Slide 1We were interested in the effects of certain environmental variables on NDVI. Based on a literature review, we found that certain variables appeared to have more effect depending on the scale at which the variables were compared. This led to some interesting questions. Are the results of studies on these variables skewed due to the scale at which they are being studied? Is it possible to predict which scale should be used when examining the relationship between certain variables? Can the relationship of the variables be examined across scales and compared statistically?
In ecological studies, the scale at which phenomenon are studied affects the relationships that are observed. Yet in most studies involving remote sensing techniques, the scale of observation used is completely anthropocentric; it is based on the scale of available data. The scale at which the ecological processes occur, and the scale at which the relationship of these processes may be observed is often ignored.
J.A. Wiens (1989) observed that To understand the drama, we must view
it on the appropriate scale...many ecologists have behaved
as if patterns and the processes that produce them are insensitive
to differences in scale and have designed
their studies with little explicit attention to detail.
The same is true in the field of remote sensing. Many studies use Landsat TM
data and therefore accept the patterns that they find at that scale,
not recognizing that the dynamics observed may be evident only
due to the scale of the study and may not be pervasive at all scales.
Many ecologists have argued for multiscale studies, both spatially and
temporally to understand the scale of ecological processes (Turner, 1989).
In particular, the relationships between certain processes could be the
most sensitive to changes in scale (Wiens, 1989).
The problem in choosing a particular scale, is that you may
observe a process that is nothing more than an artifact
of the scale you have chosen.
(Belovsky, pers.comm.)
The point is not that one process is dominant in the definition of
a variable, for example, the occurance of vegetation, because
vegetation is influenced by all the variables of its environment. Rather,
the concern is that the methods of our observation may cause us to
observe one dominant process at one scale and a different dominant
process at another scale. Turner et. al. observed that The variables
influencing a process may or may not change with scale, but a shift in
the relative importance of variables often occurs.
To test the change in the relationship, we held the extent of our study at the same size and changed the pixel size of the data. We expected a decrease in spatial variance with an increase in pixel size due to the changes in within-pixel variation. With increasing pixel size, more data is held within one pixel and is lost. (Wiens, 1989). We ran a variance on each nonemperical variable and then a covariance for each variable against NDVI comparing both across scales (Bian, 1993).
The problem with studying this type of variation is that the results may vary based on the type of data being studied and its spatial variation. For example, data that is aggregated will not get lost, whereas data that is spread out evenly could easily be lost due to the within-pixel averaging. Another problem is that we were forced to aggregate pixels to change scales of study instead of using data that was developed at particular scales. The aggregation of the pixels by Imagine introduces some error and may cause some of the statistics to show not only the relationship between the variables but also the outcome of the pixel aggregation process. This effect should be included in future calculations (adding error onto the variance side of the equation) or data that is gathered at different scales should be used, such as on site observations, aerial photos, SPOT and Landsat data. Even in this case, you may also be testing the differences in the data collection process itself, rather than the data. However, calculating this error may be difficult.
We expected the physical attributes to have a stronger relationship at larger pixel sizes, based primarily on the observation of other researchers who found thast climate, atmosphere and geological effects were only observable at small scales, because biological influences became dominate at large scales (Clark, 1985; Greig-Smith, 1979; Woodward, 1987; Levin, 1989; Carpenter, 1989; Wiens, 1989).
Slide 1 illustrates the amount of variance gained by an increase in scale.
On the left is the NDVI of a 30m TM image.
On the right, the NDVI from 2m resolution airphoto data.
Many variables affect the differences in vegetation shown below.
Next Page
Back to Front Page