You are on page 1of 11

Modulation-transfer-function analysis for sampled image

systems

Stephen K. Park, Robert Schowengerdt, and Mary-Anne Kaczynski

Sampling generally causes the response of a digital imaging system to be locally shift-variant and not directly
amenable to MTF analysis. However, this paper demonstrates that a meaningful system response can be
calculated by averaging over an ensemble of point-source system inputs to yield an MTF which accounts for
the combined effects of image formation, sampling, and image reconstruction. As an illustration, the MTF
of the Landsat MSS system is analyzed to reveal an average effective IFOV which is significantly larger than
the commonly accepted value, particularly in the along-track direction where undersampling contributes
markedly to an MTF reduction and resultant increase in image blur.

1. Introduction the location of the point source relative to the sampling


It is widely recognized that the modulation transfer (i.e., pixel) grid. This sample-scene phase phenomenon
function (MTF) of an imaging system is of fundamental is familiar to anyone who has tried to analyze system
importance in both the initial specification and design performance by observing the appearance of small
of the system and in any subsequent detailed analysis high-contrast objects in a digital image. For such sys-
of the images it produces. Because the MTF is the tems the basic question then is how can the MTF ap-
modulus of the system optical transfer function (OTF), proach to system performance analysis be rigorously
which is, in turn, the Fourier transform of the system modified to account for the particular kind of shift
point-spread function (PSF), the MTF of an imaging variance that sampling introduces? This paper pro-
system can be determined, at least in principle, by vides an answer to that question.
generating the modulus of the Fourier transform of the In this paper, the combined effects of imaging, sam-
image of a point source. Implicit in all of this is the pling, and reconstruction are analyzed in terms of the
mathematical assumption that the system is both linear spatial frequency response of a general sampled imaging
and shift-invariant, i.e., that the location (and strength) system. The (linear) system model used is 2-D and
of the point source can be chosen arbitrarily. If the sufficiently general so that the results of this MTF
shift-invariant (isoplanatic) assumption is not valid, the analysis apply to most line-scan and sensor-array
image of a point source will depend on its location, and imaging systems, particularly those used for remote
the system MTF will have a spurious location-depen- sensing. Sampling effects are accounted for by using
dent phase dependence. a point-source system input randomly located relative
Unfortunately, the MTF approach to system per- to the sampling grid and by averagingthe system output
formance analysis is not directly applicable to many over all possible point-source locations. This random
contemporary sampled imaging systems (e.g., elec- sample-scene phase approach has been used success-
trooptical line-scan and sensor-array devices) because fully by the authors recently in related studies of the
sampling causes these systems to have a particular kind radiometric error in remote-sensing imagery introduced
of local shift variance. That is, whenever these systems by the combined effects of sampling and reconstruc-
form the (sampled) image of a point source, the ap- tion.1,2
pearance of the (reconstructed) image will depend on A similar sample-scene phase approach to system
analysis was used recently by Wittenstein et al. 3
However, their approach is distinctly different from
ours in that it is nonrandom and is restricted to a class
of systems for which sample-scene phase dependence
is negligible. Our results, as well as those of Ref. 3, are
Robert Schowengerdt is with University of Arizona, Office of Arid based on the basic research of many others. 4-11 In
Lands Studies, Tucson, Arizona 185721;the other authors are with particular, the paper by Peterson and Middleton4 and
NASA Langley Research Center, Hampton, Virginia 23665. the survey articles by Schade5 and Legault6 have been
Received 22 December 1983. most influential. However, we believe that our sto-

2572 APPLIED OPTICS/ Vol. 23, No. 15 / 1 August 1984


chastic approach to MTF analysis is unique to this f(x,y) Imaging glXy) Sampling g5 (x.y) Reconstruction
trX*YI
paper and that our results are new. O- usystem - s usteSape
Scene *h X. yl Image srix
subsystem
y
-Reconstructed
The results of this paper are illustrated by performing image image
an MTF analysis of the Landsat multispectral scanner Fig. 1. General sampled image system.
system (MSS).12 This example was chosen because the
MSS is a familiar and important (line-scan) imaging
system which has been in operation more than ten years,
and yet a comprehensive system analysis of its fre-
quency response, including the combined effects of
sampling and reconstruction, has never been published.
We show that (1) in the along-track direction, the MSS
response is seriously degraded by the undersampling
inherent to a system design in which consecutive in-
stantaneous fields of view (IFOVs) are nonoverlapping;
(2) in the along-scan direction, undersampling is not a
problem, instead the MSS response is dominated by the
scanning aperture and electronic filter; (3) the MSS
response is somewhat asymmetric with superior per-
formance in the along-scan direction; and (4) the com-
bined effects of the image-forming optics, scanning
aperture, electronic filter, 2-D sampling, and bilinear
image reconstruction, produce an effective IFOV
(EIFOV) which is 104 X 148 m.
II. Formulation
A general sampled imaging system is illustrated in
Fig. 1. All the quantities in this figure are referenced
b)
to a common orthogonal spatial coordinate system (xy)
normalized so that the sampling interval in both di- Fig. 2. (a) Illustration that the sampling process is not shift-in-
rections is unity. Because of this normalizing conven- variant. A (binary) image g(x,y) with the sampling grid and sampled
tion, when the system in Fig. 1 is analyzed in the Fourier imagegs(x,y) superimposed. The sampled image is b(xy). (b) The
(binary) image g(x - u, y - v) shifted relative to the sampling grid;
domain, the associated spatial frequencies (v) will the sampled image is not (x - u, y - v).
have units of cycles/sample interval, and the Nyquist
(or folding) frequency will be 0.5.
The input scene (radiance distribution) is f(x,y) and rectangular grid. Typically, the image is both sampled
the PSF of the (presampling) imaging subsystem is and quantized for digital transmission or storage prior
h(xy). The imaging subsystem is assumed to be linear to image reconstruction. In this paper the effects of
and shift invariant; thus its output image g(xy) is the quantization are ignored.
spatial convolution of the input scene and the subsys- Because of the continuous input/discrete output
tem PSF. The PSF of the imaging subsystem must nature of the sampling subsystem, it is not shift-in-
account for all the system processing prior to sampling. variant. For example, Fig. 2 depicts a simple binary
Thus, in the analysis of an electrooptical scanner like (i.e., two-level) input image and demonstrates that
the Landsat MSS, for example, h (xy) accounts for the continuous subpixel shifts in the position of the input
combined effects of the image-forming optics, scanning image can produce significant discrete changes in the
aperture, and electronic filter. In general, the effect of output sampled image. This shift-variant nature of
the imaging subsystem is to low-pass filter the spatial image sampling has significant system implications; it
frequency components of the input scene, producing an means that the sampling subsystem does not have a
output image which is a smoothed (i.e., blurred) copy transfer function, and so the transfer functions for each
of the scene. of the three subsystems in Fig. 1 cannot simply be cas-
Image sampling is the processing which occurs be- caded as one would do in a traditional system MTF
tween the continuous output of the imaging subsystem analysis.
and the discrete input to the reconstruction subsystem. The final output of the system in Fig. 1 is a continu-
Image sampling is accomplished mathematically by ous reconstructed image gr (xy) which is formed by the
multiplying the output of the imaging subsystem by the spatial convolution of the discrete sampled image
2-D comb function g' (X,y) with the reconstruction (i.e., interpolation)
function r(xy). The reconstruction function can be
comb(x,y) = 6(x-my - n), (1) thought of as the (continuous) output response corre-
m n sponding to a (discrete) reconstruction subsystem input
where the notation indicates that the double sum is over which is one at the origin of the sampling grid and zero
all (positive, zero, and negative) integers m, n, and it is elsewhere. Because the reconstruction subsystem is
assumed that sampling occurs on an equally spaced completely characterized by its PSF, in an analysis of

1 August 1984 / Vol. 23, No. 15 / APPLIED OPTICS 2573


(a) (c)

(b) (d)

Fig. 3. Illustration of the combined effects of imaging, sampling, and reconstruction. The targets in (a) (upper left) are identical as are their
images in (b) (upper right). Image (c) (lower left) is a reconstruction of a sampled version of (b). Image (d) (lower right) is the difference
of (b) and (c); it illustrates the shift-variant image degradation associated with sampling and reconstruction.

a digital image rectification system, for example,' 3 shift-variant image degradation introduced by the
r(x,y) must account for the combined effects of all sampling and reconstruction subsystems. Two effects
postsampling operations such as resampling and dis- are evident in Figs. 3(c) and (d): (1) the extent of the
play. image degradation introduced by sampling and recon-
Figure 3 is a digital simulation of the system in Fig. struction is very much a function of sample-scene phase;
1 which provides an illustration of the combined effects and (2) on the average, the effect of sampling and re-
of imaging, sampling, and reconstruction. In this construction is to further blur (low-pass filter) the
simulation the size of the targets, the effective size of the output of the imaging subsystem The intent of this
imaging subsystem PSF, and the spacing (i.e., pixel size) paper is to rigorously develop an approach to MTF
of the sampling grid are all equal. Figure 3(a) is a scene analysis which will account for the type of shift variance
consisting of identical (square) targets randomly located illustrated in Fig. 3 and correctly characterize on the
on a uniformly black background, and Fig. 3(b) is the average the net loss of resolution produced by the
imageof this scene prior to sampling and reconstruction. combined effects of imaging, sampling, and recon-
Because the imaging subsystem is shift-invariant, each struction.
object in Fig. 3(b) is blurred identically independent of The output of the system in Fig. 1 is given by
its location. Figure 3(c) is the reconstructed image
formed by first sampling the image in Fig. 3(b) and then gr(x,y) = ilf(x,y)*h(xy)J comb(x,y)J*r(x,y), (2)
reconstructing the sampled image using bilinear inter- where * denotes spatial convolution. This equation is
polation. Figure 3(d) is the difference between Figs. the basis for all the analysis that follows, and, conse-
3(b) and (c), i.e., Fig. 3(d) illustrates the additional quently, the results of this paper are applicable to the

2574 APPLIED OPTICS/ Vol. 23, No. 15 / 1 August 1984


Il(x,yl r(x,y)
MTF analysis of any sampled imaging system whose
performance is characterized by Eq. (2). This equation
is also consistent with the system model used by Wit-
tenstein et al. ,3 except that their model only included
a 1-D sampling subsystem.
From Eq. (2), the system response to an impulsive
input scene 5(x - u,y - v) is the
(i.e., point-source)
system PSF (SPSF) given by (a) (b)
SPSF(x,y;u,v) = [h(x - uy - v) comb(xy)]*r(x,y). (3) ASPSF

As the notation indicates, the system is shift-variant,


and the SPSF is not simply a function of the difference
(x - uy - v). As in Refs. 1 and 2, the parameters u and
v which locate the point-source input relative to the
sampling grid are the sample-scene phase parameters.
Because the sampling grid is periodic, it is sufficient to
restrict these parameters to the range 0 < u < 1, 0 < v
< 1, i.e., to restrict the point sourceto the sampling grid
cell defined by the vertices (0,0), (0,1), (1,1), and (1,0). (c)
Furthermore, it is natural to assume that both u and v Fig. 4. Average system PSF (c) corresponding to an idealized system
are uniformly distributed random variables so that the with a square-imaging subsystem PSF (a) and a bilinear image re-
point source is equally likely to lie at any position within construction PSF (b). All three PSFs are scaled relative to a common
the grid cell. Equation (3) describes the reconstructed sampling grid.
image of the point source for each possible value of (u,v)
and accounts for the combined effects of imaging, (0.5,0.5); (2) sample the result of step (1) at all integer
sampling, and reconstruction. This stochastic ap- values of (xy); (3) reconstruct the discrete result of step
proach simulates the process of imaging a scene which (2) by convolving with r(x,y).
is an ensemble of point sources randomly located in the Figure 4 illustrates the ASPSF for an idealized system
(x,y) plane and pairwise separated by at least several in which the imaging subsystem PSF [Fig. 4(a)] is a
sampling cells to avoid response overlap. An average square aperture function (i.e., the detector aperture
over all the point-source reconstructions yields a sta- dominates the response of the imaging subsystem), and
tistical estimate of the average system point-spread image reconstruction [Fig. 4(b)] is bilinear interpo-
function (ASPSF) which includes the combined effects lation. In the situation illustrated, the IFOV of the
of imaging, sampling, and reconstruction. imaging subsystem is matched to the sampling grid so
From Eq. (3) and the assumption that the sample- that consecutive sampled IFOVs are just contiguous, a
scene phase parameters are uniformly distributed, the common design for some line-scan systems and most
ASPSF is given by sensor array systems. As illustrated in this example,
the ASPSF [Fig.4(c)] is considerably broader than ei-
ASPSF(x,y) = fS JS SPSF(x,y;u,v)dudv. (4)
ther the imaging subsystem's PSF (whose width is de-
Since the system in Fig. 1 is linear, as is the averaging termined by the IFOV) or the reconstruction subsys-
process in Eq. (4), and since the ensemble of impulsive tem's PSF (whosewidth is determined by the sampling
inputs are equally likely to lie anywhere within a sam- grid). This broadening is caused not only by the inev-
pling grid cell, it follows that itable blurring introduced by the imaging and recon-
struction subsystems but also by the combined effects
ASPSF(x,y) = {Frect (x- 2' Y -2) *h(xy)] of sampling and sample-scene phase averaging.
Although considerable insight into system response
can be gained from Eq. (5), this equation is more easily
X comb(xy)l *r(x,y). (5)
analyzed and understood in the frequency domain. In
where Sec. III of this paper we derive the Fourier transform
of Eq. (5), thereby using the concept of sample-scene
rect 1)=(1 O<x<1, 0 < y <1 (6) phase averaging to provide an MTF analysis of the
~ - 2, t0 elsewhere.
(x
rec system in Fig. 1.
Equation (5) is fundamental in that it represents the
Ill. Fourier Analysis
average PSF of the entire system and, therefore, ac-
counts for the combined effects of imaging, sampling, We define the system optical transfer function
and reconstruction. Provided the PSF of the imaging (SOTF) as the Fourier transform of the system PSF
subsystem, h(x,y), and the PSF of the reconstruction [Eq. (3)], i.e.,
subsystem, r(x,y), are known in a common coordinate = S
SOTF(A,v;u,v) ¢JSPSF(x,y;u,v)
system normalized to the sampling grid, the average X exp[-2iri(,ux + vy)]dxdy, (7)
system PSF can be generated as follows: (1) convolve
h(x,y) with rect(xy) and center the result at (xy) = where the spatial frequencies ,,v have units of cycles/

1 August 1984 / Vol. 23, No. 15 / APPLIED OPTICS 2575


Imaging-sampling 1.0
I nput imaging I Reconstruction Response
I_Ima-ingo- Sampling ---- a

I I i j i)l
i(p.
Fig. 5. General sampled image system is amenable to MTF analysis
provided the imaging and sampling subsystems are combined to form
an imaging-sampling subsystem whose sample-scene phase-averaged
MTF is Ii(! t ,v)I.

"
sample interval. From the convolution theorem and M1F
the transform properties of the comb functions it fol- .5 1.0
lows that
SOTF(Mv;u,v) = P(,,v) L E exp -22ri[u(u - m)
m n

+ (v- n)]&lh(- m,v-n), \\\ \ tV)


(8)
where Ih and P are the Fourier transforms of h and r, \\1V-
respectively, i.e., i (u,v) is the OTF of the imaging sub- (-
(V 1) sic ( - 1) \\
system and P(,,v) is the reconstruction filter.
Just as the average system PSF [Eq. (4)] is the sam-
ple-scene phase-averaged system PSF, we define the
average system OTF (ASOTF) as the sample-scene -1.0'
phase-averaged system OTF, i.e., Fig. 6. Comparison of the image subsystem OTF 4(v) and the av-
erage MTF of the imaging-sampling subsystem It(v)l. Below the
ASOTF(,u,v) = ' Jo STF(ku,v;uv)dudv. (9) Nyquist frequency, 0.5 cycles/sample interval, the combined effects
Equation (9) can be reduced by combining Eqs. (8) and of imaging,sampling, and sample-scene phase averaging reduce the
MTF from (v) l to ji(v) j. Above the Nyquist frequency, spectrum
(9) and making use of the identity replication causes a spurious MTF increase.
0'exp(-2iruti)du = exp(-rti) sinc(Q), (10a)
where as usual nificance; it is a consequence of the coordinate con-
iismr ~(10b)
sinc(Q) -
ventions used in this paper which yield an ASPSF [see
Fig. 4(c)] centered at (xy) = (0.5,0.5)rather than at the
The result is the expression origin.1
In practice, the modulus of the system frequency re-
ASOTF(p,v) = P(!t,v) E L sponse is the quantity of primary interest. Therefore,
m n we define the sample-scene phase-averaged systems
X exp[-7r( - m + v - n)i sinc( - m) MTF (ASMTF) as the modulus of the average system
X sinc( - n) ( - m, v - n), (11) OTF [Eq. (12a)], i.e.,
which characterizes completely the sample-scene ASMTF(,4,v) = Ii(,u,v)jjP(,,v)j. (13)
phase-averaged system OTF in terms of the imaging Equation (13) is particularly significant because it re-
subsystem OTF and reconstruction filter. It can also veals that a sampled imaging system is amenable to
be verified directly that the Fourier transform of Eq. (5) MTF analysis provided the imaging and sampling
is Eq. (11). subsystems are combined into a single imaging-sam-
It is important to recognize that Eq. (11) can be pling subsystem whose sample-scene phase-averaged
written as MTF is It (y,v)I as illustrated in Fig. 5. This MTF is
ASOTF(uv) = i(,uvW(,e,v) exp[-r(,u + v)i], (12a) cascaded with the MTF of the reconstruction subsystem
to form the MTF of the entire system.
where
IV. Imaging-Sampling Subsystem MTF
= (-l)-+n sinc( - m)
m n Because the coordinate system in this paper is nor-
X sinc( - n) ( - m,v - n)
(12b)
malized so that both sampling intervals have unit length
(and so the Nyquist frequency is 0.5), the MTF of the
accounts for the combined effects of imaging, sampling, imaging-sampling subsystem Ii(yu,v) is determined by
and sample-scene phase averaging. As Eq. (12) reveals, the OTF of the imaging subsystem 1(4,v). In an un-
a frequency response analysis of the system in Fig. 1 normalized coordinate system, I t (yu,v)I would also de-
consists of cascading t (,uv) with the reconstruction filter pend explicitly on the size of both sampling intervals.
P(u,v) to yield the average system OTF. The phase Figure 6 illustrates (in 1-D)an imaging subsystem OTF
term exp[-7r( + )i] in Eq. (12a) is of no practical sig- and the corresponding average MTF of the imaging-
2576 APPLIED OPTICS/ Vol. 23, No. 15 / 1 August 1984
1.0 define the effective sampling passband as 0 < v < 0.5,
where 0.5 is the Nyquist frequency. For the OTFs il-
lustrated in Fig. 7, the effective passband of the imaging
subsystem 0 < v < vi and the effective passband of the
sampling subsystem are matched when v, = 0.5. As
OFF .5 - 0.2 0.3 0.4 0.5 0.6 0.7 vc =0.8
Figs. 7(b) and (c) illustrate, the average imaging-sam-
rl(v) pling subsystem MTF is relatively insensitive to changes
in the optical subsystem OTF if v, > 0.5. In general,
Ii (,v) I must be evaluated using Eq. (2b); this is par-
ticularly true when vP 0.5. However, it is instructive
to consider the two cases vc < 0.5 [Fig. 7(b)] and vc > 0.5
0 .5 1.0
[Fig. 7(c)], for in each case an approximation to Eq.
v (12b) can be derived which provides insight into the
-cos (lr )
MTF of the imaging-sampling process.
1.0 It is intuitive that, if vPis significantly <0.5, the fre-
quency response of the imaging subsystem will domi-
nate the frequency response of the composite imag-
ing-sampling subsystem, i.e.,
MIF it(,uv) -; h (,'v) (14)
Clvi for u < 0.5, v < 0.5. In this case it can be seen that Eq.
(12b) reduces to Eq. (14), because at low frequencies the
sideband terms [correspondingto (m,n) is (0,0)]can be
ignored and sinc(A) sinc(v) 1. The dashed lines in
Fig. 7(b) illustrate (in 1-D) the validity of Eq. (14) for
0 .5 0 .5 two values of vc. If v, = 0.2, Eq. (14) is nearly exact, but
v v if v, is much larger than 0.2, Eq. (14) is not valid except
Fig. 7. Comparison of the optical subsystem OTF (a) with the as- at very low frequencies. Of course, for any value of v,
sociated imaging-sampling subsystem average MTF [(b) and (c)]. Eq. (14) is not valid at frequencies above 0.5 because of
The dashed lines indicate in (b) the approximation lI(v)l (v)I for the periodicity of I (/Lv) I.
v = 0.2, 0.3 and in (c) the approximation I (v) I I cos(rv) . In (b) It is also intuitive that, if v, is significantly >0.5, the
and (c) only the frequency band 0 S v 0.5 is illustrated; at higher frequency response of the composite imaging-sampling
frequencies I (v) is periodic as indicated in Fig. 6. subsystem will be dominated by what can be thought
of as the average effective MTF of the sampling sub-
sampling subsystem. As this figure illustrates, Itk(Mv)I system. This response can be determined from Eq.
and h(,i,v)I are nearly identical only at very low (12b) by first recognizing that an equivalent represen-
frequencies; at higher frequencies they differ signifi- tation for t(g,v) is
cantly because of three effects: (1) The infinite series 1 1
i(,u) = exp[7r(,u + v)il Z Z[rect(x - /2,y - 12 )*h(X,y)]mn
in Eq. (12b) represents frequency spectrum replication,
which is the inevitable consequence of periodic sam- X exp[-2r(mp + nv)i], (15)
pling. 4 Because of this replication, I (pt,v)I is periodic,
while Ih (/,,v) I is not, resulting in a spurious undesirable where [rect(x - 1
/2,y - 2)*h(xy)]mn denotes the value
MTF increase at frequencies beyond 0.5. (2) If, as il- of the convolution at x = m,y = n. The equivalence of
lustrated, the effective cutoff for Ih (i,,v) I is typically not Eqs. (12b) and (15) can be verified by taking the inverse
sharp, and some significant image subsystem response Fourier transform of each. If vPis significantly larger
is present at frequencies above 0.5, adjacent terms in Eq. than the Nyquist frequency 0.5, the size of the imaging
(12b) will overlap. Since these adjacent terms have subsystem PSF will be so small that [rect(x - 1 /2,y -
opposite signs, cancellation will occur causing significant 1/2)*h(xY)]mn will be effectively nonzero only at (m,n)
modulation suppression, particularly at frequencies = (0,0), (0,1), (1,0), and (1,1). In addition, if h(x,y) is
close to 0.5. This suppression is a mixed blessing: symmetric in x and y, these four nonzero values will be
above the Nyquist frequency it is desirable; below the equal with a common value of 0.25. Thus Eq. (15) will
Nyquist frequency it is not. (3) The sinc terms in Eq. reduce to
(12b) are the result of the combined effects of sampling
and sample-scene phase averaging; these terms are the i(uv) 1/4exp[7r(,u + v)i][1 + exp(-27rii)][1 + exp(-27rvi)],
primary source of frequency suppression at interme- (16a)
diate frequencies.
To what extent does It(u,v)J depend on h(,uv)? which reduces further to
Figure 7 illustrates this dependence (in 1-D) for a family i(yv) ~cos(7rb) cosOrv). (16b)
of imaging subsystem OTFs [Fig. 7(a)] with effective
cutoff frequencies v, [i.e., h(v) = 0.5] between 0.2 and Equation (16b) is an approximation to Eq. (12b), which
0.8. In the discussion that follows it is convenient to is valid provided vi is significantly larger than 0.5.

1 August 1984 / Vol. 23, No. 15 / APPLIED OPTICS 2577


1.0 1.0

MIF .5 ASMIF .5

0 .5 1.0
.5 1.0
v
v
Fig. 9. Comparison of the average system MTF for two recon-
1.0 struction filters: linear and parametric cubic convolution (PCC) with
a = -0.5. For comparison, the ASMTF associated with ideal re-
construction [r(x) = sinc(x)] is also shown. Below the Nyquist fre-
quency (0.5), PCC is clearly superior to linear interpolation.

MIF 5 Figure 8(b) illustrates the average system MTF (in


1-D) which is the product of I(Eu,v)i and IP(/,,v)I.The
nonideal response of the reconstruction filter leads to
two distinctly different effects: (1) within the sampling
passband the nonideal (z 1) response has the net effect
of reducing system MTF and creating additional blur
0 .5 1.0 in the reconstructed image; (2) beyond the sampling
Fig. 8. (a) Comparison of the average imaging-sampling subsystem passband the nonideal (iz 0) response fails to suppress
MTF Ii(v)I and the reconstruction system MTF I (v)I corresponding any spurious system response created by sampling and
(in 2-D) to bilinear interpolation. (b) Resulting average system MTF, out-of-band imaging subsystem response. It is this
which is the product of I(v)I and IP(v)I. latter effect which produces aliasing in, the recon-
structed image if the scene has significant energy at
frequencies outside the sampling passband.
Figure 7(c) illustrates the accuracy of this approxima- In a recent article,1 5 the authors discussed a new al-
tion; for vP= 0.8 the approximation is nearly exact. gorithm, parametric cubic convolution (PCC), which
Again, it should be emphasized that, in the situation can be implemented efficiently and yet contains a pa-
when the effective passbands of the imaging and sam- rameter a which can be adjusted to tailor the MTF of
pling subsystems are nearly matched, neither Eq. (14) the reconstruction filter. For a particular choice of this
nor Eq. (16b) is a particularly good approximation. parameter, a = -0.5, the MTF is nearly flat at low
Instead, Eq. (12b), or equivalently Eq. (15), should be frequencies and effectively zero for frequencies beyond
used to accurately calculate the average frequency re- 1.0. Figure 9 illustrates the ASMTF improvement
sponse of the imaging-sampling subsystem. which is possible if PCC (with a = -0.5) is used in place
V. Reconstruction Subsystem MTF
of bilinear interpolation for image reconstruction. The
superior ASMTF achieved with PCC has been shown
It is well-known that ideal image reconstruction is to produce reconstructed images with less radiometric
accomplished with a low-pass filter whose response is error than those produced with bilinear interpolation
one within the sampling passband and zero elsewhere. and standard bicubic convolution.2
However, the reconstruction functions commonly used
in practice are spatially limited, and at best their re- VI. Example: Landsat MSS
sponse is only a crude approximation to the ideal. That The results of this paper are applicable to any line-
is, they do not have either a flat response at low scan or sensor-array imaging system provided it is
frequencies or a sharp frequency cutoff, and they have characterized by Eq. (2). As an illustrative example,
a significant response at frequencies above 0.5. This these results will be applied to the Landsat multispec-
is illustrated in Fig. 8 for a reconstruction filter which tral scanner system (MSS). This example was chosen
corresponds (in 1-D) to bilinear interpolation. Al- because the MSS is a familiar and important line-scan
though there are other practical reconstruction filters imaging system whose multispectral images have been
with somewhat better frequency characteristics,1 5 the used in a multitude of remote-sensing applications, and
use of bilinear interpolation in this illustration is ap- yet, to date, a comprehensive system analysis of the
propriate because it is a standard reconstruction tech- MSS frequency response has not been published.
nique which enjoys widespread use in remote-sensing Fortunately, the paper by Slater1 6 and the four re-
applications where reconstructed images are routinely sponses that followed by Colvocoresses 17
Thompson, 18
resampled onto different coordinate grids. Slater,19 and Friedmann20 provided all the information
2578 APPLIED OPTICS/ Vol. 23, No. 15 / 1 August 1984
1.0 Butterworth filter produces significant low-pass fil-
tering and thus additional image blur, but only in the
Imageformingoptics
along-scan direction. (2) Unlike the PSFs for the
image-forming optics and the scanning aperture, the
PSF for the Butterworth filter does not have spatial
MIF .5 i …_ -- Scanning aperture Along-track symmetry, and, as a result, its Fourier transform is
I (a) complex; that is, the three-pole Butterworth filter has
Nyquist I I MTF
Imagingsubsystem an amplitude response and a phase response. At low
frequency I
lI spatial frequencies the phase response is linear, but at
higher spatial frequencies it is not, and, as a result, the
Iw I ~ Butterworth filter introduces an undesirable non-neg-
0 .01 .02 ligible scene-dependent lag in the along-scan direction
v' cycles/n as illustrated by Thompson. 1 8 (3) The analog-to-digital
1.0 conversion (sampling) in the along-scan direction re-
quires some nonzero integration time and thus intro-
eforming
optics
duces some additional image blur. However, this time
(80 nsec) is so small that the resultant image blur is
negligible, i.e., it corresponds to <0.5 m in the along-
MTF .5
Along-scan scan direction. 1 6
Il)
In summary, the spatial-frequency response of the
MSS imaging subsystem is given by
2
h(,,v) = exp[-(kh.,) ] sinc(s.,)fib(kbA)
2
/ Butterworth
filter
X exp[-(kv) ] sinc(syv), (17a)
.005 .01 .015 .02 where ib is the (complex) three-pole Butterworth
i' cycles/m transfer function defined by21
Fig. 10. Imaging subsystem MTF of the Landsat MSS. In the 2) - i(2 - W2)
along-track direction, the frequency response is significantly
(1- 2W
Iib(@)= 6 (c = kbat), (17b)
I undersampled. 1+ W

and the constants are determined from those published


and data needed to prepare this example. As is com- by Friedmann 20 :
mon in remote sensing, all the MSS system parameters 76.2
31.1 76.2 138 31.1
are referenced to equivalent ground-projected dimen- 58 58 58 81.5 81.5
sions at nadir. Thus, for example, 76 X 76 m is the
ground-projected area of the IFOV. The first three terms in Eq. (17a) account for the spa-
Scene radiance entering the MSS is first low-pass tial-frequency response of the imaging subsystem in the
filtered by the imaging-forming optics whose PSF can along-scan direction, and the last two terms account for
be approximated by a Gaussian with a 30-m blur cir- the response in the along-track direction.
cle.1 6 The image-forming optics is followed by a scan- Figure 10 illustrates the MTF (I F(,u,v)I)of the MSS
ning aperture (formed by optical fibers in the focal imaging subsystem in unnormalized spatial frequencies
plane) which has the effect of convolving the output of 1u',v', where it' = A158 [cycles/m] and v' = v/81.5 [cy-
the imaging imaging-forming optics with the IFOV. cles/m]. For reference, the Nyquist frequencies (163)-i
Thus the combined effect of the imaging-forming optics = 0.00613 cycles/m and (116)-i = 0.00862 cycles/m in
and the scanning aperture is to low-pass filter (i.e., blur) the along-track and along-scan directions, respectively,
the input scene equally in both the along-scan (x) and are also indicated. In the along-track direction [Fig.
the along-track (y) directions. 10(a)], the effective cutoff frequency of the imaging
The scanning aperture has the effect of sampling in subsystem (0.00765 cycles/m) is larger than the Nyquist
the along-track direction with an 81.5-m sampling in- frequency corresponding to the case illustrated in Fig.
terval. It is a consequence of the MSS design that ad- 7(c). In contrast, for the along-scan direction [Fig.
jacent sampled IFOVs in the along-track direction 10(b)], the effective cutoff frequency (0.0065cycles/m)
cannot overlap and, in fact, they are not even contiguous is smaller than the Nyquist frequency corresponding to
because of the cladding that surrounds the optical fi- the case illustrated in Fig. 7(b).
bers. 16 Figure 11 illustrates the average imaging-sampling
As the scan mirror of the MSS oscillates, the analog subsystem MTF (It(,i,v)l) of the MSS. In the along-
output of the detectors is electronically processed with track direction [Fig. 11(a)], sampling produces a sub-
a low-pass three-pole Butterworth filter and then stantial MTF reduction (the effective cutoff frequency
sampled at a rate corresponding to a ground-projected is 0.0041 cycles/m) because the sampling passband is
58-m sampling interval in the along-scan direction. significantly smaller than the imaging subsystem
The electronic signal processing in the along-scan di- passband. This mismatch of passbands is, of course,
rection has led to some confusion in the remote-sensing an inevitable consequence of the MSS design; in retro-
literature, apparently for the following reasons: (1) The spect, it could have been avoided by shaping the scan-

1 August 1984 / Vol. 23, No. 15 / APPLIED OPTICS 2579


1.0
Average
systemMTF,along-track
Averageimaglng-saipllirg
A subsystem
MTF Average
systemMTF,along-scan

i X I: \ / - I magingsubsystemMTF
MIF .5 ASMTF
I \ Nyquist Along-track
I requency

. -.I . .I _I| ve , |. @ | |
JJJI § I
---
0 .01 .02 0 .005 .010 .015
v' cyrles/m Spatialfreauency.cycles/m
1.0 Fig. 12. Average system MTF of the Landsat MSS, including the
effects of bilinear image reconstruction. Frequency response is sig-
Averagelinaging-sampling
subsystem
MTF nificantly better in the along-scan direction.
,,x Imagingsubsystem
MIF

MIF .5
:\ \ Along-scan
I \~~~~~~~~b

Nyquist
frequency

0 .01 .02
Fig. 11. Average imaging-sampling subsystem MTF of the Landsat
MSS. In the along-track direction, undersampling has produced
large MTF reductions; in the along-scan direction that is not the
case.

ning aperture in the along-track direction and by


staggering the optical fibers in the focal plane to yield Fig. 13. Simulated image of a target as it would be formed by a
overlapping consecutive IFOVs (see Refs. 22 and 23). system whose EIFOV is as indicated. The additional blur associated
In the along-scan direction [Fig. 11(b)] sampling with the larger asymmetric EIFOV is much more evident than the
produces very little MTF reduction (the effective cutoff associated asymmetry.
frequency is 0.0058 cycles/m) because the sampling
passband is larger than the imaging subsystem pass- 0.93 samples per IFOV in the along-track direction.
band. This more desirable relation between passbands However, this asymmetric blurring is not readily ap-
is the result of the low-pass electronic filter and the parent in visual examinations of Landsat images.17
overlapping of consecutive IFOVs along the line scan. Why? To answer this question, we need to relate the
As Fig. 5 illustrates, to determine the MTF of the MTF to spatial dimensions which are a measure of the
MSS system (including ground processing), it is nec- averagesystem PSF size. Such a measure was proposed
essary to cascade the MTFs in Fig. 11 with the MTF of in 1973 by a NASA working group1 0 and later refined
the filter used for image reconstruction (i.e., resam- by Slater"1 ; it is the effective IFOV (EIFOV). In the
pling); for this example, we have chosen bilinear inter- notation of this paper,
polation. The 2-D reconstruction filter associated with
bilinear image resampling is (EIFOV)aiong-scan= I-- = 104 m, (19a)
2A'c
2 2
P(liv) = sinc (Ae)sinc (v). (18)
(EIFOV)' 1 = 148 m,
=g-track (19b)
The result of cascading this filter (see Fig. 8) with the x~~along-track 2v~
MTF in Fig. 11 is the average system MTF illustrated where Ac and v are the effective system cutoff
in Fig. 12. For comparison the ASMTF in both direc- frequencies, i.e., where Muc and 'c are the spatial
tions is plotted on a common frequency axis; this com- frequencies (cycles/m) at which the ASMTF (Fig. 12)
parison illustrates the superior MTF of the MSS in the is reduced to 0.5. Thus the combined effects of imag-
along-scan direction. ing, sampling, and resampling produce on the average
Figure 12 reveals that the Landsat MSS system a rectangular EIFOV which is 104 X 148 m. For most
produces images which are somewhat more blurred in targets this rectangular EIFOV is not sufficiently
the along-track direction. That is to be expected be- asymmetric to produce an evident asymmetric target
cause, even though the IFOV is square, there are 1.31 blur. We demonstrate this with Fig. 13 using a target
samples per IFOV in the along-scan direction and just suggested by Colvocoresses.1 7 The target is in the

2580 APPLIED OPTICS/ Vol. 23, No. 15 / 1 August 1984


Alony-track graph and illustrated in Fig. 14 can be meaningfully
compared in a relative sense, but some might argue that
all are too small by some common multiplicative
factor.
If it is assumed that the system response is just the
image subsystem response (Fig. 10), the (erroneous)
conclusion would be that the EIFOV is 77 X 65 m.
Comparing this EIFOV with the scanning aperture 63-
148 In X 63-m EIFOV, we see that the imaging-forming optics
cause very little increase in EIFOV. However, the
electronic filter causes (on the average) an increase of
-12 m but only in the line-scan direction. The EIFOV
corresponding to the average imaging-sampling sub-
system response (Fig. 11) is 86 X 122 m. Thus the
104 in (average) effect of sampling is to increase the EIFOV
by just 9 m in the along-scan direction. However, in the
Fig. 14. (Sample-scene) phase-averaged EIFOV of the Landsat
MSS: (a) based on only the scanning aperture; (b) based on the along-track direction the EIFOV increase is 57 m, il-
scanning aperture, image-forming optics, and electronic filter; (c) lustrating dramatically the consequences of under-
including also the effects of sampling; (d) the system EIFOV with sampling. Finally, since the system EIFOV is 104 X 148
bilinear image reconstruction; (e) the system EIFOV using PCC m, we see that bilinear image resampling yields an ad-
(a = 0.5) in place of bilinear image reconstruction. ditional EIFOV increase of 18 m in the along-scan di-
rection and 26 m in the along-track direction. This
additional EIFOV increase can be reduced by using
upper left, and the object in the upper right is the image reconstruction filters with superior frequency response.
of this target as it would be formed by convolution with For example, using PCC with a =-0.5 in place of bi-
the square EIFOV indicated. The two objects in the linear interpolation (see Fig. 9) yields a smaller system
lower left and right are more blurred images of this EIFOV of 93 X 132 m. However, image resampling
target as they would be formed by convolving with the with PCC involves computations with a 4- X 4-pixel
larger rectangular 104- X 148-m EIFOV. Two orthog- mask; for bilinear resampling the pixel mask is just 2 X
onal orientations of this rectangular EIFOV are shown 2. Thus the superior ASMTF (and smaller system
to illustrate, when compared with the object in the EIFOV) associated with PCC image reconstruction is
upper right, just how little asymmetry is present in MSS achieved at the cost of increased digital processing.
images of this target. In addition, the actual appear- VII. Summary
ance of this object in MSS images would be further de-
graded by system noise, and, as a result, any asymme- This paper has presented a general approach to de-
tries would be even less evident than that illustrated. termining the frequency response of a sampled imaging
It should be possible to demonstrate the asymmetric system. The approach accounts for the local system
blurring of the MSS by imaging a point source, for ex- shift variance induced by sampling and the combined
ample, a small highly reflective mirror, as was done in effects of image formation, sampling, and image re-
the Shadow Mountain Eye Project.2 4 However, it is construction by using an ensemble of point-source
misleading to consider just one or two point source im- system inputs and averaging over all possible point-
ages because of the sample-scene phase phenomena il- source locations. The resultant expression for the
lustrated in Fig. 3. If an ensemble of properly regis- average system MTF has general applicability and
tered point-source images were averaged, the result provides a numerical measure of system performance
would be the ASPSF, and the magnitude of its Fourier which is easily evaluated in either of two equivalent
transform would be the ASMTF. This ASMTF should forms.
be asymmetric as indicated in Fig. 12. Unfortunately, As an example, the average system MTF of the
to the best of our knowledge, such an analysis has never Landsat MSS was analyzed. The results of this analysis
been published. provide for the first time a comprehensive examination
The results of this system MTF analysis for the of all the interacting factors which contribute to the
Landsat MSS are summarized in Fig. 14 in terms of total MSS frequency response. The results were
ground-projected EIFOVs using Eq. (19) and the vari- summarized in terms of a system EIFOV which turned
ous effective cutoff frequencies illustrated in Figs. out to be significantly larger than the commonly ac-
10-12. In each case, in the discussion that follows, the cepted IFOV, particularly in the along-track direction
EIFOV dimension in the along-scan direction will be where the effect of undersampling contributes markedly
listed first. However, it is first necessary to recognize to an MTF reduction and resultant increase in image
that, because Slater's definitional of the EIFOV is based blur.
on a 50% MTF reduction level and because the solution
to sinc(v) = 0.5 is v 0.60 rather than v = 0.50, the This paper is based on one presented at the Second
EIFOV of the 76- X 76-m scanning aperture is just 63 X Topical Meeting on Coherent Laser Radar, Aspen,
63 m. Thus, all the EIFOVs listed in the next para- Colo., 1-4 Aug. 1983.

1 August 1984 / Vol. 23, No. 15 / APPLIED OPTICS 2581


References
13. R. J. Arguello, Proc. Soc. Photo-Opt. Instrum. Eng. 271, 86
(1981).
14. J. D. Gaskill,Linear Systems, Fourier Transforms, and Optics
1. S. K. Park and R. A. Schowengerdt, Appl. Opt. 21, 3142 (1982). (Wiley, New York, 1978).
2. R. A. Schowengerdt, S. K. Park, and R. Gray, "Topics in the 15. S. K. Park and R. A. Schowengerdt, Comput. Vision Graphics and
Two-Dimensional Sampling and Reconstruction of Images," Int. Image Process. 23, 258 (1983).
J. Remote Sensing, to be published. 16. P. N. Slater, Photogramm. Eng. Remote Sensing 45, 1479
3. W. Wittenstein, J. C. Fontanella, A. R. Newberry, and J. Baars, (1979). .,
Opt. Acta 29, 41 (1982). 17. A. P. Colvocoresses, Photogramm. Eng. Remote Sensing 46,765
4. D. P. Peterson and D. Middleton, Inf. Control 5, 279 (1962). (1980).
5. 0. H. Schade, in Perception of Displayed Information, L. M. 18. L. L. Thompson, Photogramm. Eng. Remote Sensing 46, 766
Biberman, Ed. (Plenum, New York, 1973), pp. 233-278. (1980).
6. R. Legault, Perception of Displayed Information, L. M. Biber- 19. P. N. Slater, Photogramm. Eng. Remote Sensing 46, 767
man, Ed. (Plenum, New York, 1973), pp. 279-312. (1980).
7. W. K. Pratt, Digital Image Processing (Wiley, New York, 20. D. E. Friedmann, Photogramm. Eng. Remote Sensing 46, 1541
1978). (1980).
8. W. D. Montgomery, J. Opt. Soc. Am., 65, 700 (1975). 21. D. E. Johnson, Introduction to Filter Theory (Prentice-Hall,
9. D. F. Barbe and S. B. Campana, in Advances in Image Pickup Englewood Cliffs, N.J., 1976).
and Display, Vol. 3, B. Kazan, Ed. (Academic Press, New York, 22. S. J. Katzberg, F. 0. Huck, and S. D. Wall, Appl. Opt. 12, 1054
1977), pp. 171-296. (1973).
10. Advanced Scanners and Imaging Systems for Earth Observations, 23. F. 0. Huck, S. K. Park, and N. Halyo, Appl. Opt. 19, 2174
NASA Spec. Publ. 335, (1973), pp. 104-109. (1980).
11. P. N. Slater, Opt. Acta 22, 277 (1975). 24. D. S. Simonett, Manual of Remote Sensing (American Society
12. J. C. Lansing and R. W. Cline, Opt. Eng. 14, 312 (1975). of Photogrammetry, 1983), Chap. 1, pp. 21, 22.

Jack Parmley
McPherson Instruments

John Weis
Northrup Corp.

OSA 1983
New Orleans
Stanley S. Ballard
Photos: F. S. Harris, Jr.
U. Florida

2582 APPLIED OPTICS/ Vol. 23, No. 15 / 1 August 1984

You might also like