Download as pdf or txt
Download as pdf or txt
You are on page 1of 8

Landscape Ecology vol. 10 no.

2 pp 113-120 (1995)
SPB Academic Publishing bv, The Hague

Effects of sensor spatial resolution on landscape structure parameters

Barbara J. Benson 1 and Mark D. MacKenzie 1,2


1Center for Limnology, University of Wisconsin-Madison, 680 N. Park Street, Madison, Wisconsin
53706, USA; 2Environmental Remote Sensing Center, University of Wisconsin-Madison, 1225 W.
Dayton Street, Madison, Wisconsin 53706, USA

Keywords: remote sensing, grain, landscape pattern, texture, scale

Abstract

We examined the effects of increasing grain size from 20 m to 1100 m on landscape parameters characterizing
spatial structure in the northern Wisconsin lake district. We examined whether structural parameters remain
relatively constant over this range and whether aggregation algorithms permit extrapolation within this range.
Images from three different satellite sensors were employed in this study: (1) the SPOT multispectral high
resolution visible (HRV), (2) the Landsat Thematic Mapper (TM), and (3) the NOAA Advanced Very High
Resolution Radiometer (AVHRR). Each scene was classified as patches of water in a matrix of land. Spatial
structure was quantified using several landscape parameters: percent water, number of lakes (patches), aver-
age lake area and perimeter, fractal dimension, and three measures of texture (homogeneity, contrast, and
entropy). Results indicate that most measures were sensitive to changes in grain size. As grain size increased
from 20 m using HRV image data to 1100 m (AVHRR), the percent water and the number of lakes decreased
while the average lake area, perimeter, the fractal dimension, and contrast increased. The other two texture
measures were relatively invariant with grain size. Although examination of texture at various angles of ad-
jacency was performed to investigate features which vary systematically with angle, the angle did not have
an important effect on the texture parameter values. An aggregation algorithm was used to simulate addition-
al grain sizes. Grain was increased successively by a factor of two from 20 m (the HRV image) to 1280 m.
We then calculated landscape parameter values at each grain size. Extrapolated values closely approximated
the actual sensor values. Because the grain size has an important effect on most landscape parameters, the
choice of satellite sensor must be appropriate for the research question asked. Interpolation between the grain
sizes of different satellite sensors is possible with an approach involving aggregation of pixels.

Introduction and broad-scale landscape measures offer new


methods to "scale u p " or extrapolate (Hall et al.
Ecologists are being confronted with scale prob- 1988; O'Neill 1988) as ecologists attempt to use
lems as they address global issues (O'Neill 1988, knowledge based on analyses at finer spatial scales
Woodmansee 1988). H u m a n effects on the environ- to meet this challenge.
ment have resulted in a globalization of environ- Remote sensing satellite sensors provide synoptic
mental problems entailing disturbances occurring image data that are useful for ecologists working at
on broader spatial scales over shorter temporal global scales. In this study we use imagery from
scales (di Castri and Hadley 1988). Remote sensing three satellite sensors which have different levels of
114

spatial resolution. The high resolution visible mul- grain from 5-m aerial photography to 80-m Land-
tispectral sensor (HRV) onboard the French Sys- sat multispectral scanner (MSS) data. They assessed
tkme P o u r l'Observation de la Terre (SPOT) satel- the degree of homogeneity vs. heterogeneity of
lite has a spatial resolution of 20 m, the Thematic study sites in the Konza Prairie Research Natural
Mapper (TM) sensor onboard the US Landsat - 4 Area and evaluated the appropriateness of mea-
and - 5 satellites has a spatial resolution of 30 m, sures at different scales for this assessment. Their
and the Advanced Very High Resolution Radiome- results suggested that heterogeneous areas must be
ter (AVHRR) sensor onboard the US NOAA satel- measured at finer grain sizes.
lites has a spatial resolution of 1.1 km (Lillesand Turner et al. (1989b) framed their study as a con-
and Kiefer 1994). Our objective was to investigate tribution to the science of scale. They attempted to
the effect of the spatial resolution of satellite sen- derive some general rules for effects of scale change
sors on landscape structure as revealed by various by examining the effects of grain and extent on
landscape parameters. three measures of landscape structure: diversity,
Much attention has been focused on the impor- dominance, and contagion. Analytical results,
tance of spatial and temporal scale in studying eco- based on simple artificial maps of a two-phase land-
logical phenomena (Allen and Starr 1982; Delcourt scape (both random and with contagion, i.e.,
et al. 1983; O'Neill et al. 1986; Meentemeyer and clusters of the same land cover type), were generat-
Box 1987; Urban et al. 1987). Spatial scale has two ed for the proportion of each cover type in a land-
components: grain, the finest resolution possibile scape as grain increased. They also examined the ef-
within a data set, and extent, the size of the study fects of scale change by aggregating actual land-
area (Turner et al. 1989a). Turner et al. (1989a) give scape data from USGS land use data. Diversity
as one definition of extrapolation " t o transfer in- decreased linearly with increasing grain whereas
formation from one scale to another (either grain dominance and contagion were sensitive to the
size or extent)." Landscapes are generated by decrease in the number of cover types as grain in-
processes occurring at different spatial and tem- creased.
poral scales (e.g., microtopography, edaphic varia- Much remains to be explored about the effect of
tion, major landforms). Hence, when the scale changes in scale. In this study, we examined the ef-
of analysis is changed, different structure may fects of changing grain on parameters characteriz-
emerge. Extrapolation of structural measures will ing spatial structure over the range of grain sizes
only be possible over limited ranges of scale. Sever- represented by HRV, TM, and AVHRR satellite
al previous studies have examined the effect of scale imagery. We investigated whether structural
change on measures of spatial structure (Woodcock parameters remain relatively constant and whether
and Strahler 1987; Nellis and Briggs 1989; Turner aggregation algorithms permit extrapolation within
et al. 1989b). this range.
Woodcock and Strahler (1987) depicted a change
in local variance as grain increased for several
different environments (forested, agricultural, and Methods
urban/suburban). Each environment was viewed at
two different initial grains (fine resolution aerial Images from three satellite sensors, each with a
photography and Landsat-TM) and then the images different spatial resolution, were analyzed to study
were degraded to coarser grain. Their primary con- the effects of scale change on measures of land-
clusions were directed toward specifying, based on scape structure. The grain of the sensors (i.e., pixel
the amount of local variance in a scene, which in- size) varies from 20 m for HRV, 30 m for TM, and
formation extracting technique is appropriate for 1.1 km for AVHRR. Images from each satellite
thematic classification. which contained the North Temperate Lakes Long-
Nellis and Briggs (1989) applied textural contrast Term Ecological Research (LTER) site (located in
algorithms to a series of digital images ranging in northern Wisconsin) were acquired. The dates for
115

each image were: June 10, 1988 for HRV; June 9,


1988 for TM; May 30, 1990 for A V H R R . These im-
ages were processed using the Erdas PC-based soft-
ware package (Erdas 1990).
The areal extent of the study area was held
constant by using the H R V image to define the
boundaries for subscenes of T M and A V H R R
imagery. The H R V image used represented an area
of 3,614 km 2. Each of the images, HRV, TM, and
A V H R R , was georectified to the Universal Trans-
verse Mercator (UTM) coordinate system using
nearest neighbor resampling.
We then classified each subscene into two
categories: land and water (Fig. 1). The choice of
these two categories was dicated by our broader
research goals. These two cover categories repre-
sent a simple but ecologically meaningful system
for considering landscape structure. We used the
near infrared band of each sensor because this band
is appropriate for separating land from water
(Lillesand and Kiefer 1994). All three satellites
record image data in this portion of the electromag-
netic spectrum. The wavelengths for each near in-
frared band are: 0 . 7 9 - 0 . 8 9 ~tm for HRV, 0 . 7 6 -
0.90 ~tm for TM, and 0 . 7 2 - 1 . 1 0 ~m for A V H R R
(Lillesand and Kiefer 1994). The separation of land
from water was based on a histogram of the raw
(i. e., no atmospheric or radiometric corrections) in-
C. frared band digital numbers (DN). The histogram
for this band is bimodal (Fig. 2) with the first mode
' ' == i" .=" corresponding to water, the second to land. The
m= in m
~. 9 =.= cutoff value between water and land was deter-
mined subjectively based on examination of in-
dividual DN values.
lI m Accuracy of classification based on the near in-
frared (Band 4) of TM was assessed using a digi-
, tized coverage of lake boundaries f r o m USGS 7.5'
topographic quadrangles of a 44,467 ha area within
the T M subscene. The agreement between the digi-
tized coverage and the classified subscene was de-
termined using an overlay module in Erdas. Using
the digitized coverage as a reference, 92 percent of
0 10 20 30 40 50 KM
the water and 99 percent of the land were correctly
Fig. 1. Patches of water (shown in black) in a matrix of land classified by the TM band 4 cutoff value.
(white) based on classification of satellite imagery with the near After classifying an image into land and water
infrared band from A) HRV; B) TM; and C) AVHRR. The
categories, various landscape parameters were cal-
scenes represent a 3,614 km2 area surrounding the North Tem-
perate Lake LTER site in northern Wisconsin and illustrate the culated using Erdas routines or Fortran programs
effect of satellite resolution on spatial structure. written for this study. The landscape parameters in-
116

2 land

T
1.5lwater
g
"" 1
9
o

0.5

0
0 0 0 0 0 0 0 0 0 0 0 0 0 0
r CO ~" LO r I~- CO O~ 0 ~ 04 CO ~"

Digital Number (DN)

Fig. 2. Histogram of the near infrared band from the HRV image showing the bimodal distribution of digital number values.

cluded percent water, number of patches (i.e., of adjacency can be done for various angles ena-
lakes), average area of patches, average perimeter bling one to look for features which vary systemati-
of patches, fractal dimension, and three measures cally with angle.
of texture (homogeneity, contrast, and entropy). We chose to use three of these texture measures
For all calculations except percent water, water because they were interpretable as aspects of land-
patches which intersected the boundary of the sub- scape heterogeneity. Our use of these measures was
scene were eliminated f r o m the analysis. based on class values (water, land) rather than DN
The fractal dimension (Mandelbrot 1982, Bur- values. Let p(i,j) represent the frequency of ad-
rough 1986) for area/perimeter relations is given by jacency of a pixel of value i with one of value j in
the spatial dependence matrix normalized to a
P = k A D/2 proportion. Homogeneity (or the second angular
moment) measures the extent to which there are few
where P is the perimeter, A is the area, D is the frac- dominant gray-tone transitions.
tal dimension and k is a constant. For areas and
perimeters based on raster data, the fractal dimen- homogeneity = ).7, 12 {p(i,j) }2
sion is two times the slope of the regression of log i j
(perimeter/4) against log(area) (Turner 1987). Contrast is a measure of local variation in the
Haralick et al. (1973) presented fourteen meas- image.
ures of texture derived f r o m a spatial dependence
matrix which summarizes the adjacency of all pos- Ng-1 INg Ng 1
contrast = n2 ~ ~ p(i,j)
sible DN values. To calculate texture, one starts n=0 i=l j=l
with an image where DN values are expressed on a li-jl =n
gray scale and constructs the gray-tone spatial de-
pendence matrix by looking at the frequency of where Ng equals the number of distinct gray levels.
values adjacent to one another. This examination Note that the sum is weighted by n 2 which indicates
117

Table 1. Comparison of landscape parameters from HRV, TM, cent water decreased by 44 percent and only two
and AVHRR. percent of the lakes remained. The patterns of in-
crease or decrease in parameters with increasing
Parameter HRV TM AVHRR
grain were not monotonic, however. Unexpected
resolution (m) 20 30 1100 deviations from monotonic changes included the
percent water 11.9 10.9 6.7 average patch area and perimeter decreasing as
number of patches 3428 2829 63
grain increased from HRV to TM.
average patch area (ha) 11.7 7.8 360
average patch perimeter (m) 1324 1123 8600
Extrapolation from HRV to coarser grain size,
fractal dimension 1.23 1.25 1.36 based on aggregation, predicted relations between
homogeneity 0.77 0.79 0.80 parameter value and grain size (Fig. 3). For exam-
entropy 0.43 0.42 0.46 ple, the number of lakes decreases rapidly as grain
contrast 0.013 0.017 0.079
size increases while the percent water declines in an
approximately linear fashion. The mean lake area
increases in an approximately linear fashion with
how much the gray tone i differs from the gray tone increasing grain size while the mean lake perimeter
j. Thus this measure is appropriate for images with increases more gradually. In this way, aggregation
gray-tone values rather than cover types. In our based on a "majority rules" algorithm provided a
case with just two values, land and water, the mea- reasonable method for extrapolating landscape
sure reduces to the proportion of transitions of land parameters from fine to coarse grain. The actual
to water times two. Entropy is a measure of the landscape parameter values based on the A V H R R
complexity of the image. image support the extrapolated values with the best
correspondence (based on percent difference) being
entropy = - ]~ ]~ p(i,j) log (p(i,j)) for percent water (Fig. 3).
i j Two texture measures were relatively invariant
Homogeneity and entropy are negatively related. over changes in grain (Fig. 4). Homogeneity varies
Contrast is a more independent measure. It is possi- over the range of grain sizes by less then four per-
ble to construct landscapes in which decreasing cent; entropy varies by nine percent. However, con-
homogeneity can either decrease or increase con- trast of the A V H R R subscene was more than six
trast. times greater than for HRV. Texture parameters
In order to examine the feasibility of extrapolat- are calculated from a matrix of frequencies of
ing across scales, an aggregation algorithm was values being adjacent to one another. It was possi-
used to extrapolate from the fine grain HRV image ble to examine adjacency along horizontal and ver-
(20 m) to coarser grains of 40, 80, 160, 320, 640, tical directions as well as 45 degrees and 135
and 1280 m pixel sizes. The aggregation algorithm degrees. Angle did not have an important effect on
("majority rules") assigns the value of the majority the values of texture parameters (Fig. 4).
pixel values to the aggregated pixel. We compared
the extrapolated values of various landscape para-
meters to the computed parameter values based on Discussion
classified TM and A V H R R images.
Our results support the view that scaling issues must
be addressed when using satellite imagery to derive
Resulls landscape parameters since most landscape para-
meters are sensitive to changes in grain. We ex-
Most, but not all, landscape parameters were sensi- plored the feasibility of estimating parameters at
tive to changes in grain over the range in spatial different scales and found the estimation of para-
resolution from HRV to A V H R R (Table 1). As the meters at coarser grain sizes using an aggregation
pixel sixe increased from 20 m to 1.1 kin, the per- algorithm was reasonably accurate. The reverse
118
a)
4000 14-

9 aggregated value
0 satellite value
12
3000 ' 9 aggregated value
0 satellite value

10

"6
2000 9

~
8
E
E

1000 "

~ - - - - - .' -. - - .'~----~-1. . ,
0 . . . . j . .
4 . . . . i . . . . f . . . . i

0 500 1000 1500 0 500 1000 1500

grain (m) grain (m)

c) d)
4 0 0 84 10000'

0
0
~. 8000 '
300" v

*~ 6000 '
._E
200"

-~ 4000 '

E ted value
100-
0 satellite value E 2000.
alue
0 satellite value

. . . . i . . . . i . . . . i 0
500 1000 1500 500 1000 1500

grain (m) grain (m)

Fig. 3. The relation of a) number of lakes, b) percent water, c) mean lake area, and d) mean lake perimeter to grain (pixel size) for ag-
gregated images and actual satellite images.

estimation, however, from coarse-grain to fine- meters. Our results suggest that due to the depen-
grain may be more problematic unless an empirical dence of landscape parameters on grain size, use of
relation has been established since fine-scale fea- satellite sensors with coarser spatial resolutions is
tures m a y not be recoverable from coarse grain in- not a simple solution to investigating questions
formation. related to broader spatial scale phenomena.
While the spatial extent, increased temporal fre- The decrease in the number of lakes and the per-
quency, and accessibility of A V H R R images make centage of water in the landscape as grain increases
them useful for regional and global scale research is consistent with the results of Turner et al. (1989b)
(Goward et al. 1985; Iverson et al. 1989; Running that non-dominant cover types decrease as grain in-
and Nemani 1988), scaling issues must be consid- creases. This decrease was less rapid for our land-
ered in their use for estimating landscape para- scape than for the r a n d o m two-phase landscape
119

a 9o peared to remain relatively constant. Because tex-


120 60 ture measures quantify landscape and regional spa-
D spot
A tm
tial heterogeneity, such measues may be important
avhrr parameters for regional or global process modeling.
We considered the effect of the angle used for
proximity of pixels in texture parameter calcula-
18o +-- o8 - o.~ - 04 - o.2- o - o.2 - o.4- 06 -[~ --I- o
tions. Angle did not appear to have an effect. In the
homogeneity
study area, the pattern of glacial moraines is in two
orthogonal directions (north-south and east-west).
b The two distinct directions seem to cancel each
9O
other in the examination of an effect of angle.
Aggregation results which used HRV imagery to
[]
150 120 l 60 3O
A
spot
tm extrapolate to the grain of A V H R R scenes provided
avhrr reasonable estimates of landscape parameters at
coarser grains. Wiens and Milne (1989) discussed
the concept of domains of scale. A domain is a
o
range of scale values over which extrapolation is
entropy adequate. For the landscape parameters we studied,
the range of spatial resolutions from HRV to
A V H R R appeared to be within such a domain.
~; 90
120 60 Some caveats are in order, however. The land-
scape we analyzed was relatively simple. Variability
[] spot
within the landscape will be a major determinant of
150 30 =.
tm
avhrr the possibility of "scaling u p " . For landscapes with
greater local variability, the results of aggregation
might be less adequate.
1 8 0 ~ 0
contrast In addition, our aggregation results are depen-
dent on the aggregation algorithm we used which
Fig. 4. Texture values for HRV, TM, and A V H R R as a function was based on "majority rules". Milne (1992)
of angle for three texture parameters: a) homogeneity, b) entro- describes properties of various aggregation al-
py, and c) contrast.
gorithms and suggests that important properties of
landscapes are not preserved under some aggrega-
Turner et al. (1989b) analyzed because of the ag- tion algorithms. An extension of our research
gregated arrangement of water pixels in the land- would be to modify the aggregation algorithm
scape. Because the A V H R R image was not taken in either to more closely approximate the actual physi-
the same year as the HRV and TM images, differ- cal aggregation of radiance by a satellite sensor or
ences between the images could be due to an inter- to explore aggregation algorithms such as Milne in-
annual difference in the proportion of land to vestigates which preserve properties of the land-
water. However, when one considers the 1.1 km scape (such as fractal dimension).
grain size of the AVHRR, we feel this possible An additional caveat as we explore aggregation
inter-annual difference had no significant effect on as a method of scaling up is that there may be a sig-
the A V H R R classification results. nificant difference between extrapolating and scal-
Our expectation was that texture parameters ing up. Extrapolation is based on a function which
would be scale dependent. However, two texture allows one to predict values outside the measured
parameters were relatively invariant with changes in range of scale. Scaling up can involve qualitative
grain size. Over the range of grain represented from shifts when the dominant processes structuring pat-
HRV to AVHRR, homogeneity and entropy ap- terns change as scale changes.
120

Studies which examine how landscape para- Haralick, R.M., Shanmugam, K. and Dinstein, I. 1973. Textur-
meters change with scale (e.g., the effect of satellite al features for image classification. IEEE Transactions on
systems, man, and cybernetics. SMC-3" 610-621.
resolution) can provide a better understanding of
Iverson, L.R., Cook, E.A. and Graham, R.L. 1989. A tech-
the usefulness of remote sensing and large scale nique for extrapolating and validating forest cover across
landscape measures. However, satellite imagery large regions: calibrating AVHRR data with TM data. Int. J.
predetermines grain and, therefore, the choice of Rem. Sens. 10: 1805-1812.
satellite sensor must be appropriate for the research Lillesand, T.M. and Kiefer, R.W. 1994. Remote sensing and im-
age interpretation. 3rd ed. John Wiley and Sons, New York.
question asked. We show that grain has an impor- Mandelbrot, B.B. 1982. The fractal geometry of nature. Free-
tant effect on most landscape parameters. There is man, New York.
a large gap between the grain size of a TM image Meentemeyer, V. and Box, E.O. 1987. Scale effects in landscape
and that of AVHRR. We demonstrate that an ap- studies. In Landscape heterogeneity and disturbance, pp.
15-34. Edited by M.G. Turner. Springer-Verlag, New York.
proach involving aggregation of pixels can be used
Milne, B.T. 1992. Indications of landscape condition at many
to interpolate between sensor grains. scales. In Ecological indicators. Vol. 2. pp. 8 8 3 - 8 9 . Edited
by D.H. McKenzie, D.E. Hyatt, and V.J. McDonald. El-
sevier Applied Science, London and New York.
Nellis, M.D. and Briggs, J.M. 1989. The effect of spatial scale
Acknowledgments
on Konza landscape classification using textural analysis.
Landsc. Ecol. 2: 93-100.
We appreciate the comments on drafts of the O'Neill, R.V, 1988. Hierarchy theory and global change. In
manuscript by John Magnuson and two anony- Scales and global change, pp. 29-45. Edited by R. Rosswall,
mous reviewers. The research is part of the North G. Woodmansee and P. Risser. John Wiley and Sons, New
York.
Temperate Lakes Long-Term Ecological Research
O'Neill, R.V., DeAngelis, D.L., Waide, J.B. and Allen, T.F.H.
project funded by the National Science Founda- 1986. A hierarchical concept of ecosystems. Princeton
tion, Grants BSR9011660. University Press, Princeton, New Jersey.
Running, S.W. and Nemani, R.R. 1988. Relating seasonal pat-
terns of the AVHRR vegetation index to simulated pho-
tosynthesis and transpiration of forests in different climates.
References Rem. Sens. Env. 24: 347-367.
Turner, M.G. 1987. Spatial simulation of landscape changes in
Allen, T.F.H. and Starr, T.B. 1982. Hierarchy: perspectives for Georgia: a comparison of 3 transition models. Landsc. Ecol.
ecological complexity. University of Chicago Press, Chicago. 1: 29-36.
Burrough, P.A. 1986. Principles of geographical information Turner, M.G., Dale, V.H. and Gardner, R.H, 1989a. Predicting
systems for land resources assessment. Clarendon Press, across scales: theory development and testing. Landsc. Ecol.
Oxford. 3: 245-252.
Delcourt, H.R., Delcourt, P.A. and Webb, T. 1983. Dynamic Turner, M.G., O'Neill, R.V., Gardner, R.H. and Milne, B.T.
plant ecology: the spectrum of vegetational change in space 1989b. Effects of changing spatial scale on the analysis of
and time. Quaternary Science Reviews 1: 153-175. landscape pattern. Landsc. Ecol. 3: 153-162.
di Castri, F. and Hadley, M. 1988. Enhancing the credibility of Urban, D.L., O'Neill, R.V. and Shugart, H.H., Jr. 1987. Land-
ecology: interacting along and across hierarchical scales. scape ecology. Bioscience 37:119-127.
GeoJournal 17" 5-35. Wiens, J.A. and Milne, B.T. 1989. Scaling of 'landscapes' in
Erdas, Inc. 1990. Erdas User's Manual. Erdas, Inc. Atlanta, landscape ecology, or landscape ecology from a beetle's per-
GA. spective. Landsc. Ecol. 3:87 96.
Goward, S.N., Tucker, C.J. and Dye, D.G. 1985. North Ameri- Woodcock, C.E. and Strahler, A.H. 1987. The factor of scale
can vegetation patterns observed with the NOAA-7 advanced in remote sensing. Remote Sensing of Environment 21:
very high resolution radiometer. Vegetatio 64: 3-14. 311-332.
Hall, F.G., Strebel, D.E. and Sellers, P.J. 1988. Linking Woodmansee, R.G. 1988. Ecosystem processes and global
knowledge among spatial and temporal scales: vegetation, change. In Scales and global change, pp. 11-27. Edited by R.
atmosphere, climate and remote sensing. Landsc. Ecol. 2: Rosswall, G. Woodmansee and P. Risser, John Wiley and
3-22. Sons, New York.

You might also like