2. DEFINITIONS AND NOTATION
Since we are interested in
(mathematically) about
CPs, we shall
notation.
saying something
the spatial distribution of
introduce some basic ideas and
Our main tool will be the comparison of the spatial
distribution of the (say "n") given CPs with
respect to the uniform distribution of (say "n",
again) points in the unit square (that we will
write "S"); whenever data do not come in this
scale, the transformation is obvious.
The n CPs are said to be uniformly distributed in S
iff their coordinates (x,, ys x y.) are
the outcomes of 2n independent identically
distributed random variables with the uniform
distribution in the [0,1] interval (the reader
should notice that these are not n evenly spaced
points). We will test the (empirical) distribution
of this set of n coordinates, against the
hypothesis of uniform distribution over S, using
simulation. The use of simulation is needed since
there are no theoretical results about the tests we
shall use.
Let d, be the distance of the point j to its
nearest neighbour; let d''-(d,,...,d ), with m-n(n-
1)/2 be the n(n-1)/2 dimensional vector of the
nearest neighbour distances among the n points.
Denote d' the vector d'' with its elements sorted
in ascending order, and d the vector formed by the
first n elements of d'. We will compare the values
of the n components of d with other values obtained
by the sampling from the hypothesized distribution;
this sampling is repeated a fixed number (M) of
times ("iterations"), usually with M=19 or M=99.
3. THE ALGORITHM
The algorithm is as follows (Diggle, 1983):
1) For j » 1 to n do
min (j) =2
max (j) = 0
mean (j) = 0
2) For k= 1 to M do
2.1) Observe u,,...u 3 V,...V : 2n independent
outcomes of a random variable uniformly
distributed over [0,1].
2.2) Form the coordinates set for the current
iteration CDAPPELEE (ub v).
2.3) Calculate e - (e,,..., e ), where e is a
n-dimensional vector defined in the same
manner as d above, for the current set of
coordinates,
2.4) For j » 1 to n do
mean (j) = mean (j) -* e,
if e.€ min(j) then
lét min(j) = e,
if e.>max(j) then
let max(j) = e
3) For j » 1 to n do
mean(j) = mean(j)/n
4) Draw the plot
mean(j), min(j), d., max(j)
for every j between 1 dnd n.
110
5) End
4. INTERPRETATION OF RESULTS
For every j, min(j) and max(j) form an approximated
confidence interval of level M/(M-1) for d,
obtained by simulation. mean(j) gives a (roughl
though useful) approximation of the mathematical
expectation of the random variable E., whose
outcomes e, were observed. The plot “shows a
confidence dnvelope for the input data: if the data
lie in the region above the minima 8nd below the
maxima we accept the hypothesis of uniform
distribution over the sets. If the data lie in the
region below the minima, there is evidence
favouring an "attraction" model for the location of
CPs, i.e. they are more spatially clustered than
independent (uniformly distributed) points. If the
data lie in the region above the maxima, there is
evidence favouring a "repulsion" model for the
location of CPs., i.e. they tended to be further
appart from each other than independent (uniformly
distributed) points.
5. METHODOLOGY
The algorithm above is very simple to implement,
and we feel that it might be helpful in the
diagnostic and identification of problems regarding
the spatial distribution of CPs. (Listings in the
CH programming language available from the authors
upon request).
Given n, the number of CPs, and once obtained the
values mean(j), min(j), max(j) for every j, these
envelopes could be understood as frontiers between
the following regions: (i) the "acceptable"
([min(j), max(j)]); (11) the "optimal" ar
2 ,max(j))) and (iii) the "unacceptable"
((0,min(j))).
Regions (i), (ii) and (iii) correspond to the
labels we would attach to the data sets whose
vectors d fall into, with respect to their spatial
quality.
6. ANALYSIS OF THREE DATA SETS
In the following we used M=19 obtaining,
aproximated 957 confidence intervals for d
thus,
The first data set (W1) was used in the
registration of an image of Tucurui, Para State,
Brazil. The whole image matching (including the
evaluation of its accuracy) involved seven CPs
whose spatial distribution exhibits a strong
tendency to the "attraction", as previously defined
(see Fig, la). The plot 1b shows that most of these
points are located in region (iii), considered as
"unacceptable",
(W2) comes from the area of
Roncador, Mato Grosso State, Brazil. Ten
used in this case, where an "atraction"
model is also suited to describe their spatial
distribution (weaker, though, than the Wl set; see
Fig. 2a). The plot 2b shows that most of the points
fall into region (i), considered as "acceptable",
with a few shifted towards region (iii).
The second data set
Serra do
CPs were
The Wl and W2 data sets correspond to regions in
the brazilian legal Amazonas, and are typical cases
of the conditions found in that area: few or no
features like crossroads, bridges, etc. preclude
the se
distri
that a
The W3
Paulo
select
spatia
model
qualit
"accer
Thus,
assess
distri
sugges
Hardwe
Langue
Specie
Pseudc
1990).
CPU ti
result
The ai
Moraes
supply
Roncac
BUSTO:
GFSR[(
de Mai
DIGGLI
point
FORD,
Quant:
Corre:
Engin:
MASCAI
Proce:
Kapeli
de In:
ORTI,
point.
error
Sensi