ocedure named
tor the gating
In Gating-Lrn,
k are shown on
ly show fairly
; networks start
ity to different
to oscillate as
ted. The other
was used to
luring network
ut of the gating
raph, each bar
local network.
1e network, the
tribution of the
the composite
er function was
xtocol, 100,000
ed MANN was
ds Landsat TM
e reserved test
| accuracy of
appa index of
yd classification
y) was achieved
iension TM and
dimension data
ave almost the
| eighteen band
xcept that each
teen PEs in the
Es represent the
d data, and one
ta. The MANN
"training data as
(s but also had
ed. The same
ed. The trained
eteen-dimension
ation. Based on
all classification
ed. Visually, all
red successfully
3. Classification of eighteen bands TM and
three-illumination model data
For this data combination, the topological
structure of the local BPANN consists of
twenty-one PEs in the input layer, which
represent the eighteen band Landsat TM data,
and three illumination model data corresponding
to the May, August, and October TM data,
respectively. The MANN was trained by the
same group of training samples as used in the
eighteen dimension case but had the illumination
model data involved. The RMS error, network
weight, Gating-Lrn, and Gating-Rcl were
monitored during the MANN training and
classification. The trained MANN was applied in
twenty-one dimension multisource spatial data
classification. An overall classification accuracy
of 91.79% was achieved. A balanced
classification accuracy was achieved for all of the
land cover categories.
Classification Comparison and Neural
Network Paradigms Evaluation
1. Classification performance in eighteen band
Landsat TM data
In MANN, two "local networks"
compete to learn the patterns inside the input
vectors. Therefore, in high dimensional spatial
data cases, land cover patterns inside the
multispectral, multitemporal Landsat TM data
could be effectively identified. The Gaussian
maximum likelihood (GML) classification, on the
other hand, will perform well if the input data are
in perfect normal distribution and the prior
probabilities for each land cover category are
known. However in remote sensing data
classification, Landsat TM data may not be
perfectly normally distributed for certain land
cover categories. Also, it is difficult to know the
prior probability for each category before the
classification has been conducted. Usually, an
equal prior probability assumption is adopted in
the application of GML. It is obvious that this
assumption is not always a good one, particularly
when the area is dominated by certain land cover
categories, such as large forest coverage and
small residential areas. Those factors may effect
the classification performance of GML.
2. Classification performance in nineteen-
dimension multisource spatial data
In nineteen dimension multisource spatial
data cases, the transportation data were involved
in pattern recognition and classification. The
transportation data, in nominal data type, has no
meaningful statistical mean and covariance
matrix. Therefore, the transportation data could
not be effectively classified by GML. The
distribution-free capability of artificial neural
networks, on the other hand, made the MANN
flexible in input data selection. The classification
assessment indicated that MANN paradigm
achieved good classification results. Visually, all
of the details of transportation lines were
successfully identified.
3. classification performance in twenty-one
dimension multisource spatial data
The twenty-one dimension multisource
spatial data include eighteen multispectral,
multitemporal Landsat TM data and three
illumination models. Accuracy assessment results
(Figures 3) show that by having the illumination
data involved, the variability of classification
accuracy on different categories was reduced. In
particular, a relatively stable and balanced
classification accuracy was achieved for all of the
categories by MANN.
CONCLUSIONS
1. MANN has the abilities to handle multisource
spatial data. In MANN, there is no constraints
about statistical distribution and measurement
scale of different source data.
2. MANN apply the combination of local
networks and the gating network in pattern
recognition. MANN assign input data to
different local network and adjust the output of
local network by the gating network. This
mechanism is effective in high dimensional
multisource spatial data classification. Since two
or more local networks compete to learn the
input data patterns, more details among closely