In: Wagner W., Székely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B
from a meteorological station located inside the test area. The
test was made inside Vale do Rio dos Sinos (UNISINOS)
campus situated on Sao Leopoldo/RS city. The sample points
for test were located on places with vegetation, concrete and
asphalt paving.
Each collected point had the necessary information to be
inserted in the trained neural model (UTM coordinates, altitude,
temperature and averages of air relative humidity). The model
supplied for each point a ST value that was compared with the
value obtained in field by laser thermometer.
The statistical analysis used on research results presentation
were based on a comparison between ST values modeled
through ANN and ST values considered true by obtainment by a
laser thermometer. Were used the statistical t-student test and
coefficient of determination (R 2 ) analysis with linear regression.
Being x the size of certain elements attribute from an A
population (ST modeled by ANN);
Being y the size of the same elements attribute from a B
population (ST obtained by laser thermometer);
Being x and y ordinarily distributed with unknown variances;
Being the hypothesis: px = py which px = average of x and py
= average of y.
For testing the hypothesis of averages equality from the two
populations was utilized the t test, but for that was necessary to
initially test if the two populations presented equal variances
using the F test from Fischer:
F = SQD * (09)
SQD y
2 S Q D z (10)
* ■(»,-!)
S> V^) (11)
F ( 12 >
calculated 2
where: SQDx and SQDy correspond, respectively, to the
sums of square deviations from x and y;
and correspond, respectively, to sample variances
from x and y;
nx and ny correspond, respectively, to the number of
variables from x and y.
The tested hypothesis (HO) was that the population variance
from x is equal to the population variance from y. If Prob > F is
less than 5% then HO is accepted. If Prob > F is bigger than 5%
then HO is refused. If the population variances were statistically
equals, then a common variance is calculated ():
2 {SQD x +SQD y )
Ik - IM«, -i)j
(13)
[(*M
,-1))]
r
(14)
Afterward was tested the HO for population average equality
using the t random variable, defined by:
-M y
y/v-fax-My)
(15)
Being:
v-ta -A y ) =v CO +v W=—+-
»x "j
where: V correspond to the average variance.
2 2 2
Accepting S=S -S there are:
(16)
n.
f 1 1'
1
\ n * n yj
t =
Mx-My
2
il
n
s c •
— + -
”y)
In degrees of freedom (ri) = (Yl x + Yl y — 2)
In case of different variances there are:
M x -M v
t =
, 2 si
+ —
72 n
(17)
(18)
(19)
(20)
The degree of freedom is calculated with the following
equation:
n =
n yj
(¿1
2
K n y J
(21)
n —1 n-1
3. RESULTS AND DISCUSSIONS
The ANN that presented the best performance was composed
with an input layer (5 variables), an intermediate layer (with 4
neurons) and an output layer (with one neuron), as show on
Figure 4. The fact that the selected network, with the best
performance, has only one intermediate layer, is in accord with
the results found by Kumar et al. (2002) and Zanetti et al.
(2008), because these authors have modeled the
évapotranspiration and concluded that an ANN with only one
intermediate layer was sufficient to represent a nonlinear
relation between the climatic elements and the modeled
variable.
1
4
615