Geomorphology, 8 (1993) 245-262 245

Elsevier Science Publishers B.V., Amsterdam

Fractals, fractal dimensions and landscapes a review

Tingbao Xu*, Ian D. Moore and John C. Gallant

Centre for Resource and Environmental Studies, The Australian National University, GPO Box 4, Canberra, ACT 0200, Australia
(Received May 1, 1992; revised July 22, 1993; accepted July 28, 1993 )

ABSTRACT

Mandelbrot's fractal geometry is a revolution in topological space theory and, for the first time, provides the possibility
of simulating and describing landscapes precisely by using a mathematical model. Fractal analysis appears to capture some
"new" information that traditional parameters do not contain. A landscape should be (or is at most) statistically self-
similar or statistically self-affine if it possesses a fractal nature. Mandelbrot's fractional Brownian motion (fBm) is the
most useful mathematical model for simulating landscape surfaces. The fractal dimensions for different landscapes and
calculated by different methods are difficult to compare. The limited size of the regions surveyed and the spatial resolution
of the digital elevation models (DEMs) limit the precision and stability of the computed fractal dimension. Interpolation
artifacts of DEMs and anisotropy create additional difficulties in the computation of fractal dimensions. Fractal dimen-
sions appear to be spatially variable over landscapes. The region-dependent spatial variation of the dimension has more
practical significance than the scale-dependent spatial variation. However, it is very difficult to use the fractal dimension
as a distributed geomorphic parameter with high "spatial resolution". The application of fractals to landscape analysis is
a developing and immature field and much of the theoretical rigour of fractal geometry has not yet been exploited. The
physical significance of landscape fractal characteristics remains to be explained. Research in geographical information
theory and fractal theory needs to be strengthened in order to improve the application of fractal geometry to the geosciences.

Introduction matical description of complex natural land-

scapes and provides a systematic way of char-
Over the last decade, fractal models and re- acterizing landscapes in quantitative terms.
lated analysis techniques have shifted from Milne (1991 ) stated that, "fractal models of
being matters for speculative coffee break dis- landscape structure provide the elements of a
cussions between sessions at geoscience meet- calculus for quantifying and predicting the
ings, to being primary topics for numerous multiscale dynamics of landscape processes."
technical sessions (Snow and Mayer, 1992). Meakin ( 1991 ) also stated that fractal geome-
The term "fractal" was coined by Mandelbrot try provides a much more complete and real-
more than fifteen years ago. At that time it was istic description of most structures in geology
not used primarily to describe landscapes. It and geophysics than does Euclidean geometry
has since become the most successful mathe- and will revolutionize the study of geomor-
matical model for describing real landscapes phology. So, although fractal analysis is a rela-
because the fractal dimension appears to cap- tively new field, it has been used in a large
ture the essence of the surface topography of number of biological and geoscience studies,
the earth in a way that other geomorphological particularly for the characterization of land-
attributes do not. scapes. The concepts loosely associated with
Fractal geometry makes possible a mathe- the term "fractal" have broad general appeal
(Goodchild and Mark, 1987 ).
*Corresponding Author. Fractal analyses for characterizing the spa-

0169-555X/93/$06.00 © 1993 Elsevier Science Publishers B.V. All rights reserved.

246 T.-B. XU ET AL.

tial relationships of the earth's surface, as well topographic attributes that can be used to de-
as methods for calculating the fractal dimen- scribe landforms. Klinkenberg ( 1992 ) assem-
sion, have been reported by Armstrong and bled 24 traditional landscape morphometric
Hopkins ( 1983 ), Shelberg et al. ( 1983 ), Mark attributes when he analysed the relationships
and Aronson (1984), Brown and Scholz between the fractal parameters and these attri-
(1985), Clarke (1986), Kennedy and Lin butes. They included the mean, standard de-
( 1986 ), Brown ( 1987 ), Goodchild and Mark viation, skewness and kurtosis of elevation,
(1987), Laverty (1987), Roy et al. (1987), slope gradient, aspect and plan and profile cur-
Culling (1988, 1989), Milne (1988), Andrle vature and some special morphometric pa-
and Abrahams (1989), Dubuc et al. (1989), rameters such as elevation range and the coef-
Elliot (1989), Fox (1989), Gibert (1989), ficient of dissection. All of these attributes can
Hough (1989), Huang and Turcotte (1989), be extracted from a digital elevation model.
Jones et al. ( 1989 ), Longley and Batty ( 1989 ), In using landscape attributes in modelling,
Unwin (1989), Whalley and Orford (1989), questions can be raised, such as: - - are these
Kumar and Bodvarsson (1990), Piech and parameters good enough to characterize land-
Piech (1990), Klinkenberg ( 1992 ), Klinken- scape properties? - - which of them is most ef-
berg and Goodchild (1992), Ouchi and Mat- ficient? - - is it reasonable to use these param-
sushita (1992), Gallant et al. ( 1993 ). Hjelm- eters in the models encompassing them? - - are
felt ( 1988 ), Tarboton et al. ( 1988 ), Gan et al. there new landscape attributes that perform
(1992), Hatano and Booltink (1992), and better? Evans (1984) carried out a principal
Karlinger and Troutman (1992) have used component analysis of 23 landscape mor-
fractal analysis techniques to investigate the phometric variables and his results indicated
structure of river networks, and the river- that the first seven components only incorpo-
length/catchment-area ratios. Their work has rate 89% of the information of the original
advanced our understanding of the geometry morphometric variables and the first compo-
and composition of river networks (Tarboton nent did not dominate the results. A landscape
et al., 1988). is therefore a complex surface and no single or
Fractal analysis is now a growing field of re- simple linear combinations of these mor-
search and an extensive literature concerned phometric variables can describe it compre-
with the fractal nature of landscapes is devel- hensively. However, it has also been implied
oping. This paper summarizes fractal and re- that the information content of landscape
lated theories and reviews the last decade's re- properties contained in each morphometric
search on the fractal analysis of landscapes. variable is low, with each one only able to de-
scribe a restricted aspect of landscape proper-
Landscapes and landscape parameters ties. New parameters that better characterize
the variability of landscapes are continually
"We are now in an era of spatial modelling" being sought. The fractal dimension may be
(Moore et al., 1991 ). Most geo-processes oc- one of them.
cur on the land surface and they have close re- There are explicit differences between the
lationships with landscape properties. The rel- theoretical basis of fractals and those of tradi-
ative magnitudes of many hydrological, tional landscape morphometric parameters.
geomorphological and biological processes are The main characteristic is whether or not the
sensitive to topographic position (Moore et al., parameter displays a homogeneous or contin-
1991 ). Many parameters have been developed uous characteristic. Fractals treat the land-
for characterizing landscapes and related scape region being investigated as a distinct
properties. Speight (1974) described over 20 entity and the fractal nature of the region is
FRACTALS, F R A C T A L D I M E N S I O N S A N D L A N D S C A P E S - - A R E V I E W 247

initially assumed to be homogeneous. There- in topological space theory (this revolution did
fore, one can use mathematical functions (i.e., not originate from Mandelbrot), but also, for
models) to describe certain characteristics of the first time, provides the possibility of de-
the landscape region (fractal characteristics) scribing and simulating landscapes precisely by
and one should obtain a uniform (consistent) using a mathematical model.
fractal dimension. Traditional landscape pa- The key idea introduced into applied math-
rameters do not behave in the same way. For ematics by Mandelbrot is that rugged and in-
example, slope gradient depends on the char- determinate systems can often be described by
acteristics of a small local area surrounding the extending the classical concept of dimensional
point of measurement. Generally, the charac- analysis to include a fractional number that
teristic value at individual points can be ex- describes the ruggedness of the system in the
tracted from the elevation values of a neigh- space spanned by the whole number dimen-
bouring point subset in the DEM matrix using sions encompassing its fractional magnitude
a variety of algorithms. The computed slope (Kaye, 1989). It means that we can use the
gradient is scale-dependent and pixel-depen- fractal dimension to describe the irregularity
dent and the entire slope "surface" of the in- and roughness characteristics of natural
vestigated region is not continuous in a math- landscapes.
ematical sense, but is perhaps continuous in a The manifestation of a fractal is via the frac-
visual sense. Obviously other traditional land- tal dimension, usually expressed by the symbol
scape parameters also display similar proper- D. The fractal dimension indicates the ability
ties to slope gradient. of a set to fill the Euclidean space in which it
resides, and is the quantitative description for
Fractals and fractal dimensions of landscapes the fractal characteristics of the investigated
objects. In other words, the fractal dimension
"Fractal" is a word invented by Mandelbrot is generally a measurement of the irregularity
( 1977 ) to bring together under one heading a of the object in a mathematical sense. Mandel-
large class of objects that have certain struc- brot called the Hausdorff-Besicovitch dimen-
tural features in common, although they ap- sion the "fractal dimension" principally be-
pear in diverse contexts in astronomy, geog- cause the values of Hausdorff-Besicovitch
raphy, biology, fluid dynamics, probability dimensions are non-integer real numbers
theory, and pure mathematics (Dyson, 1978 ). (could be integer) in most cases. Precise
According to Mandelbrot (1977, 1982), the mathematical definitions of the Hausdorff-
term "fractal" comes from the Latin adjective Besicovitch dimension can be found in the
"fractus", which has the same root as "frac- works of Mandelbrot ( 1977, 1982), Falconer
tion" and "fragment" and means "irregular ( 1991 ) and Tricot ( 1991 ).
and fragmented". Mandelbrot (1982) stated There are less sophisticated, but neverthe-
that he conceived and developed a new geom- less relevant, definitions of fractal dimensions.
etry of nature (fractals) and implemented its "In landscape research, fractal measurements
use in a number of diverse fields. It describes provide information about the space-filling
many of the irregular and fragmented patterns properties of a mosaic of patches at all scales"
around us. Mandelbrot's ( 1977, 1982) math- (Milne, 1991 ); "D provides a characteristic
ematical definition of fractal is: "a set for which parameter whose variation can be usefully in-
the Hausdorff-Besicovitch dimension strictly terpreted in terms of the processes that have
exceeds the topological dimension." The no- influenced the entity's development" (Good-
tion of a fractal not only is quite new for tra- child and Mark, 1987 ); "... fractal dimensions
ditional applied mathematics and a revolution are real numbers in which the value of the
248 T.-B. XU ET AL.

whole number describes the nature of the data nature should be homogeneous. In fact there is
set under observation (0 being points, 1 being currently great controversy over whether or not
lines, 2 planes and so on) and the size of the self-similar landscapes exist and if fractal di-
decimal fraction represents the irregularity ex- mensions are stable in real landscapes. In the
hibited within the data" (Elliot, 1989). Intui- fractal analysis of landscapes, self-affinity is
tively, a small fractal dimension corresponds generally more common and is applicable un-
to a smooth surface, and a large value corre- der a wider range of conditions than is self-
sponds to a rough surface. Some people have similarity.
concluded that it also captures other character- Landscapes are a composite of many com-
istics of the landscape (Klinkenberg, 1992; peting and complicated geological processes,
Klinkenberg and Goodchild, 1992 ). such as faulting, folding, flexure, erosion and
sedimentation, etc. "A wide variety of natural
Self-similarity, self-affinity and statistical phenomena exhibit complicated, unpredicta-
fractals ble, and seemingly random behaviour" (Jen-
sen, 1987). Milne ( 1991 ) indicated that, "ex-
The prerequisite for the measurement of the act fractal patterns are unlikely to occur in
fractal dimension of an object is that it dis- landscapes because contemporary patterns are
plays self-similarity or self-affinity. Therefore, the results of several processes that dominated
self-similarity or self-affinity is regarded as a in the past. So far, little is known about how
fundamental characteristic of fractal objects the legacies of different processes (e.g., ero-
and is one of the central concepts of fractal ge- sion versus deposition) accumulate to pro-
ometry. Mandelbrot ( 1977, 1982) gave the duce existing fractal landscape patterns." Shel-
following definition of self-similarity: "when berg et al. (1983) also commented that,
each piece of a shape is geometrically similar "seldom in nature (crystals are one exception)
to the whole, both the shape and the cascade does self-similarity occur and therefore a sta-
(the generating mechanism for the details of tistical form of self-similarity is often encoun-
the shape) that generate it are called self-simi- tered." Under statistical self-similarity, each
lar." Self-similarity assumes that transforma- small portion of the object, when it is magni-
tions in each direction of Euclidean coordinate fied, is not exactly like a larger portion or the
space are the same, although objects may be whole, but looks like a larger portion or the
rotated. If the transformations are different in whole. In other words, each portion of the ob-
each direction, then the object is self-affine ject is statistically indistinguishable from the
rather than self-similar (Mandelbrot, 1982, whole. It is clearly impossible that each por-
1986a; Shelberg et al., 1983; Goodchild and tion of a landscape is geometrically (exactly)
Mark, 1987; Roy et al., 1987; Milne, 1991; similar to the whole in most cases. Therefore,
Ouchi and Matsushita, 1992). a landscape should be called statistically self-
Self-similarity is better termed scale-invari- similar or statistically self-affine if it possesses
ance (or scale-independence) in landscapes. a fractal nature. Stanley (1986) divided frac-
This means that for any line or surface of a self- tals into exact fractals and statistical fractals.
similar object, a portion of the whole is iden- According to his interpretation, exact fractals
tical to the whole after suitable transforma- are regular fractals that display self-similarity,
tions, or "each small portion, when magnified, are highly artificial, and are not expected to
can reproduce exactly a larger portion" (Voss, appear in nature. Statistical fractals exhibit
1985a). Therefore, the value of the fractal di- fractal characteristics when average properties
mension should be stable when the range scale are examined, are statistically self-similar, and
changes in a fractal landscape, and its fractal can appear in nature. It suffices that the parts
FRACTALS,FRACTALDIMENSIONSAND LANDSCAPES-- A REVIEW 249

and the whole reduced by similarity should

have identical (statistical) distributions
E[ (B( t 2 ) - B ( tl ) )2] = It2-fi 12/4 (2)
(Mandelbrot, 1982 ). The B(t) with H = 0 . 5 is also called the ordi-
As with statistically self-similar fractals, sta- nary Brown line-to-line function and its trail is
tistically self-affine objects have been pro- a curve. When one changes the exponent H
posed. The hypothesis of statistical self-affin- from H = 0 . 5 to any real number satisfying
ity should be more appropriate than that of 0 < H < 1, then B(t) is called the fractional
statistically self-similarity in determining the Brown line-to-line function (fractional Brown
fractal characteristics of the surface of the function ), and denoted by B/4 (t), for which:
earth. For land surfaces, the vertical size range
is usually much less than the horizontal size E[(B/4(t2)-B/4(tl))2] = It2-tl 12/4 (3)
range. Therefore, the vertical ratio under For a surface, the single variable t is replaced
transformations is clearly less than the hori- by point coordinates x and y on a plane to give
zontal ratio. B/4(x,y) as the surface altitude at position x,y
(i.e. Z). The surface that consists of these
Calculating the fractal dimension B/4(x,y) points is usually called a fractional
Brownian surface. A fractional Brownian sur-
Fractional Brownian model and landscapes face B/4(x,y) has the following properties
(Mandelbrot, 1977, 1982, 1986a):
(a) The surface is continuous, but nondif-
Estimating fractal dimensions and other ferentiable, and everything concerning this
fractal parameters of landscapes is the princi- surface is dependent on the single scaling pa-
pal way of quantitatively describing the fractal rameter H.
properties of landscapes. Until now, Mandel- (b) The plane (x,y) is isotropic.
brot's fractional Brownian motion (fBm), or (c) The fractal dimension of the surface is
fractional Brown function, which is the math- D=3-H.
ematical generalization of fBm, probably pro- (d) The section profile in any direction, i.e.,
vides the most useful mathematical model for the intersection of a vertical plane with the
the random fractals found in nature, particu- surface B/4(x,y), is self-affine and D = 2 - H ;
larly for application to landscape analysis. i.e., the surface dimension minus one. Usually,
The fundamental form of Mandelbrot's the profile is a curve of the fractional Brown
fractional Brownian model is the Brown line- line-to-line function.
to-line function B (t), which is a random func- (e) The traverse at any altitude, i.e., the in-
tion and for t~ and t2 (t~ ~ t2 ) it satisfies: tersection of the horizontal plane with the sur-
P{ [B(t2) - B ( h ) ]/It2 - tl I/4<x} = F ( x ) face B/4(x,y), which produces a family of
(1) curves, is self-similar and D = 2 - H . This
means that coastlines and contours are statis-
where t and x are real numbers, H = 0.5 and
tically self-similar when the B/4(x,y) is used to
F(x) is the cumulative contribution function simulate the surface of the landscape.
for a random variable X = [ B ( t 2 ) - B ( t l ) ] /
It2-q I/4 (or X = B ( t 2 ) - B ( f i ) , because for (f) E[ (BH(X2,Y2) --BH(X, ,y~ ) )21
fixed (t2-fi), It2-tl Inis a constant) that fol-
=C[ ((x 2-x 1)2
lows a Gaussian distribution with E [ X ] = 0
(i.e., the expected value of Xis equal to 0) and -I- (Y2 --Yl ) 2 ) 1/2] 2H
E[X2]=I ( E [ X 2 ] = Itz-t,I/4 for X=B(t2) =C[(x2--Xl)2-l-(y2-Yl)2] H (4)
- B ( fi ) ). This means that:
250 T.-B.XU ET AL.

where C is a constant that replaces the original Variogram method

constant 1 in the single variable fractional The variogram method (Mandelbrot, 1977,
Brown function. 1982; Goodchild, 1980; Mark and Aronson,
(g) The fractional Brownian surface is self- 1984; Pentland, 1984; Voss, 1985a,b; Roy et al.,
affine. 1987; Gallant et al., 1993) is a well-known
method to directly measure the surface fractal
Methods of computing fractal dimensions dimension of landscapes. It is based on the as-
sumption that the landscape surface is a frac-
Mandelbrot ( 1977, 1982), Shelberg et al. tional Brownian surface and uses eq. (4) with
(1983), Mark and Aronson (1984), Voss C and H as constants for a given surface.
(1985a,b), Clarke (1986), Roy et al. (1987), Therefore, for a given surface, the mean square
Culling (1988, 1989), Milne (1988), Fox altitude difference E [ (BI-I(xa,y2)
( 1989 ), Yokoya et al. ( 1989 ), Ouchi and Mat- --BH(xl,yl ) )2 ] depends on the horizontal dis-
sushita (1992), and others have developed a tance h between points (xl,y~) and (x2,Y2).
range of methods for estimating the fractal di- The term E[ (BH(X2,Y2) --BH(XI,yl ) )2] is the
mension of a landscape. The principal meth- variance of the random variable
ods are presented below. X=BI-I(x2,y2) --BH(xl,yl ) because the mean,

Fig. 1. Shaded relief diagram of the 512 × 512 pixel DEM from the northern section of the Brindabella Range, southeast-
ern Australia ( 10 m × l0 m grid spacing).
FRACTALS,FRACTALDIMENSIONSAND LANDSCAPES-- A REVIEW 251

could be fitted, with a correlation greater than

0.9, to the first linear segment in the variogram
10000
J
plot.
Figure 1 i s a 5 1 2 × 5 1 2 p i x e l , 1 0 m × 1 0 m
1000
grid DEM of a 26 km 2 area in the northern sec-
tion of the Brindabella Range of southeastern
f Australia (35°22'S, 148°48'E). The total re-
100
lief is 720 m (mean elevation= 1090 m) and
the slope gradient ranges from 0 to 200% with
f Slope = 1 80
a mean slope gradient of 34%. The DEM was
10 100 1000
produced by ANUDEM (Hutchinson, 1989),
Lag (m)
which uses a spline interpolation of contour
Fig. 2. Log-log semi-variogram plot of the D E M of Fig. 1. and spot height data while preserving sensible
The slope at short lags ( 10 to 200 m ) is 1.80 ( D = 2 . 1 0 ) .
drainage properties input as digitized stream-
For lags over 200 m, the slope is 1.35 ( D = 2 . 3 3 ) .
lines. Figure 2 is a log-log variogram plot of
the DEM based on eq. (5) and illustrates the
E[BH(X2,Y2)--BH(Xl,Yl )], is equal to zero. If
results that can be obtained with the vario-
one can determine the parameter H from eq.
gram method using traditional data sources.
(4), then, the fractal dimension D can be ob-
The slope of this function is the coefficient term
tained from D = 3 - H.
(2H) in eq. (5), and two distinct slopes are
Changing eq. (4) into logarithmic form, the
evident: 2H=1.80 ( D = 2 . 1 0 ) at short lags
following simpler relationship is obtained:
( <200 m) and 2 H = 1.35 ( D = 2 . 3 3 ) at larger
Iog{E[ (BI-I(x2,Y2)-BH(Xl,yl) )2]} (5) lags. The reasons for the variable fractal di-
mension are discussed latter.
= 2 H log(h) + l o g ( C )
The parameters H and C are usually estimated Divider method
as the slope and intercept of a least-squares line Contour lines are regarded as statistically
fitted to a log-log plot of self-similar when fractional Brownian surfaces
E[ (BH(x2,y2) -BI-I(xl,yl) )2] versus h. In or- are used to simulate the landscape surface. The
der to make the estimate of surface fractal dimension Ds can be obtained
E[ (Bn(x2,Y2) -B,v(xl,yl))Z] meaningful, a from the contour fractal dimension Dc using the
sufficient number of random sampling point relationship:
pairs are needed for each distance h. Oliver and Ds=Dc+I (6)
Webster's (1986) semi-variogram is one of The divider method (Mandelbrot, 1977, 1982;
several methods of calculating eq. (5) (Gal- Shelberg et al., 1983; Kennedy and Lin, 1986;
lant et al., 1993). A grid-based DEM is re- Laverty, 1987; Hayward et al., 1989) can be
quired for the variogram method. used to measure De. The divider method, also
Klinkenberg and Goodchild (1992) pre- called the Mandelbrot method or ruler method,
sented two auxiliary fractal characteristics, is a well-known measurement technique for
gamma and break-distance, in addition to the linear fractal characteristics. It determines the
fractal dimension. Gamma is the ordinate in- fractal dimension of linear phenomena such as
tercept of the straight line (equation 5 ) in the contours by measuring their length using a "di-
variogram plot and represents the expected vider". The general mathematical form of di-
difference in elevation for point pairs a unit vider method is:
distance apart. The break-distance is the max-
imum distance to which a least-squares line L(r)=Cr 1-o (7)
252 T.-B. XU ET AL.

where D is the fractal dimension, L(r) is the ilar linear objects. The primary data form used
length of the linear object measured, r is the with this method is a grid-based DEM.
divider step and C is a constant. Equation (7)
can also be written in log-log form as: Power spectral method
The power spectral method (Voss, 1985a,b;
log[L(r) ] = ( 1 - D ) l o g ( r ) +log(C) (8)
Fox and Hayes, 1985; Mandelbrot, 1986a; Fox,
Similarly to the variogram method, D and C 1989; Huang and Turcotte, 1989 ) can be used
can be estimated from the slope and intercept to measure the fractal dimension of a self-af-
of a linear regression equation fitted to a log- fine linear object. The section profile of the
log plot of L(r) versus r. This method yields landscape is statistically self-affine under the
meaningful results only when the linear object assumption of fractional Brownian motion and
measured is self-similar. For a self-similar lin- for which:
ear object, the smaller the step r, the longer the S v ( f ) = C f 5-2D (11)
length L(r), and the L(r)--,oo as r - 0 . The
format of the primary data used with the di- where D is the fractal dimension of the profile,
vider method is usually a vectorial one with a Sv(f) is the Fourier power spectral density of
string of point coordinate pairs on a plane. the profile data set measured, f is the fre-
quency and C is a constant. The log-log form
of eq. ( 11 ) is:
Box counting method
The box counting method (Mandelbrot, log[Sv(f) ] = ( 5 - 2D)log(f) +log(C) (12)
1977, 1982; Voss, 1985b; Milne, 1988; Jones,
1991; Meakin, 1991; Klinkenberg and Good-
Oprofae is obtained from eq. (12 ) using least-
squares regression, similar to the procedures
child, 1992), also called the cell counting
described above. The surface fractal dimen-
method, is another useful method for deter-
sion of a landscape should be Dpron~e+ 1. This
mining the fractal dimension of linear phe-
method includes the one-dimensional Fourier
nomena. It can be used to measure the fractal
transform calculation.
dimension of contour lines, Dc, and hence the
Huang and Turcotte (1989) generalized eq.
surface fractal dimension using equation (6).
( 11 ) into two-dimensions based on the work
The general mathematical form of the box
of Voss (1985a,b) and directly estimated the
counting method is:
fractal dimension of a self-affine surface.
N(r) =Cr -D (9) Huang and Turcotte (1989) proposed that in
two-dimensional form:
where D is the fractal dimension, N(r) is the
number of boxes (cells) that cover the linear S v ( f ) = C f 7-2D (13)
object measured, r is the side length of the and:
square box and C is a constant. The log-log
form of eq. (9) is: log [Sv(f) ] = ( 7 - 2D)log(f) +log(C) (14)

log[N(r) ] = - D log(r) +log(C) (10) where D is the fractal dimension of the surface,
Sv(f) is the Fourier power spectral density of
To obtain a believable value of D, one needs to the surface data set measured, and C and f a r e
count the N(r) for different side lengths, r, then the same as defined previously.
obtain the D from the data pairs N(r) and r The maximum entropy method (MEM) of
using least-squares regression, in the same way the spectral analysis technique was developed
as the methods mentioned above. The box by Burg ( 1967, 1968) in response to the prob-
counting method is usually used with self-sim- lem of truncation of the sample autocorrela-
FRACTALS, FRACTAL DIMENSIONS AND LANDSCAPES - - A REVIEW 253

tion function. Burg's method essentially ex- ing a statement to say that it characterizes only
trapolates the sample autocorrelation function "irregularity" or "roughness". The meanings of
to infinite length by choosing successive values traditional morphometric parameters, such as
that maximise the entropy of the power spec- altitude, slope, aspect, curvature, etc., which
trum (Burg, 1967 ). This method was later im- are called landform elements by Speight
proved (Burg, 1968) by avoiding the step of (1974), are intuitive and easy to understand.
calculating the autocorrelation estimates and, But fractal research is still too immature, and,
instead, directly estimating the coefficients of "the links between the fractal dimensions and
the prediction error filter from the data. The the physical processes which produce that
spectrum is then easily calculated from the fil- characteristic of form captured by D have not
ter coefficients. The details of the computa- been identified as of yet, although some stud-
tional method are described by Denham ies have begun to explore those links" (Klin-
( 1975 ). The number of filter coefficients is not kenberg, 1992 ).
defined by the method but must be chosen by
experimentation, fewer coefficients giving Are landscape surfaces self-similar?
smooth spectra with few spectral peaks and
more coefficients giving noisier spectra with There is insufficient definitive evidence to
many spectral peaks. Gallant et al. (1993) be able to conclude that landscapes are self-
found that twenty coefficients produced rea- similar or self-affine over all ranges of scales.
sonable spectral estimates. The MEM spectral Klinkenberg and Goodchild (1992) con-
method gave more precise estimates of D for cluded that, "it is not possible to provide blan-
profiles than the traditional direct Fourier ket statements about the overall fit of the frac-
transform method (Gallant et al., 1993 ). tal model. Outright rejection of the self-similar
fractal model does not appear to be warranted,
Questions and problems but neither does a blind application." Their
conclusions were from the very mixed results
Do fractals capture the essence of a landscape? of a study of the fractal characteristics of 55
DEMs from seven different United States phy-
Fractal analysis, as a mathematical method siographic provinces using seven methods of
for describing landscapes, seems unlikely to calculating the fractal dimension. Research to
capture the essence of complex and diverse date indicates that self-similarity is exhibited
landscapes. However, it appears that the frac- only in limited regions and over limited ranges
tal dimension includes "new" information that of scale in real landscapes, although one can
traditional parameters do not contain because find statements in the literature like: "the ap-
of its mathematical basis and methods of cal- plicability of fractal concepts to the Earth's to-
culation are different from those of traditional pography should not be surprising since it only
morphometric parameters. "It differs from requires scale invariance, and it has long been
other measures of 'roughness' in that it pro- recognized that topography is scale invariant
vides a description theoretically independent in a variety of geological terrains" (Huang and
of the sample m most other measures of rough- Turcotte, 1989). Unfortunately, the unstable
ness are sample dependent" (Brown and values of fractal dimensions and the non-lin-
Scholz, 1985 ). The key issue is how to explain ear log-log plots have demonstrated that cate-
the significance of the information contained gorical statements about the fractal properties
within the fractal dimension. In other words, of landscapes cannot be made. Moreover, "al-
how do we understand the physical signifi- though fractional Brownian landscapes with
cance of the fractal dimension? It is too sweep- dimension of 2.2-2.3 achieve realistic repre-
254 T.-B. XU ET AL.

sentations of the surface of the earth, little is visual appearances. "The ability to generate a
known about the dimensionality of natural ter- visually appealing display based on a frac-
rains" (Roy et al., 1987). tional Brownian model does not necessarily
Generally, the prerequisite for the measure- mean that the landform being represented is
ment of the fractal characteristics of land- actually fractal, or that it obeys quantitative
scapes is that the landscapes possess statistical predictions of the model" (Piech and Piech,
self-similarity or self-affinity. Measurements of 1990). It is also premature to reject the fractal
the fractal nature of landscapes are unbelieva- model of landscapes based on some spatially
ble if one fails to prove the self-similarity or variable fractal dimensions that have been ob-
self-affinity of the landscapes. However, to date tained to date. It appears that self-similarity is
there are no methods for proving this except exhibited only in limited regions and over lim-
by trying to calculate a constant fractal dimen- ited ranges of scale in landscapes. In the next
sion of complex landscape surfaces. few years it will be the main task of research to
Indeed, many people are infatuated with test the self-similarity property of landscapes
"magical" computer generated fractional (and other natural phenomena ).
Brownian surfaces. "The fractional Brownian There is a view that spatial variations of D
process has been used as a convenient way of over landscape surfaces, namely scale-depen-
generating self-similar surfaces, and certainly dent variations and region-dependent varia-
such surfaces more closely resemble some types tions, are perhaps reasonable. Burrough
of real topography than do the results of any ( 1981 ) indicated that, "although many natu-
other available methods of simulation" ral phenomena do display certain degrees of
(Goodchild and Mark, 1987). However, "the statistical self-similarity over many spatial
publicity that fractals and related mathemati- scales, there are others that seem to be struc-
cal concepts have received is largely due to their tured and have their levels of variability clus-
strong visual impact" (Goodchild and Mark, tered at particular scales. This behaviour does
1987). Mandelbrot (1977, 1982) held that not exclude them from the fractal concept."
visual appearance is the most important test of Mandelbrot ( 1977, 1982 ) considered that it is
stochastic models of natural phenomena and quite acceptable to have a series of zones of
that on this basis fractional Brownian model distinct dimensions connected by transition
surfaces must be accepted as models of terrain. zones. If this is reasonable, it means that the
Pentland (1984) also stated that "... the natu- examination of the variability of D would be
ral appearance of fractals is strong evidence useful for trying to separate scales of variation
that they capture all of the perceptually rele- that might be the result of particular natural
vant shape structure of natural surfaces." We processes. Roy et al. ( 1987 ) observed that, "in
are unable to deny the resemblance of fractal fact, one should anticipate that the dimen-
model surfaces to some natural landscape sur- sionality of most natural terrains should vary
faces. Furthermore, the human eye (and brain) spatially. Variations in processes and/or struc-
has the unsurpassed ability to observe and in- tures may be responsible for these changes in
terpret complex phenomena. So it is easy to in- dimensions .... Systematic variations in the di-
fer that landscapes possess self-similar mension within the surface bear important
properties. cartographic consequences .... more attention
There appears to be an inherent connection should be given to the fractal signature of char-
between fractional Brownian processes, land- acteristic terrains." Most studies more or less
scapes, and other natural phenomena. How- confirm the spatial variation of D over land-
ever, it is inappropriate to define the self-sim- scape surfaces, particularly the scale-depen-
ilarity of landscapes only based on intuitive dent variation. Theoretically, the spatial vail-
FRACTALS, FRACTAL DIMENSIONS AND LANDSCAPES - - A REVIEW 255

ation does contradict the hypothesis of self- Brownian profiles also produce log-log plots
similarity of landscapes. But, as yet there is no with multiple linear segments.
explicit criterion for defining how large the re- Several researchers have raised questions
gion with homogeneous fractal properties for about the fractal dimensions of non-self-simi-
which the landscape is (statistically) self-sim- lar landscapes. Kennedy and Lin ( 1986 ) stated
ilar needs to be, and how to delimit the region. that being non-fractal in nature is no limita-
The challenge in the future is to relate the spa- tion to the fractal dimension technique and it
tially variable fractal dimensions to spatially could be simply seen as a method for extract-
variable landscape formation processes, and ing useful information from a log-log plot. As
climate and geographical characteristics. an analogy, normal distribution models have
We have only measured D in limited regions been applied in the geo-sciences for a long time,
of landscapes, and over restricted ranges of and many natural phenomena appear to follow
scale. There is also a large number of methods the normal distribution. Normal distribution
of calculating D. Therefore, it is not yet appro- models are often applied to objects without
priate to define the regularity of the spatial proving the distributive characteristic of the
variation of the fractal nature. However, one objects. Furthermore, often one can obtain
should be aware that the spatial variation of D reasonable results from normal distribution
might imply a more significant usefulness of models even when the objects do not strictly
follow the normal distribution. Some models
fractals. In particular, the spatial variation of
for discrete data sets can not even be proved
D across a landscape surface with different
mathematically. We use these models because
topographic characteristics may be more sig-
they usually maintain stable and reasonable
nificant than the spatial variation with the
relationships between them and the objects
change of the measurement distance h of the
being investigated.
variogram or divider step r of the divider
method. In other words, the region-dependent
How to explain the differences in fractal
spatial variation has more practical signifi-
dimensions?
cance than the scale-dependent spatial
variation. Landscapes with different surface character-
The log-log plots used to define D often have istics are expected to have different fractal di-
multiple linear segments (Mark and Aronson, mensions, and, many studies have demon-
1984; Roy et al., 1987; Klinkenberg and Good- strated this. Table 1 is a brief summary of
child, 1992; Ouchi and Matsushita, 1992), or Klinkenberg and Goodchild's (1992) results.
are divergent at one end of the plot for the dif- Differences in the characteristics of landscape
ferent methods of calculation. Are these rea- surfaces are not, however, the only reason to
sonable or not? There is still insufficient defin- lead to differences in the computed fractal di-
itive evidence to answer this. Mandelbrot mensions. The following factors should be
( 1986b, 1989 ) considered that the presence of considered:
multiple linear segments of a log-log plot may (a) The fractal dimension should distin-
indicate that the land surface is in fact a mul- guish between landscapes with different
tifractal surface. However, this conclusion may "roughness" characteristics and agree with the
be dubious. Sometimes the object from which visual observations. "If the fractal dimension
one extracted the log-log plot is non-fractal and is to be a useful parameter, the measurement
the log-log plot is actually an arc and the so- methods used to determine the dimensions
called "linear segment" is just a segment of the must be robust, consistent, and have the capa-
arc. Gallant et al. ( 1993 ) found that synthetic bility of differentiating between visibly dis-
256 T.-B. XU ET AL.

TABLE 1

Fractal dimensions for seven physiographic regions in the United States computed using seven methods (from Klinkenberg and
Goodchild, 1992)

Physiographic region D~al D~ Dans Dec Ds.d. Dt.d. De

Coastal Plain 2.65 2.66 2.62 2.67 2.63 1.20 1.29

Blue Ridge 2.32 2.42 2.31 2.19 2.23 1.18 1.26
Valley and Ridge 2.36 2.52 2.34 2.23 2.32 1.15 1.23
Appalachian Plateaus 2.30 2.38 2.31 2.16 2.23 1.17 1.24
Interior Low Plateaus 2.41 2.45 2.42 2.21 2.28 1.19 1.27
Colorado Plateau 2.34 2,32 2.31 2.16 2.21 1.18 1.27
Basin and Range 2.29 2.33 2.25 2.10 2.21 1.12 1.18

D,~t: Entire-surface variogram dimension (first segment values only).

Ds : Mean of the dimensions of four (or less) sections.
Dang : Mean of the dimensions of six (or less) angles.
Dc~ : Mean of the dimensions of five cell counting values.
Ds.d.: Mean of the dimensions of five surficial dividers values.
Did : Mean of the five (or less) traditional dividers dimensions obtained from the contours.
D, : Mean of the five (or less) equipaced polygon dimensions obtained from the contours.

similar surfaces" (Klinkenberg and Good- such as differences in their mathematical basis
child, 1992). and mechanisms and understanding of the
(b) The fractal dimensions computed using physical significance of fractals. Klinkenberg
different methods in the same region of a land- and Goodchild (1992) reported that the vari-
scape are often different. Except for problems ability in the computed fractal dimensions is
with the method of calculation itself, it also more a function of the methods used than it is
raises the question of whether landscapes are a reflection of any theoretical inadequacy of the
self-similar and have fractal properties. Theo- self-similar fractal model. We could, perhaps,
retically, "in a fractal and self-similar terrain, see some differences between methods such as
the values of D should be in agreement regard- the oriented (angled) variogram methods in a
less of method used" (Roy et al., 1987 ). In ad- new light. They must reflect the oriented dis-
dition, some methods are strictly based on the tribution characteristics of valleys and ridges
hypothesis that the landscape possesses fractal of landscapes, i.e., the anisotropy of the large
properties. For example, the variogram method terrain features. The distributional structure of
assumes that landscapes have statistical prop- valleys and ridges are mainly controlled by
erties similar to those of fractional Brownian geological tectonism. The tectonic stress field
surfaces, namely that the altitude is a Brown- controlling the tectonic processes has strong
ian function of latitude and longitude. "Unless oriented characteristics.
the fractional Brownian model is first vali- (d) The limited size of the regions surveyed
dated for surfaces under consideration, the and the spatial resolution of the DEM data
variogram analysis would not necessarily be limit the precision and stability of the com-
expected to yield a correct fractal (surface) di- puted fractal dimension. Most of the investi-
mension" (Piech and Piech, 1990). gations only use a limited DEM matrix of 256
(c) There are many methods for determin- by 256 pixels or less (e.g. Shelberg et al., 1983;
ing the fractal dimensions of landscapes. Fur- Huang and Turcotte, 1989; Klinkenberg, 1992;
thermore, every method can be divided into Klinkenberg and Goodchild, 1992; Ouchi and
several subclasses. Having so many methods is Matsushita, 1992). Theoretically, Scale invar-
unfortunate because it could conceal problems iance of fractal measures allows one to "ex-
FRACTALS, F R A C T A L DIMENSIONS AND LANDSCAPES - - A REVIEW 2 57

trapolate from properties observed at one scale and resolution of source data used to generate
to the properties of a scale which has not been the DEM, and the interpolation method of
observed" (Gibert, 1989). That is the case in generation. Usually, fine-scale terrain of inter-
an exactly self-similar object. However, land- polated DEMs is smoother than the real one,
scapes are not exactly self-similar and, at most, particularly when using linear interpolation
are statistically self-similar. Thus, one needs methods and sparse source data. The interpo-
sufficient pixels to obtain the effective statis- lation process has a smooth filter function, as
tical parameters for the fractal models. mentioned above, and lowers the value of small
The DEMs used can only represent finite de- scale D while there is no smooth function for
tails of landscape surfaces because of the lim- the large scale terrain. This results in a discrep-
ited size range and the resolution of the DEM ancy between small scale and large scale D.
data. For example, according to the "Data User Polidori et al. ( 1991 ) stated that the discrep-
Guide 5 of DEMs of the U.S. Geological Sur- ancy may be interpreted as a consequence of
vey (USGS)" (1987), "within a standard the smoothing interpolation process. More-
DEM, most terrain features are generalized by over, based on knowledge that the directional
being reduced to grid nodes spaced at regular interpolation methods (e.g., bilinear interpo-
intersections in the horizontal plane. This gen- lation) and the uneven distribution of source
eralization reduces the ability to recover posi- data causes local artificial anisotropy of inter-
tions of specific features less than the interval polated DEMs, Polidori et al. (1991) stated
spacing during testing and results in a de facto that the variations of D with direction for short
filtering or smoothing of the surface during distances may be interpreted as a consequence
gridding." It is obviously impossible that "big of the anisotropic interpolation process. That
rills have little rills, and little rills have smaller is, it may be possible to use fractal analysis to
rills, until spatial saturation" (Thornes, 1990 ) say something about the scale at which arti-
under the finite resolution of DEMs. There are facts produced by the interpolation method end
not enough hierarchically recursive subdivi- and the true behaviour of the landscape be-
sions (or not enough levels of cascade) to ob- gins. The interpolation artifact of DEMs is not
tain relative steady "statistical averages" of the the only factor to cause the scale-dependent
fractal dimension. The limited size range and variation and directional bias of D, but is likely
resolution are two reasons why steady fractal to be a considerable one. In fact, many tradi-
dimensions have not been obtained. tional topographic attributes are closely re-
(e) The variability of the estimated D is lated to the accuracy of the DEMs, particularly
quite large, even for synthetic fractal data. those based on second derivatives of the sur-
Gallant et al. ( 1993 ) show that the 95% con- face such as plan and profile curvature.
fidence interval of D for the variogram method The 7.5-minute DEMs from the USGS are
is about __0.1. This places limits on the use of used to calculate the surface D of landscapes
D as a discriminator of different landscape by many researchers (Mark and Aronson,
types. 1984; Clarke, 1986; Roy et al., 1987; Huang
(f) Some methods of calculating the fractal and Turcotte, 1989; Klinkenberg and Good-
dimension of landscape surfaces, particularly child, 1992 ). These DEMs also contain the ar-
the variogram methods, use grid DEM data. tifacts described above. The USGS has used
The accuracy of DEM data could be an influ- four processes to generate 7.5-minute DEMs:
ential factor of scale-dependent variation and the Gestalt Photo Mapper II (GPM2), man-
directional bias of D (Polidori et al., 1991; ual profiling from photogrammetric stereo-
Klinkenberg and Goodchild, 1992). The con- models, stereomodel digitizing of contours, and
trolling factors of DEM accuracy are the scale derivation from digital line graph hypsography
258 T.-B. XU ET AL.

and hydrography categories. Among them, the statistical models (parameters) often mask the
latter three processes generate the DEM from detailed characteristics of the objects being in-
the source profile, contour or spot data by us- vestigated. Kaye (1989), for example, re-
ing interpolation (usually bilinear interpola- ported that carbonblack profiles with different
tion). For example, in order to obtain the final local structures have the same fractal dimen-
DEM with a 30 m grid spacing, the process of sions and that the fractal dimension does not
manual profiling from photogrammetric ster- tell us anything about the overall gross shape
eomodels produces two new profiles between of the profile.
every two source profiles with 90 m spacing, (h) Using only the fractal model with a sin-
using the bilinear interpolation method. The gle parameter, the fractal dimension, to repre-
process results in different degrees of the sent landscapes is not sufficient. "Despite the
smoothness of new profile pixels in different success of single parameter fractal models in
directions and results in the directional bias of creating visually realistic images representing
D (Klinkenberg and Goodchild, 1992). Ob- mountains, coastlines, the distribution of
viously, for the DEMs generated by this pro- landmasses, moonscapes, etc., it seems that a
cess, the reliability of the fine scale D is description of these systems in terms of a frac-
reduced. tal model with a single fractal dimensionality
Polidori et al. ( 1991 ) concluded that fractal is not completely realistic" (Meakin, 1991 ).
geometry can contribute to DEM quality as- Research on fractals has mainly concentrated
sessment through a simple technique that does on the fractal dimension and little attention has
not require a reference DEM. However, their been paid to other fractal features, although
conclusion was based on the hypothesis that the Fox and Hayes (1985) and Klinkenberg and
landscape surface is a fractional Brownian sur- Goodchild (1992 ), appear to be aware of this
face. Although scale-dependent variations of D problem. Klinkenberg and Goodchild (1992)
have been found by many people, because of stated that, "following Fox and Hayes ( 1985 ),
the differences between the fractional Brown- it is anticipated that the combination of the
ian model and real landscapes and the com- fractal dimension and the Gamma value will
plex factors that cause scale-dependent varia- capture the essential characteristics of the land
tion of D, Polidori et al.'s conclusion is tenable surfaces. Thus, although two land surfaces may
only under very limited situations. have similar forms (i.e. values of D) when
(g) We usually assume that a landscape is considered from a scale-independent view, the
statistically self-similar or self-affine when cal- magnitude of that roughness may well be very
culating its fractal dimension. The methods of different, and that difference will be captured
calculating the dimensions are based on the by the value of Gamma." A multi-parameter
statistical characteristics of landscape proper- fractal model should have a stronger capacity
ties. Sometimes the statistical characteristics of to describe and differentiate the properties of
a region do not describe the local properties and landscapes than a single parameter fractal
the detailed structure of the region very well, model. Fourier power spectral analysis has
thereby failing to distinguish the actual differ- been successfully applied in the geo-sciences.
ences between the objects being investigated. There are three related parameters, the fre-
For example, the "mean," a common statisti- quency, the power spectral energy intensity and
cal parameter, often cannot differentiate dif- their orientation, which characterize the Four-
ferent data sets. The different permutations ier spectral characteristics of objects in a two-
{ 1,2,3,4,5}, { 5,4,3,2,1} and { 3,1,2,5,4} have dimensional Fourier spectral domain.
the same mean, 3, but the mean can not be used (i) There may be actual physiographic dif-
to distinguish these permutations. In fact, some ferences between landscapes. However, these
FRACTALS, F R A C T A L DIMENSIONS AND LANDSCAPES - - A REVIEW 259

differences may not be distinguishable by vis- Although researchers have estimated the re-
ual and conventional means. Fractals capture gion-dependent spatial variation of D, they
the differences and yield different fractal di- have not yet found the border between regions
mensions and so apparently similar land- with different fractal properties. Also, they
scapes have different fractal characteristics. cannot explain whether the spatial variation is
These differences may reflect the new infor- continuous or not. Most of the results reported
mation of landscapes that are not reflected in to date are from regular grid DEM matrices,
the conventional morphometric parameters of and each matrix is usually treated as a unit with
landscapes. However, we cannot confirm this homogeneous fractal characteristics. In view of
yet and we do not really know the reasons why the definition of fractals and the present meth-
fractal analysis of landscapes yields different ods of calculating D, it is difficult to determine
fractal dimensions. the spatial variation of D and the pattern of this
variation within a DEM matrix. Perhaps we
Fractal dimension as a spatially variable could assume that the fractal nature matches
parameter the geomorphological unit or hydrological unit?
Perhaps we should calculate the fractal char-
As mentioned earlier, the fractal nature of acteristics inside the irregular boundary of the
landscapes appears to vary spatially and this hydrological or geomorphological unit, i.e.,
spatial variation has potentially practical sig- within a catchment?
nificance. Furthermore, the spatial variation of We postulate that the fractal properties of
fractal parameters could classify the land sur- landscapes should be more sensitive to spatial
face into fractally homogeneous regions (Klin- location. We therefore need a fractal model
kenberg 1992 ). Despite this, there is still some with multi-parameter (characteristics). Klin-
resistance to using D as a spatial parameter in kenberg ( 1992 ) indicates that, "the fractal di-
geo-models. Indeed, it is still very difficult to mension and associated variogram parameters
use D as a distributed parameter with high have potential as useful general geomorphom-
"spatial resolution", even though fractals could etric parameters and, like the results of any
be used to describe the profile of a fine parti- other general geomorphometry study, they
cle. The fractally homogeneous regions pro- could be used to bring to light those spatial
posed by Klinkenberg ( 1992 ) were quite large variations and structures in the land surface
and the spatial resolution of them is very coarse that geomorphology attempts to explain."
and far from that of conventional landscape
parameters like altitude, slope gradient, as- Developing trends in the fractal analysis of
pect, etc. In fact, only differentiating physio- landscapes
graphic provinces is of more limited practical
significance. The applications of fractals to landscape
To obtain relative steady and believable val- analysis is a developing and immature field and
ues of the fractal dimension, one has to have a much of the theoretical rigour of fractal geom-
DEM data matrix with a very large number of etry has not yet been exploited. We are at-
cells. For example, in the variogram methods tempting to find better methods with which to
outlined previously, one needs a large number estimate the fractal characteristics of land-
of random samples within a large area under scapes. New methods or techniques need to be
each distance increment in order to calculate tested and improved before they can be proved
E[ (BI4(xz,y2)-BH(Xl,yl))2]. Therefore, at to be effective and reliable. Remote sensing, a
present D should be regarded as an areally technology that arose in the early 1970s, has
lumped parameter. provided a good example of its developing
260 T.-B.XU ET AL.

course. We quote a paragraph from Molenaar's tion theory (geo-information t h e o ~ ) used to

paper in 1991: "that [remote sensing ] too was describe objects, phenomena or processes at the
a promising field in the seventies, the expected earth's surface. Improvement in geoinforma-
potentials were great, it was almost treated as tion theory will lead to improvements in the
a kind of magic with its own priesthood. The application of fractal geometry to landscapes.
deception came when it seemed that the poten- Until now studies of the fractal analysis of
tials could not be realised as easily and quickly landscapes and the fractal analysis of river net-
as expected. Now remote sensing people have works, the main two branches of the fractal
to fight hard to get their research funds in com- analysis of landscapes, have developed almost
petition with others. They are no wonder chil- in parallel. They need to be integrated with one
dren anymore. This does not mean that re- another to describe the surface structure of
mote sensing is not a powerful tool, but rather landscapes effectively. River networks are es-
that it takes much more time and hard work to sential features for structuring the land sur-
realize its potentials, to extract from remote face. We can not verify whether or not surface
sensing data the information required in the fractals contain definite information about
different application fields." The problem with river networks. Spatial structures of river net-
remote sensing was that it "took much more works exhibit some difference from the spatial
time and hard work", and many people did not structures of profiles or contours, and should
have the scientific basis and methods to study have different fractal characteristics.
and apply remote sensing technologies.
The fractal analysis of landscapes, as a part Acknowledgements
of the developing field of fractals, is also in its
preliminary and descriptive stage, and lacks The study was funded in part by grant 1991 /
theoretical explanation and rigour. The physi- 92-ANU3 from the Land and Water Resources
cal significance of landscape fractal character- Research and Development Corporation and
istics remains to be explained. It seems that the by the Water Research Foundation of Aus-
development of fractal theory has lagged be- tralia. The authors thank Adam Lewis for his
hind many of the potential applications. Man- valuable discussions and suggestions.
delbrot (1986b) argued that this situation is
reasonable and is the usual pattern of science. References
Obviously, measuring the fractal dimensions of
landscapes is not our final goal. We hope that Andrle, B. and Abrahams, A.D., 1989. Fractal techniques
fractal analysis will provide effective tools for and the surface roughness of talus slopes. Earth Surf.
characterizing landscape properties, and that Process. Landforms, 14: 197-209.
fractal parameters, which possess definite Armstrong, M.P. and Hopkins, L.D., 1983. Fractal en-
hancement for thematic display of topologically stored
physical significance, can be integrated with data. In: B.S. Wellar (Editor), Proc. Automated Car-
diverse geo-models. In order to do this, we need tography: International Perspectives on Achievements
to strengthen research in geographical infor- and Challenge, Ottowa, Canada, pp. 309-318.
mation theory as well as fractal theory. In the Brown, S.R. and Scholz, C., 1985. Broad bandwidth study
of the topography of natural rock surfaces. J. Geophys.
age of mathematics and computers, the con- Res., 90(B14): 12,575-12,582.
ventional geographical and geomorphological Brown, S.R., 1987. A note on the description of surface
theories have performed poorly, and geo- roughness using fractal dimension. Geophys. Res. Lett.,
models have developed independently of each 14: 1095-1098.
Burg, J., 1967. Maximum entropy spectral analysis. In:
other without any unified and systematic the- D.A. Hansen (Editor), 37th Meeting Soc. of Explora-
oretical basis. Therefore, it is necessary to tion Geophysicists, Oklahoma City, 31 Oct. 1967.
structure a systematic geographical informa- Burg, J., 1968. A new analysis technique for time series
FRACTALS, FRACTAL DIMENSIONS AND LANDSCAPES - - A REVIEW 261

data. NATO Advanced Study Institute on Signal Pro- Hayward, J., Orford, J.D. and Whalley, W.B., 1989. Three
cessing, Enschede, Netherlands, Aug. 12-13, 1968. implementations of fractal analysis of particle out-
Burrough, P.A., 1981. Fractal dimensions of landscape and lines. Comput. Geosci., 15: 199-207.
other environmental data. Nature, 294: 240-242. Hjelmfelt, A.T., 1988. Fractals and the river-length catch-
Clarke, K.C., 1986. Computation of the fractal dimension ment-area ratio. Water Resour. Bull., 24: 455-459.
of topographic surfaces using the triangular prism sur- Hough, S.E., 1989. On the use of spectral methods for the
face area method. Comput. Geosci., 12:713-722. determination of fractal dimension. Geophys. Res.
Culling, W.E.H., 1988. Dimension and entropy in the soil- Lett., 16: 673-676.
covered landscape. Earth Surf. Process. Landforms, 13: Huang, J. and Turcotte, D.L., 1989. Fractal mapping of
619-648. digitized images: application to the topography of Ar-
Culling, W.E.H., 1989. The characterization of regular/ izona and comparisons with synthetic images. J. Geo-
irregular surfaces in the soil-covered landscape by phys. Res., 94(B6): 7491-7495.
Gaussian random fields. Comput. Geosci., 15: 219- Hutchinson, M.F., 1989. A new procedure for gridding el-
226. evation and streamline data with automatic removal
Denham, C., 1975. Spectral analysis of paleomagnetic time of spurious pits. J. Hydrol., 106:211-237.
Jensen, R.V., 1987. Classical chaos. Am. Sci., 75: 168-
series. J. Geophys. Res., 80( 14): 1897-1901.
181.
Dubuc, B., Zucker, S.W., Tricot, C., Quiniou, J.F. and
Jones, H., 1991. Fractals before Mandelbrot: A selective
Wehbi, D., 1989. Evaluating the fractal dimension of
history. In: A.J. Crilly, R.A. Earnshaw and H. Jones
surface. Proc. R. Soc. Lond., A425:113-127.
(Editors), Fractals and Chaos, Springer, New York,
Dyson, F., 1978. Characterizing irregularity. Science, 200: pp. 7-33.
677-678. Jones, J.G., Thomas, R.W. and Earwicker, P.G., 1989.
Elliot, J.K., 1989. An investigation of the change in sur- Fractal properties of computer-generated and natural
face roughness through time on the foreland of Austre geophysical data. Comput. Geosci., 15: 227-235.
Okstindbreen. North Norway. Comput. Geosci., 15: Karlinger, M.R. and Troutman, B.M., 1992. Fat fractal
209-217. scaling of drainage networks from a random spatial
Evans, I.S., 1984. Correlation structures and factor anal- network model. Water Resour. Res., 28:1975-1981.
ysis in the investigation of data dimensionality: Statis- Kaye, B.H., 1989. A Random Walk through Fractal Di-
tical properties of the Wessex land surface, England. mensions. VCH, Weinheim, 19 pp.
Preprint of paper presented at the International Sym- Kennedy, S.K. and Lin, W., 1986. FRACT - - A Fortran
posium on Spatial Data Handling, Aug. 20-24, 1984. subroutine to calculate the variables necessary to de-
Falconer, K.L., 1991. Dimensions - - their determination termine the fractal dimension of closed forms. Corn-
and properties. In: J. Belair and S. Dubue (Editors), put. Geosci., 12:705-712.
Fractal Geometry and Analysis. Proc. on Fractal Ge- Klinkenberg, B., 1992. Fractals and morphometric meas-
ometry and Analysis. Kluwer, Dordrecht, pp. 221-254. ures: is there a relationship ? In: R.S. Snow and L Mayer
Fox, C.G. and Hayes, D.E., 1985. Quantitative methods (Editors), Fractals in Geomorphology. Geomorphol-
for analyzing the roughness of the seafloor. Rev. Geo- ogy, 5: 5-20.
phys., 23: 1-48. Klinkenberg, B. and Goodchild, M.F., 1992. The fractal
Fox, C.G., 1989. Empirically derived relationships be- properties of topography: a comparison of methods.
tween fractal dimension and power law from fre- Earth Surf. Process. Landforms, 17:217-234.
quency spectra. Pure Appl. Geophys., 131:211-239. Kumar, S. and Bodvarsson, G.S., 1990. Fractal study and
Gallant, J.C., Moore, I.D. and Gessler, P., 1993. Estimat- simulation of fracture roughness. Geophys. Res. Lett.,
ing fractal dimension: A comparison of methods. Math. 17: 701-704.
Laverty, M., 1987. Fractals in karst. Earth Surface Pro-
Geol., submitted.
cesses and Landforms, 12: 475-480.
Gan, K.C., McMahon, T.A. and Finlayson, B.L., 1992.
Longley, P.A. and Batty, M., 1989. Fractal measurement
Fractal dimensions and lengths of rivers in south-east
and line generalization. Comput. Geosci., 15: 167-183.
Australia. Cartographic J., 29:31-34. Mandelbrot, B.B., 1977. Fractals: Form, Chance, and Di-
Gibert, L.E., 1989. Are topographic data sets fractal ? Pure mension. W.H. Freeman, San Francisco, 365 pp.
Appl. Geophys., 131: 241-254. Mandelbrot, B.B., 1982. The Fractal Geometry of Na-
Goodchild, M.F., 1980. Fractals and the accuracy of geo- ture. W.H. Freeman, San Fransisco, 468 pp.
graphical measures. Math. Geol., 12: 85-98. Mandelbrot, B.B., 1986a. Self-affine fractal sets. In: L.
Goodchild, M.F. and Mark, D.M., 1987. Review Article Pietronero and E. Tosatti (Editors), Fractals in Phys-
- - The fractal nature of geographic phenomena. An- ics. Proc. Sixth Trieste Int. Symp. on Fractals in Phys-
nals Assoc. Am. Geogr., 77: 265-278. ics, ICIP, Trieste, Italy, July 9-12, 1985. North-Hol-
Hatano, R. and Booltink, H.W.G., 1992. Using fractal di- land, Amsterdam, pp. 3-28.
mensions of stained flow patterns in a clay soil to pre- Mandelbrot, B.B., 1986b. Multifractals and fractals. Phys.
dict bypass flow. J. Hydrol., 135:121-131. Today, 39(9): 11-12.
262 T.-B. XU ET AL.

Mandelbrot, B.B., 1989. Multifractal measures, especially Shelberg, M.C., Lam, N. and Moellering, H., 1983. Meas-
for the geophysicist. Pure Appl. Geophys., 131: 5-42. uring the fractal dimensions of surfaces. Proc. Auto-
Mark, D.M. and Aronson, P.B., 1984. Scale-dependent mated Cartography, 6, pp. 319-328.
fractal dimensions of topographic surfaces: an empir- Snow, R.S. and Mayer, L., 1992. Introduction. In: R.S.
ical investigation, with applications in geomorphology Snow and L. Mayer (Editors), Fractals in Geomor-
and computer mapping. Math. Geol., 16:671-683. phology. Geomorphology, 5: 1-4.
Meakin, P., 1991. Fractal aggregates in geophysics. Rev. Speight, J.G., 1974. A parametric approach to landform
Geophys., 29: 317-354. regions. In: E.H. Brown and D.L. Linton (Editors),
Milne, B.T., 1988. Measuring the fractal geometry of The Progress of Geomorphology. Br. Geogr., 7:213-
landscapes. Appl. Math. Comput., 27: 67-79. 230.
Milne, B.T., 1991. Lessons from applying fractal models Stanley, H.E., 1986. Form: an introduction to self-simi-
to landscape patterns. In: M.G. Turner and R.H. larity and fractal behaviour. In: H.E. Stanley and N.
Gardner (Editors), Quantitative Methods in Land- Ostrowski (Editors), On Growth and Form: Fractal
scape Ecology: the Analysis and Interpretation of and Non-Fractal Patterns in Physics, Martinus Nijhoff,
Landscape Hetergeneity, Springer, New York, pp. 199- Boston, 308 pp.
235. Tarboton, D.G., Bras, R.L. and Rodriguez-Iturbe, I., 1988.
Molenaar, M., 199 l. Status and problems of geographical The fractal nature of river networks. Water Resour.
information systems. The necessity of geoinformation Res., 24: 1317-1322.
theory. J. Photogram. Remote Sensing, 46: 85-103. Thornes, J., 1990. Big rills have little rills... Nature, 345:
Moore, I.D., Grayson, R.B. and Ladson, A.R., 1991. Dig- 764-765.
ital terrain modelling: a review of hydrological, geo- Tricot, C., 1991. Rectifiable and fractal sets. In: J. Belair
morphological, and biological applications. Hydrol. and S. Dubue (Editors), Fractal Geometry and Anal-
Process., 5: 3-30.
ysis. Proc. on Fractal Geometry and Analysis. Kluwer,
Ohver, M.A. and Webster, R., 1986. Semi-variograms for
Dordrecht, pp. 367-404.
modelling the spatial patten of landform and soil prop-
Unwin, D., 1989. Fractals and the geosciences: Introduc-
erties. Earth Surf. Process. Landforms, 11: 491-504.
tion. Comput. Geosci., 15:163-165.
Ouchi, S. and Matsushita, M., 1992. Measurement of self-
U.S. Geological Survey (USGS), 1987. Digital Elevation
affinity on surfaces as a trial application of fractal ge-
Models, Data Users Guide 5, Department of the Inte-
ometry to landform analysis. In: R.S. Snow and L.
Mayer (Editors), Fractals in Geomorphology. Geo- rior of United States, Reston, VA, pp. 1-38.
morphology, 5:115-130. Voss, R.F., 1985a. Random fractal forgeries. Fundamen-
Pentland, A.P., 1984. Fractal-based description of natural tal Algorithms for Computer Graphics. NATO ASI Se-
scenes. IEEE Trans. Pattern Analysis and Machine In- ries, F17: 805-835.
telligence, PAMI-6:661-674. Voss, R.F., 1985b. Random fractals: characterization and
Piech, M.A. and Piech, K.R., 1990. Fingerprints and frac- measurement. Scaling Phenomena in Disordered Sys-
tal terrain. Math. Geol., 22: 457-485. tems, NATO ASI Series, 133:1-11.
Polidori, L., Chorowicz, J. and Guillande, R., 1991. De- Whalley, W.B. and Orford, J.D., 1989. The use of fractals
scription of terrain as a fractal surface, and application and pseudofractals in the analysis of two-dimensional
to digital elevation model quality assessment. Photo- outlines: review and further exploration. Comput.
gramm. Eng. Remote Sensing, 57: 1329-1332. Geosci., 15: 185-197.
Roy, A.G., Gravel, G. and Gauthier, C., 1987. Measuring Yokoya, N., Yamamoto, K. and Funakubo, N., 1989.
the dimension of surface: a review and appraisal of dif- Fractal-based analysis and interpolation of 3D natural
ferent methods. In: N.R. Chrisman (Editor), Proc. surface shapes and their application to terrain model-
Automated Cartography, Baltimore, Maryland, pp. 68- ing. Comput. Vision Graph. Image Process., 46: 284-
77. 302.