You are on page 1of 12

Laser Radar Detection Statistics:

A Comparison of Coherent and Direct Detection Receivers


Philip Gatt and Sammy W. Henderson
Coherent Technologies, Inc.
ABSTRACT
Detection statistics for a coherent laser radar are substantially different from those of a direct detection laser radar. Direct

detection ladar detection statistics vary depending upon the detection mode (i.e., photon-counting vs. continuous direct
detection). Speckle noise also impacts the detection statistics. For a single-pixel single-frequency single-polarization
coherent detection transceiver, speckle noise can only be suppressed through temporal averaging. Some degree of speckle
averaging can also be achieved in coherent detection systems by using a multiple frequencies or dual polarizations. In
addition to these, a direct detection receiver can exploit spatial diversity to suppress the effects of speckle. This paper
develops theory useful for describing the performance of these three receiver architectures against diffuse and glint targets
and provides example performance comparisons. We show that a photon-counting direct detection receiver can, in principle,
provide superior performance, however practical limitations of current detector technology particularly in the near IR spectral

region reduces the performance margin and for many applications a coherent detection receiver provides superior
performance.

Keywords: Laser radar detection statistics

1.

Introduction

There are many applications, which require precision measurements of laser radar signal intensity. Some of these include,

range detection for target detection and ranging, differential absorption (DIAL) lidar for trace gas detection, differential
scattering (DISC) for particle size discrimination, and polarametric imaging for discrimination of manmade and natural
targets. In this paper we develop intensity detection statistics theory for both coherent and direct detection receivers and
compare and contrast their performance in terms of the minimum number of accumulated photoelectrons required to achieve
a given performance level. We consider both diffuse and glint targets and performance in terms of the number of temporal
averages. Targets with glint characteristics are often encountered in the mid to far IR. Glint targets represent the diffuse
target signatures under high levels of speckle diversity (spatial, temporal, polarization and/or frequency). The effect of
spatial, frequency and polarization speckle diversity is not specifically addressed by this work. However, a direct analogy
between these types of diversity and the number of independent pulses averaged (temporal diversity) exists and is discussed
in Section 4. This work ignores the impact of refractive turbulence, which can produces slightly higher intensity fluctuations
than those produced by a diffuse target.
The literature contains numerous publications covering focused aspects of this generalized topic. For example, Goodman's
work published by Bachman1 derives the intensity statistics for a heterodyne and photon-counting laser radar sensors for
diffuse and glint targets with single pulse averaging. In another work2, Goodman takes into account the effect of aperture
averaging of speckles for a photon-counting direct detection receiver. Youmans3 has published works on the performance of
an avalanche photodiode direct detection receiver with single pulse averaging assuming a diffuse target and aperture
averaging of speckle. Many others have contributed to the field but none provide a complete comparison of the three receiver
(coherent, continuous direct and photon-counting direct detection) architectures described herein. This work provides, for the
first time in the published literature, a unified presentation of the material, covering all three receiver types for both diffuse
and glint targets as a function of the number of temporal averages. Relative performance is compared in Section 3 and
conclusions are drawn in Section 4.

2.

Receiver Theory

In this section we develop the theory which describes the performance of the three intensity receiver architectures. The
performance is characterized by the probability of detection as a function of the signal strength, the false alarm probability,
the number of search bins, and the level of averaging or accumulation.

2.1.

Coherent Detection Intensity Statistics

A block diagram of the coherent detection intensity receiver is shown in Figure 1.


Laser Radar Technology and Applications VI, Gary W. Kamerman, Editor,
Proceedings of SPIE Vol. 4377 (2001) 2001 SPIE 0277-786X/01/$15.00

251

vs+ vn

is
PS

+ii;r:o1

Figure 1 . Coherent detection intensity processor block diagram. The detected photocurrent is narrowband filtered using a

matched filter bank (or FFT processor) then passed through a square modulus operator to obtain a signal proportional to the
intensity of the return field. This signal is then averaged and compared to a prescribed threshold level. The threshold level is
adjusted to achieve a desired probability offalse alarm.

In this receiver, the signal return strength is characterized by the received optical power Ps. This return signal is mixed with
a continuous wave optical local oscillator beam with power PLO. A non-multiplying (i.e., p-i-n) detector senses the intensity
of the combined beams. The detector photocurrent consists of four terms.

1d (t) = RPLO +RP5 (t) +2RJ7PLoP5 (t) cos(cnt + 0(t))+ i, (t).

(1)

In this expression, ci is the carrier frequency, 0(t) is the signal phase and R is the detector responsivity
(R = iiqq/hV), 11q is the detector quantum efficiency, q is the electron charge, h is Planck's constant and V is the optical
frequency. The heterodyne mixing efficiency, is the efficiency with which the local oscillator and signal fields interfere on

the detector. The first two terms are direct detection terms. They are baseband currents, which are proportional to the LO
and signal power or intensity. The third term is an intermediate frequency (IF) photocurrent, corresponding to the mixing of
the local oscillator field and the signal field. This term is the coherent detection signal current, which contains the target
amplitude and phase information. The last term corresponds to the noise current. In a direct detection receiver the signal
intensity is linearly related to the baseband photocurrent (2nd term in Eq. 1). In a coherent detection receiver, the signal
intensity is related to the square of the IF photocurrent envelope.

The coherent detection photocurrent is obtained by AC coupling the total current. This signal contains the IF signal
photocurrent and a noise photocurrent

i(t) = i (t) + n (t)

(2)

i(t) = 2R.J FLOPS (t) cos(ot + 0(t)).

(3)

Where the signal current is given by

In a properly designed coherent detection receiver the dominant noise source is the shot-noise generated by the localoscillator beam. This noise is well modeled as a zero-mean Gaussian random process whose variance is given by

Y =c(ifl(t))=2qRPLOB.

(4)

Where, <.> is the expected value operator andB is the receiver bandwidth. One of the more important performance metrics
of a coherent detection receiver is the carrier-to-noise power ratio (CNR) given by.

CNR = (is (t)2

)/( (t)2)

= Y'flqPS / h'B.

(5)

For a matched filter receiver PIB is the energy of the return signal, E, and E5/hv is the number of detector-plane signal
photons. Therefore, the matched filter CNR is the number coherently detected signal photoelectrons (K = y11qEsIhV). If we
define the CNR as the number of signal photoelectrons divided by the effective number of noise electrons, we see that the
effective number of noise electrons in an integration time ('c = 1/B) is unity.
252

Proc. SPIE Vol. 4377

2.1.1. Noise Model and Probability of False Alarm


The coherent detection photocurrent, i(t) is a narrowband random process. As such, it is can be modeled by its analytic signal

representation i(t) (i.e. a complex photocurrent).

i(t) =

(t) +i,. (t))exp(jot) .

(6)

Where i is and i are the complex amplitudes ofthe signal and noise photocurrents andj= 'NJ-i . In this formalism, the noise
can be written as a complex sum of inphase and quadrature noise components (n, and nq), each a zero mean Gaussian random
process with variance (i.e., power) equal to
(7)

in (t) = fl1 (t) + Jflq (t).


Therefore, the joint statistics of the noise currents are circular complex Gaussian
1

2it
2

exp(z +q )/2).

(8)

Where p(i,q) is the joint probability density function (PDF) of the two noise components. For a coherent detection receiver,

it is the square magnitude of the photocurrent that is proportional to the optical field intensity, (J = I2). Therefore the
intensity noise PDF is an exponential density function4.

(9)

b = 22 is the total noise power. When averaging is included in the receiver, the noise statistics change from
exponential to gamma (or Chi22N) provided the averaged signals are independent5. For averaging as defined in the block
diagram of Figure 1 (i.e., sum/N) the appropriate form of the Gamma density function is
Where

p (I) = F(N) '(NI b)N 1N -l exp( NI/b)

(10)

Where N is the number of signals averaged and F(z) is the gamma function. The mean and variance of the gamma density
function, as defined above, are

(I)_b and v[I]_((I_(I))2)b2/N.

(11)

The probability of false alarm depends on the threshold level and the noise statistics. For Gamma distributed noise, it is
given by

PFA =

p (I)dI = ]T(N,NIth /2)/F(N).

(12)

Where 'th is the threshold level, set to achieve a specified PFA, and F(z,x) is the incomplete gamma function defined by

F(a,x) = je_tta_ldt.

(13)

For a single pulse averaging the false alarm probability reduces to PFA = exp(-Ih/22).

2.1.2. Target Detection Statistics


The probability of detection, PD, is the probability that the detected signal (signal plus noise) exceeds the threshold level

defined by the specified PFA. The primary factors, which impact the probability of detection, are the PFA, the CNR, the
number of independent signals averaged and the target model. In this section we develop expressions for the PD for a fully
diffuse (i.e., rough) and perfectly specular (i.e., smooth or glint) target.
Diffuse Target

For a diffuse target, the complex signal photocurrent is, because of laser speckle, a circular complex Gaussian process5.
Therefore, the joint PDF of the signal is given by Eq. 8 where the variance of each quadrature component is Y2 and the total
Thus the signal plus noise is also a circular complex Gaussian random process. The
variance (i.e. signal power) is
variance of each component is the sum of the noise and signal variances. Hence the intensity PDF of the signal plus noise,
Proc. SPIE Vol. 4377

253

Psn('), is a gamma distribution (c.f. Eq. 10) with a mean, b = (2r2 + 2o,2) or b = 2o,2 (CNR +1) since the
CNR = 22/2o2.
For an intensity threshold detector, the probability of detection is defined as

PD=J7pSfl(I)dI

(14)

PD = F(N,NIth I2i(1+ CNR))/F(N).

(15)

The detection probability is, from Eq. 12, given by

For single pulse averaging the probability of detection reduces PD = exp(-Ih/2fl2(1+CNR)) and the following simple relation
applies

PD=PFA1.

(16)

Glint Target

For glint targets, the signal photocurrent complex amplitude, A, is assumed to be a constant and is arbitrarily chosen to have
zero phase. Therefore, the photocurrent complex amplitude is described by

i(t)A+fl,(t)+Jflq(t).

(17)

Where the quadrature noise components are the same as they were for the diffuse target case and thus the statistics of the
intensity noise and the probability of false alarm remain unchanged. However, the signal plus noise statistics are different.
For single pulse averaging, a Rician-Square or noncentral Chi22 density function6 applies.

p5(I) =(1/2)exp(_(A2 +I)I2)Io(A-/7R).

(18)

Where as before, 2 is the noise power of each quadrature component. The density function of the averaged intensity is
known to follow the noncentral chi2N2 density function6. The correct form of this density function, which applies to an
average rather than accumulation, is given by

= (2& )(I/ A2)_2 exp(_N(A2 + I)/2 )IN_l(NA/)

(19)

Its mean and variance are given by

(i)= A2 +2 2(CNR+1) and var[I]= 4o(A2 +)/N.

(20)

Since, for a constant amplitude signal, CNR = A2/2o2. This form of the noncentral chi2N2 distribution is directly related to
the form often found in the literature (ch2)

p(x) =(1/2)(xIX)_24exp(_(x+)I2))Iv,2_l()

(21)

Where v is the number of degrees of freedom and X is the noncentral parameter. It can be shown that these two distributions

are the same with the following substitutions, v = 2N, ?= A2N/o2 and x = I NI2. The form presented in Eq. 19 is
intuitively advantageous since it provides direct access to the constituents to the random process (i.e., signal amplitude and
noise power and the number of averages). The probability of detection is, from Eq. 14, given by

PD = f

) (22)

(x)dx = J(1/2)(x/ )(v_2) '4exp ((x + X) I2))I ,2_1()dx = QN (,

Xth

Xth

Where QN(a,13) is the generalized Marcum Q-function7'8 defined as

QN(0c,1 Jz(zi a)/V_l exp(_(z2 +a2)/2)IN_l(otz)dz

254

Proc. SPIE Vol. 4377

(23)

Eq 23 follows from Eq. 22 with the substitutions v =2N, x = z2,

7t =

a2 and Xth = P2 Returning back to first form of the


noncentral chi2N2 distribution (i.e., Eq. 19), the probability ofdetection is given by

PD= QM(INA2 2 ,INIth n) QM(I2NCNRdNIth n)

(24)

The Marcum Q-function can be evaluated using the series expansion published by Robertson9.

Continuous Direct Detection Intensity Statistics

2.2.

A block diagram ofthe continuous direct detection intensity receiver is shown inFigure 2.

PS

IN
7 j::1

linear APD

detector

'

Figure 2. Continuous direct detection intensity processor block diagram. The detected photocurrent is lowpass filtered to remove

excessive noise bandwidth, then averaged, and compared to a threshold level set to produce a specified PFA.

The adjective "continuous" is used to distinguish this receiver from a "photon-counting" direct detection receiver. In this
receiver, the detector is either a p-i-n or a linear-gain detector like a conventional avalanche photo-diode with photoelectron
gain M. The signal photocurrent is therefore given by

i(t)=RP(t)f.

(25)

This photocurrent is a baseband random process, which is directly proportional to the optical field intensity. In this type of

receiver the dominant noise term is typically preamplifier thermal noise (referred to its input as an equivalent noise
photocurrent i(t)). This thermal noise is modeled as a zero-mean Gaussian random process whose variance is given by

=NB.

(26)

Where N0 is the "white" noise spectral density (Amp 2IHz). For high levels of signal power, (i.e. short range) signal shot
noise can dominate preamplifier thermal noise, however for this to be the case the number of signal photoelectrons must be
large and the noise is still well approximated by a zero mean Gaussian random process, with a larger N0.
Continuous direct detection receiver performance is characterized by its signal-to-noise ratio (SNR), which is defined as

SNR =

(i5(t))2/(i (t)2 ) = (RP(t)M)2/N0B.

(27)

The notation SNR is used here rather than CNR because the signal photocurrent in direct detection receiver is a baseband
signal and there is no carrier frequencyper Se. The numerator of the SNR is defined as the square of the mean signal, unlike
the coherent detection CNR whose numerator is defined as the mean of the squared signal, because for the direct detection
receiver the signal is not a zero mean process.

2.2.1.

Noise Model and Probability of False Alarm

The PDF of the averaged noise is Gaussian because the sum of a series of Gaussian random variables remains Gaussian.

However the noise power (variance) is reduced by a factor of N.

(I) =

(28)

exp(12N/

J2ico/N
For the general case of a non-zero mean Gaussian random process, the probability that x > x0 is given by

Pr(x > x0)=

exp((xu)2 /22) dx =

J2ic

ec((xo

_u)I)/2
Proc. SPIE Vol. 4377

(29)

255

Where u is the mean and is the standard deviation and erfc is the complementary error function. Therefore, for zero mean
Gaussian noise, the probability of a false alarm is given by
(30)

PFA= Jp(I)dI =erfc(IthTN/2Y )12.


2.2.2.

Detection Statistics

Diffuse Target

For the case of single pulse averaging and unit speckle diversity, the density function of the signal intensity follows an
exponential density function with mean <I> = b (c.f. Eq 9). Hence, the signal power is b2, the SNR is b2/o2 and 1> =
iSNR. The impact of higher levels of speckle diversity will be addressed in Section 4. The corresponding density
function for the signal plus noise can be evaluated from the convolution integral'0. That is, the signal plus noise PDF is the
convolution the Gaussian noise PDF with the exponential signal PDF. The resultant PDF can be shown to be given by

Psn (I) = exp(( 2b1) I2b2)erfc[(

bI)/-/i1x]/2b.

(31)

This expression reduces to the exponential distribution when the noise variance is zero. The averaged intensity, again
assuming unit speckle diversity, is Gamma distributed (c.f., Eq 10) with mean b and variance of b2/N. The density function
of the averaged signal plus noise can be evaluated from the convolution integral. This results in the following expression

(I) = 2_+1 2(b/ N)_N exp(N12 I 2

)(N /N)"32

N 1 l(RY J_2
Cj)lFl 2'2'2t b
- 2(2N_l)/2(2

(32)

_bI)) lFl[;;![ ]2 }(br(N))

Where, ,F,(a;b;x) is the Hypergeometric function. From a numerical standpoint, this expression is unstable for large values
of the noise variance o,2 or small values of the mean signal current or equivalently for SNRs <-15 dB. This is because
numerical evaluation of the Hypergeometric function, under these conditions, fails to provide sufficient accuracy to yield
reliable results. Because of this numerical stability problem, an alternate technique was developed to evaluate the probability
density function. This numerical technique employed the Fast Fourier Transform to invert the PDF's characteristic function
(i.e. Fourier Transform of the PDF), which is given by

= exp(w2 /2N)
(1+j&o/N)

(33)

The probability of detection was then calculated using numerical integration techniques. The mean and variance of the
averaged signal plus noise can be calculated from the characteristic function but are more readily calculated through a
moment analysis, which produces

(I) =

b--j SNR and var[IJ= (b2 +)IN = (SNR+1) IN

(34)

Thus, the mean of the signal plus noise intensity is equal to the mean signal intensity, which is proportional to ISNR.
Glint Target

For a glint target (or for high levels of speckle diversity with high receiver SNR), the signal intensity probability density
function is a delta function located at JSNR. The noise probability density function is Gaussian with a variance equal to
o2/N. Therefore, the PDF of the signal plus noise is a Gaussian distribution with a mean of JSNR and a variance of o2/N.
The probability of detection is evaluated by direct application of Eq. 29.

256

Proc. SPIE Vol. 4377

(35)

PD=erfc((Ith SNR)JN/2))/2.
Therefore for

a glint target, the impact ofaveraging is equivalent to an increase in the SNR by a factor ofN, since for a fixed
PFA, 'th scales by 1/\JN. That is to say, the PD obtained for a given PFA, SNR and N is the identical to the single pulse
averaging PD when the SNR is increased by a factor of N. This conclusion is demonstrated by the performance curves
presented in Section 3.

2.3.

Photon-counting Direct Detection Intensity Statistics

The structure of the photon-counting direct detection intensity receiver is shown in Figure 3.

n(t) _________

_________

Figure 3. Photon-counting direct detection intensity processor block diagram. The detected photocurrent triggers a high speed
gated counter or alternatively an accumulated charge is sampled by the gated counter and converted to a discrete count level. The
counts are then accumulated over N counting intervals and compared to a prescribed threshold level.

An extremely high-gain detector capable of detecting single photon events is assumed. This detector is followed by a gated

counter, which reports the number of events (counts) generated during a counting interval (range gate). Counts from
successive lidar pulse returns are accumulated. The accumulated count is then compared to a threshold level, adjusted to
produce a specified probability of false alarm.
For a photon-counting direct detection receiver, the number of signal counts is given by

ksTIqEs/hV.

(36)

Where E is the signal energy or integrated signal power, which is directly proportional to the optical field intensity. Added
to the signal are noise counts, k. The additive noise counts arise from several sources, which include background light and
dark current. The mean number of additive noise counts are not of direct concern because it can be characterized and
subtracted from the number of detected counts. It is the fluctuation of the noise and signal counts that results in receiver
noise. This noise is called shot-noise and is assumed to be Poisson distributed. An interesting property of Poisson random
variable is that its variance is equal to its mean. Therefore, we define the SNR for a photon-counting direct detection receiver
as

SNR =

K /(K + Ku).

(37)

Where K and K are the mean number of signal and dark counts accumulated in the counting interval. When K =0, the
receiver is signal photon-limited and, like a coherent detection receiver, the SNR is proportional to the number of signal
photoelectrons (SNR Ks). When K >> K, the receiver is dominated by dark noise and, like the continuous direct detection
receiver, the SNR is proportional to the signal power squared (SNR K2/K).

2.3.1. Noise Model and Probability of False Alarm


For a photon-counting receiver the noise signal (i.e. the number of noise photoelectron counts in an integration time) is a
discrete random variable. The distribution that describes the number of photoelectron counts k, in a given time interval for a
signal with a stationary mean, is the Poisson Distribution' .

p,. (k) =

K exp(K )/ k!

(38)

Where, K is the mean number of noise counts in the integration interval. The variance of a Poisson process is equal to its
mean. Like a Gaussian process, the Poisson distribution is preserved under accumulation. That is to say that the sum of
independent Poisson processes results in a random variable, which is also Poisson. The mean of the accumulated process is
the sum of the constituent means or simply NK for the sum of N identically distributed Poisson Processes. Here the symbol
N represents the order of accumulation rather than the number of averages. The distinction is that an averaged quantity is the
Proc. SPIE Vol. 4377

257

accumulated quantity divided by the order of accumulation N. Thus the PDF of the accumulated noise in a photon-counting
receiver is given by

p(k) = (NK) exp(NK)/k! .

(39)

The probability that the Poisson-distributed noise exceeds a threshold represents the PFA. This probability is given by
00

PFA =

(k) = 1
k=Kth

Kth-l

p (k) = 1 T(Kth , NK,. ) /(Kth

1)!

(40)

k=0

When the threshold is set to its minimum (i.e., Kth = 1) this probability reduces to

(41)

PFA=1exp(NK).

Because of the discrete nature of this counting receiver's threshold level, precise PFAs can not be realized. For example
when NK = 0.1 the PFA is [0.095, 0.0047, and 0.00015] for Kth = [0, 1, and 2].

2.3.2.

Detection Statistics

Diffuse Target

The probability density function for a discrete random process can be calculated from its continuous counterpart using the
Poisson Transform6.

p(k)=f00(xkexp(_x)/k!)p(x)dx.

(42)

Where, p(x) is the continuous density function. The Poisson transform of a continuous distribution is the distribution of a
discrete Poisson random variable whose mean, x, is conditioned on the statistics of the continuous distribution p(x). For
example for a glint target p(x) is a delta function and the discrete PDF is Poisson. For a diffuse target, the averaged signal
intensity follows a gamma distribution (Eq. 10) and the resultant discrete signal PDF is known as the negative-binomial
distribution

p(k) (N

+k

1)Kk I(K5 + 1)N+k

(43)

Where K is the mean number of signal counts in a single accumulation interval. The mean of this distribution is NK and its
variance is NKS(KS +1). If the order of accumulation is unity the negative binomial distribution reduces to the geometric or
Bose-Einstein distribution.

The PDF of the signal plus noise can be calculated from any one of a number of techniques (e.g. a convolution of the noise
PDF with the signal PDF, or by inverse transforming the product of their characteristic functions, or by the Poisson transform
integral using the continuous counterpart forp(x)).

Goodman2 derives the appropriate expression using the conditional probability integral. In his expression he assumes a
spatial diversity (averaging) parameter M, whereas here we assume a temporal accumulation N. Therefore, we obtain our
and
expression by substituting N for Goodman's M parameter, and NK and NK for his
parameters. Using these
substitutions we arrive at

pSn(k)=

exp(NK)
(K +1)N(N 1)!

(NK)3 (k+Nf1)!( K5 k-J


(kf)! K5 +1)

The probability of detection is just the infinite summation of the discrete density function starting from Kth. Which is equal to
1 minus the finite sum from zero to the 1-Kth.
00

Kth-l

kKth

k=0

PD = p (k) =1

(k).

When the threshold is a minimum (i.e., NK <PFA such that Kth = 1), the detection probability reduces to
258

Proc. SPIE Vol. 4377

(45)

PD = 1 exp( NK ) /(K + 1)N

(46)

Furthermore, when NK << 1 the detection probability reduces to

PD = 1 (K + 1)_N

(47)

For very large N this expression is equivalent to the threshold crossing probability for a Poisson random process. That is

PD = 1 exp(NK ) .

(48)

Therefore, for large N (i.e., little speckle noise) and low dark noise (NK << 1) the number of accumulated photoelectrons
required to achieve a given probability of detection is given by

NK = Jii(1 PD) .

(49)

Glint Target

When the target is a glint target, the signal is Poisson distributed with mean NK. Therefore, the density function for the
signal plus noise is also Poisson with mean and variance given by N(K + K) because the sum of two Poisson processes is
still a Poisson process. Therefore, the probability of detection is given by Eq. 40 using N(K+K) for the mean.

3.

Performance Comparison

In this section we present performance plots in terms of the probability of detection as a function of the receiver CNR (SNR
for direct detection), the probability of false alarm and the number of signals averaged (accumulated for a photon-counting
direct detection receiver). The total required number of accumulated photoelectrons to achieve a specified PFA and PD is
also calculated and presented as a function of N.
Diffuse and glint target performance curves for a coherent detection receiver and a false alarm probability of 0.1 % are shown
in Figure 4. These curves demonstrate the benefit of averaging. For example to obtain a 90% PD with a 0.1% PFA a CNR of
18 dB is required. However only 4 and 3 dB is required for 10 and 100 averages. Therefore a gain of 14 and 21 dBis
obtained for 10 and 100 pulse averaging. The diffuse target performance converges to the glint target performance as the
number of averages tends towards infinity. This is expected because the square-mean normalized signal intensity variances is
1i'IN. At low CNR (CNR < 0 dB), the diffuse target performance is slightly better than the glint target due to diffuse target
signal surges (anti-fades).
99.9

--io
30

-50

100

S--- 10

' I
s

100

L F/I
I7 t

'i

I
-20

-10

CNRdB

iir4 i:
: /I
::
ii ill

30
..-' 50

/ /

TT1

_______

NAg

-- 5

l.;.

uonerent Detection 'aunt: Log(i'ia) = -i.u

Coherent Detection: Loq(Pta) = -3.0


urJJuI
111111111, flIIIII uIIuIuJ.II
NMg
I
! '
13
I
! ' !

9c

'/

SC

,!

I
10

20

Sc
2C

-20

'--f------l ,

-10

10

20

CNRdB

Figure 4. Coherent detection performance for a diffuse (left) and glint (right) target. Probability of detection as a function of the
single-pulse CNR, parametric in the number of independent averages for a PFA = i03. The heavy dashed curve is a direct
calculation of Eq 16. For high levels of averaging the performance against a diffuse target approaches that of the glint target.

Similar data is shown for the continuous direct detection receiver in Figure 5. For a 0.1% PFA, a direct detection receiver
with SNR equal to the coherent detection CNR, performs better than a coherent detection receiver when the CNR is less than
Proc. SPIE Vol. 4377

259

approximately 4 dB. For a glint-target, as the level of averaging increases the performance curves shift to lower SNRs
according to SNRIN. Additionally, the diffuse target performance converges to glint target performance of high levels of
averaging.
Continuous Direct Detection: Log(Pla) =
99.9

uuiiiiiiiiii,- ir
NAvg

9C

0
S

0S
0
.0
.0

3.

-3.0

I I I I I I I

I
1

::

/I

L__L__

6C

/___i-_____

:
-f

,'

I'

:,

,-------

7(

"/

I'

8C

I/
I I
-4i--/

3C

I I I

:.::
C

irr t

/L

// I' if:

I
/i/ II,IAI:IIIIIIIIIII
.

2C

10

2O

-10

II

10

20

SNR(dB)

Figure 5. Same as Figure 4 except for continuous direct detection. For the glint target, the effect of averaging is an equivalent to
an increase in the single-pulse SNR by a factor equal to the number of pulses averaged. For example, 0 dB performance at 10
pulse averaging is the same as 1 0 dB performance for single pulse averaging.

Similar plots for a photon-counting direct detection receiver are shown in Figure 6. These curves demonstrate the superior
performance of a photon-counting direct detection receiver. Because the threshold level is discrete, each curve corresponds
to a different false alarm probability, which is less than the prescribed values of 0.1%. This explains the peculiar horizontal
spacing between the curves as the order of accumulation is increased.
i'noton uounting Du: Log(I'la; <-i.u

99

Photon Counting DD Glint: Log(Pla) <-3.0


IIlIIIIII IIIuI uIIIrIII

9
C

}
2C

Os

f /7
I // - ////:'I i
I I I'
/I/ i I 'Ii'f//
./

4-I--

Ol..L...J_._IIJ..IIIJ I
-10
-20

/'

-/--i--

7C

N Kth LagiPtal

1 3 .3.31

3 4 .350

5 5 .37

--

106-3.23
0 11 -3.53
5014.3.16

./

50
40
30

i' /

__

/-

00 22 .3.16

1111111 111111111 111111111


10

t I ______
.
/11:,
..__/
'
I Kth LogiPtal
20 --410
-.-. 1055.3.7
/ /f,y---6 .323
I
I
nc I__/
3011.3.53
// / i 1 /
50 14 .3.16
00 22 .3.16
I, /
01 V _____ iI.III_1I#IIIII IllillIlt 111111111
8C

."

'I

_f-I--ac

/; I

Lt_; -i-i

20

-20

-10

10

20

SNR = Ka2/(K&#K (dB)

SNR = K2I(K+K,) (dB)

Figure 6. Same as Figure 4 except for a photon-counting direct detection parametric in the order of accumulation, N, with the
mean number of noise electrons equal to 0.1. The threshold level, for which the PFA is less than the prescribed PFA, and the
corresponding actual PFA is indicated on the plot.

A useful figure of merit for comparing the three receiver models is the total number of accumulated photoelectrons Wpe)
required to achieve a given detection probability. For a matched filter coherent receiver (E = Ps/B), this figure of merit is
given by

Nped = NCNR = NYTIqPS/hVB.


260

Proc. SPIE Vol. 4377

(50)

Whereas, for a matched filter continuous direct detection receiver (E =P/2B) the number ofphotoelectrons required is

Npedd

N ISNRNo
2qlqMV B

(51)

This result demonstrates the well know result the non-photon-limited performance improves with shorter laser pulse widths
(i.e., increased bandwidth). For a photon-counting direct detection receiver we have

Npedd NK .

(52)

Comparison plots of this figure of merit are presented in Figure 7 for both diffuse and glint targets. For the coherent
detection receiver, the FOM is minimized for levels of averaging between 5 and 1 0. The curves suggest that a total of 26
photoelectrons distributed over 6 independent returns (-4 photoelectrons/shot) achieves the desired performance (90% PD
with a 0.1 % PFA). At lower levels of averaging speckle noise mandates a higher number of photoelectrons. At higher
levels of averaging performance is degraded because the noise standard deviation, and hence the threshold level to achieve a
given PFA, decreases at a rate ofonly 1I'JN. In this regime, the FOM asymptotically converges to a square-root dependence
on the number of pulses averaged. For a glint target, there is no speckle noise and optimal performance is achieved with no
averaging. Here the single-shot number ofrequired photoelectrons is approximately 12.
For the continuous direct detection receiver, the performance is substantially worse than a coherent detection receiver. Here
298 (86 for a glint target) photoelectrons are required to achieve the desired detection statistics. This analysis assumed a 2
ns pulsewidth ('-250 MHz BW) a near-JR (i.e., InGaAs) avalanche photodiode receiver in with an APD gain of 10, and an
optimistic preamplifier noise spectral density of 1 pA/\JHz. The analysis ignored the effects of signal shot-noise and APD
excess noise, which would degrade performance even further.
The photon-counting performance can be superior to both coherent detection and continuous direct detection. Photon-noise
limited (K 0) performance requires only 2.3 accumulated distributed across many ('100 or more) returns. When K is as
high as 0.1, only 9.4 photoelectrons are required over 7 returns. When K is approximately equal to 1, the photon-counting
performance is almost the same as a coherent detection receiver. This results from the fact that in a coherent detection
receiver the equivalent number of dark counts, due to shot-noise, is unity. The jagged nature of the non-photon limited (K >
0) receiver is a result of the discrete threshold levels required to achieve a given PFA.

---------

10,

Total Number of Photoelectrons


ml:__ Coh Det, 25.2,8
Log(Pa3C)

CohDet,11.9,1

.Y..CDD,2.1,4
- - - DDLog(Nn)'=-If,2.3, 100

to

COD, 88.2,1
- - - DD
Log(Nn) =-Int,23,1

-. - OD Loq(Nn) =-1 ., 9.4,7


I

DO Nn) .00, 20.3,4

Total Number of Photoelectrons


Log(PFa.40O

DD LoNn)=-1.00, 5.2,1
--.DDLg(Nn) =0.00,82, 1

CI

t00

o
o

0
0

10

10I
-

io0
1

-, -

CDD Pamnieters:
BW: 250 MHz, APD GaIn: 10, NolseDen: 1.00 pAirt(Hz)

10
Navg

io
I

100

--

GOPararae4erw
BW: 250 MHz, APD GaIn: 10, NolseDen: 1.00 pAirt(Hz)

10

100

Navg

Figure 7. Total number of accumulated photoelectrons required to achieve a 90% PD with a 0.1 % PFA versus the number of
averages for a diffuse (left) and glint (right) target. The curves conespond to a coherent detection receiver (solid), a continuous
direct detection receiver (dot) and three photon-counting receivers with varying number of dark-counts per sample interval (1
dot-dot-dot-dash), (0.1 dot-dash) and (0 dash). The classical direct detection receiver assumes 250 MHz BW an APD gain of 10,
and an optimistic preamplifier noise spectral density of 1 pAt'JHz.

4.

Conclusions and Summary

Direct detection receivers can more easily exploit speckle diversity (aperture, frequency, and polarization averaging). The

effective level of diversity is analogous to pulse averaging except for the fact that there is no reduction in noise variance.
Approximate performance for a given level of speckle diversity can be estimated from the plots presented in this paper, by
Proc. SPIE Vol. 4377

261

rescaling the SNR axis appropriately to account for the lack of decreased noise power with averaging. For example, the
performance of a direct detection receiver with a speckle diversity of ten is approximately the same as that shown in the
curves with the ten pulse averaging and an SNR ten times larger. For very high levels of speckle diversity the performance
approaches that of a glint target. For a coherent detection receiver, it is difficult to obtain speckle diversity levels greater than
unity unless array receivers or multiple polarizations or transmitted frequencies are employed'2.
The results presented above suggest that the low-noise photon-counting receivers provide superior performance over coherent
detection receivers, in terms of the number of required photoelectrons. In general this is true, however the actual energy-

aperture product required to achieve equal or greater performance may be significantly higher because of poor quantum
efficiency of practical detectors. In the UV to visible regime, the quantum efficiency of PMTs is typically on the order of
10%. In the near-to far IR PMT efficiency is substantially worse. New "photon-counting" devices for near IR operation are
emerging. However, to date no commercial devices are available with substantial quantum efficiencies. Research grade
Geiger-mode APDs have problems with low quantum efficiency, after-pulsing and relatively long reset times associated with
quenching. Thus before a conclusion can be drawn for a specific application, a detailed analysis needs to be conducted.
This analysis needs to consider all aspects of the sensor including "effective" aperture diameter, energy, detector performance
and availability, pulse width, background light, etc.
5.

References

Bachman, C. G., Laser Radar Systems and Techniques, Chapter 2, Artech, Wayland, Mass 1979

2 Goodman, J. W., "Some Effects of Target-Induced Scintillation on Optical Radar Performance," Proceedings of the IEEE,

53,111965

'

Youmans, D. "Avalanche Photodiode detection statistics for a direct detection laser radar," Proc. SPIE, 1633. 1992
Goodman, J. W., Statistical Optics, Chapter 2, Wiley and Sons, New York, 1985

Goodman, J. W., "Statistical Properties of Laser Speckle Patterns," Chapter 2 of Laser Speckle and Related Phenomena,
Springer-Verlag, New York, 1984
6 B. Saleh, Photoelectron Statistics, Table 2.1 with ji = 2o2/N and jt = A2, Springer-Verlag, New York, 1978
Whalen, Anthony D., Detection of signals in Noise, Chapter 8, Academic Press, Inc., 1971

8 Marcum, J. I. and P. Swerling, "A statistical theory of target detection by pulsed radar", IRE Transactions on Information
Theory, April, 1960
Robertson, G. H., "Computation of the noncentral chi-square distribution", Bell Syst Tech Journal 48, No 1, 201-207, 1969

' The PDF of a random process that is the sum of two independent random processes is the convolution of the constituent
PDFs.

' Papoulis, A., Probability, Random Variables, and Stochastic Processes, McGraw-Hill, New York, 1984
12 Gatt, P. et. al. "Coherent optical array receivers for the mitigation of atmospheric turbulence and speckle effects" Applied
Optics, 35 No. 30, 5999-6009, 1995....

262

Proc. SPIE Vol. 4377

You might also like