Appendix revission UW TL

advertisement
1
Continuity and Contrast: parameters in neutral point-pattern landscapes.
2
Spectral density has often been used for time series and lattice data (for relevant references, see
3
Mugglestone & Renshaw 1996). Here, we use a spectral representation to calculate two measures
4
that capture essential characteristics in point-pattern landscapes. The basis of the spectral method
5
follows from the Fourier theorem, which states that continuous time series, or surfaces and higher
6
dimensions, can be represented by a combination of sine and cosine terms with different
7
frequencies and amplitudes. Thereby, analysis of the time series, or surfaces, can be performed
8
on the sine and cosine terms. From physics, we use the notation power spectral density function
9
(PSDF) for this representation of a time serie or surface by the amplitudes of sine and cosine
10
terms of different frequency.
11
Continuity
12
Large values of Continuity (γ) in point patterns reflect that nearby areas have similar density, and
13
hence γ is a measure of autocorrelation. It is defined from the spectral representation of the
14
landscape such that if the spectral density is dominated by the amplitudes of low frequencies,
15
then there is a higher degree of autocorrelation. This is a point-pattern measure, akin to a two-
16
dimensional measure of 1/f noise (Halley et al. 2004, Keitt 2000). Thus a higher value of
17
Continuity is analogous to a redder 1/f noise. As is done when determining the colour of 1/f
18
noise, we analyse the autocorrelation as a linear regression fitted to the log(frequency) vs.
19
log(amplitude) and thereby also infer self-similarity. The slope of the regression line is the
20
parameter Continuity (γ). In our analysis, we assume isotropy, which implies identical patterns in
21
all directions. Hence, we can collapse the spectral analysis into a one-dimensional assessment
22
(i.e. a regression line instead of a surface) after the Fourier transform of the two-dimensional
23
point pattern. To calculate the PSDF of the point pattern distribution, denoted PSDFpp, we use the
24
efficient method that was developed by Mugglestone and Renshaw (1996):
2
25

 

PSDFPP    cos(C p xi )     sin( Cq yi ) 
 i
  i

2
(A.1)
26
where (μp,μq) are frequencies for Fourier transform of grid data, (xi,yi) are coordinates in the point
27
pattern, and C is a constant that determines the number of frequencies. Mugglestone and
28
Renshaw stated that the number of frequencies used should not exceed the number of points in
29
the point pattern, Np, in order to keep them independent. Therefore, given a quadratic landscape
30
we chose C to be
31
number of components in the two dimensional spectral representation.
N p 2 rounded down to the nearest integer, and hence nF =2C×2C is the
32
33
Contrast
34
As a point-pattern equivalent of coefficient of variation, we introduce Contrast,  , to measure
35
the difference in density between sparse and dense areas. To formulate an equation for that
36
measure, we first have to relate it to how variance in the continuous case can be measured using
37
spectral representation. Note that when time series and surfaces are represented by sine and
38
cosine terms after the Fourier transform, it is the amplitudes of the frequencies that determine the
39
variance. If M is the number of components in the PSDF, the variance,  2 , is calculated as
2 
40
1
mean( PSDF  PSDF (origin ))
M
(A.2)
41
The mean is represented by the amplitude in the origin1 of the spectral representation,
42
PSDF(origin), and hence the coefficient of variation (standard deviation divided by the mean) is
CV 
43
1
1
mean( PSDF  PSDF (origin ))
PSDF (origin ) M
44
Substituting PSDF with PSDFpp from equation A.1 where nF is the number of spectral
45
components we end up with an equation of the point-pattern measure Contrast (  ):
CVPP 
46
47
1
1
mean( PSDFPP  PSDFpp (origin ))
PSDFPP (origin ) nF
(A.3)
(A.4)
Generating neutral point-pattern landscapes
1
Note that if implemented in Matlab using built-in functions of Fourier transform, for example
the fft function, the amplitude in the PSDF(origin), represents the sum instead of the mean.
48
To obtain neutral point-pattern landscapes (NPPLs) with given characteristics, we generated
49
lattice landscapes of size m×m. The density defines the probability of a point in the landscapes.
50
We first generated two-dimensional 1/f-noise (denoted LG) using a method similar to that
51
presented by Halley et al. (2004), and hence this is still a representation of a lattice landscape and
52
not a point pattern. The values in LG are normally distributed and thus not suitable for describing
53
probabilities, because they may include negative values. Although it might be possible to solve
54
this by truncating, we found that that would not allow for generating sufficiently high values of δ.
55
Therefore, we transformed LG using spectral mimicry. This method was introduced by Cohen et
56
al. (1999) and has been employed to construct stationary time series with different Fourier
57
spectra. Cohen et al. mainly used their technique to achieve transformation to normally
58
distributed values with specific mean and variance. We instead transformed LG to LΓ using a
59
gamma distribution (which contains no values <0) with mean=1/m2 and a coefficient of variation
60
δL. Point locations were distributed according to the probabilities given by LΓ.
61
The γ and δ of the spectral point pattern are determined by γL and δL, but they are altered by the
62
distribution of points as this adds some randomness. Accordingly, we measured these quantities
63
in the generated point pattern landscape (see the method described above). The relationship
64
between the spectral point-pattern values of γ and δ used in our study and the γL and δL required
65
to generate them was found iteratively. Furthermore, we observed that the linear relationship in
66
the power spectra was maintained better for large grids (values of m), and hence we used
67
m=2000. The autocorrelation parameter Continuity generates a general aggregation pattern,
68
whereas the variance within the system is reflected by the Contrast parameter. A high Contrast
69
parameter will impose more isolated clusters of aggregated points on the aggregation structure
70
defined by the Continuity parameter.
71
References
72
73
74
Cohen J. E., Newman C. M., Cohen A. E., Petchey O. L. & Gonzalez A. 1999 Spectral mimicry:
A method of synthesizing matching time series with different Fourier spectra. Circ. Syst. Signal
Process. 18, 431-442.
75
76
Halley, J. M., Hartley, S., Kallimanis, A. S., Kunin, W. E., Lennon, J. J. & Sgardelis, S. P. 2004
Uses and abuses of fractal methodology in ecology. Ecol. Lett. 7, 254-271.
77
Keitt, T. H. 2000 Spectral representation of neutral landscapes. Landscape Ecol. 15, 479-493.
78
79
80
Mugglestone, M. A. & Renshaw, E. 1996 A practical guide to the spectral analysis of spatial
point processes. Comput. Stat. Data An. 21, 43-65.
Download