144
T.-S. Chon . Y.S. Park' I.-S. Kwak . E.Y. Cha
patternized and March is the target month for training, for example, weights for
the two previous months were appended in front of March (Le. January - February
- March). In total T x n weights were used for inputs where T and n respectively
represented for the number of months and that of variables for training as inputs
for ART (Figs. 8.7a and 8.7b). This was similar to creating a window of a
specified input period (three months in this case) and scanning all through the
target sampling times through the survey period. The detailed process could be
referred to Chon et al. (2000c).
The self-organizing Kohonen network with the array of T x n artificial neurons
maps the data feature in a reduced dimension as previously demonstrated. In this
case two-dimensional array with 9 x 9 neurons was used. The weights in the
Kohonen network were represented as w k ( " ') (t). Since j* was determined as a
• mJ I
winner in ART, and all the winner nodes were selected for input to the Kohonen
network, j* was set to a constant for training. Among designating digits for input
node, sampling month in the sequential period, m, and input node for ART, i, were
varied in this case. When the input vector was sent through the network, each
output neuron, k, computed the summed distance, d' k (t), between input vector
and weights, and subsequently training is conducted in the similar manner
explained previously for the Kohonen network.
The field data used for ART training (Fig. 8.8a) was also provided as input to
the network for grouping cornrnunity changes. Figs. 8.lOa and 8.lOb show the
mappings for two and three months after the training by the Kohonen work. The
trained results showed general characteristics as observed in the results from onemonth sampies (Fig. 8.8b). Grouping was mainly based on pollution levels and
topographical conditions. In the two-month sequences (Fig. 8.lOa), several large
groups appeared. A large number of sampies from polluted sites of TCL and THP
were grouped together (Group A) at neuron (6, 0). This was also the case in the
one-month mapping (neuron (0,3) in Fig. 8.8b). Many sampie communities from
YCK and TSD formed another large group (Group B) at neuron (3, 7).
Communities collected at YIG in the early part of 1994 also forrned a group
(Group C) at neuron (6, 3). Communities from TKC were spread on the map with
small groups, similar to Fig. 8.8b. In contrast to one-month sampling, however, a
slight difference was observed in the two-month map (Fig. lOa). Sampie data
from TSD were absorbed into Group B in the two-month map. Communities
collected from YSC, which were mostly located dose to, or inside the group
mainly consisting ofYCK (neurons (2,7) and (3,7) in Fig. 8.8b), tended to drop off
at group B in the two-month map in some cases. However, the YSC communities
did not move far way from Group B (Fig. 8.10a).
In the maps describing community pattern of the period longer than two
months, the characteristics shown in the two-month map were generally preserved.
Most sampie communities in Groups Band C were consistently found inside the
groups as the input period increased. Sampie communities from TKC and YSC
also showed a similar tendency as seen in the two-month map. In Group A,
however, the size of sampies was gradually decreased. In the 3-month map (Fig.
8. lOb), for example, THP4-1, THP4-6, TCL4-2, TCL4-6 were separated from
Group A. Detailed discussion and
T.-S. Chon . Y.S. Park' I.-S. Kwak . E.Y. Cha
patternized and March is the target month for training, for example, weights for
the two previous months were appended in front of March (Le. January - February
- March). In total T x n weights were used for inputs where T and n respectively
represented for the number of months and that of variables for training as inputs
for ART (Figs. 8.7a and 8.7b). This was similar to creating a window of a
specified input period (three months in this case) and scanning all through the
target sampling times through the survey period. The detailed process could be
referred to Chon et al. (2000c).
The self-organizing Kohonen network with the array of T x n artificial neurons
maps the data feature in a reduced dimension as previously demonstrated. In this
case two-dimensional array with 9 x 9 neurons was used. The weights in the
Kohonen network were represented as w k ( " ') (t). Since j* was determined as a
• mJ I
winner in ART, and all the winner nodes were selected for input to the Kohonen
network, j* was set to a constant for training. Among designating digits for input
node, sampling month in the sequential period, m, and input node for ART, i, were
varied in this case. When the input vector was sent through the network, each
output neuron, k, computed the summed distance, d' k (t), between input vector
and weights, and subsequently training is conducted in the similar manner
explained previously for the Kohonen network.
The field data used for ART training (Fig. 8.8a) was also provided as input to
the network for grouping cornrnunity changes. Figs. 8.lOa and 8.lOb show the
mappings for two and three months after the training by the Kohonen work. The
trained results showed general characteristics as observed in the results from onemonth sampies (Fig. 8.8b). Grouping was mainly based on pollution levels and
topographical conditions. In the two-month sequences (Fig. 8.lOa), several large
groups appeared. A large number of sampies from polluted sites of TCL and THP
were grouped together (Group A) at neuron (6, 0). This was also the case in the
one-month mapping (neuron (0,3) in Fig. 8.8b). Many sampie communities from
YCK and TSD formed another large group (Group B) at neuron (3, 7).
Communities collected at YIG in the early part of 1994 also forrned a group
(Group C) at neuron (6, 3). Communities from TKC were spread on the map with
small groups, similar to Fig. 8.8b. In contrast to one-month sampling, however, a
slight difference was observed in the two-month map (Fig. lOa). Sampie data
from TSD were absorbed into Group B in the two-month map. Communities
collected from YSC, which were mostly located dose to, or inside the group
mainly consisting ofYCK (neurons (2,7) and (3,7) in Fig. 8.8b), tended to drop off
at group B in the two-month map in some cases. However, the YSC communities
did not move far way from Group B (Fig. 8.10a).
In the maps describing community pattern of the period longer than two
months, the characteristics shown in the two-month map were generally preserved.
Most sampie communities in Groups Band C were consistently found inside the
groups as the input period increased. Sampie communities from TKC and YSC
also showed a similar tendency as seen in the two-month map. In Group A,
however, the size of sampies was gradually decreased. In the 3-month map (Fig.
8. lOb), for example, THP4-1, THP4-6, TCL4-2, TCL4-6 were separated from
Group A. Detailed discussion and