» MANN, as 
hat the gating 
rks are fully 
he number of 
is the same as 
> output values 
ed to sum to a 
ut of the gating 
ut vector from 
rk. The final 
" the weighted 
yating network 
earning rule is 
jetition among 
ut vector, the 
| single local 
fore, the input 
; so that each 
for a different 
|| networks is 
  
ks. 
| of error. In 
be solved with 
as well by a 
lowever, since 
rks learning 
computationally 
) implemented a 
n based on the 
Experts" as 
). According to 
NeuralWare (1993), the output of MANN is the 
combination of output from each individual local 
networks. Let's define L as the number of PEs in 
the gating network output layer, g=(g,, ... , g;) as 
the activation vector of gating network output 
layer, y,7 (y. ... ; yy) as the activation vector of 
k^ local network output layer, and y-(y,, ... , y) 
as the MANN output. g, adjust the output of ith 
local network by the gating network output. 
L 
y= à gi 
MANN IN MULTISOURCE SPATIAL 
DATA CLASSIFICATION 
Study Area 
The study area was in the USGS 7" 
minute quadrangle of South Canaan in the Great 
Mountain Forest in Northwestern Connecticut. 
Figure 2 shows the geographical location of the 
study area. The area is dominated by different 
types of forest ecosystems and agricultural 
practices. Because of the dramatic seasonal 
spectral change for different tree species, 
multitemporal Landsat TM data can provide the 
information to identify the vegetation categories 
which possess similar spectral features in one 
season, but different ones in another. 
  
  
  
  
  
  
  
dcs 
BS he x Rs 
AN) p 
  
  
  
  
Figure 2. Location of the Study area 
in the state of Connecticut 
Multisource Spatial Data 
The multisource spatial data used in this 
research include geo-referenced Landsat TM 
data from May 2, 1993; August 30, 1990, 
October 6, 1992; DEM (digital elevation model) 
95 
data and the derivative of slope; Digital Line 
Graph (DLG) transportation; and three 
illumination model data. The illumination model 
data gave the solar-terrain geometry at the time 
of each TM scene acquisition. The purpose of 
having the illumination model data involved was 
to estimate the illumination effects and the 
influence of slopes and shadows. Above 
multisource spatial data represent spatial data 
from different distributions and measurement 
scales. The Landsat TM data represent spectral 
features of ground covers in different seasons. 
The slope and transportation data not only 
provide useful spatial information in different 
distribution format and measurement scales, but 
also can be applied to test the tolerance and 
performance of different neural network 
paradigms and statistical classifier on different 
types of input data. 
Training/Test Samples 
In order to providing sample pixels for 
MANN training and test, sample pixels were 
selected from multispectral, multitemporal 
Landsat TM data. The sample pixels were evenly 
divided into two groups, while one group was 
used for MANN training, the other group was 
reserved for MANN test and classification 
accuracy assessment. 
The architecture of the MANN designed 
in this research include two local BPANNs. The 
local BPANN has eighteen PEs in the input 
layer, and twenty-four PEs in the output layer. 
The eighteen input PEs represent six reflective 
bands of Landsat TM data for each of the three 
seasons. The twenty-four output PEs represent 
twenty-four land cover categories. Different 
numbers of PEs in hidden layer were tested. 
Choosing an appropriate number of PEs in the 
hidden layer is important, and often difficult to 
predetermine. Generally, the number of hidden 
PEs is determined by experiment for each 
specific problem. There are different rules to 
estimate the number of PEs in the hidden layer. 
One of the rules is geometric pyramid rule 
(Masters, 1993), which specifies the number of 
hidden PEs based on the number of input PEs 
and the number of output PEs. With Nue N,,,,, 
input