i, 1995. Changes 
Lake Michigan 
ym regional lake- 
graphy. Vol 40, 
, 1987. Lake ice 
icator of climate 
g using remote 
F Climate Change 
ydrologic Regime 
of the Vancouver 
mber 168. 
<. Clayton, T.M. 
ss. Determinants 
ite-derived 1987- 
n the Laurentian 
d, 1993. Satellite 
> indicator: Initial 
i in Wisconsin. 
! Remote Sensing. 
L. Scarpace, and 
-derived lake ice 
Shield as a robust 
mnual Convention 
Charlotte, North 
MODULAR ARTIFICIAL NEURAL NETWORKS: A NEW PARADIGM IN 
MULTISOURCE SPATIAL DATA MODELING AND CLASSIFICATION 
Yeqiao Wang 
Laboratory for remote Sensing and GIS 
Department of Anthropology (M/C 183) 
University of Illinois at Chicago 
Chicago, IL 60607-7138, USA 
. Daniel L. Civco 
Laboratory for Earth Resources Information Systems 
Department of Natural Resources Management & Engineering 
University of Connecticut 
Storrs, CT 06269-4087, USA 
ABSTRACT 
Backpropagation artificial neural networks (BPANN) have been applied in remote sensing and 
GIS data processing in the past several years. In this research, a new paradigm, modular artificial 
neural network (MANN), was designed for multisource spatial data modeling and classification. The 
modular design partitioned artificial neural networks into several subnets (individual BPANNs). Each 
subnet act as a local expert responsible for interpreting a subset of multisource spatial data. A gating 
network adjusts the output from each local network. While the unique information from different 
data sources could be employed by individual subnets, the confusion and mixing within a single 
source data could be filtered by the MANN. Therefore, the information from each data source could 
be efficiently applied in spatial data modeling and land cover classification. 
INTRODUCTION 
Applications of artificial neural networks 
(ANN) in remote sensing and multisource spatial 
data classification have been frequently reported 
in the past several years (Benediktsson, et al., 
1990, Ritter and Hepner, 1990, Liu et al, 1991, 
Heermann and Khazenie, 1992, Civco, 1993, 
Salu and Tilton, 1993, Hara ef al., 1994, Peddle 
et al, 1994, Chen, et al, 1995, Foody et al, 
1995). In the previous research, backpropagation 
artificial neural network (BPANN) is the most 
widely adopted paradigm. This popularity 
primarily revolves around the ability of 
backpropagation paradigm to learn complicated 
multidimensional ^ mapping (Hecht-Nielsen, 
1989). The advantageous of applying artificial 
neural network in multisource spatial data 
modeling and classification had been 
demonstrated (Benediktsson, ef al., 1990, Ritter 
and Hepner, 1990, Heermann and Khazenie, 
1992, Civco, 1993). Other ANN paradigms, 
however, are seldom applied and reported. At 
current stage, BPANN are treated as black-box 
tools. All of the initial or preprocessed 
multisource data are submitted to BPANN for 
solutions. Compared with using single date 
remote sensing data, the input vectors are 
expanded when multisource spatial data are 
involved. Although improved classification 
accuracy can be achieved (Hepner ef al., 1990; 
Bischof et al., 1992, Civco and Wang, 1994), the 
contribution from each individual data source, 
however, is neither observed nor fully 
93