Dummy case representations were also inserted prior
to network initialization in order to assess degra
dation of final convergence and/or the measure of
freak case over-ride available. All the cases con
verged on or below threshold setting for weighted
convergence.
The learned model was progressively tested, first
for interactions within the subgroups and then
against random selections from specified case
characteristics. The subgroup convergence results
were converted to percentages and then ascertained
across classification categorizations SI through
N2. Subgroups accounting foi <40 % score against
their respective classifications were noted for
their departure tendencies to other classifica
tions. The results of subgroup interactions and
random selections were tabulated and graphed for
presentation as figs. 1 through 7. The model, clas
sifications were transferred to the polygon deline
ations within 1:20.000 scale sample area maps to
represent irrigation suitability assessments (figs.
8 & 9).
Suffixes for classification symbology were not used
because of incompatibility with I lie 1:20,000 wink
ing scale. However, it would be possible to int in
duce such elemental descriptions within the neuial
network against more refined field da.ta that, would
lead to net incremental irrigation benefit studies
under a project development plan.
RESULTS
Within-group interactions reveal mixed results.
Some interactions were very stiongly borne by mini
mal of descriptions, e.g. in Si category, leaving
aside < 24 % within-case accountability for water
quality, the soil characteristics and ionic con
centrations rated at 60-75 % consistency , and in
many cases > 70 % definition accuracy with only one
parameter description. Water quality definitions
for this category are heavily influenced by the
like descriptions within S2 category. This is also
true to some extent for S3 category. The slow take
off of the curve in fig. 3 explains this ten-deucy
and is contrasted well against the otherwise excel
lent showing of >80% in water quality in tig. 4.
The strong showing by the remaining subgroups in SJ
category is primarily' contributed by an explicit
set of descriptions for the best quality id land
with no fuzzy overlap with the degrading .land char
acteristics within the remaining categorizations.
The less than optimum segregations for water quali
ty 7 parameters are partly influenced by the diffi
culty iu adapting the recommendations of University
of California CommitLee of Consultants. 1974, im
SAR vs ECw interactions to the oites more n innionJy
known ppm descriptions forwarded Lr.v the United
States Department: of Agriculture.
Figures 4 through 7 dwell on the heuristics of land
evaluation with subgroup borrowings ot parameter
descriptions, for every slop of the convergence
model.. The central tendency is mound the maigin-
nally suitable land when* tin*, "good" characteris
tics arc balaiicing-oi f the less-l han-dcs i i nL 11
characteristics.
Category S2 had the most, number of training models
prepared lor network learning, and the tesull in
fig. 4 illustrates tire sharp convergence and con
sistency of the scores. "S3" curve is compaiat ive.ly
slow but sut e in approach ing convergence. Iho point
of inflection iu t i g. 6 is because of the bias dep
ress jug. i he .1 i valent cations resulting in their pre-
vjpit.al inn by the high IICf)3 presence. This loss in
potency of cationic behavior is actually a sore
thumb like indication to the generated model to
progress towards a sharper convergence. Unlike the
regression er.t imat ion, where the emphasis is on the
least squares fit , t he neural network model explores
maximum divergency and vat iation, and wayward beha
vior is most conducive to problem identification.
However, this waywardness is desirable only within
the subgioupings established for the model.
Figure. 7 is a hallmark of drastic swings in the
model led appropriations. The rise is sharp because
of favorable endorsements from water quality, how
ever I lie near convergence loses momentum arid starts
l.o collapse because of ionic concentration borrow
ings from S3 and N1 to portray saline-alkali bo-
ha v ioi .
A fantastic ol-servat ion was made concerning the use
of .lummy case doscri.pl ions; these cases had been in
serted to observe the error propagation in the neu
ral network, in one of the cases, the water quality
for S2 category had been changed (piioi to network
learning) to the one appropriate for S3. The within
subgroup analysis repeatedly failed to give any cre
dit to the S2 category, indicating the over-ride in
the model for inojdentaJ errors in input.
While seveial characteristics were used towards sam
ple case definitions, only 12 selections were avail
ed (three from each subgroup) for studying the cumu
lative rate of convergence the ultimate test of
the model, for each suitability categorizations.
While the line of convergence is not bordering on
exactitude, it may be considered, in the ’neural
sense' of the way, a best: fit. This is because the
model oversees the number of times a case charac
teristics had been used during the formula!ion of
samples. If a certain paiamefer, e.g. response to
organic matter amendment, has been highlighted re
peatedly. then its weighted influence will propor
tionately increase. It s omission from model test ing
would indicate depressed into of convergence, and
likewise its inclusion may be the next ihing to full
convergence. Since each subgroup contributed only
3 characteristics, there were 1 i.kely to be omissions
that influenced the .sl.oj.re of the line of conver
gence. However, for. (lemonsl:rat ion purposes, the
sLope of the line has been averager! over several
characteristic combinations for each subgroup.
(,'ONUI US IONS
This study attempted an adaptation of software emu-
la led neuial net wo l k mod« 1 to the prescribes of
irrigated land suitability ass.* -.smeii t . Though the
model was not: tweaked to its optimum because of
coarser pre-set orroi eiiteiin. lack of oxperimenta-
I ion with in( ci node selections. fewer number of sam
ple cases, tin* results have been promising. In all
fairness, an evaluator who is already prone to his
intuit ton is mostly interested in “stabl i siting zones
of significance lot match-up of land cover and land
use dim aet.pr i si Ls. I'o he .»hie to accomplish the
a!rove decision-making in an automated sense of the
way with resilience ami consistency would b«_> a pla
cid achievement . Foi Hie case study. the neural
neluoik lias not only facilitated subjective classi
fication of data lull .•»1 >i» sei veil a--, t he cornerstone
tor [ in ju.se|a 1 advice lending to land rehabilitation
.1! t ivit ¡is. all governed by the equity cons idera-
t ions. In a country, wiu'ii- lert ¡ary level of infor
mation is s* j hi. mi or da i tied iu map-making, the element
■ ! suggest ibilit.y for 'reasoned' ut il.izatjon of i r-
i ¡gated land is a signll it. ant development.
315
