1.2. The experimental radar employs matched filter pulse
compression, where pulse compression codes cpc are designed to yield a specific pulse
compression gain, sidelobe levels and blind range. In the calibration procedure
the height and range of the helicopter carrying the calibration sphere over the
sea were restricted. The above-mentioned restrictions result in not all codes
being calibrated. It is possible however, to estimate Ccal for the uncalibrated codes from the calibrated
codes by adding the relative pulse compression gain for the uncalibrated
codeGpc=∑n=1N|cpc(uncal,n)|2∑k=1K|cpc(cal,k)|2,where N is the uncalibrated and K the calibrated code lengths. Equation (2) is
valid for a matched filter with unit noise gain. Similarly, the Doppler
processing gain can be estimated for the uncalibrated Doppler processing
coefficients cdp,Gdp=[∑l=1Lcdp(cal,l)2][∑m=1Mcdp(uncal,m)]2[∑l=1Lcdp(cal,l)]2[∑m=1Mcdp(uncal,m)2],where M is the uncalibrated and L the calibrated Doppler processing coefficient
lengths. For the experimental radar, the image rejection and spurious response
is sufficiently below the noise floor. Pulse compression codes utilized during
the measurement trial yielded range sidelobe levels in the order of −35dB.
2.2.3. Trial Summary
Sea clutter
datasets were recorded on eight different days over a period of thirteen days.
The predominant wind direction was northwestern, but with southeastern
intervals. The average wind speed varied between 0kts and 40kts,
with a maximum gust of 60kts.
The significant wave height ranged in 1–4.5m,
whilst the swell direction varied between 230∘N and 270∘N.
Datasets of the instrumented boats depicted in Figure 8 were recorded on five different days.
The Nadine Gordimer is a 10 m Class A deep sea rescue vessel with two MTU 1000 turbo
diesel inboard motors and a range of communication antennas. The Rotary
Endeavour is a Class 3 5.5m RIB with two 60hp Yamaha outboard motors and a single VHF
antenna. The South African National Parks (SANParks) RIB is a 4.8m RIB with a 60hp Yamaha outboard motor. The 4.2m pencilduck has a single 50hp outboard motor with no antennas. In addition,
datasets were recorded for a large variety of noncooperative boats of
opportunity. Recordings were made using a range of fixed frequency and stepped
frequency waveforms.
3. Sea Clutter Analysis
Various
statistical properties are evaluated in this section for seven sea clutter
datasets recorded during the OTB 2006 measurement trial. These datasets
represent low and high sea states at grazing angles 1∘ and 0.5∘ for a single Tx frequency and pulse widths of 100 nanoseconds and 300 nanoseconds.
The range-time intensity plot for the high sea state, 1∘, 100 nanoseconds dataset CFC16-001 is presented in
Figure 9.
Range-time intensity plot for dataset CFC16-001.
The strong underlying modulation caused by the well
developed waves is clearly visible for this up-swell configuration dataset. For
cross-swell configurations and further ranges coupling between the waves and
the underlying modulation becomes less pronounced as multiple waves are
contained within a resolution cell, which is defined by the azimuth beamwidth
and radar range resolution at low grazing angles. The underlying modulation
also becomes less pronounced in weaker swell conditions.
3.1. Mean Reflectivity and Amplitude Statistics
The mean
reflectivity σ0 and clutter-to-noise ratio (CNR) for the
different OTB 2006 datasets are tabulated in Table 2 and compared to the GIT
and the hybrid (HYB) models [4]. In both models the sea state S was derived from the mean wind speed vwind using the empirical relationvwind=3.16S0.8.
Empirical reflectivity [dBm2/m2] and CNR [dB].
Sea state
Low
Low
Low
Low
High
High
Grazing angle
1∘
0.5∘
1∘
0.5∘
1∘
0.5∘
Resolution
[m]
15
15
45
45
15
15
CNR [dB]
11
−3
17
5
21
7
Reflectivity [dBm2/m2]
−48
−51
−51
−52
−39
−41
GIT
model [dBm2/m2]
−101
−115
−101
−115
−35
−39
HYB
model [dBm2/m2]
−47
−53
−47
−53
−36
−39
From Table 2 it can be concluded that there is good
agreement between empirical σ^0 and the HYB model, with values generally
between that of the HYB and the GIT. The GIT typically underestimates σ0 at low grazing angles for low sea states by a
significant margin, as discussed in detail in [4]. Of particular interest is
the good fit found by matching sea state to local wind speed rather than Hs in transient sea conditions (not fully
developed).
The 2nd to 4th normalised intensity moments I2−I4 and the estimated shape parameter ν^ are tabulated in Table 3, assuming a k-distributed
envelope process and compared to the shape parameter model νmodel [15]. In this estimation, the theoretical relationship
between the actual shape parameter ν^ and the effective shape parameter ν^eff in the presence of noise is used [18]:ν^eff=ν^(1+1CNR)2.
Empirical versus model amplitude statistics.
Sea state
Low
Low
High
High
Grazing angle
1∘
0.5∘
1∘
0.5∘
I2
3.57
2.27
6.05
4.61
I3
24.88
9.47
72.32
63.75
I4
261.9
66.88
1238.6
1629.6
ν^
1.18
2
0.57
1.22
νmodel
1.88
1.79
0.92
0.88
The non-Rayleigh envelope statistics is evident.
Theoretically spikiness should increase with a decrease in grazing angle.
However, this is contradicted in the empirical analysis where a decrease in
spikiness is observed with a decrease in grazing angle. This may be due to
well-developed waves located closer to the radar (3km) at the high grazing angles yielding
increased spikiness, with less developed waves at the far-out ranges (8km) where the low-grazing-angle sea clutter data
was recorded.
These observations are indicative of the highly
complex scattering environment and illustrate a still incomplete understanding
of sea clutter.
3.2. Average Doppler Characteristics
As the
capillary waves are the main scattering mechanisms at X-band and they are directly
influenced by the near-surface local wind [2], it can be expected that the Doppler and
autocorrelation properties of sea clutter are directly influenced by the local
wind. The sea clutter speckle autocorrelation r(τ) [14] is plotted in Figure 10 for four different datasets. From the magnitude response it is evident
that the speckle decorrelation time is 10–20 milliseconds,
which is consistent with literature [14]. It also indicates that the decorrelation time is
affected by sea state, where the decorrelation time decreases as the sea state
(roughness of the sea) increases. An explanation for this may be the more rapid
deformation of the capillary waves in rough seas. The complex autocorrelation
is strongly coupled to the Doppler characteristics of the sea clutter.
Evaluation of the real and imaginary components of r(τ) reveals that the second zero-crossing of ℑ{r(τ)} approximates 1/2 of the mean projected Doppler
period,fd(ϕwind)≈[2τ|(τ>0,ℑ{r(τ)}=0)]−1.This together with the empirical
model [3]fd(ϕwind)≈2vwindcos(ϕwind)f04cenables the estimation of the
local projected wind speed from an analysis of the estimated autocorrelation.
With a complete azimuth scan of the radar it would be possible to infer both
wind speed and direction. This lies outside the scope of this paper and will be
the subject of future research.
Speckle autocorrelation for different datasets: (a) magnitude, (b) real, and (c) imaginary.
3.3. Spectrally Inhomogeneous Sea Clutter
Section 3.2
provided empirical evidence that the average wind speed can be inferred from
the sea clutter speckle autocorrelation, which is strongly correlated to the
average Doppler response thereof. Experimental data suggests however that the
sea clutter spectrum is inhomogeneous in both range and time in general. High
Doppler resolution spectrograms of three different geometrical configurations
and environmental conditions are presented in Figures 11–13. In addition to
the spectrograms I2 is plotted as a function of Doppler frequency,
yielding an indication of the spikiness per Doppler resolution cell. The
spectrogram in Figure 11 is representative of an up-swell sea (Hs=3.4m) with strongly developed waves and up-wind (vwind=15kts) configuration at a range of 3.8km. From the spectrogram the different
individual waves can be distinguished as they are propagating through the given
range cell. The different individual waves have very different Doppler spectra,
which results in a significantly raised I2 at the Doppler velocities associated with
localised wind gusts. Thus in the Doppler domain these echoes will compete with
those of real targets and hence may have an adverse effect on the false alarm
rate. Since the individual waves are resolved, I2 is also higher than 2 at the mean Doppler frequency. The spectrogram
in Figure 12 is representative of a 70∘ cross-swell sea (Hs=2.8m) and down-wind (vwind=15kts) configuration at a range of 5.3km.
The sea was more representative of open sea conditions. From the spectrogram it is
clear that the short-time Doppler spectrum is much more homogeneous and it is
impossible to distinguish individual waves or events. At the mean Doppler
frequency I2 tends to the theoretical value of 2 and is only slightly raised at the average
Doppler spectrum edges. The spectrogram in Figure 13 is representative of a 20∘ up-swell sea (Hs=2.5m) and up-wind (vwind=7kts) configuration at a range of 5.6km.
Once again the sea was representative of open sea conditions. For this dataset the
short-time Doppler spectrum is inhomogeneous compared to the previous dataset,
but not as severe as the first dataset analysed in this subsection. For this
up-swell configuration there is evidence of
individual waves propagating through the range cell, but it is clear that more
than one wave are contained within the resolution cell. The events associated
with the broadened Doppler response may be associated with whitecaps blown off
the top of the waves by the higher wind and/or gusts. The spikiness at the
Doppler frequencies associated with the local maximum wind speed is confirmed
by a significant raise in I2.
Spectrogram and I2 for OTB
dataset CFC16-001.
Spectrogram and I2 for
spectrally homogeneous Signal Hill dataset.
Spectrogram and I2 for
spectrally inhomogeneous Signal Hill dataset.
This brief analysis of the sea clutter Doppler
spectrum and I2(fd) suggests that it is possible to also infer the
existence and severity of whitecaps from the sea clutter.
3.4. Frequency Agility
It is generally
accepted that sea clutter speckle decorrelates with frequency agility when the
frequency step size exceeds the pulse bandwidth, Δfc≥B [3]. The correlation coefficient ρ(f0,fn) is plotted for a coherent processing interval
(CPI) of 100 milliseconds at a fixed range cell over a period of 60 seconds–-depicting the correlation between the base
Tx frequency f0 and an offset frequency of up to f0+130MHz for a pulse bandwidth of 10MHz.
From Figure 14 it can be
concluded that in general the sea clutter speckle decorrelates whenever Δfc≥B.
However, there are a number of CPIs where the
speckle only decorrelates after a step size of 40MHz.
Discrete spike events can also be identified where there is strong correlation
for a step size of up to 130MHz.
Most radar detection mechanisms will declare these spikes as targets. It is
important to note that these discrete spike events have a typical lifetime of 0.5–2 seconds.
Sea clutter frequency agility decorrelation in the presence of
discrete spikes.
Thus, overall these observations show broad agreement
with those reported elsewhere and with the GIT and HYB models. They also
re-emphasise the complexity of the scattering environment in which a target
is required to be detected. This aspect is examined
further in the following two sections.
4. Small Boat Reflectivity Analysis
This section
presents the results of the analysis of a range of small instrumented boats
deployed during the two trials detailed in Section 2.
4.1. Small Boat RCS and Amplitude Statistics
The mean RCS of
the small boats, averaged over aspect angle, have been estimated from the
measured data and tabulated in Table 4. Isolation of the boat signature from
the sea clutter was obtained by Doppler filtering, using the GPS-estimated
Doppler frequency as input. The responses of the three-range cells closest to
the GPS range were coherently added to counter range-gate straddling losses.
The effects of multipath fading from smooth sea surfaces and of shadowing of
the boat in higher sea states on the RCS values were not corrected for in the
results presented in Table 4.
Small boat RCS [dBm2].
Boat description
Length
Width
Engines
RCS (m2)
Pencilduck
4.2m
1.6m
1×50hp
≈1
SANParks
RIB
4.8m
1.8m
1×60hp
1–3
WaveRider
RIB
6.5m
2m
2×85hp
1–5
Ski
Boat
5.2m
1.8m
2×85hp
4–15
Chokka
Fishing Vessel
6.2m
2m
1×inboard
5–16
From Table 4 it is clear that the RCS of
an RIB is related to its physical size and ranges
between 1–5m2.
The RCS of solid boats are larger in general, with the mean RCS of the chokka
fishing vessel up to 16m2.
In addition to the mean RCS of the small boat it is also critical for detection
performance calculation to have a good model for its RCS fluctuation. The
isolated boat RCS signature for the chokka fishing vessel is plotted in Figure 15, with the empirical probability density function (PDF) plotted for both the
chokka fishing vessel and the WaveRider RIB in Figure 16.
Empirical PDF of chokka fishing vessel and WaveRider RIB compared to
Swerling models 1 and 3.
Slight differences between the PDFs of the two
boats are visible as well as the inability of the Swerling models to accurately
describe their amplitude distributions. A characteristic of small boats in
heavy sea is the fading of the RCS as the boat steers into the troughs of the
waves, as indicated by the close-up view about 25–29 seconds in Figure 15. This suggests a strong
correlation of boat RCS with the local sea waves and explains the poor fit of
the empirical PDFs to the Swerling
models at low RCS values. The noncoherent autocorrelation of a boat steering
directly into the waves (Figure 17) shows periodicity with the same period as
the mean sea wave period. This correlation of boat RCS and sea clutter echo
strength further complicates an already challenging detection problem.
RIB non-coherent autocorrelation showing strong correlation with sea
waves.
4.2. Small Boat Doppler Bandwidth
The Doppler bandwidth of a target is an important
design parameter for optimal coherent detection. Coherent processing gain is
only achieved by an increase in the CPI whilst it still remains less than or
equal to the inverse of the target Doppler bandwidth. High Doppler resolution
spectrograms have been computed for the different small boats, with the results
for CPI's of 10 milliseconds and 200 milliseconds plotted in Figure 18 for the WaveRider RIB.
High Doppler resolution spectrograms for different CPI's: (a) 10 milliseconds, and (b) 200 milliseconds.
The Doppler bandwidth for the WaveRider RIB can be
observed from the high-resolution spectrogram. As for most of the other small
boats, this is approximately 10Hz.
This yields an optimal CPI ranging between 100–200 milliseconds.
As the CPI increases beyond this value, the boat energy will start to spread
over multiple Doppler resolution cells, yielding no additional coherent
processing gain. This is a key consideration in the design of an optimal
detector.
4.3. Highly Manoeuverable Small Boats
Due to its
small size, light weight and powerful engines, the RIB class of small boats is
highly manoeuvrable with the ability to reach speeds of up to 40kts for even the small 4.2m pencilduck. Especially in the design of
coherent detection and tracking algorithms, it is important to have a good
understanding of the anticipated manoeuvrability of the target. In addition,
the disturbance of the manoeuvring boat on the local sea surface may also
greatly influence its detectability either adversely or positively. The high
Doppler resolution spectrograms of two different RIB's are plotted in Figure 19–21 for three different manoeuvres. The narrow Doppler response of the drifting
pencilduck (Figure 19) is evident, as well as the slight movement of the
pencilduck due to the local waves. The drifting pencilduck caused little
disturbance on the local sea surface. The WaveRider RIB steering radially
outbound at a speed of about 10kts (Figure 20) had a narrow Doppler response, with
a local disturbance of the sea surface visible when the RIB was crashing
through the crests of the waves. This local disturbance is observed as quite
broad Doppler bandwidth noise with Doppler velocities ranging from slightly
higher than the speed of the boat down to the Doppler velocity of the local sea
clutter speckle. The spectral density of the disturbance is 20dB lower than the boat signature. The pencilduck
racing at 40kts radially outbound (Figure 21) still had a
narrow Doppler response for the body of the boat, but caused a significant
local disturbance of the sea water (e.g., splashing waves and water
spray by the propeller), decreasing the localised signal-to-interference ratio
(SIR) to less than −10dB.
With such a low SIR, it becomes increasingly difficult to detect the boat with
clutter suppression algorithms, even though the local disturbance of the sea
surface may be detected by a basic envelope thresholding detector. However,
there is still ample Doppler separation between the boat and interference, and
in principal a long dwell time range-Doppler detector could be constructed that
will consistently declare detections for this fast moving boat. Also of
interest is the case where the boat is racing cross-range. This still yields
strong self-induced interference, but the Doppler response of the body of the
boat will be buried within the interference and it will become extremely
difficult to detect.
High Doppler resolution spectrogram for drifting pencilduck.
High Doppler resolution spectrogram for radial outbound WaveRider
RIB.
High Doppler resolution spectrogram for pencilduck racing at
40kts radially outbound.
From this subsection it can be concluded that the
exact manoeuvre of the small boat has a great influence on its detectability,
especially due to its potential disturbance to the local sea surface. It is
also clear that not only the speed, but also the heading of small boats has to
be modelled for accurate performance prediction.
4.4. Frequency Agility Correlation for Small Boats
Theoretically the correlation coefficient ρ(f0,fn) for the RCS of a point scatterer for different
frequencies should be unity. For a target that can be approximated as a point
scatterer it is assumed that |ρ(f0,fn)|→1.
In the presence of clutter and multipath fading, the correlation will be
adversely effected. ρ(f0,fn) is plotted in Figure 22 for a CPI of 100 milliseconds for the two range cells containing most of the
energy for the chokka fishing vessel.
Frequency agility correlation for chokka fishing vessel: (a) gate 10, and (b) gate 11.
The frequency agility correlation for the boat
exceeded 0.5 with Δfc=100MHz for 84% of the total time period, compared to only 10% for sea clutter only as represented in Figure 14. There were time periods when the correlation coefficient dropped to similar
levels as the sea clutter, which coincided with low levels of SCR and/or SNR.
For a 10MHz pulse bandwidth empirical evidence suggest
that small boats exhibit significant frequency agility correlation for
frequency step sizes up to and beyond 130MHz.
It is possible to design a detection algorithm that uses ρ(f0,fn) as the basis for its test statistics. This
detector may still detect discrete spikes, since they have similar
characteristics as the small boats (e.g., at t∈(6,51)s in Figure 14). Once again, these discrete
spikes only have a limited lifetime of less than 2 seconds.
5. Detectability of Small Boats
This section
presents the detectability of small boats undergoing different manoeuvres using
the ALQ detector as an example of the asymptotically optimal class of
detectors.
5.1. Overview of the ALQ Detector
The ALQ detector
is designed by extending the generalized likelihood ratio test approach, as
suggested by Kelly [9]
for Gaussian interference, to the spherically invariant random process model
for non-Gaussian interference [19]. Assume that the radar transmits a coherent train of m pulses. The associated m received complex samples can be constructed as
a vector z=[z(1),…,z(m)]T.
Under the assumption that M is known exactly, the ALQ detector can be
expressed mathematically as|pHM−1z|2(pHM−1p)(zHM−1z)≷H0H1χt,where p is the steering vector typically constructed
with elements pi=ej2πifdT [20], T the radar PRI and fd the target Doppler frequency [11]. It is generally accepted
that M is highly dependent on the radar
configuration, geometry, and the environmental conditions and has to be
estimated from adjacent range gates that are not contaminated by the boat itself.
Various estimation techniques have been proposed [21]. Gini and Greco [11] describe one such technique
that makes a good compromise between detection losses and hardware processing
requirements:M^AML(i+1)=1K∑k=1Km⋅zkzkHzkHM^AML(i)−1zk,for i=0,1,2,…,Nit.
During each iteration the approximately maximum likelihood (AML) estimation is
normalized such that its trace is equal to m.
Since the ALQ detector involves inversion of M,
care has to be taken to ensure that the matrix does not become singular. This
can be ensured by setting the number of independent sea clutter time vectors at
different range gates k equal to at least the length of the test
vector m, k≥m [9]. Detectability can be improved by increasing this
ratio, but at the expense of increased hardware processing requirements.
5.2. ALQ Performance for Different RIB Manoeuvres
In the first dataset evaluated, the 4.2 m pencilduck
was floating close the southwestern shore of Robben Island, with Hs=3m and vwind=6ktsNE.
The radar look angle was 343∘N at range R=11km with grazing angle θ=1.5∘.
SCR and CNR were 6dB and 24dB, respectively. Figure 23(a) plots the
spectrogram with a dwell time equal to that of the ALQ detector with the
Doppler-dependent thresholding detections overlaid for PFA=10−4.
Figure 23(b) plots the test statistic χ.
Figure 23(c) plots the sliding window Pd with window length L=31,
with E{Pd}=23%.
For such a low SIR this is rather significant. The ability of the detector to
whiten the sea clutter is clear in Figure 23(b), whilst the boat signature
shows very little evidence of decorrelation. The
fading in target signature and the subsequent fading in detectability may very
well be due to shadowing of the boat by the sea
waves.
ALQ performance for drifting pencilduck: (a) spectrogram with
detections, (b) test statistic χ, and (c) sliding window Pd.
In the second dataset, the WaveRider RIB was steering
away from the radar into the well-developed waves at a speed of 10kts at range R=3.3km.
The radar look angle was 166∘N,
whilst Hs=3.2m and vwind=16ktsSSE.
SCR and CNR were 4dB and 17dB, respectively. The performance of the ALQ
detector for PFA=10−4 is plotted in Figure 24, with E{Pd}=62%.
Even though the SCR is lower than in the previous
case, a significant increase in Pd is observed. This is most probably due
to the increased Doppler
separation. Figure 24(c) indicates that low Pd is in general associated with low SCR and/or
low SNR.
ALQ performance for 10kts WaveRider RIB: (a) spectrogram with
detections, (b) test statistic χ, and (c) sliding window Pd.
With the observed increase in sensitivity with an
increase in Doppler separation, it can be expected that detectability will
increase even further for the high-speed pencilduck. The performance of the ALQ
detector is plotted in Figure 25 for the pencilduck racing at 40kts radially outbound at a range of 21.5km.
Even though the boat can be distinguished in the high Doppler resolution
spectrogram (Figure 21) the ALQ detector only manages very intermittent
detections. The sea clutter and localised disturbance are whitened over all
Doppler, effectively masking the boat. In this case the ALQ cannot be
classified as an asymptotically optimal, since a range-Doppler cell-averaging
CFAR detector can be configured to steadily detect the boat due to the
separation in Doppler of the interference and the boat signature and its narrow
Doppler spectrum.
ALQ masking of fast moving pencilduck.
5.3. Detection of Seagulls and Angels
With improved
subclutter visibility, the problem arises that first detections are declared
not only for small boats, but also for large birds such as seagulls. ALQ
detections for the last dataset in Section 5.2 are
overlaid on the range-time intensity plot in Figure 26. Even though only intermittent detections were declared for the racing
pencilduck, a large number of detections were declared for the entire dataset.
Examination of these detections reveals that they coincide with scatterers
yielding an RCS of approximately 0.01–0.1m2 and a narrow Doppler spectrum with very fast
acceleration. Comparing this to the Doppler signatures of birds in [13] and observation by the
copilot of the pencilduck led to the conclusion that these scatterers are
indeed seagulls.
Detection of seagulls and an angel with the ALQ detector.
Of particular interest in Figure 26 is the consistent
detections declared from a range of 21.9km at t=0 second closing in to a range of 21.4km.
Closer inspection revealed that this was a flock of about 6 seagulls flying in
formation. The combined RCS of this “angel” was 10dB lower than the pencilduck, but still yielded
significantly higher Pd.
The main reason for this is the Doppler separation and that the seagulls caused
no local disturbance of the sea surface. The resultant SIR for the angel was
indeed higher than for the pencilduck. As radar sensitivity is increased, this
will become a more and more significant problem.
6. Conclusions
Current
commercial products provide near real-time estimation of basic wave and surface
current parameters using the video output of standard X-band marine radar. This
paper investigated sea clutter and small boat reflectivity in the littoral and
proved that sea clutter reflectivity is related to vwind rather than Hs.
Temporal characteristics of sea clutter were investigated, with empirical
results suggesting that vwind and ϕwind can be estimated from the sea clutter speckle
autocorrelation. The spectral inhomogeneity of sea clutter was investigated for
different sea conditions. The brief analysis of the sea clutter Doppler
spectrum and I2(fd) suggested that it is possible to also infer
the existence and severity of whitecaps from the sea clutter. Discrete spikes
in sea clutter were clearly visible when the frequency agility decorrelation
was estimated.
For safe navigation, it is pertinent that the
detection capabilities of marine radar in adverse conditions are improved,
especially for small boats. This requires an in-depth understanding of the
dynamics and associated reflectivity of these boats. The absolute RCS,
amplitude statistics, and temporal characteristics of a range of small boats
have been analysed using a comprehensive set of recorded datasets. Of
particular interest were the dependency of the boat reflectivity on the local
sea, deviation from the Swerling RCS models, the perceived persistence of
reflectivity for short periods of time and the distinguishable pulse-to-pulse
frequency agility correlation properties of small boats. It was shown using
real data that the ALQ detector can, under certain conditions, be subject to
self-masking. A definite contribution to the knowledgebase is the importance of
not only modelling the sea clutter and boat reflectivity accurately, but also
to model the local disturbance caused by small boats, especially during fast
manoeuvring.
Acknowledgments
The authors acknowledge the CSIR team for their
contribution in the planning and execution of the trial; the South African
Department of Defence, the South African National Parks Board, the National Sea
Rescue Institute, and the South African Weather Services for their contribution
to the execution and funding of the measurement trials; as well as the
Armaments Corporation of South Africa and the South African Air Force for
permission to use the Fynmeet radar. The research was cofunded by the South
African Department of Science and Technology, the South African Department of
Defence and the Royal Society.
ReichertK.kreichert@oceanwaves-pacific.co.nzHessnerK.hessner@oceanwaves.deLundB.lund@oceamvaves.deHessnerK.Coastal applications of X-band radar to achieve spatial and temporal surface wave monitoringProceedings of the OCEANS Europe ConferenceJune 2007Aberdeen, Scotland1610.1109/OCEANSE.2007.4302276BriggsJ. N.200416London, UKThe Institution of Electrical EngineersIEE Radar, Sonar and Navigation SeriesNathansonF. E.1999Mendham, NJ, USASciTechReillyJ. P.DockeryG. D.Influence of evaporation ducts on radar sea return199013728088WardK. D.ToughR. J. A.WattsS.200620London, UKIETIET Radar, Sonar and Navigation SeriesHerselmanP. L.BakerC. J.Analysis of calibrated sea clutter and boat reflectivity data at C- and X-band in
South African coastal watersThe IET International Conference on Radar Systems (RADAR '07)October 2007Edinburgh, UK5Report on the investigation of the loss of the sailing yacht ouzo and her
three crew south of the isle of wight during the night of 20/21 august
2006April 20077/2007Southampton, UKMarine Accident Investigation BranchWadeB.Sharpeye: a new technology marine radarThe IET International Conference on Radar Systems (RADAR '07)October 2007Edinburgh, UKKellyE. J.An adaptive detection algorithm1986222115127GiniF.GrecoM. V.SangstonK. J.FarinaA.Coherent adaptive radar detection in non-Gaussian clutter1Proceedings of the 31st Asilomar Conference on Signals, Systems & ComputersNovember 1997Pacific Grove, Calif, USA25525910.1109/ACSSC.1997.680181GiniF.f.gini@ing.unipi.itGrecoM.Covariance matrix estimation for CFAR detection in correlated heavy tailed clutter200282121847185910.1016/S0165-1684(02)00315-8McDonaldM.mike.mcdonald@drdc-rddc.gc.caLycettS.Fast versus slow scan radar operation for coherent small target detection in sea clutter2005152642943510.1049/ip-rsn:20050003FlockW. L.GreenJ. L.The detection and identification of birds in
flight, using coherent and noncoherent radars1974626745753FarinaA.GiniF.GrecoM. V.VerrazzaniL.High resolution sea clutter data: statistical analysis of recorded live data19971443121130WardK. D.BakerC. J.WattsS.Maritime surveillance radar—part 1: radar scattering from the ocean surface199013725162AndersonF.Awarenet: persistent, ubiquitous surveillance technologies for enhanced
national securityBrochure, CSIR, Pretoria, South Africa, October 2007, http://www.csir.co.za/dpss/brochures.htmlHamannR.Wind climatology of the south western capeJanuary 2006,
http://www.windsurfingafrica.org/Articles/tabid/285/Default.aspxWattsS.Radar detection prediction in K-distributed sea clutter and
thermal noise19872314045ConteE.LongoM.Characterisation of radar clutter as a spherically
invariant random process19871342191197GiniF.Sub-optimum coherent radar detection in a mixture of K-distributed and Gaussian clutter199714413948GiniF.gini@iet.unipi.itPerformance analysis of two structured covariance matrix estimators in
compound-Gaussian clutter200080236537110.1016/S0165-1684(99)00135-8