discovered. While more information can be
supplied by multisource spatial data, noise,
redundancy and confusion may also be
introduced. If artificial neural network paradigms
with modular functions or competition
mechanism could be developed, the information
process for each data source could be
decomposed, and the contribution of each data
set could be separately evaluated. Therefore, the
advantages from each data source could be
discovered and efficiently employed to perform
the classification. To reach this goal a
modularized ANN (MANN) is a paradigm
requiring further development. Based on the
above discussions, the objectives of this research
are to:
1. Develop MANN to handle
multispectral, multitemporal remote
sensing data and spatial data from
other sources;
2. Evaluate the capabilities of
MANN in multisource spatial data
classification, and develop
strategies for multisource spatial
data analysis;
3. Apply MANN to produce land
cover classifications from
multisource spatial data.
MODULAR ARTIFICIAL NEURAL
NETWORK
In the main stream of neural network
research, artificial neural network are viewed as
a black box tool. Hrycej (1992) argued that the
black box state could not be expected to persist
if artificial neural networks grow and the
applications become more complicated and
difficult. Therefore, a modularization of neural
network is desirable. Modular ANN (MANN) is
to decompose a complex task into several
relatively small and independent subtasks.
MANN consists of a group of local networks
competing to learn different aspects of a
problem. A gating network controls the
competition and assign different regions of the
data space to different local networks. Each local
network can be an individual backpropagation
network. The architecture of MANN, as
illustrated in Figure 1, shows that the gating
network and the local networks are fully
connected to the input layer. The number of
output PEs of the gating network is the same as
the number of local networks. The output values
of the gating network are normalized to sum to a
value of 1.0. The normalized output of the gating
network is used to weight the output vector from
the corresponding local network. The final
output of MANN is the sum of the weighted
output from each local network.
Local networks and the gating network
are trained simultaneously. The learning rule is
designed to encourage the competition among
local networks. For a given input vector, the
gating network will choose a single local
network to handle the data. Therefore, the input
space is partitioned into regions so that each
local network takes responsibility for a different
region. Training of gating and local networks is
Global output layer
Figure 1. Architecture of MANN
with two local networks.
achieved using backpropagation of error. In
general, any problem that can be solved with
BPANN can be solved at least as well by a
MANN (NeuralWare, 1993). However, since
there are several networks learning
simultaneously, MANN is computationally
intensive. NeuralWare Inc. (1993) implemented a
modular neural network paradigm based on the
"Adaptive Mixtures of Local Experts" as
proposed by Jacobs ef al. (1991). According to
94
NeuralWa:
combinatic
networks.
the gating
the activa!
layer ? ye
k? local ne
as the MA
local netw
MANN
Study Are
The
minute que
Mountain
Figure 2 sl
study area
types of
practices.
spectral c
multitempc
informatioi
which pos
season, but
Figui
Multisour:
The
research ii
data from
October 6,