IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. PAMI-4, NO.

2, MARCH 1982 157

[3] L. Devroye and T. J. Wagner, "Distribution-free consistency re- [9] C. J. Stone, "Consistent nonparametric regression," Ann. Statist.,
sults in nonparametric discrimination and regression function esti- vol. 8, pp. 595-645, 1977.
mation,"Ann. Statist., vol. 8, pp. 231-239, 1980.
[4] L. Devroye and G. L. Wise, "Consistency of a recursive nearest
neighbor regression function estimate," J. Multivariate Anal.,
vol. 10, pp. 539-550, 1980. RNMO Luc Devroye was born in Tienen, Belgium, on
[5] L. Gordon and R. A. Olshen, "Asymptotically efficient solutions ...
August 6,1948. He received the Ph.D. degree
to the classification problem," Ann. Statist., vol. 6, pp. 515- from the University of Texas, Austin, in 1976.
533, 1978. In 1977 he became an Assistant Professor at
[61 L. Gyorfi, "Recent results on nonparametric regression function the School of Computer Science, McGill Uni-
estimate and multiple sclassification," in Problems of Control and versity, Montreal, P.Q., Canada. He is inter-
Information Theory, 1981. ested in various applications of probability
[7] R. A. Olshen, "Comment on a paper by C. J. Stone," Ann. Statist., theory and mathematical statistics such as
vol. 5, pp. 632-633, 1977. nonparametric estimation, probabilistic algo-
[8] C. Spiegelman and J. Sacks, "Consistent window estimation in rithms, the computer generation of random
nonparametric regression," Ann. Statist., vol. 8, pp. 240-246, numbers, and the strong convergence of ran-
1980. dom processes.

A Model for Radar Images and Its Application to

Adaptive Digital Filtering of Multiplicative Noise
VICTOR S. FROST, STUDENT MEMBER, IEEE, JOSEPHINE ABBOTT STILES, STUDENT MEMBER, IEEE,
K. S. SHANMUGAN, SENIOR MEMBER, IEEE, AND JULIAN C. HOLTZMAN, MEMBER, IEEE

Abstract-Standard image processing techniques which are used to en- image restoration and enhancement techniques requires a
hance noncoherent optically produced images are not applicable to mathematical model of the imaging process. This paper pre-
radar images due to the coherent nature of the radar imaging process.
A model for the radar imaging process is derived in this paper and a sents a model for the noise in radar images and uses the model
method for smoothing noisy radar images is also presented. to develop an adaptive algorithm to smooth noisy nonstation-
The imaging model shows that the radar image is corrupted by multi- ary images.
plicative noise. The model leads to the functional form of an optimum Imaging radars, specifically the synthetic aperture radar
(minimum MSE) filter for smoothing radar images. By using locally (SAR), are beginning to make use of the digital techniques,
estimated parameter values the filter is made adaptive so that it pro-
vides minimum MSE estimates inside homogeneous areas of an image and digitally correlated SAR images are now becoming avail-
while preserving the edge structure. It is shown that the filter can be able. However, optimum techniques for digitally processing
easily implemented in the spatial domain and is computationally effi- radar images are not fully developed due to a lack of under-
cient. The performance of the adaptive filter is compared (qualitatively standing of the properties of radar images from a digital image
and quantitatively) with several standard filters using real and simulated processing perspective. Thus, there is an important need for
radar images.
developing statistical models for radar noise and for using
Index Terns-Adaptive filtering, image enhancement, minimum mean them in deriving appropriate algorithms for processing radar
square error (MMSE), multiplicative noise, radar image modeling, radar images.
image processing, speckle reduction, synthetic aperture radar (SAR). This paper presents a model and a model-based image en-
hancement technique which is specifically designed for active
I. INTRODUCTION microwave sensors utilizing coherent imaging techniques. The
LARGE number of image restoration and enhancement model portrays the observed radar image as corrupted by
techniques have been proposed in recent years, for re- multiplicative-convolved noise. That is, the desired informa-
moving a variety of degradations in recorded images of objects tion, the terrain backscatter, is multiplied by a stationary ran-
and scenes. These degradations result from the nonideal na- dom process which represents the effects of coherent fading
ture of practical imaging systems. The design of optimum [1]-[3]. The product signal is then processed (convolved)
with the point spread function of the radar system to produce
Manuscript received September 8, 1980; revised September 21, 1981. the observed image.
This work was supported by the California Institute of Technology This model can be applied to the design of digital image en-
President's Fund, NASA under Contract NAS 7-100, and the U.S. Army hancement algorithms through several approaches. The proce-
Research Office under Contract DAAG29-77-G-0075.
The authors are with the Remote Sensing Laboratory, Center for Re- dure used here was to develop a minimum mean square error
search, Inc., University of Kansas, Lawrence, KS 66045. (MMSE) filter to estimate the terrain backscatter from the

0162-8828/82/0300-0157$00.75 © 1982 IEEE

158 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. PAMI-4, NO. 2, MARCH 1982

image data. The design of an MMSE filter is predicated on the i.e., nonoverlapping subareas of the spectrum of the complex
assumption of stationarity of both the signal and the noise, image are used to form N independent images. Continuous
but the image is nonstationary on a global basis. For radar the scanning of the spectrum is also used [13], [17], [18]. The
noise can be modeled as being stationary but the signal is non- net effect is to reduce the bandwidth (degrade the resolution)
stationary since the mean backscatter changes with the type of while improving the signal-to-noise ratio S/N. Several studies
target being sensed. Therefore, it is necessary to adapt the [12], [14], [15] have shown the advantages of noncoherent
filter to changes in local properties of the terrain backscatter. processing for the interpretation of SAR imagery. However,
Analysis of the system model shows that only the local ob- there are several disadvantages to this approach: 1) the tech-
served mean and standard deviation are required to properly nique is spatially invariant and thus does not account for the
adapt the filter so that it produces the MMSE estimates in- nonstationarity of the signal; 2) the technique was developed
side homogeneous (stationary) areas while preserving edge for coherent optical processing and thus it is easily imple-
structure. mented with such a processor but is not necessarily optimal
The performance of the adaptive filter is compared with for digital processing. A technique is thus needed to take full
other filters using simulated [4] and actual Seasat-SAR digital advantage of digital image processing for radar. Several differ-
imagery. A number of qualitative and quantitative perfor- ent existing approaches have been considered.
mance measures are used for comparison and the results show Linear spatially invariant filters have been used extensively
that the adaptive filter performs better than many of the com- in digital image processing [19] -[21 ]. These techniques are
monly used filters. primarily designed for recovering "ideal" images distorted by
Even though this adaptive enhancement technique has been linear spatially invariant operations in the presence of additive
developed and evaluated for radar, it is applicable for other noise [20] . The nonstationarity of the ideal images is not ac-
types of images. Obviously, it is applicable for coherent counted for in most of these techniques and thus these meth-
speckle reduction in general, as the noise processes are similar ods tend to blur edges. Furthermore, these techniques are not
for all coherent sensors [2] . Further, the approach described suited to radar imagery because of the multiplicative nature of
in this paper can be used for developing adaptive filtering algo- the noise.
rithms for other applications in which the image is corrupted Homomorphic filtering [5] is applicable to images corrupted
by multiplicative noise. Because of its simple spatial domain by multiplicative noise. Homomorphic filtering refers to a
implementation the adaptive filtering algorithms provide an technique of preprocessing the observed image to transform
alternative to homomorphic filtering [5]. nonadditive noise into additive noise using some nonlinear
The contents of the paper are organized as follows. Section memoryless operator. Then standard techniques are applied
II contains the background of the modeling and filtering prob- for additive noise reduction. The enhanced data are then
lem. In Section III the model for noise in radar images is de- formed by applying the inverse nonlinear operator. For multi-
veloped from a digital image processing perspective. The plicative noise logarithmic and exponential operators are re-
MMSE filter is derived in Section IV and the criteria for the quired. The model for radar images (which will be derived in
adaptation of the filter based on local statistics is discussed. the next section) represents the observed data as being multi-
The utility of this adaptive processing algorithm is evaluated in plicative noise operated on by a linear system, or viewed from
Section V using both real and simulated radar images. It is another perspective, each output pixel (picture element) is a
shown qualitatively and quantitatively that the adaptive filter weighted sum (convolution) of a product. A logarithmic op-
improves the ability of both human and machine interpreters eration will not properly separate the signal from the noise for
to extract pertinent information from radar images. Recom- this case. A possible solution for this would be to deconvolve
mendations and conclusions are presented in Section VI. [20] the observed data to remove the effects of the linear sys-
tem and then to apply homomorphic filtering. Unfortunately,
IL. BACKGROUND the performance of deconvolution techniques degrades rapidly
A SAR image is constructed by processing the two-dimen- as the signal-to-noise (S/N) ratio decreases [20] (especially if
sional complex waveform received by the sensor (when stored there are zeros in the system function). Radar images already
photographically these raw data are called the signal film). have a low S/N ratio, and hence the inverse filtering does not
This processing is performed either by a coherent optical sys- produce images of acceptable quality.
tem [6] or a digital computer [7] . The output of this process- There have been some other approaches to image processing
ing is a complex image whose properties are well known [8]. for multiplicative noise. Most of these have been concerned
For a region within the sensed terrain which is homogeneous with image restoration in the presence of film-grain noise.
and which is classified as a distributed target (i.e., it is com- Multiplicative noise is treated by Walkup and Choens [22]
posed of a single target class, e.g., corn or trees, and contains where a grain noise model is presented and a Wiener filter is
many scatters no one of which predominates), the complex designed for this particular model. As in the case of linear fil-
image is statistically characterized by a narrow-band white tering, the nonstationarity of the signal characteristics was not
Gaussian random process. incorporated. Wiener filtering for multiplicative noise was also
The standard technique for enhancing SAR images is to use treated by Kondo et al. [23]. A system model similar to the
noncoherent integration [ 11 -[13] . Enhancement implies in- one developed here for radar noise is presented for film granu-
creasing the signal-to-noise ratio at the expense of spatial reso- larity. Again, a Wiener filter is derived and the ideal signal is
lution [14], [15]. Most SAR systems use frequency diversity, assumed to be stationary. Naderi and Sawchuk [24] follow
 FROST etal.: RADAR IMAGES AND DIGITAL FILTERING 159

the same basic approach of the previous papers, i.e., a model Pr (xo, yo) = observed power at position xo, Yo
for film grain noise is presented and a discrete Wiener filter is
developed. Here the nonstationarity of the signal is treated P, = E [Pr (xO, yo)] = the expected value of the ob-

and an adaptive technique is presented. This adaptive filter served power at position x0, Yo
assumes a constant form for the variance and second-order N =number of independent images averaged by
properties of the ideal image and the filter only changes rela- the sensor.
tive to the local mean. Thus, the nonstationary effects as
manifested by variations in the local mean are accounted for, The observed power can be rewritten as [28], [29]
while all second-order properties are assumed stationary. In
addition, the approach of Naderi and Sawchuk requires exten- Pr (xo, yO) =Prn(xo,
Pr(o,Y)= 2N
yo) (2)
sive computations. Each of these techniques treats multiplica- where the fluctuations of Pr(xo, yo) are governed by the ran-
tive noise; however, none were seen as being directly applica- dom variable n(xo, yo). This random variable represents the
ble to radar image data even though they do provide some fading variations. The pdf of n(xo, yo) is given by
insight as to the appropriate direction to pursue.
Nonlinear and adaptive image enhancement techniques have N-1 e n(xO,YO)]
been developed in a wide variety of forms. A simple nonlinear
Jn (n(xo, yo)) =
It ..,sNN
n (x0,_ yo) exp
, L- 2 In
operator is the median filter [19]. This filter replaces each (N- 1)!2N (55)
pixel value with the median of its neighbors. The median filter
is heuristic and is not based on any specific image model. The which is a standard chi-square distribution. Equation (2) sepa-
performance of this operator was evaluated with respect to rates the observed power into: 1) a signal independent fading
radar image processing and the results are presented in Sec- component n(xo, YO) which is stationary, non-Gaussian with
tion V of this paper. Other adaptive algorithms have been nonzero mean, and 2) a mean value component. That is, the
proposed, many of which are derived using a Kalman filter moments of n(xo,y,) are solely a function of N which is a sys-
approach. Chen [25] presented one such algorithm. The tem parameter, while the moments of Pr(xO, yo) are depen-
major disadvantages of these techniques are computational in- dent upon the desired signal parameter Pr.
efficiency and the assumption of additive noise. Equation (2) can be generalized to represent the first-order
A heuristic adaptive filter is proposed by Panda [26]. This properties of the observed power by allowing x0, yo to be-
algorithm is not applicable to radar images because of its addi- come variables, i.e.,
tive noise assumption and the criterion for filter adaptation is (x, y) n(x, y)
not suitable. However, it is interesting to note that the basic Pr (X,
Pr(XY)=
Y =
2N(4 (4)
filter shape is very similar to the one developed here. Re-
cently, Lee [27] suggested the use of local statistics to adapt Pr(x, y) represents the expected value of the observed power
image enhancement algorithms for both additive and multipli- as a function of position. If stationarity is assumed, then
cative noise. Pr(x, y), which is defined as an expected value, is a constant
for an ensemble of similar homogeneous areas. Pr(x,y) is pro-
portional to the terrain backscatter coefficient a' for that class
IIL. A MODEL FOR PROCESSING DIGITAL of targets [30].
RADAR IMAGES However, actual measurements of a single area (i.e., one sam-
The purpose of this section is to present a statistical descrip- ple function) will exhibit random variation in ao. Real terrain
tion for radar images which is suitable for use in the develop- scenes are composed of many homogeneous areas each having
ment of digital image enhancement and feature extraction different average backscatter such that field boundaries exist.
algorithms. Many sophisticated models [6] have been devel- It is reasonable to model Pr as a slowly varying function of
oped for SAR systems; however, these were derived for system position and to define a new random process r(x, y), which is
design and not for image processing. actually the desired image, to represent these variations. The
As previously discussed, the complex radar image of a homo- observed power can thus be written as
geneous area can be represented by a narrow-band Gaussian
process. The first-order statistics of the received power of Pr (x, y) = r(x, y) * n(x, y). (5)
such a signal are known to have an exponential probability Clearly, the random process Pr(x, y) is not stationary because
density function (pdf) [1]. If noncoherent averaging is per- its expected value is a function of position. However, it is rea-
formed, then the pdf for the power follows a gamma distribu- sonable to assume that n(x, y) is stationary and r(x, y) is sta-
tion [2], i.e., tionary in homogeneous areas.'
The recorded radar image I(x, y) is not simply proportional
p -l(x0, yo)exp to the observed power on a point-by-point basis. There are
several components of a SAR system, e.g., the antenna, re-
fpr(pr (xo, Yo))= e[ I/N ] (1) ceiver, and correlator which introduce a spatial correlation.
(N- 0!
IThe above discussion assumes a power display format; however, this
model is applicable for other display formats or multiplicative noise
where models. Only the form of (3) is changed.
160 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. PAMI-4, NO. 2, MARCH 1982

These components can be represented by a single linear spa- Rr(T) = ae + 2

tially invariant transfer function [6]. Therefore, the recorded (lOa)
radar image is modeled by
Sr (fJ) a2 2 -2' (lOb)
I(x,y) = [r(x,y) n(x,y)] * h(x,y)
-
(6)
where h(x, y) is the system impulse response and * denotes where the parameters a2, r2, and "a" have different values for
convolution. Equation (6) is the model of a SAR image suit- different terrain categories. The model for the multiplicative
able for use in developing digital processing algorithms. The white noise is
estimation technique presented next attempts to remove the R 25) = a6(r) + fi2 (Ila)
fading noise (the dominant source of randomness in radar
images) and generate an estimate of the "ideal image" r(x,y) Sn(f) = a2 + n 6(f) (llb)
from the observed image I(x, y). where the parameters n and n2 are sensor dependent but are
not scene dependent. Substituting the power spectral densities
IV. DEVELOPMENT OF AN ENHANCEMENT TECHNIQUE of r(t) and n(t) into (9), it can be shown that the impulse re-
BASED ON THE IMAGE MODEL sponse of the filter is given by [33]
The image enhancement technique to be developed in this
section is based on the model described in the previous sec- m'(t) = K1 ae-° Itl (12)
tion. The approach that was followed was to first derive the with
functional form of the minimum mean square error (MMSE)
filter to estimate r(x, y) from I(x, y) under the assumptions [n 1
(13)
of stationary image data. The nonstationary aspects of real Qt =
-nJI-
2a1- 1
+ F 21+a
SAR data are then incorporated by identifying the important LVJr
filter parameters which vary with position, and by determining
a method for estimating the parameter values from the ob- where K1 is a normalizing constant.
served image. The result of this analysis is an adaptive filter The MMSE filter described by (12) has some interesting
which is computationally efficient, provides the MMSE esti- properties. We have tacitly assumed that both r(t) and n(t) are
mate in homogeneous areas, and tends to preserve the edge wide-sense stationary random processes. This assumption re-
structure of the image. garding n(t) is valid over an entire radar image because both fi
The impulse response m(t) and the transfer function M(f) and a' are functions of system parameters which can be as-
of the minimum mean square error (MMSE) filter that pro- sumed to be constant. But r(t) is stationary only in individual
vides an estimate of r(t) from I(t) is obtained by minimizing homogeneous regions, and thus the filter is theoretically appli-
the mean square error e given by cable in only those areas. However, if oa is varied continuously
6 2 = E [(r(t) -JI() * M(t))2 ]
(adapted) with respect to scene conditions, then the filter will
(7) provide the MMSE estimate in all homogeneous areas.
where t = (x, y) is the spatial coordinate. The MMSE solution To evaluate its performance at edges consider two homoge-
[31] leads to a transfer function neous (stationary) areas A1 and A2 with F1 =r2 and a2 >
o2, then from (13) we find that
iSr(f) 1 for f# O
Ca1 > a12 . (14)
M(fS) r(f) * Sn(f)J H*(f)
{[nr(f (8)
This result indicates that the impulse response of the MMSE
for f =O filter for A I is narrower than the filter for A2 - If r(t) has a
large variance, then a wide impulse response would excessively
where fi = E {n(t)} and f = (f, fy) is the spatial frequency average the desired variation in backscatter. Thus, for areas
coordinate. Sr(f) and Sn(f) are the power spectral densities with a2 large the impulse response of the filter should be
of the terrain reflectivity and the noise process, respectively. narrow.
The filter given in (8) is valid for processing image data within Next consider an area A3 which contains a boundary be-
homogeneous areas inside of which r(t) can be modeled as a tween two stationary areas A1, A2. First, note that A3 is not
stationary random process. In (8) H *(f ) is the complex con- a stationary area so theoretically this MMSE filter does not
jugate of the transfer function of the system which is not data provide the minimum mean square estimate. However, the
dependent, and hence can be assumed constant over some presence of an edge will result in a large variance for r(t) in
finite bandwidth. The data dependent part of the filter is A3. That is,
M'(f), which can be written as 2 > r2 (15)
nSr (f) (9) and
Sr(f ) *Sn(f )

The standard model for r(t) is an autoregressive process with a2 >.ua2 (16)
an autocorrelation function Rr (r) and a two-sided power spec- So for an area encompassing an edge, the MMSE filter will av-
tral density Sr (f) [32] of the form erage less and therefore preserve edge structure.
FROST et al.: RADAR IMAGES AND DIGITAL FILTERING 161

If a is estimated from the observed data within some local

neighborhood of each position (x, y), the MMSE filter could Sr(Z) dZ U
= 2 + r2 (23)
then be adapted to local changes. This enhancement tech-
nique would then exhibit two very important characteristics. the desired ratio is found to be
First, it provides the MMSE estimate of r(t) in homogeneous (y2 U o2 +kb
areas. Second, it tends to preserve edge structure.
-12 =;ark~ s +b
The decay constant a is a function of all three signal param-
eters U2, r, and "a." As consecutive homogeneous areas are where
examined, au2 and r will change significantly but "a" will vary
only by a small amount. Thus, for the purposes of adapting
the filter a2 and r can be assumed to vary while "a" can be
ks = kha n +1
treated as a constant. It is left to determine what local proper- kh 2
ties of the observed image can be used to estimate (I/ar)2 , and b= (24)
thus to determine how a is to be varied.
The ratio of the observed image variance to the square of the The importance of (24) is that it shows that to estimate the
mean will be shown to provide the required information to changes in a2 /2 and thus a, only the observed mean and vari-
properly adapt the filter. The expected value of the observed ance need to be found, i.e., to properly adapt the filter the
image in a homogeneous area is given by local sample mean and variance are used. The decay constant
is now written as
I = E [I(t)] = fnFH(O). (17) a2 j2 /I2 (25)
H(O) is assumed to be unity. The variance of the observed To process the image at location (xo,yo), the parameters 2
image is defined as and qI are estimated using data from a local neighborhood
uI = E [I (t)] - I2. (18) (say a 5 X 5 window) centered at (xo,yo). The adaptive filter
performs a weighted average of data in the neighborhood of
The mean square E [2 (t)] can be found by (xO, yo), the weights being determined from the local statistics
of the data using (25) and (12). Note that the window used to
gather the local statistics can be larger than the one used to
E[V2(t)] = S1(f) df (19) perform the filtering. The results presented in this paper were
generated by continuously adapting ax to changes in the local
where SI(f) is the power spectral density of the observed statistics. These statistics were gathered using a 5 X 5 window;
image. From (6) it can be shown that also, the filter operated on a 5 X 5 neighborhood.
The idea of using local statistics as a basis for spatially vary-
SI(f) = Sr(f) * Sn(f) |H(f)12 (20)
ing image enhancement is not new [27] and its advantages,
So e.g., computational efficiency, are well known. However, the
technique presented here has a firm theoretical basis and it
exhibits several additional beneficial properties, i.e., it provides
EV[2(t)] = JH(f) 2 Sr(z) Sn(z - f) dz df the MMSE estimate in homogeneous areas and preserves edge
structure. Even though this algorithm has been developed spe-
cifically for radar images, the technique is generally applicable
Sr(Z)f_~~00 H(f)2Sn(z- f)dfdz- (21) to images corrupted by multiplicative noise and should provide
_00
an alternative to homomorphic filtering because of its simple
Substituting the specific form of Sn(f) into (21) and perform- implementation.
ing the integration the mean square value becomes V. RESULTS
r00 ,00 Radar images processed by the MMSE adaptive filter are pre-
E [V2(t)] =kha4 Sr(Z)dZ + f2 Sr(z) IH(z)I2 dz sented and evaluated in this section. This assessment is per-
formed in two ways. First, a qualitative evaluation is con-
where ducted using real sensor images (in this case digitally correlated
Seasat-A SAR data). Even though the adaptive filter was de-
veloped to minimize the mean square error which does not
kh=J H(f)2 df. (22) necessarily correlate to any improvement in interpretability
(either human or machine [40]), these results demonstrate
The dependence on the system transfer function H(f) can be that the algorithm does in fact enhance the utility of radar
removed from the rightmost term of (22) because r(t) has images for important applications, e.g., target discrimination,
been assumed to be a slowly varying function of position. The geologic analysis, and agricultural assessment. For quantitative
effect of Sr(f) thus predominates, and using the following evaluations of the image enhancement algorithm we use an
definition: edge figure of merit proposed by Pratt [1 91 .
162 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. PAMI-4, NO. 2, MARCH 1982

A. Qualitative Evaluation of the Adaptive Filter

Using Digital Seasat-A SAR Imagery
The adaptive filter derived for multiplicative noise was ap-
plied to several digitally processed synthetic aperture radar
images which were provided by the Jet Propulsion Labora-
tory. Seasat-A SAR images, some processed to signal-to-noise Aaaptive Filter
(S/N)= 1, 50 m resolution and some processed for S/N=
14.6, 25 m resolution, were operated on by the adaptive filter
as well as with several standard image noise reduction tech-
niques well known in the digital image processing field. These
filters were: 1) the median filter, 2) a 3 X 3 box filter (uni- Digitally Correlated
Synthetic Aperture
formly weighted averaging over a 3 X 3 window), and 3) a Radar Imagery
1 Look 50 m Resolution
5 X 5 box filter (uniformly weighted averaging over a 5 X 5 xi Box Filter V) tso r
window). Fig. 1. Comparison of processing algorithms on Seasat-A SAR imagery.
The first digital SAR image processed by the noise reduction
filters is a scene containing the target responses of nine (non-
equal cross section) corner reflectors used to represent point
targets for the sensor at Goldstone, California (see Goldfinger
[34] for a complete description of the test site). Fig. 1 con-
tains the original image (S/N= 1, 50 m resolution) and four
filtered versions. The digital display from which these scenes
were photographed has a 512 X 512 pixel, 256 gray level for-
mat (each square pixel represents a 36 X 36 m area). The
original image has a poor signal-to-noise ratio, which makes it
difficult to unambiguously locate the corner reflectors and to
Aaptive riiter
describe the surrounding terrain. The 3 X 3 and the 5 X 5 box
filters smooth the image without regard to neighborhood sta-
tistics, thus spreading the point target returns, although these
filters might facilitate gross scale structural analysis of the
scene. These filters perform a simplistic tradeoff of resolution uriginai mage-A
SAR I mage
for averaging. The median filter produces a visible signal-to-
noise improvement, and it does not blur the image as do the
box filters. However, the response to the point targets be-
Digitally Correlated
Radar Imagery
comes much dimmer as the high intensity pixel values are re-
placed by their respective median values. Since these target 1Look 50 m Resolution
returns are of interest to us, the median filter's performance 3x3 Box Filter 5x5 Box Filter
is inferior to the box filter's. The adaptive filter preserves the Fig. 2. Comparison of processing algorithms on Seasat-A SAR imagery
(isometric display).
small spatial extent of the point targets and the brightened
foreslopes of the terrain, and averages as does a box filter in re- Original Processed by the Adaptive MMSE Filter
gions which are homogeneous.
Focusing our attention next on the region containing the
point targets, we compare the responses of the filters in detail.
Fig. 2 is the composite of photographs of the point targets
taken from an isometric display. The original image, with
its various intensities displayed as "heights," demonstrates
the extreme noisiness of the SAR image. The point target
retention characteristics of the adaptive filter are easily seen.
Eight of the nine point targets are detectable at this level of
magnification.
A primarily agricultural Seasat-A scene (S/N= 14.6, 25 m
resolution, 17 m pixel spacing) is shown in Fig. 3 along with SEASAT-A Synthetic Aperture Radar Images
Frequency: L-Band
an adaptively filtered version of the site. The effects of averag- Resolution: 25m
Numbers of Looks: 4
ing in extensive homogeneous regions are visible, and field Fig. 3. Example of adaptive filter on agricultural features.
boundaries are shown to be well preserved in the adaptively fil-
tered version. A third Seasat-A SAR image which contains both lower left-hand corner of the photographs). The primary im-
agricultural elements and geologically interesting structures provements noticed in this comparison study are the more
was also adaptively processed. Fig. 4 shows a location in clearly defined field boundaries near the bend and greater
Tennessee through which the Holston River flows (in the definition of the ridges between the foreslopes and backslopes
FROST et al.: RADAR IMAGES AND DIGITAL FILTERING 163

Original Processed by the Adaptive MMSE Filter mean radar backscatter in dB for each side of the edge differed
by the given amount) and two signal-to-noise ratios 1 and
14.6. The resolution and dynamic range were set to approxi-
mate that of the Seasat-A SAR, i.e., a resolution of 25 m with
a pixel spacing of 17 m. The image size was 145 X 145 pixels.
Each of the images described above was filtered first and
then further processed by a Robert's gradient. The selection
of the threshold greatly affects the nature of the edge maps.
An optimum threshold (i.e., the threshold which maximizes
R) was found experimentally for each processed scene assum-
ing the 3 dB edge is a worst case. That is, for all images (either
SEASAT-A Synthetic Aperture Radar Images
Frequency: L-Band
S/N = 1 or S/N = 14.6) processed by a particular algorithm, a
Resolution: 25 m
Number of Looks: 4
single optimum threshold was applied to the 3, 6, and 9 dB
Fig. 4. Example of adaptive filter on geologic and agricultural features.
images. This treats the 3, 6, and 9 dB edges as having occurred
in the same image.
of the terrain above the bend. The adaptive processing of The edge quality results are summarized in Table I. There
radar images of this type does facilitate a more rapid analysis are several interesting observations to be made from these data.
for geologic and hydrologic purposes because strong linear First, note for the S/N = 1 case that R decreased slightly from
features are emphasized as Fig. 4 illustrates [41] . the 3 dB edge to the 6 dB edge in several cases. This occurred
because of the basic multiplicative nature of radar data. Even
B. Quantitative Evaluation of the Adaptive Filter though there is a greater edge contrast, the edge quality (based
Using an Edge Map Quality Criterion on this experiment) does not improve because the variance on
In this section the performance of the filters are quantita- one side of the edge has also increased. This is true until
tively compared using edge maps generated from enhanced enough averaging has been performed, e.g., S/N = 1, 5 X 5 box
images. This approach was selected because: 1) a quantitative and adaptive filter, and all the S/N= 14.6 images except the
figure of merit for edge maps has recently been reported [19], original image.
[351, and 2) edge maps are used as an intermediate step in Another important observation is that the Robert's gradient/
many image segmentation and classification algorithms. Five
global threshold technique does not provide adequate edge
processing algorithms were evaluated: the adaptive, median, maps for any of the S/N= 1 images and only the adaptive fil-
3 X 3 box and 5 X 5 box filters, and noncoherent integration. ter produces adequate edge maps for the S/N = 14.6 data.
To evaluate these processing algorithms the following proce- Further, note that the adaptive filter clearly provides superior
dure was used: 1) produce the input images, 2) apply the edge maps with respect to the original 3 X 3 box, and median
enhancement operators, 3) apply a Robert's gradient operator filter for both the S/N= 1 and S/N= 14.6 cases. The 5 X 5
[19] to each enhanced image, 4) threshold the gradient images, box filter does provide comparable results except for the
and 5) apply the figure of merit. S/N = 14.6, 3 dB case, where the adaptive filter produces a
An edge map is a binary image where a "1" at location (i,j) figure of merit which is 20 percent higher. This last case is
indicates that the pixel at (i,j) is on an edge, while a "0" indi- significant because it shows that the adaptive filter is less sensi-
cates the absence of an edge. The edge figure of merit used tive to the edge height as expected from its adaptive nature.
here is the one proposed by Pratt [19] Define two congruent Thus, the adaptive technique performs well on low contrast
edges; these are the ones where the most enhancement is
.

images I, and IA, representing ideal and actual edge maps of a

single step edge of height h. The ideal edge map contains NI needed. Even though the 5 X 5 box filter performs as well as
edge pixels, while the actual edge map containsNA. Further, the adaptive filter when there is sufficient edge contrast, the
define d as the perpendicular distance from an actual edge adaptive technique is still seen as superior because of resolu-
pixel to the ideal edge. The edge figure of merit is then de- tion considerations. Also, the adaptive filter operating on the
fined by S/N = 1 images produced better edge maps than were pro-
duced noncoherently averaging the S/N = 1 image to a S/N =
R= 1 NA 1 14.6, i.e., using this criterion the adaptive filter is slightly bet-
max(NA,NI) j=1 l+ d2 ter than noncoherent averaging, especially for the larger edge
where contrasts. Combining these results with the response of each
technique to the corner reflectors and a S/N versus bandwidth
B = scaling constant. analysis [33], the adaptive filter does provide superior results.
The scaling constant ,3 allows the user to weight the-penalty Results are shown for the S/N= 14.6 and 6 dB edge height
for offset localized edges (for this experiment ,B = 1, following in Fig. 5(a)-(c). The original simulated image and the four
Pratt [19]). Further details concerningR are found in [35]. enhanced images are given in Fig. 5(a). The output of the
Input images for this experiment were generated using a Robert's gradient are shown in Fig. 5(b). These gradient
radar image simulation package available at the Remote Sens- images demonstrate an important statistical property of radar
ing Laboratory [4]. Six simulated images were produced for images, i.e., any form of high pass filtering on a radar image
three different edge step heights of 3, 6, and 9 dB (i.e., the generates a noisier image with radar "like" properties, and thus
164 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. PAMI-4, NO. 2, MARCH 1982

TABLE I
RESULTS OF EDGE FIGURE OF MERIT STUDY. FIGURE OF MERIT IN
PERCENTAGE VERSUS PROCESSING ALGORITHM
Processing S/N=1 S/N=14.6
Algorithm 3dB 6dB 9dB 3dB 6dB 9dB

Original 6.0 5.5 6.4 6.3 6.2 7.0

Median 6.8 6.2 8.4 7.1 11.3 17.8
3x3 9.5 7.8 13.2 9.1 14.4 17.8
5x5 8.2 13.0 15.3 23.3 52.9 67.7
Adaptive 9.2 11.9 17.3 43.2 56.4 62.6

ta)

(b) (c)
Fig. 5. Edge experiment results for 6 dB edge height with S/N = 14.6.
(a) Input images. (b) Robert's gradient images. (c) Edge maps.

there is no benefit for edge detection. This is a direct result of VI. CONCLUSIONS
the bandpass nature of radar images [36], [37]. The resulting
edge maps for each processing scheme for this case are pre- A system model for imaging radars was developed and the
sented in Fig. 5(c). model was used to design an adaptive filter algorithm to
In summary, an experiment has been conducted to quanti- smooth noisy radar images. It was shown that the radar image
tatively evaluate the adaptive filter as applied to SAR imagery. is corrupted by multiplicative noise due to fading. The multi-
Four other processing algorithms were also evaluated as a basis plicative model was used to derive the functional form of an
for comparison. The edge map quality was used as the perfor- optimum filter for enhancement of radar images. An adaptive
mance criterion. The adaptive filter provided better edge version of the filter was designed and its performance was
maps than the 3 X 3 box and median operators for both the evaluated both quantitatively and qualitatively using simulated
S/N= 1 and S/N= 14.6 cases. The 5 X 5 box filter generated and actual radar images.
edge maps of similar quality to the adaptive filter except for The filter presented in this paper is easily implemented in
low contrast edges. However, when resolution is also consid- the spatial domain and is computationally very efficient. In
ered the adaptive filter is considered to be the best processing addition to its use for smoothing noisy radar images, the filter
algorithm of those used in this study. can also be used for processing other images that are degraded
FROST et al.: RADAR IMAGES AND DIGITAL FILTERING 165

by a multiplicative noise process. For example, the degrada- [21] K. R. Castleman, Digital Image Processing. Englewood Cliffs,
tion due to illumination changes in an optically produced NJ: Prentice-Hall, 1979.
image can be removed using a similar adaptive filter. The [22] J. F. Walkup and R. C. Choens, "Image processing in signal de-
pendent noise," Opt. Eng., vol. 13, pp. 250-266, May/June 1974.
problems associated with digital radar image processing have [23] K. Kondo, Y. Ichioka, and T. Suzuki, "Image restoration by
just begun to be addressed. The radar system model presented Wiener filtering in the presence of signal dependent noise,"Appl.
here can be used for deriving feature extraction and classifica- Opt., vol. 16, Sept. 1977.
[24] F. Naderi and A. A. Sawchuk, "Detection of low-contrast images
tion schemes for analyzing the contents of radar images. The in film-grain noise," Appl. Opt., vol. 17, pp. 2883-2891, Sept.
edge quality experiment clearly indicates a need for the devel- 15, 1978.
opment of specialized edge detection algorithms for radar. [25] C. H. Chen, "Adaptive image filtering," in Proc. Pattern Recog.
and Image Processing Conf., Chicago, IL, Aug. 1979, pp. 32-37.
Other optimization criteria for the filter design should be con- [26] D. P. Panda, "Nonlinear smoothing of picture," Comput. Graph-
sidered. For example, criteria more closely related to specific ics and Image Processing, vol. 8, pp. 259-270, Aug. 1978.
applications, either human interpretation [40] or machine [27] J. S. Lee, "Digital image enhancement and noise filtering by use
of local statistics," IEEE Trans. Pattern Anal. Machine Intell.,
analysis, should be used in the development of advanced pro- vol. PAMI-2, pp. 165-168, Mar. 1980.
cessing algorithms. [28] D. P. Meyer and H. A. Mayer, Radar Target Detection. New
With the increasing availability of digital radar imagery, digi- York: Academic, 1973.
[29] V. S. Frost, "Development of statistical models for radar image
tal image processing techniques will be called upon to improve analysis and simulation," M.S. thesis, Univ. of Kansas, Lawrence,
both the quantity and quality of the extracted information. 1978.
Techniques which incorporate the unique characteristics of [30] R. K. Moore, "Microwave remote sensors," in Remote Sensing
Manual, R. G. Reeves, Ed. Falls Church, VA: Amer. Soc. of
radar imagery, like the one presented here, can be used to de- Photogrammetry, 1975, ch. 9.
rive the optimum processing algorithm for a given application. [31] L. E. Franks, Signal Theory. Englewood Cliffs, NJ: Prentice-
Hall, 1969.
[32] A. Habibi, "Two-dimensional Bayesian estimation of images,"
REFERENCES Proc. IEEE, vol. 60, pp. 878-884, July 1972.
[33] V. S. Frost, J. A. Stiles, and J. C. Holtzman, "Radar image pro-
[1] T. F. Bush and F. T. Ulaby, "Fading characteristics of panchro- cessing," Remote Sensing Lab., Univ. of Kansas, Lawrence, Tech.
matic radar backscatter from selected agricultural targets,"IEEE Rep. TR 420-1, Jan. 1980.
Trans. Geosci. Electron., vol. GE-13, pp. 149-157, Oct. 1976. [34] A. D. Goldfinger, "SEASAT SAR processor signatures: Point tar-
[2] J. W. Goodman, "Some fundamental properties of speckle," J. gets," JHU/APL Tech. Rep. CP 078, Apr. 1980.
Opt. Soc. Amer., vol. 66, pp. 1145-1150, Nov. 1976. [35] I. E. Abdou and W. K. Pratt, "Quantitative design and evaluation
[3] -, "Noise in coherent optical processing," in Optical Informa- of enhancement/thresholding edge detectors," Proc. IEEE, vol.
tion Processing. New York: Plenum, 1976. 67, no. 5, pp. 753-763, 1979.
[4] J. C. Holtzman, V. S. Frost, J. A. Stiles, and V. H. Kaupp, "Ra- [36] F. M. Dicky, "Image oriented analysis of synthetic aperture radar
dar image simulation," IEEE Trans. Geosci. Electron., vol. GE- systems," Ph.D. dissertation, Univ. of Kansas, Lawrence, 1974.
16, pp. 296-303, Oct. 1978. [37] W. M. Brown and C. J. Palermo, Random Process, Communica-
[5] A. V. Oppenheim, R. W. Schafer, and T. G. Stockham, "Nonlin- tion and Radar. New York: McGraw-Hill, 1969.
ear filtering of multiplied and convolved signals," Proc. IEEE, [38] R. N. Bracewell, The Fourier Transform and Its Applications,
vol. 56, no. 8, 1968. 2nd ed. New York: McGraw-Hill, 1978.
[6] R. 0. Harger, Synthetic Aperture Radar Systems. New York: [39] R. E. Morden and R. L. Withman, "An analysis of SAPPHIRE
Academic, 1970. image parameters as a function of processing parameters," in
[7] J. R. Bennett and I. G. Cumming, "A digital processor for the Proc. NAECON 1979, vol. 2, 1979, pp. 818-824.
production of Seasat synthetic aperture radar imagery," inProc. [40] W. A. Pearlman, "A visual system model and a new distortion
SURGE Workshop, ESRIN, Frascati, Italy, July 1979, ESA Re- measure in the context of image processing," J. Opt. Soc. Amer.,
print SP-154. vol. 68, pp. 374-386, Mar. 1978.
[8] E. J. Kovaly, Synthetic Aperture Radar. Dedham, MA: Artech [41] M. Perry, "Correlation analysis of SEASAT-SAR in the Southern
House, 1976. Appalachians," M.S. thesis, Dep. Geology, Univ. of Kansas, Law-
[9] W. B. Davenport and W. L. Root, Random Signals and Noise. rence, Dec. 1980.
New York: McGraw-Hill, 1958.
[10] J. W. Goodman, Introduction to Fourier Optics. New York:
McGraw-Hill, 1968.
[11] W. A. Penn, "Signal fidelity in radar processing," IRE Trans. Mil.
Electron., vol. MIL-6, pp. 204-218, Apr. 1962.
[12] L. J. Porcello, "Speckle reduction in synthetic-aperture radars,"
J. Opt. Soc. Amer., vol. 66, no. 11, pp. 1305-1311, 1976.
[13] J. S. Zelenka, "Comparison of continuous and discrete mixed- Victor S. Frost (S'75) was born in Kansas City,
integrator processors," J. Opt. Soc. Amer., vol. 66, no. 11, pp. MO, on March 6, 1954. He received the B.S.
1295-1304, 1976. and M.S. degrees in electrical engineering from
[14] R. K. Moore, "Trade-off between picture element dimensions the University of Kansas, Lawrence, in 1977
and noncoherent averaging in sidelooking airborne radar," IEEE and 1978, respectively.
Trans. Aerosp. Electron. Syst., vol. AES-15, pp. 697-708, Sept. From 1974 to 1977 he was a Research Tech-
1979. nician at the University of Kansas Remote Sens-
[15] G. R. Dicaprio and J. E. Wasielewski, "Radar, image processing, ing Laboratory (RSL). From 1977 to 1978 he
and interpreter performance," Photogrammetr. Eng. and Remote was a Research Engineer at RSL engaged in
Sensing, vol. 52, pp. 1043-1048, Aug. 1976. radar simulation and modeling research with
[16] J. C. Dainty, Laser Speckle and Related Phenomena. New York: emphasis on imaging systems. Currently, he is
Springer, 1975. a Project Engineer at RSL pursuing the Ph.D. degree in electrical engi-
[17] P. Hariharan and Z. S. Hegedus, "Reduction of speckle in coher- neering, with research interests in stochastic modeling and digital pro-
ent imaging by spatial frequency sampling, Part II," Opt. Acta, cessing of radar images.
vol. 21, no. 9, 1974. Mr. Frost has received an Honors Undergraduate Research Grant from
[19] W. K. Pratt,Digital Image Processing. New York: Wiley, 1978. the University of Kansas. He was a Student Branch Chairman of the
[20] H. C. Andrews and B. R. Hunt, DigitalImageRestoration. Engle- IEEE at the University of Kansas. He is a member of Eta Kappa Nu
wood Cliffs, NJ: Prentice-Hall, 1977. and Tau Beta Pi.
166 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. PAMI-4, NO. 2, MARCH 1982

Josephine Abbott Stiles (S'75) was born in Wichita State University, Wichita, KS, where he taught and conducted
Lyons, KS, on March 12, 1952. She received research in systems theory, communication systems, and image process-
the B.S. degree in engineering physics in 1974 ing. From December 1978 to January 1980 he was a Visiting Scientist
and the M.S. degree in electrical engineering in at Bell Laboratories, Holmdel, NJ, where he worked on problems in the
1977, both from the University of Kansas, modeling and analysis of satellite communication systems. He joined
Lawrence., the University of Kansas, Lawrence, in January 1980 as an Associate
In 1976 she joined the Remote Sensing Labo- Professor of Electrical Engineering. His current interests are in the area
ratory of the University of Kansas, Lawrence, as of image processing and communication theory. He has publlshed a
a Research Assistant, at which time she was in- number of articles in the above areas and is the author of a textbook,
volved in the design and testing of small an- Digital and Analog Communication Systems (New York: Wiley, 1979).
tennas for mobile traffic control. Recently, she Dr. Shanmugan is a member of the American Society for Engineering
has been engaged in radar simulation research for the purposes of both Education and the Society for Automotive Engineers (SAE). He re-
image analysis and radar guidance. At present, she is a Research Engi- ceived the Outstanding Young Engineering Faculty Award from SAE in
neer at the Remote Sensing Laboratory and is working toward the 1979.
Ph.D. degree in electrical engineering.
Mrs. Stiles has been a Student Branch President of IEEE at the Uni-
versity of Kansas and has been actively involved in a number of univer-
sity service organizations. She is a member of Eta Kappa Nu. Julian C. Holtzman (S'63-M'66) was born in
Staten Island, NY, on August 14, 1935. He re-
ceived the B.S.E.E. degree from the Polytechnic
Institute of New York, Brooklyn, NY, in 1958,
the M.S. degree from the University of Califor-
K. S. Shanmugan (M'70-SM'75) was born in nia, Los Angeles, in 1962, and the Ph.D. degree
India on January 6, 1943. He received the froim Cornell University, Ithaca, NY, in 1967.
B.E. degree from Madras University, Madras, From 1958 to 1962 he worked for the
India, in 1964, the M.E. degree from the Indian Hughes Aircraft Company in the areas of radar,
Institute of Science, Bangalore, India, in 1966, antennas, and communications satellites; he
and the Ph.D. degree from Oklahoma State Uni- also served as a Consultant to Lockheed Missiles
versity, Stillwater, in 1970, all in electrical and Space Company and Applied Technology. After teaching at San
engineering. Jose State University, San Jose, CA, from 1965 to 1969, he joined the
From 1970 to 1973 he worked as a Post- faculty of the University of Kansas, Lawrence, where he presently is a
Doctoral Fellow at Oklahoma State University Professor, Staff Member of the Remote Sensing Laboratory, and Chair-
and the University of Kansas Center for Re- man of the Department of Electrical Engineering. His current fields of
search, where he worked on problems in pattern recognition, image pro- research encompass tele- and data communications, radar systems,
cessing, and modeling and analysis of communication systems. From simulation, image processing, and analysis.
1973 to 1978 he was with the Department of Electrical Engineering at Dr. Holtzman is a member of Eta Kappa Nu and Sigma Xi.