Beruflich Dokumente
Kultur Dokumente
Abstract
Traditional spectral classi®cation of remotely sensed images applied on a pixel-by-pixel basis ignores the
potentially useful spatial information between the values of proximate pixels. For some 30 years the spatial
information inherent in remotely sensed images has been employed, albeit by a limited number of researchers, to
enhance spectral classi®cation. This has been achieved primarily by ®ltering the original imagery to (i) derive texture
`wavebands' for subsequent use in classi®cation or (ii) smooth the imagery prior to (or after) classi®cation. Recently,
the variogram has been used to represent formally the spatial dependence in remotely sensed images and used in
texture classi®cation in place of simple variance ®lters. However, the variogram has also been employed in soil
survey as a smoothing function for unsupervised classi®cation. In this review paper, various methods of
incorporating spatial information into the classi®cation of remotely sensed images are considered. The focus of the
paper is on the variogram in classi®cation both as a measure of texture and as a guide to choice of smoothing
function. In the latter case, the paper focuses on the technique developed for soil survey and considers the
modi®cation that would be necessary for the remote sensing case. 7 2000 Elsevier Science Ltd. All rights reserved.
0098-3004/00/$ - see front matter 7 2000 Elsevier Science Ltd. All rights reserved.
PII: S 0 0 9 8 - 3 0 0 4 ( 9 9 ) 0 0 1 1 7 - X
362 P.M. Atkinson, P. Lewis / Computers & Geosciences 26 (2000) 361±371
1 X
p
h
univariate feature space. In a simple empirical
gn
h zn
xa ÿ zn
xa h2 :
1
2P
h a1 approach presented by Carr (1996) this allocation is
executed through the prior construction of a simple
Curran (1998) and Woodcock et al. (1988a,b) pro- look-up table such that any pixel value will have as-
vide readable introductions to the variogram while sociated with it the destination class without the need
Dungan (1998) reviews geostatistical techniques for to compare the distances to all class means. This pro-
estimation and simulation in a remote sensing context. cedure is readily extended to multivariate feature space
Geostatistical techniques that utilise spatial infor- where the distances to class means may be obtained by
mation in classi®cation can be split into two distinct Euclidean geometry.
groups. In the ®rst, spatial information is used to pro- A second popular method of supervised classi®-
vide data on texture. It is implicit in such approaches cation is maximum likelihood (ML) classi®cation
that texture varies spatially across the image, and par- based on Bayes' Theorem. ML classi®cation proceeds
ticularly between the classes of interest, so that data by selecting the largest posterior probability rather
on texture can be used to inform classi®cation. In the than the minimum distance. In the univariate case, the
second group, spatial information is used to smooth simple empirical method for ML classi®cation
the classi®ed image. The rationale for smoothing is described by Carr (1996) proceeds as for c-means
that inaccuracies that arise from simple spectral classi- classi®cation except that for each training site the com-
®cation applied on a pixel-by-pixel basis can be plete distribution of values is retained and a histogram
reduced using the spatial dependence between neigh- computed in place of the mean. For 8-bit remotely
bouring pixels. Proximate pixels are likely to be similar sensed imagery, each histogram (per class) has 256
(where the spatial resolution is ®ne relative to the scale bins. The number of occurrences in each bin relative to
of variation) and this dependence can be formalised the total number of occurrences determines the con-
(e.g., in a modelled variogram) and utilised to increase ditional probability distribution (conditional because it
classi®cation accuracy. The goal is to choose a smooth- is per class). Once the conditional probability distri-
ing function based upon this variogram model. butions are computed, the simpli®ed equation for
The two approaches (texture and smoothing) to determining class membership probability, as derived
using the variogram in multivariate classi®cation are from Bayes' Theorem, is:
considered separately in the following sections. First,
the principles of supervised spectral classi®cation are p
z j c
p
c j z
2
revisited to provide a foundation and context within X
t
p
z j r
which to regard geostatistical approaches.
r1
standardized statistical distance from which the a (obtained from training data) using simple geometry.
posteriori probability of class membership can be This could provide an image of t distances to each
obtained using Bayes' Theorem. The maximum like- class mean. Since the distributions of the training data
lihood classi®cation is achieved by allocating each are likely to form ellipsoids in feature space the Eucli-
pixel to the class with which it has the highest a dean metric is often replaced with the Mahalanobis
posteriori probability of membership. distance Mci which takes into account the variance±co-
It is useful to view multivariate, and in particular variance matrix Vc associated with a given class c:
ML, classi®cation as a series of stages from the initial
Mci
zk
xi ÿ ukc T Vÿ1
c
zk
xi ÿ ukc
3
remotely sensed image to the classi®ed image, each of
which contains potentially useful spatial information where zk (xi) is the vector of {k = 1, . . . ,K } waveband
for classi®cation purposes (see Fig. 1). For example, values at pixel locations xi and ukc is the vector of
start with the remotely sensed image of K wavebands. means in K wavebands for class c. In Fig. 1, the var-
The information in the image can be viewed in (spec- iance±covariance matrices per class are represented
tral) feature space (each waveband providing a dimen- graphically by probability contours. It is clear that
sion or axis) instead of geographical space. In this whereas the pixel to be classi®ed is closer to the Veg-
feature space it is possible to compute Euclidean dis- etation class mean, it is more likely to be classi®ed as
tances from each pixel to each of the class means Tarmac2 when the variance±covariance matrix is
Fig. 1. Stages in maximum likelihood classi®cation: (1) original imagery, (2) Mahalanobis distances, (3) probabilities derived from
probability density function (pdf), (4) a posteriori probabilities derived from Bayes theorem and (5) ®nal classi®ed image.
364 P.M. Atkinson, P. Lewis / Computers & Geosciences 26 (2000) 361±371
bands' (features), one for each lag. To these features, where h is the lag distance and c0 is the nugget var-
the original image wavebands could be added allowing iance, c1 is the structured variance and a' is the non-
a combined spectral and textural classi®cation. linear parameter: the parameters of the exponential
Several dierent methods of supervised classi®cation model. Given the theoretical value of the covariance C
could be used to allocate pixels on the basis of the tex- (equal to c0+c1) where the variation is second-order
ture information provided by the semivariances at stationary (Eq. (7)):
dierent lags. Miranda and Carr (1994) used a paralle-
lepiped classi®er because it was simple and computa- 1h i
C lim g
h lim mean
z
xi h ÿ
z
xi 2
tionally ecient. They used the standard deviation to h41 h412 x2V
particular, several coecients which exploited class optimally an unknown value. In kriging, the variogram
speci®c periodicity in the variogram. Herzfeld and Hig- is used to determine optimal weights li to apply to n
ginson combined linear and non-linear coecients to sample data {z(xi), i = 1,2,3, . . . ,n ) to form a
form a feature vector for each pixel upon which classi- weighted linear combination z(x0) which estimates
®cation was based. some unknown value z(x0), thus:
The uses of variogram model coecients noted
above all relate the texture measure empirically to class X
n
z
x0 li z
xi :
9
properties to provide a classi®cation. It is worth noting i1
that such measures have also been related both empiri-
cally, and using physically-based scene models, to con- The advantage and main attraction of kriging is that it
tinuous variables such as tree size and density in estimates optimally by referring to the variogram
forested areas. St-Onge and Cavayas (1995), for (model) estimated from the data themselves. It would
example, related forest structural parameters empiri- be attractive also if this advantage could be transferred
cally to directional variogram properties. Jupp (1997) to multivariate classi®cation. Two sets of authors have
describes a physically-based model for variance and used the variogram in the kriging sense as a spatial
variograms as a function of viewing and illumination weighting function in unsupervised, multivariate classi-
angles based on geometric optics. Any such mapping ®cation (Bourgaullt et al., 1992; Oliver and Webster,
from variogram characteristics to continuous variables 1989). While supervised classi®cation is the primary
could be applied to classi®cation. There are two par- goal, these two papers are reviewed brie¯y before con-
ticular points of note in such work: (i) directional sidering the potential problems of applying the
(rather than omnidirectional) variograms are typically approach to supervised classi®cation in remote sensing.
used and (ii) physically-based models can describe how
the variogram changes as the viewing and illumination 4.1. Oliver and Webster's original idea
angles change.
Some 10 years ago Oliver and Webster (1989) pro-
3.3. Problems with variogram-based texture posed and demonstrated a geostatistical basis for the
classi®cation spatial weighting of multivariate classi®cation for ap-
plication in soil survey. Their approach was particu-
The main problem with using the variogram as a larly suited to sample data provided as a regular lattice
measure of texture is that the homogeneous regions of or complete cover in the form of a raster array. It
dierent texture within the image must be suciently would seem, therefore, that their approach might also
large to allow computation of the variogram up to a be suitable for classi®cation in remote sensing.
reasonable number of lags. In many cases, the parcels Oliver and Webster's (1989) proposal built on the
of interest in the image are too small relative to the work of others, notably Webster and Burrough (1972)
spatial resolution of the imagery. Berberoglu et al. who ®rst introduced the idea of modifying the dissimi-
(2000) address this issue by computing texture within larity matrix by a non-linear function of separating
parcels de®ned a priori using vector data. distance in geographical space. However, whereas
The main problem with variogram model-based Webster and Burrough chose non-linear functions
approaches to texture classi®cation is that automatic (inverse distance square and exponential) arbitrarily,
®tting of (non-linear) models to variograms is unreli- Oliver and Webster's proposal was to use a function
able. Thus, the choice of variogram model may be obtained from the data themselves. The method is
inappropriate for certain regions of the image or for described brie¯y as follows.
certain classes and the coecients of the model ®tted The similarity matrix may be constructed for all
to the local variogram may be misleading. pairs of observations i and j (pixels) on which K prop-
erties (wavebands) have been measured using a simi-
larity coecient such as Gower's (1971) coecient
(Eq. (10)):
4. Smoothing the classi®cation
X
K
In the preceding section the variogram was used to
1ÿ j zik ÿ zjk j =rk wijk
provide information on texture for use in classi®cation. k1
Sij
10
This is dierent to the common use of the variogram X
K
Wijk
in geostatistics, that is, as a spatial weighting function k1
in kriging (Matheron, 1965, 1971). In kriging, the
objective is to smooth or average local values based on where sij is a measure of the similarity between pixels i
the variogram (or other structure function) to estimate and j, zik is the pixel value at i for class k, rk is a class-
P.M. Atkinson, P. Lewis / Computers & Geosciences 26 (2000) 361±371 367
speci®c constant and wijk is a weight. This matrix may K
h E
Z
x ÿ mM
Z
x h ÿ mT
15
then be converted to dissimilarity by (Eq. (11)):
where, Z(x) is a row vector of p second-order station-
dij 2
1 ÿ sij 1=2
:
11 ary random functions, m=E[Z(x)], and M is a p by p
positive de®nite symmetric matrix used as metric in the
The objective is to modify this dissimilarity matrix to calculation of (dis)similarities.
take account of both the geographical proximity Bourgaullt et al. also based the dissimilarities matrix
between pixels and the form of spatial variation. This on the Mahalanobis distance (Eq. (3)). Unfortunately,
may be achieved by multiplying the dissimilarity by a despite oering what would appear to be a less ambig-
function of geographical distance (Eq. (12)): uous method for selecting a variogram model, Bour-
gaullt et al. (1992) introduced a further ambiguity in
d ij dij f
xi ÿ xj :
12 that the multivariate variogram and covariogram,
despite being computed from common data, produced
If the function were the exponential model (Eq. (6)) dierent classi®cation results.
then the modi®cation would take the following form Although not discussed by Bourgaullt et al., the
(Eq. (13)): above can be attributed to what might be termed the
c1 c0 `local' and `global' operation of the two measures.
d ij dij 1 ÿ exp
ÿhij =a 0 dij :
13 Essentially, modi®cation of the dissimilarity by the
c0 c1 c0 c1
(multivariate) variogram does not modify the contri-
The modi®ed dissimilarity matrix may be used in unsu- bution of points that lie beyond the range of spatial
pervised classi®cation whether that be based on hier- correlation. In the method of Bourgaullt et al., a mean
archical clustering or non-hierarchical dynamic spatially modi®ed dissimilarity is calculated: if all
clustering (Oliver and Webster, 1989). points considered in the calculation of this were further
Oliver and Webster applied their method to three than the range from a candidate point then the
small data sets on soil properties to demonstrate its weighted dissimilarity would be the same as the
utility as a tool in classi®cation, particularly for soil unweighted value. If a mean spatially-modi®ed simi-
management purposes. The unsupervised algorithm larity measure was used, however, with the spatial
chosen was a form of non-hierarchical dynamic cluster- modi®cation performed using a multivariate co-vario-
ing that operates on orthogonal principal coordinates gram which tends to zero beyond the range then only
rather than directly on the original dissimilarity matrix. points within the range would contribute positively to
Because of the need to extract a single spatial weight- the estimate of the mean. These measures are non-line-
ing function from the multivariate feature space the arly related, so no linear transformation of modi®ed
modelled variogram of the ®rst (or ®rst few) principal similarity to modi®ed dissimilarity will produce equiv-
component(s) was used in the modi®cation (Eqs. (12) alent results using the two methods in the general case.
and (13)). This seems to undermine somewhat the The former is `global' in the sense that it includes con-
unbiasedness desired for such an approach. Further tributions from all points while the latter is `local' in
ambiguity was introduced by the suggestion of the that only points within the range contribute. Further,
authors that the range a of the model could be varied in the `global' case, the eect of spatial modi®cation
to achieve dierent amounts of smoothing as desired. will depend on the image extent considered, as it will
For example, a larger range (which yields smoother vary according to the proportion of observations
results) may be more appropriate for management pur- which lie within the range.
poses. Again, this falls somewhat short of the objective
of an unbiased and even optimal solution as is the case 4.3. Alternative smoothing approaches
with kriging.
Many alternative approaches to smoothing in super-
vised classi®cation have been employed. These range
from simple low-pass ®ltering of remotely sensed ima-
4.2. Modi®cations to the method gery to more intricate processing (e.g., see the graph-
theoretic approach adopted by Barr and Barnsley,
Bourgaullt et al. (1992) proposed to replace the var- 2000). Some of these alternative approaches are dis-
iogram of the ®rst (few) principal component(s) with cussed in this section.
the multivariate variogram (Eq. (14)) and the multi- An important alternative to the spatial weighting of
variate covariogram (or covariance function) (Eq. (15)) multivariate classi®cation based on the variogram is
de®ned as follows: the Gibbs sampler. Based on Bayes' Theorem, it allows
one to incorporate information in neighbouring pixels
2G
h E
Z
x ÿ Z
x hM
Z
x ÿ Z
x hT
14 into the spectral classi®cation procedure. The Gibbs
368 P.M. Atkinson, P. Lewis / Computers & Geosciences 26 (2000) 361±371
sampler updates iteratively the predicted class of a for dierences in class homogeneity. There is much
pixel (chosen at random) conditional upon the pre- overlap between such an approach and those discussed
vious values of all other pixels (and particularly neigh- above.
bouring pixels). The iteration is stopped when the There are many algorithms available for segmenting
sequence converges to a (suciently) stable solution. remotely sensed images (Haralick and Shapiro, 1985).
In some sense then, this solution may be regarded as While a review of this literature is beyond the scope of
optimal (SchroÈder et al., 2000). However, it is only op- the present paper, it is worth pointing out that segmen-
timal in that the ®t achieved using the Gibbs sampler tation is fundamentally dierent to traditional spectral
attains maximum pseudo likelihood for the given classi®cation because spatial contiguity is an explicit
choice of spatial weighting function and kernel size goal of segmentation whereas it is only implicit in
(Augustin et al., 1996). In most instances, it is necess- classi®cation. There are many approaches to segmenta-
ary to experiment with dierent choices for the weight- tion including those based on edge detection (most
ing function and kernel size to search for a generally often based on some high-pass ®lter) and those based
optimal solution. on region-growing (based on the growth of hom-
van der Meer (1996, 1999) used indicator kriging ogenous regions conditional upon similarities between
(Goovaerts, 1997) applied to multivariate data to the pixel to be merged and previously merged pixels).
obtain a classi®cation for all pixels in a remotely There exist many examples of segmentation applied to
sensed image. The approach involved de®ning indi- remotely sensed imagery (Janssen and Molenaar, 1995;
cator variables for each feature (waveband) in an Khodja and Mengue, 1996; Lemoigne and Tilton,
image and obtaining variograms for each indicator. 1995; Lobo et al., 1996; Ryherd and Woodcock, 1996).
These variograms were then used in block indicator Segmentation routines involving region-growing al-
kriging to estimate the average value of each indicator gorithms may be based on the similarities between var-
for a block or area of pixels centred on the pixel to be iograms or the coecients of variogram models (Lloyd
classi®ed. This amounts to smoothing of the tra- and Atkinson, 1998) resulting in segmentation based
ditional classi®er. However, importantly the spatial in- on variogram texture.
formation (weighting function) incorporated into the
classi®cation via the variogram is derived from the
form of spatial variation in the variable itself as is
desired. The ambiguity in this approach would appear
to come from the initial selection of indicator cut-os 5. Discussion: spatial weighting for remote sensing
and the size of blocks (i.e., the amount of smoothing). classi®cation
Further, extrapolation at the tails of the distribution
can alter the results substantially.
The above indicator approach is similar to an 5.1. Selecting an appropriate space
approach known as regionalised classi®cation (Bohling,
1997; Harf and Davis, 1990; Moline and Bahr, 1995). One of the ®rst decisions facing the analyst wishing
Regionalised classi®cation is, in fact, nothing more to use a spatial weighting in smoothing a classi®er for
than the interpolation to unobserved sites of the inputs remote sensing is in which space should the spatial
to or the outputs from some traditional classi®er (geographical) weighting be determined and applied?
applied to sparse data. However, papers on regiona- The work on regionalised classi®cation of Harf and
lised classi®cation have been useful for pointing out Davis (1990), Moline and Bahr (1995) and Bohling
that dierent stages in the classi®cation process can be (1997) (among others) illustrates clearly that there are
interpolated (or in the present case smoothed). For many possible spaces in which the smoothing can take
example, one could interpolate the feature vectors and place. These include (see Fig. 1):
then proceed with classi®cation at all sites. Alterna-
tively, one could interpolate the Mahalanobis dis- 1. multivariate feature space (for example, the actual
values in the wavebands),
tances, or the probabilities for use in ML classi®cation.
This choice, which is entirely general, represents an im- 2. Euclidean distance-to-means in multivariate feature
portant decision for investigators wishing to devise space,
3. Mahalanobis distance-to-means in multivariate fea-
strategies for incorporating spatial information into a
classi®cation. ture space,
4. probabilities (per class),
Following on from the above theme, Palubinskas et
5. a posteriori probabilities (per pixel) and
al. (1995) applied a smoothing algorithm to the fuzzy
(for example, a posteriori Bayesian probabilities) out- 6. the classi®ed image.
puts from the classi®cation of a remotely sensed The spatial weighting function could be estimated in
image. The algorithm was modi®ed locally to account any one of these spaces and used to aid classi®cation.
P.M. Atkinson, P. Lewis / Computers & Geosciences 26 (2000) 361±371 369
5.2. A geostatistical basis for supervised classi®cation? depend on a stationary variogram. It should be noted
that the Gibbs sampler does not necessarily help in
this regard. The Gibbs sampler is a generally appli-
For several reasons Oliver and Webster's (1989)
cable ®tting procedure most useful for Markov chains.
approach without modi®cation is unlikely to be of
It does not necessarily allow for non-stationarity
much utility for the classi®cation of remotely sensed
although it can be used to ®t models which do so to
imagery. First, in remote sensing unsupervised classi®-
some extent (for example, Brunsdon, 2000). In most
cation is used less frequently than supervised classi®-
implementations, the result of the Gibbs sampler will
cation and for the latter there is no dissimilarity
be a maximum pseudolikelihood ®t for a selected
matrix. The equivalent to the dissimilarity matrix, the
smoothing function and kernel size ®xed over the
distances between the pixel values and the class means
entire image. A potential solution is provided by a per-
expressed in terms of some distance metric in feature
parcel classi®er that utilises a priori knowledge in the
space, do not have a spatial dimension (they are not
form of digital vector boundaries to segment the region
expressed as a function of lag). Each class mean is
of interest into distinct parcels prior to classi®cation.
de®ned over the whole image, not for a single pixel
Within each parcel the spatial weighting could be
located in geographical space. These distances are,
applied independently. The per-parcel classi®er would
therefore, less readily modi®ed by a function de®ned in
also limit naturally the number of data to a tolerable
geographical space and the variogram cannot be
value.
employed in the same way as for unsupervised classi®-
cation. The most straightforward approach would be
to use the (multivariate) variogram as a smoothing
function directly on the images of distances-to-class-
means (Fig. 1). However, the simple application of the
variogram as a smoothing function (that is, convolu- 6. Conclusions
tion of the images of distances with a kernel-based
weighting function) is unlikely to result in optimal use Most techniques that use the variogram to classify
of the spatial correlation in the image. The Gibbs sam- remotely sensed images have their shortcomings. For
pler discussed above is attractive in this regard because example, texture classi®ers based on the variogram
of the lack of an explicit structure function. work only where the homogenous regions of each class
Second, the data sets must be small for the algor- in the image are suciently large, homogenous and
ithm to be eective (for example, one of the data sets dierent texturally between classes. Even then there is
analysed by Oliver and Webster consisted of 6 14 no guarantee that the extra data on texture will yield
cells) and in remote sensing the data sets are typically useful information (above that provided by the original
large (commonly in excess of 1000 1000 pixels). imagery). Approaches that use the variogram for
Despite the problems discussed above, supervised smoothing in classi®cation may be divided into two
classi®cation does have the advantage of removing the groups: those based on simple ®ltering of the image (at
need to compare distances between all pixels and all any stage in the classi®cation process) and the method
other pixels (involving nearly 500,000,000,000 compari- of Oliver and Webster (1989) in which the variogram
sons for an image of 1000 1000 pixels, which would is used to modify the dissimilarities in unsupervised
be prohibitive in most cases). Therefore, small data classi®cation. The former approaches do not deliver
sets are not a prerequisite for supervised classi®cation. the optimality associated with geostatistical techniques
Finally, the variogram is de®ned as a parameter of a such as kriging. The latter approach holds much prom-
RF model which is stationary in the squared dier- ise, but problems will need to be overcome if it is to be
ences between pairs of locations separated by a given applied in remote sensing.
lag h (referred to as the intrinsic hypothesis). For
many remotely sensed images the entire scene is not
readily modelled in such a way. In particular, for land-
scapes aected by human activity (for example, agri-
cultural ®elds, forest stands, urban areas and so on)
the objective of classifying an entire image using a Acknowledgements
single variogram is ultimately ¯awed because the
stationary RF model is unjusti®ed. Where a stationary This paper was written while PMA was on leave at
variogram model cannot reasonably be considered, but the School of Mathematics, University of Wales, Car-
rather the model should be allowed to vary smoothly di. The authors thank Professor Giles Foody for use-
across the image, the objective should be to adopt ful information relating to this paper and are grateful
(probably non-parametric) approaches which do not to the referees for their comments.
370 P.M. Atkinson, P. Lewis / Computers & Geosciences 26 (2000) 361±371
semivariogram textural classi®er (STC). International tional variogram. International Journal of Remote Sensing
Journal of Remote Sensing 17 (17), 3523±3529. 16 (11), 1999±2021.
Miranda, F.P., Fonseca, L.E.N., Carr, J.R., 1998. SchroÈder, M., Walessa, M., Rehrauer, H., Seidel, K., Datcu,
Semivariogram textural classi®cation of JERS-1 (Fuyo-1) M., 2000. Gibbs random ®eld models: a toolbox for infor-
SAR data obtained over a ¯ooded area of the Amazonian mation extraction. Computers & Geosciences 26 (4), 423±
rainforest. International Journal of Remote Sensing 19 (3), 432.
549±556. Smith, G.R., Woodward, J.C., Heywood, D.I., Gibbard, P.L.,
Miranda, F.P., Macdonald, J.A., Carr, J.R., 1992. 2000. Interpreting pleistocene glacial features from SPOT
Application of the semivariogram textural classi®er (STC) HRV data using fuzzy techniques. Computers &
for vegetation discrimination using SIR-B data of Borneo. Geosciences.
International Journal of Remote Sensing 13 (12), 2349± Thomas, I.L., Benning, V.M., Ching, N.P., 1987.
2354. Classi®cation of Remotely Sensed Images. Adam Hilger,
Moline, G.R., Bahr, J.M., 1995. Estimating spatial distri- Bristol.
butions of heterogenous subsurface characteristics by van der Meer, F., 1996. Classi®cation of remotely-sensed ima-
regionalized classi®cation of electrofacies. Mathematical gery using an indicator kriging approach-application to
the problem of calcite-dolomite mineral mapping.
Geology 27 (1), 3±22.
International Journal of Remote Sensing 17 (6), 1233±
Oliver, M.A., Webster, R., 1989. A geostatistical basis for
1249.
spatial weighting in multivariate classi®cation.
van der Meer, F. 1999. Geostatistical approaches for image
Mathematical Geology 21 (1), 15±35.
classi®cation and assessment of uncertainty in geologic
Palubinskas, G., Lucas, R.M., Foody, G.M., Curran, P.J.,
processing. In: Atkinson, P.M., Tate, N.J. (Eds.),
1995. An evaluation of fuzzy and texture-based
Advances in Remote Sensing and GIS Analysis. Wiley,
approaches for mapping regenerating tropical forest classes
Chichester, pp. 147±166.
from Landsat-TM data. International Journal of Remote Wallace, C.S.A., Watts, J.M., Yool, S.R., 2000.
Sensing 16 (4), 747±759. Characterizing the spatial structure of vegetation commu-
Raghu, P.P., Poongodi, R., Yegnanarayana, B., 1995. A com- nities in the Mojave desert using geostatistical techniques.
bined neural-network approach for texture classi®cation. Computers & Geosciences 26 (4), 397±410.
Neural Networks 8 (6), 975±987. Webster, R., Burrough, P.A., 1972. Computer-based soil map-
Ramstein, G., Ray, M., 1989. Analysis of the structure of ping of small areas from sample data. II. Classi®cation
radiometric remotely-sensed images. International Journal smoothing. Journal of Soil Science 23, 222±234.
of Remote Sensing 10 (6), 1049±1073. Woodcock, C.E., Strahler, A.H., Jupp, D.L.B., 1998a. The
Ryherd, S., Woodcock, C., 1996. Combining spectral and tex- use of variograms in remote sensing I: Scene models and
tural data in the segmentation of remotely sensed images. simulated images. Remote Sensing of Environment 25 (3),
Photogrammetric Engineering and Remote Sensing 62 (2), 323±348.
181±194. Woodcock, C.E., Strahler, A.H., Jupp, D.L.B., 1998b. The
St-Onge, B.A., Cavayas, F., 1995. Estimating forest stand use of variograms in remote sensing II: real digital images.
structure from high-resolution imagery using the direc- Remote Sensing of Environment 25 (3), 349±379.