Sie sind auf Seite 1von 5

Royal Swedish Academy of Sciences

Quantifying Spatial Patterns of Landscapes


Author(s): Magnus Ekstrm
Source: Ambio, Vol. 32, No. 8, Remote Sensing for the Environment (Dec., 2003), pp. 573576
Published by: Springer on behalf of Royal Swedish Academy of Sciences
Stable URL: http://www.jstor.org/stable/4315442
Accessed: 09-06-2016 06:26 UTC
REFERENCES
Linked references are available on JSTOR for this article:
http://www.jstor.org/stable/4315442?seq=1&cid=pdf-reference#references_tab_contents
You may need to log in to JSTOR to access the linked references.
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted
digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about
JSTOR, please contact support@jstor.org.

Royal Swedish Academy of Sciences, Springer are collaborating with JSTOR to digitize, preserve and
extend access to Ambio

This content downloaded from 202.43.95.117 on Thu, 09 Jun 2016 06:26:23 UTC
All use subject to http://about.jstor.org/terms

Magnus Ekstrom

Quantifying Spatial Patterns of Landscapes


In this paper we discuss information theoretical landscape
indices based on data from digitized maps in grid format:
measures based on Shannon's entropy, e.g. the measures
of diversity and contagion; and measures based on
conditional entropy, e.g. a new index which can be seen
as an alternative to the measure of contagion that does
not have the disadvantage of being highly correlated to
the measure of diversity. We also introduce a measurement
on how much information is contained in a coarse-scale
map about a fine-scale map.

values indicate that the landscape is dominated by one or a


few cover categories.

In O'Neill et al. (4) the deviation from the maximum


possible entropy was introduced as an index, the measure of
domination:
K

DK = 10g2 K +k1 Pk 092 PkMoreover, in the same paper another landscape index based
on the entropy was introduced, the measure of contagion.
Here we consider the (corrected) version discussed by Li and

INTRODUCTION

Reynolds (5):

The spatial pattern of ecosystems influences many ecological


phenomena. An underlying assumption of many environmental

decisions is that some patterns are more desirable than


others (1), and 'before the interaction between landscape
structure and ecological processes can be understood,
landscape patterns must be identified and quantified in
meaningful ways' (2). Therefore, tools are needed to
understand and describe the spatial structures of landscapes.

One idea is to use a set of indices that captures important


aspects of landscape structure in a few numbers.

Since the late 1980s many landscape indices have been


proposed; see, e.g. the review by Haines-Young and Chopping

(1) where a selection of 56 different indices, introduced


since 1988, are listed. The choice of which indices to use
seems overwhelming, and it is clear that there is a good deal
of redundancy in the indices available.

In this paper the focus will be on indices, based on data


from digitized maps in grid format, which are derived from
information theory.

Indices Based on Shannon's Entropy


Consider a raster map, where each picture element, or pixel,
can take any one of K distinct land-cover types. For example,
the map might be a Thematic Mapper satellite image with a
30 m x 30 m pixel resolution, classified into different landcover categories, e.g. land and water.
The first index is Shannon's entropy:

K K

CK= 2 10g2 K + E E Pij0g2 Pij 1


where pj is the probability that a randomly chosen pixel is of
category i, and that a pixel, chosen randomly from the pixels

adjacent to the first pixel, is of category j. Here, a pixel is


considered to be adjacent only to the neighboring pixels on
the 4 sides, i.e. upper, lower, left, and right (pixels touching
at corners are not considered as adjacent in this paper). CK is
a nonnegative number, and is always less or equal to 210g2 K
(which is a direct consequence of the inequality in Eq. 1).
According to O'Neill et al. (4) higher values of CK indicate
that a few large patches are found on the landscape, whereas
lower values indicate that the landscape is disserted into
many small patches. Unfortunately, with this in mind, one is
easily misled by this index. For example, compare the
simulated landscapes (c) and (g) in Figure 1; landscape (c)

clearly has fewer and larger patches than landscape (g),


but according to O'Neill et al. the values of CK indicate
the opposite. From Figure 1 it is clear that the measure of
contagion decreases as the landscape becomes more diverse,
i.e. less dominated by one or a few land-cover categories,
and as the landscape becomes more dissected into many
small patches. Note that CK and EK (or DK) do not give
independent information, i.e. CK changes when the diversity
changes. See also Riitters et al. (6) for more detailed comments
on the measure of contagion.

For any K, the function EK = EK (P1,... ,p) is a continuous


and symmetric function with respect to all its arguments.
Furthermore, for every K, we have

EK kp lok 1g02pk
where Pk iS the probability that a randomly chosen pixel is of

land-cover category k (we define xlog2 x = 0 if x = 0, thus


extending xlog2 x to the origin by continuity). This index is
also known as the measure of diversity; see e.g. Turner et al.
(3). A well-known property of the entropy is that

EK < 10g2 K, Eq. 1


with equality if and only if Pk =1/K for all k. Thus, large
values of EK indicate that the land-cover categories are

represented in approximately equal proportions, and low

EK+1 (PI' 'PK'O) = EK (PI ...'PK), Eq. 2


i.e., the adding of a nonobserved land-cover category does
not change the measure of divergence. These 2 properties
together imply that the entropy is nonsensitive to the presence
of rare land-cover categories. The measures of dominance
and contagion, however, do not satisfy a property like Eq. 2.
Thus, when K is defined as the number of observed distinct
land-cover categories on the landscape, it is clear that the
contagion index, just as the measure of dominance, is sensitive
to the presence of rare land-cover categories (3, 5), which
may cause problems in the interpretation, e.g. when comparing

Ambio Vol. 32 No. 8, Dec. 2003 ?O Royal Swedish Academy of Sciences 2003 573
http://www.ambio.kva.se

This content downloaded from 202.43.95.117 on Thu, 09 Jun 2016 06:26:23 UTC
All use subject to http://about.jstor.org/terms

I'Or
A~~~~~

Cl
1.15
Cl
= 0.20
Cl
1.52
CZ
= 0.22
ClC5
2=1.74
(a) Cl=
E5 =0.15
1.54
C5 ==W1
3.055=(b)
E5
=H2.02
C5- =H
2.56
E5
= -0.10
2.31
2.26
10252
~~~5.10
5,252
I(c)-0.2
2521
7
W,f510%.10
0.76
= -010
2=0.75
H5'
Ii =-076
,7io
1?39
j25.
=
0.38
H1
O182
t
5
0
50
5-1
2
09
0,252
056
5,10 5,252 5,10 -5,252 -5,10 5,252-

"7~~~~7

A~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~A

(d) CI.
Es==0.24C1
1.52 C5=3.04
(e) E5= 2.05 C = 2.49 (f) E5= 2.30 Cs= 2.22
= 1.40 CI = 0.32 C 2= 1.84 C/510_0.37 Cl 1ss- 2.08
-,=0.16 H5 = 0.92 W1.=00016 W125= 0.90 H = -016 i- -090

5,0128 j.,252..=
0.12 5,10
110:= 1.73
1. -5 0.21
71,10 1-5,252
93 P252 -0--23
15,10
- 5,252
-5,252
-5,10

(g) E5 = 1.55 C5 =2.95 (h) E5 = 2.04 C5 = 2.42 (i) E5 = 2.32 C5 =2.11

CI. -0=041 CZ =1.51 CI -052 CI 521.98 CI -060 CI, -220

,10. 26 _l5 I= 0.98 fl10 026 142- 9 I 11006 f252 =0.95

15'10-1.14
1'252-004 11015
2/52
-007
,110=172
=52 012
7 11 " I5 52200775,10
- - 5,252
5,10
- 5,52-5,10
-5,252Figure 1. The images above are based on 3 different vector images; one image for each row. The coloring of the

polygons in each image was done by random; with probabilities 0.6 (white), 0.25 (light grey), 0.1 (grey), 0.045 (dark
grey) and 0.005 (black), for the images to the left; with probabilities 0.4, 0.3, 0.15, 0.1 and 0.05, for the images in
the middle; and with equal probabilities for the 3 at the right-hand side. Thereafter, each image was transformed
to a raster image of 2520 x 2520 pixels, and the different indices presented above were calculated based on the pixel
data given in each raster image.

contagion indices from different landscapes and/or spatial


resolution. For instance, in Turner et al. (3) the relationship
of the measures of dominance and contagion to different
spatial resolutions, i.e. to different pixel sizes, was examined
by using simple linear regression. When the spatial resolution
becomes increasingly coarser, rare land-cover categories
disappear and K, defined as the number of observed distinct
land-cover categories on the landscape, decreases. Thus,

since neither DK nor CK satisfy a property like Eq. 2, the


benefit of using simple linear regression, i.e. of fitting a
straight line, is limited, e.g. for extrapolation of the measures
DK and CK from one level of resolution to another. However,
if K is defined as the number of land-cover categories of
interest, rather than the number of observed categories which can be less than the number of categories of interest
- then methods like linear regression could be more useful,
since K does not change as the resolution changes in this case.

574 C Royal Swedish Academy of Sciences 2003 Ambio Vol. 32 No. 8, Dec. 2003
http://www.ambio.kva.se

This content downloaded from 202.43.95.117 on Thu, 09 Jun 2016 06:26:23 UTC
All use subject to http://about.jstor.org/terms

picked by random is of category k, and let qk be the


probability that a pixel picked by random from the
coarse-scale map will be assigned land-cover category
k. Further, let P]jk be the probability that a randomly
chosen pixel from the fine-scale map is of category
j, given that this is one of a group of m x m pixels
that defines an aggregate of category k. Then the
conditional entropy between the 2 maps is
K K

(a) H 20.75 /1 = .50 (b) H =0.88 1 = 0.24 (c) H2= 0.00 I2= 2.00

H 40 75 / = 0.50 H 40 96 / = 0.07 H = 0.00 I = 2.00

H- =0.75 /4, 8= 050 H=0.99 / 8=0.02 H -8 ??? / 8=2.00

U...

48=4U

1,

H'4 1.00 I' = 0.00 H'4 =1.00 /' = 0.00 H' = 0.98 14' = 0.04

(d) =4=0 /20.00 (e) H4=.4= 2.00 (f) H = 0.58 12= 0.83
H 8=1.0/ ?-? H = 1.00 / = ?-?? H 8 = 0-99 14 = 0.01

CIK m =- XI qk Pjlk 10_2 Pjlk In the definition of CIKmX i.e. of qk and PjlkI' we do not
assume that the values of all data pixels on the coarsescale map are known; that is, it is not assumed that
the outcomes of the assignments in the multimodal
cases are known. One could, however, define CIKm
from one realization of the coarse-scale map, but we
do not do that here, since that would make CIKm
dependent on realizations of random variables that
have nothing to do with the landscape under study.
The conditional entropy is a nonnegative quantity

bounded from above by EK = EK (P' ... ' PK)' i.e.


K

0 < CIK m EK= - Pk1g2 Pk1 Eq.3

()H =0.67 14 = 0.60 (h) H =0.79 14 = 0.39 (i) H4 = 0.60 14 = 0.79

H' =0 67 / = 0.61 H' =0 83 / = 0.32 H = 0.76 / = 0.48

H- =0.66 /4 8= 0.63 H- =0.80 / 8 0.37 H- 8=0.96 / 8= 0.08


Figure 2. Nine different images represented on a 16 x 16 grid.

INDICES BASED ON EXTENSIONS OF ENTROPY


Studies into the way landscape analysis is affected by scale
considerations are numerous in the literature, and it has
been recognized that scale has a significant effect on most
measurements of different aspects of spatial patterns.
Therefore, one should seek for measurements that can
incorporate a range of scales. In Johnson and Patil (7) one
such measurement is discussed, the conditional entropy,
which can be applied over a range of resolutions.
Again, consider a raster map composed of a large number
of square pixels of equal size, where each data pixel represents
one of K distinct land-cover categories. From this original
map of, let us say, M x N pixels, we will, in order to define
conditional entropy, derive a map at a coarser resolution by
aggregating neighborhoods of m x m pixels into new larger
pixels. Thus, if we for simplicity assume that both M and N
are divisible by m, we obtain a coarser-scale map of M/m x
N/m pixels. The assignment of a land-cover category to an
aggregate can be done in different ways; e.g. by choosing a
category from the original pixels that formed the aggregate,
which can, for instance, be done by choosing a fixed position
of the original pixels (like the middle, the upper right,
etc.), by choosing one of the original pixels by random, or
by choosing the mode of the original pixels categories (in the
case when the mode is not unique, one of the modes is
chosen by random). Below, we will use the latter technique.
For the original map, let Pk be the probability that a pixel

The lower bound is attained when for all k with qk > 0,


there exists j such that plk= and pi = 0 for all i ? k,
i.e. when each group ofm x m pixels, that defines an
aggregate, consists of one land-cover category only,
e.g. as in Figure 2 (c) for m = 2, 4 and 8. The upper
bound is obtained when each group of m x m pixels,
that defines an aggregate, contains the land-cover
categories 1,.. .,K in equal proportions (e.g. as in Fig.

2 (d) for m = 2, 4 and 8, and Fig. 2 (e) for m = 4 and


8). For more details about conditional entroDv and
related quantities in information theory, see e.g. Guiasu (8).
Here we propose a new index, a normed variant of the
conditional entropy:
CIKm

HKm- EK

which gives a measurement with quite different properties


than the conditional entropy itself. From the inequalities in
Eq. 3 we see that HK,m is bounded from below by zero, and
from above by one. When the landscape, represented by the
fine-scale map, is grouped into disjoint neighborhoods (of
m x m pixels), and if each neighborhood contains land-cover
categories in approximately the same proportions as for the
whole fine-scale map, then HKm will be large. If most
neighborhoods contain only one of the different land-cover
categories found on the map, however, then HKm will be
small. Therefore, we call HK,m the measure of homogeneity,
since the larger HKm, the more equal the neighborhoods are
to each other (with respect to proportions of land-cover
categories).
Even though we have assumed that each neighborhood
should be of the same size and shape, i.e. a square of m x m
pixels, we want to stress that this is not a necessary assumption.
The neighborhoods can, for instance, be naturally defined
regions on the landscape. Observe that we get information
about the landscape on a local level by choosing small m
(neighborhoods), and on a regional level by choosing large
m (neighborhoods). HKm will increase, as the value of m

increases, if the proportions of land-cover categories 1,...,K

Ambio Vol. 32 No. 8, Dec. 2003 ? Royal Swedish Academy of Sciences 2003 575
http://www.ambio.kva.se

This content downloaded from 202.43.95.117 on Thu, 09 Jun 2016 06:26:23 UTC
All use subject to http://about.jstor.org/terms

in the aggregates get closer and closer to the proportions

into many small patches (here "small" includes elongated

PI' 'PK (see e.g. Fig. 1 or Fig. 2 (b) and (i)). However, it

and thin patches of small area), the smaller value of IKf. In


particular, if the pixel size in the coarse-scale map is larger
than many of the patches found on the map at the fine scale,
then it is clear that substantial information is lost, e.g. see
Figure 1 where the values of I5252 are very low for images
(d)-(h), reflecting that a considerable amount of information
is lost when-transforming the raster images of 2520 x 2520

should be noted that if there is a large regional variety, i.e. if


some land-cover categories are rare, sparse or absent in some
regions, but dominate in others, HKm might have a more
irregular behavior as m increases (Fig. 2 (g) and (h)).
Further, it is clear that HKm increases as the landscape
becomes more dissected into many small patches, as illustrated in Figure 1, and therefore one may consider HKm as a
measure of fragmentation as well.
An important criterion for a set of useful indices is that
they should be relatively independent of each other (4, 9).
Figure 1 indicates that this is the case for the 2 indices EK
and HK m Also, it should be noted that the measure of
homogeneity (fragmentation), just as conditional entropy,
could be applied over a range of resolutions as described in
Johnson and Patil (7).
According to Johnson and Patil (7) the conditional
entropy (with m = 2) is closely related to the measure of

contagion. However, if one considers 210g 2 K- CK, rather


than the measure of contagion itself, the relationship is even
stronger, e.g. both measurements increase both with diversity
and fragmentation, and neither of them is sensitive to the
presence of rare land-cover categories. Therefore, it is possible
to consider the quantity

21og2K- CK

EK
as an alternative to HKm. However, this alternative measure
does not have a natural upper bound, and is therefore more
difficult to interpret. Moreover, by using HK m we can study
both local (by choosing small m) and regional phenomena
(by choosing large m).

A MEASUREMENT OF THE INFORMATION


CONTAINED IN A COARSE-SCALE MAP
Studies on how the landscape analysis is affected by scale
considerations are numerous in the literature. A well-known
problem is that important information is often lost as the
resolution of spatial data becomes increasingly coarser (1-3,
9,10). In Turner et al. (3) '... the development of methods
that will ... quantify the loss of information with changing
scales' was asked for, since 'such methods are necessary
before ecological insights can be extrapolated between
spatial and temporal scales'. Here we propose a related
quantity, i.e.

IK,m =EK CIK mK

pixels into images of only 10 x 10 pixels. In most cases, IKm is


a nondecreasing function of m, reflecting that more and more
information is lost as the resolution becomes increasingly
more coarse. When the landscape has large local/regional

differences (as in Fig. 2 (a), (c), (g) and (h)), however, IK,m
sometimes decrease with m (as in Fig. 2 (g) and (h)), and

should therefore be used with caution.

DISCUSSION
Even though it is impossible to capture the complexity of

spatial structures of landscapes in just a few numbers, the


use of landscape indices can help us understand and describe
important features of it, just like quantities like an average or a
standard deviation can describe important aspects of a data set.
We believe that indices like the measures of diversity and
contagion do provide important information about a landscape's spatial patterning. A problem however, is that these 2
indices are highly correlated (6). But the main characteristics
attached to the measure of contagion in O'Neill et al. (4) are
captured by our proposed index HKm as well. Thus, both a large
CK and a small HKm may indicate that a few large patches are
found on the landscape, while both a small CK and a large
HKm may indicate that the landscape is dissected into many
small patches (note, what is considered to be a small or a
large patch when dealing with HKm depends on the choice of
m). But in addition, HK,m has the advantage of not being highly
correlated to the measure of diversity, as indicated in Figure 1.

References and Notes


1. Haines-Young, R. and Chopping, M. 1996. Quantifying landscape structure: a review
of landscape indices and their application to forested landscapes. Prog. Phys. Geogr
20, 418-445.

2. Turner, M.G. 1990. Spatial and temporal analysis of landscape patterns. Landscape
Ecol. 4, 21-30.

3. Turner, M.G., O'Neill, R.V, Gardner, R.H. and Milne, B.T. 1989. Effects of changing
spatial scale on the analysis of landscape pattern. Landscape Ecol. 3, 153-162.

4. O'Neill, R.V, Krummel, J.R., Gardner, R.H., Sugihara, G., Jackson, B., DeAngelis,

D.L., Milne, B.T., Turner, M.G., Zygmunt, B., Christensen, SW, Dale, VH. and
Graham, R.L. 1988. Indices of landscape pattern. Landscape Ecol. 1, 153-162.
5. Li, H. and Reynolds, J.F. 1993. A new contagion index to quantify spatial patterns of
landscapes. Landscape Ecology 8, 155-162.

6. Riitters, K.H., O'Neill, R.V, Wickham, J.D. and Jones, K.B. 1996. A note on contagion
indices for landscape analysis. Landscape Ecol. 11, 197-202.
7. Johnson, G.D. and Patil, G.P. 1998. Quantitative multiresolution characterization of
landscape patterns for assessing the status of ecosystem health in watershed management
areas. Ecosyst. Health 4, 177-187.

which can be seen as a measurement of the amount of


information which is contained in the coarse-scale map
about the fine-scale map. From the inequalities in Eq. 1 and
Eq. 3 it follows that

8. Guiasu, S. 1977. Information Theory with Applications. New York, McGraw-Hill.


9. O'Neill, R.V, Hunsaker, C.T., Timmins, S.P., Jackson, B.L., Jones, K.B., Riitters, K.H.
and Wickham, J.D. 1996. Scale problems in reporting landscape pattern at the regional
scale. Landscape Ecol. 11, 169-180.

10. Benson, B.J. and MacKenzie, M.D. 1995. Effects of sensor spatial resolution on
landscape structure parameters. Landscape Ecol. 10, 113-120.

O < IK m < EK< 10g2 K


If the landscape consists of land-cover categories in
approximately equal proportions, and if the patches found
on the landscape are "large" (compact and extensive) relative

to mn, then the information contained in the coarse-scale map


about the fine-scale map is large, i.e. little information is lost
when the resolution becomes more coarse. However, the
more the landscape is dominated by one or a few land-cover
categories, or the more the landscape becomes more dissected

Magnus Ekstrom, PhD, is an associate professor at


the Center of Biostochastics, Swedish University of
Agricultural Sciences. His research interests include
estimation theory, resampling techniques, spatial
statistics, and classification theory. His address:
Center of Biostochastics, Swedish University of
Agricultural Sciences, SE-901 83 Umea, Sweden.
magnus.ekstrom@sekon.slu.se

576 ? Royal Swedish Academy of Sciences 2003 Ambio Vol. 32 No. 8, Dec. 2003
http://www.ambio.kva.se

This content downloaded from 202.43.95.117 on Thu, 09 Jun 2016 06:26:23 UTC
All use subject to http://about.jstor.org/terms

Das könnte Ihnen auch gefallen