You are on page 1of 10

05-100

6/11/07

9:41 AM

Page 841

Exploring the Geostatistical Method


for Estimating the Signal-to-Noise
Ratio of Images
P.M. Atkinson, I.M. Sargent, G.M. Foody, and J. Williams

Abstract
The signal-to-noise ratio (SNR) has been estimated for
remotely sensed imagery using several image-based methods
such as the homogeneous area (HA) and geostatistical (GS)
methods. For certain procedures such as regression, an
alternative SNR (SNRvar), the ratio of the variance in the
signal to the variance in the noise, is potentially more
informative and useful. In this paper, the GS method was
modified to estimate the SNRvar, referred to as the SNRvar(GS).
Specifically, the sill variance c of the fitted variogram model
was used to estimate the variance of the signal component
and the nugget variance c0 of the fitted model was used to
estimate the variance of the noise. The assumptions required
in this estimation are presented. The SNRvar(GS) was estimated
using the modified GS method for six different land-covers
and a range of wavelengths to explore its properties. The
SNR*var(GS) was found to vary as a function of both wavelength and land-cover. The SNR*var(GS) represents a useful
statistic that should be estimated and presented for different
land-cover types and even per-pixel using a local moving
window kernel.

Introduction
Remote sensing of properties distributed spatially within the
physical environment involves measurement with uncertainty. As a result, a pixel z(x0) (of spectral response z at
location x0) in a remotely sensed image can be viewed as
being comprised of the true or underlying pixel value u(x0)
plus some measurement error e(x0):
z(x0) = u(x0) + e(x0).

(1)

The term e(x0) usually comprises systematic es(x0) and


random er(x0) errors such that
e(x0) = es(x0) + er(x0).

(2)

Measurement error can reduce the accuracy of prediction of both continuous (e.g., nitrogen content of vegetation)
and categorical (e.g., land-cover) variables from remotely

P.M. Atkinson and G.M. Foody are with the School of


Geography, University of Southampton, Highfield,
Southampton SO17 1BJ, UK (P.M.Atkinson@soton.ac.uk).
I.M. Sargent is with the Ordnance Survey, Romsey Road,
Maybush, Southampton SO16 4GU, UK.
J. Williams is with QinetiQ, St Andrews Road, Malvern,
Worcestershire WR14 3PS, UK.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

sensed imagery (e.g., Lucht and Lewis, 2000; Valor et al.,


2002; Foody et al., 2004). Thus, it is important that remote
sensing researchers and practitioners are aware of such
uncertainty, particularly where sensors are new or little data
are available on sensor performance. However, quantitative
assessment of accuracy is often missing from the image
processing chain associated with remotely sensed data
(Schott, 1997).
A problem which has generated interest within the
remote sensing community is how best to quantify the error
e(x) given only a remotely sensed image z(x) in a given
waveband. It is generally not possible to estimate the error
e(x0) directly for pixel location x0 given only the observed
value z(x0). If the error were known, the true value would
also be known and the error could be removed. Rather,
researchers have sought to estimate the expectation of the
absolute error E[|e(x)|] over all x, that is, to estimate the
error using a statistical model. Systematic error can often be
associated with a known source within the remote sensor
system, and so often it can be removed (Wrigley et al., 1984;
Nichol and Vohora, 2004). Thus, most attention has focused
on estimating over all x the expectation of the absolute
random error, referred to here as random noise or just noise
(Curran and Dungan, 1989; Roger, 1996; Smith and Curran,
1996; Schowengerdt, 1997; Corner et al., 2003).
Random noise at a single location x0 can be estimated
by measuring repeatedly the variable of interest z(x0) and
using the mean mz(x0) of the observed values to estimate the
true value u(x0) (i.e., assuming es(x0) = 0) and the standard
deviation sz(x0) to estimate the random noise z(x0) (Duggin
et al., 1985). However, for remotely sensed images, repeated
measurement at a single location is impractical (for a given
time and date). Fortunately, several methods exist for
estimating noise from a single remotely sensed image based
on the image data alone (Smith and Curran, 1999), and some
of these are described below.
It is important to interpret an estimate of random noise
sz for a given remotely sensed image z(xi),i  1,2, . . . ,n
(where n is the number of pixels) in relation to the true or
underlying image u(xi) (Equation 1). Traditionally, the
signal, estimated by the mean spectral response mz, has been
expressed relative to the noise, estimated by the sample
standard deviation, sz using a signal-to-noise ratio (SNR)

Photogrammetric Engineering & Remote Sensing


Vol. 73, No. 7, July 2007, pp. 841850.
0099-1112/07/73070841/$3.00/0
2007 American Society for Photogrammetry
and Remote Sensing
July 2007

841

05-100

6/11/07

9:41 AM

Page 842

(Lo, 1986; Curran and Dungan, 1989; Lillesand and Kiefer,


2000) as
SNR* 

SIGNAL
m
 z .
NOISE
sz

(3)

Several image-based methods of estimating image noise


and image SNR have been developed. Smith and Curran (1999)
describe five approaches that estimate image noise by minimizing the component of the local standard deviation sz(W(x))
due to underlying variation, where W(x) is a local kernel and
 represents a specific wave band. These are the Homogeneous Area (HA) (Duggin et al., 1985), Nearly Homogeneous
Area (NHA) (Boardman and Goetz, 1991), Geostatistical (GS)
(Curran and Dungan, 1989), Homogeneous Block (HB) (Gao,
1993), and Multiple Waveband (MW) (Roger and Arnold, 1996)
methods. In each of the five approaches the SNR is estimated
by comparing some estimate of the mean mz (representing the
signal) to the standard deviation sz (representing image noise),
as seen in Equation 3. Two of these five approaches are now
described briefly: the HA method and the GS method.
The HA Method for Estimating the SNR
In the simple HA method a small window W(x0) (hereafter
abbreviated to W) is defined within which the underlying
variation in the image is expected to be homogeneous (the
same everywhere), or as close to this as possible. The signal
is estimated by the local mean mz(W) within W and the
noise is estimated with the local standard deviation sz(W).
The SNR is estimated by
SNR* 

mlz(W) .
slz(W)

(4)

To apply this method in practice it is necessary to select


a window within which the variance of the underlying
image (i.e., the variance due to the signal devoid of noise) is
exactly zero or as close to zero as possible.
The GS Method for Estimating the SNR
In comparison to the HA method, the GS method is preferable because it is not necessary to select locally homogeneous areas. It centers on estimation and modeling of the
variogram (strictly the semi-variogram), a function representing spatial dependence of measured values (Matheron, 1965
and 1971). Spatial dependence is the tendency for proximate
pixels to take more similar values than more distant ones.
The model fitted to the variogram is often composed of
several components: at least one structured component
combined with a nugget component (see below). In the GS
method, the mean estimates the signal whereas the square
root of the variance of the nugget component of the variogram estimates the noise. The advantage of this approach
over the HA approach is that the presence of spatially
structured underlying variation is not detrimental to the
estimation of image noise. In fact, the presence of spatial
structure is helpful in separating the underlying variation
from the noise as part of the process of modelling the
variogram (Curran and Dungan, 1989; Chappell et al., 2001).
Often in remote sensing, regression is used to predict
the property of interest at the ground using remotely sensed
image data. For procedures such as regression, the
researcher is interested in the deleterious effect of noise
relative to variation in the signal (as opposed to the signal
itself). For this reason researchers have proposed alternative
SNR measures. Of particular interest is the SNRvar defined by
SNR*var 

842

July 2007

VARIANCESIGNAL
s2u
 2
VARIANCENOISE
se

(5)

The SNRvar has been applied commonly to remotely


sensed images (e.g., Schowengerdt, 1997). The information
contained in the SNRvar is fundamentally different to that in
the SNRmean (from this point onwards the SNR of Equations 3
and 4 is referred to as the SNRmean to distinguish it clearly
from the SNRvar). While the denominators are basically the
same (one is the square of the other), the numerators differ.
In particular, in the SNRmean the numerator is the mean
spectral response (often within a local window), while in
the SNRvar the numerator is the variance of the spectral
response within a local window.
The SNRvar contains useful information where the objective is to use the information in the image to predict some
property of interest using a model. An example would be
where the objective is to regress the image variable (e.g.,
reflectance) on some ground variable (e.g., leaf area index
(LAI)) with which it is moderately correlated (commonly the
case in remote sensing). Then the SNRvar conveys the ratio of
the variation that may be correlated with the ground variable
to the variation that is not. The point is that the extent to
which the underlying spatial variation (and, therefore, the
information resolved by an image) is affected by image noise
may be of great interest to remote sensing researchers wishing
to evaluate the viability of a particular investigation before
actually executing it. Therefore, the SNRvar may be useful in
circumstances where the SNRmean is not.
In this paper, we propose using a modification of the GS
method to estimate a SNRvar(GS) statistic in which the signal
and noise components are estimated using the variogram.
The new SNRvar(GS) is estimated using the modified GS
method and evaluated in a range of circumstances. Atkinson
et al. (2005) found that the SNR*mean varied with land-cover
type and this variation had important implications for use
and interpretation of the statistic. That analysis is extended
here to the SNR*var(GS). In the next section, the GS method is
described and adapted to allow estimation of the SNRvar(GS).

The Geostatistical Method


Geostatistics is a set of techniques for the analysis of spatial
data (Journel and Huijbregts, 1978; Isaaks and Srivastava,
1989; Goovaerts, 1997). Central to geostatistics is the
variogram, as described in the introduction. The variogram
is itself best viewed as a parameter of the random function
model, described below.
The RF model
Continuous random variables (RVs) Z(x0) are characterized
fully by the cumulative distribution function (CDF) which
gives the probability that the RV Z(x0) at location x0 is less
than or equal to a given threshold z:
F(x0;z) = Prob{Z(x0)  z}

 z

(6)

A Random Function (RF) is the spatial equivalent of a


in which the inter-dependence between any two point
locations may be expressed as a function of separating
distance (Matheron, 1965 and 1971; Journel and Huijbregts,
1978; Isaaks and Srivastava, 1989; Goovaerts, 1997). A
realization of a RF is termed a Regionalized Variable (RV).
In geostatistics, spatial data (e.g., a remotely sensed image)
are modeled as a RV (Goovaerts, 1997).
The spatial law of the RF Z(x) is given by the n-variate
or n-point cdf:
RV

F(x1,...,xn; z1,...,zn) = Prob{Z(x1)z1,...,Z(xn)zn}

 z

(7)

defined for any choice of n and any location x. The entire
spatial law is not required for most applications. Rather, the
analysis is usually restricted to CDFs involving at most two
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

05-100

6/11/07

9:41 AM

Page 843

locations at a time. Since repeated observations are usually


not available for any fixed location x0, the analysis is
restricted further to fixed vectors of distance and direction
known as lags, h. Each pair of locations {x, x  h} for
which data are available are then treated as repeats of h.
This strategy amounts to the decision to adopt a stationary
RF model. Stationary, in this context, means that the variogram parameter is constant across space. Such a decision
has implications for the implementation of the modified GS
method to estimate SNRvar(GS).
When the expected value E{Z(x)} exists and is independent of x (invariant within a region V) we have
E[Z(x)Z(x + h)] = 0.

(8)

If the variogram exists and depends only on h then


g(h) 

1
E {Z(x) Z(xh)}2
2

(9)

and the RF is said to be intrinsically stationary. The variogram may exist given intrinsic stationarity only, but the
covariance function and autocorrelation function (which
imply second-order stationarity) may not exist.
Variogram Estimation and Modeling
For continuous variables, the sample semivariance is
defined as half the average squared difference between
values separated by a given lag h. The sample variogram
may be estimated using:
g(h) 

1
2P(h)

P(h)

 [z(xa) z(xa  h)]2

(10)

a1

where P(h) is the number of paired comparisons at a


specific lag h and z(x) represents an observation or realization of Z(x) at location x. Further, the variogram may be
obtained either as an average representing all directions
(omnidirectional) or for several different orientations where
variation is anisotropic (i.e., depends on orientation). In a
single remotely sensed image many observations are
available allowing greater confidence in the estimated
variogram and increasing the number of orientations in
which it can be estimated.
The estimated variogram obtained from Equation 11 is a
set of semivariances at a discrete set of lags only. To allow
statistical prediction, it is customary to fit a continuous
mathematical model to the sample variogram. The model
fitted must be such as to ensure that linear combinations of
the RF cannot have negative variances. This property is
referred to as conditional negative semi-definiteness (CNSD).
It is common practice to select a model from several which
are known a priori to be CNSD (McBratney and Webster,
1986; Webster and Oliver, 1990). Two such models are given
below, both of which are used in this paper.
1. The nugget effect model:

g(h)  c0

01

if h  0
otherwise

(11)

where c0 is the nugget variance; and


2. The spherical model:

1.5

g(h)  c

 

h
h
0.5

a
a

if h  a
otherwise

(12)

where c is the sill variance or structured component and


a is the range or non-linear parameter. In simple terms, the
fitting procedure is as follows. The user chooses a type of
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

variogram model and decides whether an anisotropic model


is required (based on directional variograms). A model is
then fitted to the sample variogram usually by some automatic process such as weighted least squares (e.g., Cressie,
1985 and 1991), that is, by minimizing:
WSS 

 (hk)
[g(hk) g(hk)]2
k1

(13)

where WSS is the weighted sum of squares, and (hk) is a


set of weights at k  1,2, . . . K lags hk (Goovaerts, 1997).
Interpreting the Modeled Variogram
Consider the coefficients of a fitted spherical model. The sill
c of the spherical model provides information on the
amount of variation present in V. More precisely, the sill
estimates the a priori variance D2(v,) of Z (that is the
variance obtained on a support v within an infinitely sized
region). The dispersion or sample variance D2(v,V) of Z (that
is, the variance obtained on a support v within a region V)
is also predictable from the variogram through numerical
integration. The range a of the model provides information
on the scale of spatial variation.
The variogram models described may be used singly or
in a positive linear combination (Goovaerts, 1997). The
nugget model is often fitted together with a structured
component (e.g., spherical model) to represent a discontinuity at the origin. The nugget variance c0 (the sill of the
nugget component) represents unresolved variation that
exists at a micro-scale (at an interval smaller than the
smallest distances of separation in the sampling frame) and
measurement error, but it can also arise from uncertainty
both in estimating the variogram and in fitting the model at
short lags (Atkinson, 1997). Where the amount of microscale variation is believed to be small (e.g., because observations are abutting or overlapping, as for remotely sensed
imagery), then the nugget variance can be assumed to be
due primarily to measurement error. This assumption
underlies the Curran and Dungan (1989) GS method for
estimating the SNR.
The Modified GS Method
The two key arguments supporting the GS method are as
follows. First, noise is not correlated with the signal. This
implies that the variogram of the image z(h) is simply the
sum of the variograms for the underlying variation u(h) and
the noise e(h):

z(h) = u(h) + e(h).

(14)

Second, since imagery is already convolved by the


point-spread function (PSF) of the sensor, proximate pixels
are likely to be highly autocorrelated. This means that in the
limit, that is, as h approaches zero, the contribution from
the signal to the variance will approach zero.
limit gz (h)  s2e .

(15)

(h : 0)

Estimating SNRvar
Here we present a new and simple estimator of the SNRvar.
The SNRvar may be estimated by the ratio of the two variogram model coefficients, c (representing the variance in the
underlying image, and potentially the sum of more than one
individual sill component) and c0 (representing the variance
due to noise):
SNR*var(GS) 

c
.
c0

(16)

July 2007

843

05-100

6/11/07

9:41 AM

Page 844

It is implicit in such an approach that a second-order


stationary RF model must be adopted (i.e., so that the covariance exists) and a bounded model (with a clearly defined sill,
e.g., the spherical model) must be fitted to the variogram of
the stationary RF (Myers, 1989; Goovaerts, 1997).
In the analysis that follows, the SNRvar(GS) is estimated
for six different land-cover types and for a range of different
wavebands to evaluate the behavior of the statistic. First, the
field site and data used are introduced briefly.

Field Site and Data


A Compact Airborne Spectrographic Imager (CASI) image was
acquired over Falmouth, Cornwall in July 1999 by the UK

Environment Agency. The image was acquired in Enhanced


Spectral Mode with 72 spectral bands of approximately 8.3
nm width that were contiguous across the spectrum from
400 nm to 900 nm. The increased integration time required
for these narrow wavebands results in pixels that are
marginally longer than they are wide. However, the image
was resampled so that the pixels represented an area of
approximately 4 m by 4 m.
The scene covered by the CASI image included a range
of land-cover types (i.e., varying crop types). Six areas
containing relatively homogeneous areas of specific landcover types were chosen. The spatial and spectral data for
pixels contained wholly within these areas were extracted.
The six regions are shown in Figure 1.

Figure 1. CASI sub-images, with a spatial resolution of approximately


4 m, representing areas of (a) barley, (b) cauliflower, (c) corn, (d)
grassland, (e) potato, and (f) woodland. (Atkinson et al., 2005; used
with permission from International Journal of Remote Sensing,
http://www.informaworld.com).

844

July 2007

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

05-100

6/11/07

9:41 AM

Page 845

Analysis
Example 1
Two (of the available six) land-cover types, grassland and
woodland, were selected to illustrate the SNR*var(GS). The
remaining four land-cover types exhibited variation that was

similar to that for grassland, and so they were omitted to


reduce redundancy in the presentation. Omnidirectional
variograms for the two land-cover types, and for a sample of
eight wavebands (wavebands 1, 11, 21, 31, 41, 51, 61, and
71), are shown in Figure 2. Variograms were estimated and
modeled for all wavebands, but only the sample of eight is

Figure 2. Variograms of (a, c, e, g, i, k, m, and o) woodland and (b, d, f, h, j, l,


n, and p) grassland for wavebands (a and b) 1, (c and d) 11, (e and f) 21, (g
and h) 31, (i and j) 41, (k and l) 51, (m and n) 61, and (o and p) 71 sampled
from the complete set of 72 wavebands. The models shown were fitted using
weighted least squares approximation. Note that to allow the models to be
shown fully, the scale on the abscissa for woodland reaches a maximum of
40 pixels (160 m), whereas for grassland the maximum is 14 pixels (56 m).
The scales of the ordinates on all plots differ for the same reason. (Atkinson
et al., 2005; used with permission from International Journal of Remote Sensing,
http://www.informaworld.com).

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

July 2007

845

05-100

6/11/07

9:41 AM

Page 846

Figure 2. (Continued) Variograms of woodland and grassland for designated


wavebands. (Atkinson et al., 2005; used with permission from International
Journal of Remote Sensing, http://www.informaworld.com).

shown, again to reduce redundancy. All variograms were


estimated and fitted with models using weighted least
squares approximation using the GSTAT software (Pebesma
and Wesseling, 1998). The wavelength ranges for these
wavebands are 401.6 nm to 501.3 nm (14 slightly overlapping bands with a bandwidth of 8.2 nm) and 500.1 nm to
917.7 nm (58 slightly overlapping bands with a bandwidth
of 8.4 nm). The variograms for the grassland and woodland
classes were all fitted best (in the WSS sense) with a spherical or nested spherical model combined with a nugget
component.
846

July 2007

The SNRvar was estimated using the modified GS


method (Equation 16). The sill variance, the nugget variance and the SNRvar are all plotted against wavelength for
each land-cover type in Figure 3 (note the different scales
on the ordinates of the plots). Three observations can be
made. First, the sill variance is highly dependent on
wavelength and land-cover type. This is not surprising: as
described in the Geostatistical Method Section, the overall
sill (c1, . . . ,cn and c0) represents the a priori variance
D2(v,) of the imagery per land-cover. This result simply
means that variance in remotely sensed imagery is a
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

05-100

6/11/07

9:41 AM

Page 847

Figure 3. The (a and b) sill variance, (c and d) nugget variance, and


(e and f) SNR*var(GS) plotted against wavelength for both (a, c, and e)
woodland and (b, d, and f) grassland. (Atkinson et al., 2005; used
with permission from International Journal of Remote Sensing,
http://www.informaworld.com).

function of wavelength and land-cover type. Geostatistical


descriptions of texture have been devised (e.g., Carr, 1996)
to exploit such variation to help to discriminate between
land-cover types. Second, the nugget variance is also
dependent on wavelength and land-cover type, but does
not follow exactly the same pattern as the sill variance.
Third, as a consequence of the above relations, the
SNR*
var(GS) is also dependent on wavelength and land-cover
type. Variation in the SNR with wavelength is well known
and understood. Variation of SNR*var(GS) with land-cover is
not surprising, but it is interesting. It means that the
predictive ability of models such as regression and classification models may also depend on variation in the ratio of
signal variance-to-noise variance with land-cover type and
wavelength. Thus, the SNR*var(GS) statistic might be used to
evaluate predictive models prior to their application.
Example 2
To investigate the dependence of SNR*var(GS) on land-cover type
further, we conducted a second study using a single waveband.
A single, near-infrared waveband (880.1 to 888.5 nm)
was extracted for each of the six land-cover types. This
band was chosen using feature selection (the band which
produced the greatest Mahalanobis distance between the
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

two spectrally closest bands) (Mather, 2004). Omnidirectional variograms were estimated and modeled as before.
The nested spherical model combined with a nugget
component provided the best (WSS) fit in each case except
for the variogram for the land-cover cauliflower for which a
single spherical component (plus nugget) was satisfactory
(Figure 4). The woodland variogram is different to the
others in that (a) it has a markedly different scale on the
ordinate, and (b) a sizeable short-range structure with a
range of about four pixels (or 16 m) is included, which is
most likely related to the size of the tree crowns.
The SNRvar(GS) was estimated using the modified GS
method. The sill variance c, the nugget variance c0 and the
SNR*
var(GS) are all plotted against land-cover type in Figure 5
(again, note the different scales on the ordinates of the
plots). The differences in the estimates of c and c0 and the
SNR*var(GS) between woodland and the other land-covers are
immediately apparent. This may be a function of the large
near-infrared reflectance for woodland. However, notice that
the differences in the signal variance c are much greater
than the differences in the noise variance c0. Examining
only those classes representing crops or grasses there still
exist important differences in c, c0 and the SNR*var(GS)
between land-cover types. The differences in SNR*var(GS)
July 2007

847

05-100

6/11/07

9:41 AM

Page 848

Figure 4. Variograms estimated from the sub-images representing areas of (a)


barley, (b) cauliflower, (c) corn, (d) grassland, (e) potato, and (f) woodland.
The models shown were fitted using weighted least squares approximation.
(Atkinson et al., 2005; used with permission from International Journal of
Remote Sensing, http://www.informaworld.com).

between all classes are large: the smallest value is less than
3 (potato) and the largest value is greater than 80 (woodland). This means that if the objective were to predict some
biophysical (e.g., leaf area index, biomass) or biochemical
(e.g., nitrogen content) property from the imagery, then, a
priori, the predictions for the woodland class would be
expected to be less affected by noise.

Discussion
We now summarise the findings of this paper, make some
suggestions for improving the GS method for estimating the
SNRvar(GS) and make some observations about the variogram
range coefficient.
Variation in SNR*var(GS) with Land-cover
The SNR*var(GS) varied with wavelength and land-cover type.
Variation with land-cover type arises primarily because
variance in the signal is a function of land-cover (e.g., it is
well known that different land-covers have different textures).
However, variation in the SNR*var(GS) arises to some extent as a
function of the dependence of noise on land-cover type as
848

July 2007

discussed by Atkinson et al. (2005). Specifically, components


of the noise may be due to atmospheric effects that interfere
with the signal such that the noise is partially a function of
the signal.
Variation in SNR*var(GS) with land-cover, as demonstrated
here, means that it is important that local variation in
SNR*
var(GS) is presented. The SNRvar(GS) should at least be
estimated and presented for each land-cover class. Preferably,
the SNRvar(GS) should be estimated locally and the resulting
estimates displayed as images (Atkinson et al., 2005).
Improving the GS Method
An advantage of the GS method over other methods of
estimating the SNRvar is that it can be estimated by landcover class, or locally per-pixel using a local kernel or
moving window leading to an image of SNR*var(GS) statistics.
The suggestion that the SNR*var(GS) might be presented as an
image has implications for how the GS method might be
improved. For example, it would be possible to estimate a
distribution of noise estimates. This distribution could be
used to refine the noise value used in estimating SNRvar(GS).
For example, the smallest or near-smallest (e.g., 5-percentile)
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

05-100

6/11/07

9:41 AM

Page 849

correlation should be close to one for much hyperspectral


data. However, the noise may be uncorrelated even between
adjacent bands, depending on its source. Where this is the
case, the cross-variogram (representing the cross spatial
dependence or cross-correlation between two variables)
should yield useful information on noise. The cross-variogram is due to underlying variation alone and contains no
component due to noise. This occurs where the noise terms
are independent, and thus do not contribute to the crossvariogram (e.g., van der Meer and Bakker, 1997). Thus, for
hyperspectral imagery, the cross-variogram represents a
potentially useful alternative to the variogram.
Variogram Model Ranges
Interestingly, the variograms estimated from the woodland
class produced a slightly larger total range (maximum of the
two ranges of the nested variogram models) than did the
variograms of grassland (Figure 2). The full set of model
ranges for each of the two land-cover types is shown in
Figure 6. For grassland the range was relatively constant
with wavelength. However, for woodland there was a small
increase in the range at around the position of the red-edge
(Curran, 1994). Such variation with wavelength has been
observed by others (Chavez, 1992; Atkinson and Emery,
1999) and has implications for techniques for processing
hyperspectral imagery (e.g., Atkinson and Emery (1999)
found that strategies for sampling reflectance needed to be
wavelength-dependent).

Conclusion
The SNR*var(GS) represents a useful image-based statistic that
may be interpreted as the ratio of the component of variance
that may be correlated with (or useful in predicting) a
variable of interest to the component of variance that is not

Figure 5. The (a) sill variance, (b)


nugget variance and (c) SNR*var(GS)
plotted as bar charts against
barley, cauliflower, corn, grassland, potato, and woodland.

value of SNR*var(GS) could be used to provide a cautious or


conservative estimate of image SNRvar(GS) or one that was
least contaminated by signal. Alternatively, the distribution
of SNR*var(GS) estimates might be useful to investigators. For
example, the distribution of values could be used to provide
information on the uncertainty inherent in image-based
estimation of SNRvar(GS).
The variograms estimated in this paper were all omnidirectional, meaning that isotropy was assumed. While in the
present case we checked for anisotropy, isotropy is not
guaranteed generally. Anisotropy may be useful given the
present objective: where the variogram varies with orientation
different estimates of c0 may arise. Then, it may be possible
to utilize such differences to obtain a more accurate estimate
of noise than that provided by modeling the omnidirectional
variogram. For example, given that the variogram models for
each orientation are of the same basic type it may be preferable to select the smallest c0, rather than the average, on the
basis that this reduces the possibility of underlying variation
contributing to the estimate.
Another property that may be useful for estimating
noise is the correlation between adjacent bands. This
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

Figure 6. The range plotted


against wavelength for both
(a) woodland, and (b) grassland.

July 2007

849

05-100

6/11/07

9:41 AM

Page 850

correlated with that variable (i.e., noise). Such information is


of obvious utility in remote sensing where the objective de
facto is to predict ground properties from images that are a
function of those ground properties. The GS method provides a simple means of estimating this ratio. However, care
should be exercised when presenting this ratio. This paper
has shown that the SNR*var(GS) varies as a function of both
wavelength (as expected) and land-cover class. Therefore,
the SNRvar(GS) should at least be estimated and presented per
land-cover class and preferably as an image of local values.
It would be interesting to extend this analysis to include a
quantitative interpretation of the effects of the SNR*var(GS) on
regression and classification models.

Acknowledgments
The authors thank the Environment Agencys National
Centre for Environmental Data and Surveillance for contributing data. The research was performed with the aid of
funding derived from the UK Ministry of Defence Corporate
and Applied Research Programmes.

References
Atkinson, P.M., 1997. On estimating measurement error in remotelysensed images with the variogram, International Journal of
Remote Sensing, 18:30753084.
Atkinson, P.M., and D. Emery, 1999. Exploring the relation between
spatial structure and wavelength: Implications for sampling
reflectance in the field, International Journal of Remote
Sensing, 20:26632678.
Atkinson, P.M., I.M.J. Sargent, G.M. Foody, and J. Williams, 2005.
Interpreting image-based methods for estimating the signal-tonoise ratio, International Journal of Remote Sensing,
26(22/20):50995115.
Boardman, J.W., and A.F.H. Goetz, 1991. Sedimentary facies analysis
using AVIRIS data: A geophysical inverse problem, Proceedings
of the Third AVIRIS Workshop, JPL Publication 9128, Jet
Propulsion Laboratory, Pasadena, California, pp. 413.
Carr, J.R., 1996. Spectral and textural classification of single and
multiple band digital images, Computers and Geosciences,
22:849865.
Chappell, A., J.W. Seaquist, and L. Eklundh, 2001. Improving the
estimation of noise from NOAA AVHRR NDVI for Africa using
geostatistics, International Journal of Remote Sensing,
22:10671080.
Chavez, P.S., 1992. Comparison of spatial variability in visible and
near-infrared spectral images, Photogrammetric Engineering &
Remote Sensing, 58(8):957964.
Corner, B.R., R.M. Narayanan, and S.E. Reichenbach, 2003. Noise
estimation in remote sensing imagery using data masking,
International Journal of Remote Sensing, 24:689702.
Cressie, N.A.C., 1985. Fitting variogram models by weighted least
squares, Mathematical Geology, 17:563586.
Cressie, N.A.C., 1991. Statistics for Spatial Data, John Wiley and
Sons, New York.
Curran, P.J., 1988. The semi-variogram in remote sensing: an
introduction, Remote Sensing of Environment, 3:493507.
Curran, P.J., 1994. Imaging spectrometry, Progress in Physical
Geography, 18:247266.
Curran, P.J., and J.L. Dungan, 1989. Estimation of signal-to-noise: A
new procedure applied to AVIRIS imagery, IEEE Transactions
on Geoscience and Remote Sensing, 27:620628.
Duggin, M.J., H. Sakhavat, and J. Lindsay, 1985. Sytematic and
random variations in Thematic Mapper digital radiance data,
Photogrammetric Engineering & Remote Sensing,
51(12):14271434.
Foody, G.M., I.M. Sargent, P.M. Atkinson, and J. Williams, 2004.
Land-cover classification from hyperspectral remotely sensed

850

July 2007

data: An investigation of spectral, spatial and noise issues,


International Journal of Remote Sensing, 25:23372363.
Gao, B-C., 1993. An operational method for estimating signal to
noise ratios from data acquired with imaging spectrometers,
International Journal of Remote Sensing, 11:16691694.
Goovaerts, P., 1997. Geostatistics for Natural Resources Evaluation,
Oxford University Press, New York.
Isaaks, E.H., and R.M. Srivastava, 1989. Applied Geostatistics,
Oxford University Press, Oxford.
Journel, A.G., and C.J. Huijbregts, 1978. Mining Geostatistics,
Academic Press, London.
Lillesand, T.M., and R.W. Keifer, 2000. Remote Sensing and Image
Interpretation, 4th edition, John Wiley and Sons, New York.
Lo, C.P., 1986. Applied Remote Sensing, Longman, Harlow.
Lucht, W., and P. Lewis, 2000. Theoretical noise sensitivity of BRDF
and albedo retrieval from the EOS-MODIS and MISR sensors
with respect to angular sampling, International Journal of
Remote Sensing, 21:8198.
Mather, P.M., 2004. Computer Processing of Remotely-Sensed Images:
An Introduction, 3rd edition, John Wiley and Sons, Chichester.
Matheron, G., 1965. Les Variables Rgionalises et Leur Estimation,
Masson, Paris.
Matheron, G., 1971. The Theory of Regionalized Variables and its
Applications, Centre de Morphologie Mathmatique de
Fontainebleau, Fontainebleau.
Myers, D.E., 1989. To be or not to be stationary? That is the
question, Mathematical Geology, 21:347362
McBratney, A.B., and R. Webster, 1986. Choosing functions for
semi-variograms of soil properties and fitting them to sampling
estimates, Journal of Soil Science, 37:617639.
Nichol, J.E., and V. Vohora, 2004. Noise over water surfaces in
Landsat TM images, International Journal of Remote Sensing,
25:20872094.
Pebesma, E.J., and C.G. Wesseling, 1998. Gstat, a program for
geostatistical modelling, prediction and simulation, Computers
and Geosciences, 24:1731.
Roger, R.E., 1996. Principal Components transform with simple,
automatic noise adjustment, International Journal of Remote
Sensing, 17:27192727.
Roger, R.E., and J.F. Arnold, 1996. Reliably estimating the noise in
AVIRIS hyperspectral images, International Journal of Remote
Sensing, 17:19511962.
Schott, J.R., 1997, Remote Sensing: The Image Chain Approach,
Oxford University Press, New York.
Schowengerdt, R.A., 1997. Remote Sensing: Models and Methods for
Image Processing, Academic Press, London.
Smith, G.M., and P.J. Curran, 1996. The signal-to-noise ratio (SNR)
required for the estimation of foliar biochemical concentrations,
International Journal of Remote Sensing, 17:10311058.
Smith, G.M., and P.J. Curran, 1999. Methods for estimating image
signal-to-noise ratio (SNR), Advances in Remote Sensing and
GIS Analysis (P.M. Atkinson and N.J. Tate, editors), John Wiley
and Sons, Chichester, pp. 6174.
Valor, E., V. Casselles, C. Coll, E. Rubio, and F. Sospedra, 2002.
NEDT influence in the thermal band selection of satellite-borne
instruments, International Journal of Remote Sensing,
23:34933504.
Van Der Meer, F., and W. Bakker, 1997. CCSM: Cross-correlogram
spectral matching, International Journal of Remote Sensing,
18:11971201.
Webster, R., and M.A. Oliver, 1990. Statistical Methods for Soil and
Land Resources Survey, Oxford, Oxford University Press.
Wrigley, R.C., D.H. Card, C.A. Hlavka, J.R. Hall, F.C. Mertz,
C. Archwamety, and R.A. Schowengerdt, 1984. Thematic
Mapper image quality: Registration, noise and resolution, IEEE
Transactions on Geoscience and Remote Sensing, 22:263271.
(Received 5 August 2005; accepted 28 November 2005; revised 31
January 2006)

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

You might also like