the history of science. Its domain of reference
has steadily expanded beyond the realm of cognitive
science and intelligence, and into the multiple
assimilations of neurobiological structure-function
relationships. Any major cognitive process can be
analyzed into component subprocesses. Analysis of
subprocesses will then suggest principles that can
also be used in models of other processes. The in
puts to such analyses would come through the fea
ture, or input, nodes that upon activation would
excite the response patterns of category, or out
put, nodes. In between are the internode or hidden
connections that determine the strength of the
final outcome. For a given application, the
adaptive-resonance, back-propagation, or brain-
state- in-a-box theories might be suitable in some
ways but not in others. The formulation stage es
sentially decides the algor ithm selections for
application specifics.
A network is established with a pre-set bias to
wards the features considered important for the
problem. The network's recognition of the problem
features is sustained and recursively enhanced by
the strength of the weighted connections. A "train
ed" network when presented with subsequent, (and
often incomplete) problem definition will recourse
to the most converging solution derived from its
adaptive learning. As per implementation, the
artificial neural networks use both hardware and
software emulation to achieve the desired synthesis
of information.
Adapting to Land Evaluation
Land capability assessment is an interweave of a
host of variables, with positive and negative cau
sative inheritance, that, in the evaluators judge
ment, prevail on the utility of the land. Depend
ing upon the region of interest, the sum total of
variable inheritance could change significantly
enough to require skewed decision making. The
changes, and their subsequent impact on the recom
mendations on utility of the land may be so unique
as to disfavor harnessing conventional computing
prowess. Under such circumstances, when step-wise
integrations of causative factors is likely to
yield the best results, the use of neural network
simulations is not unwarranted.
For neural network to simulate irrigation suita
bility assessment, it would have to focus on those
agronomic and management parameters that are class
determining, i.e. their significance in affecting
the evaluation of land is recognizable. Their in
teraction is best understood under local farming
conditions, and the attributes of land productivity
may not be transferable elsewhere. The suitability
assessments would, therefore, be highly responsive
to the relative productivity thresholds of the loc
al area. For a simulation to be considered comp
lete, it must not only facilitate arrival at clas
sified decision making but also expose the relative
contribution of each significant input parameter to
the overall categorization. If the network has
been trained on consistent integrity data sets, the
outcome would be devoid of invalid extractions.
This consistency gives itself to tremendous useful
ness when the operator is constrained by availabil
ity of inputs or when suitability of inputs or when
suitability trends are being assessed for large
problem cases.
METHODOLOGY
Principal component separations on geometrically
corrected multiband SPOT XS data yielded the tenta
tive locations of sampling sites whore contrast of
crop stands vs bare soil appeared optimum. October
(1989) was selected for satellite coverage in order
to expose maximum differentiation between the cotton
crop, fallow land, and the village hutments/bare
soil. Standard parallelpiped classified distin
guished between these principal land use classes
throughout the Froject area (fig. 2). The classi
fication map (at 1:20,000 scale) facilitated land
unit boundary delineations for match-up against the
land utilization types. Reference to spatially
sensitive locales, e.g. distributaries was estab
lished with the superimposition of panchromatic
coverage over the classified image.
A questionnaire, invoking the principal agronomic
and management considerations, was prepared to as
sist towards farmer interviews. Disturbed soil sam
ples were taken from pits along three horizons of 0-
15 cm, 15-30 cm, and 30-40 cm to be evaluated later
for topsoil stratification. The combines of ques
tionnaire information and soil tests were processed
through a PC-based back propagation neural network
weighted to give thresholds along pre-set B/C ratios
for equity considerations. The redeeming feature of
the neural network (NemoShell from Wards Systems
Group, USA) is the ability to afford subjective de
cision-making against pre-set error criteria. Each
suitability assessment was preempted by the econo
mics of the land use management, over: land holdings
of varying size.
Amongst the specifications for land use requirements
and limitations included in the questionnaire were
such class determining factors like crop types grown
at seasonal turn-arounds, growing periods, fallow
land requirements, seeding rates, and yield/ha for
different varietals. Soil nutrient capacity, in ad
dition to saturation extract determinations on cat
ionic and anionic concentrations, both available
and exchangeable, were some of the principal deter
minants of quality of the soil. Economic inputs,
like use of pesticides, fertilizers, crop-specific
watercourse charges, operational use of tube wells
by owners, and rent charges for on-farm/locale-
specific mechanization fm laud preparation, in
addition to labor supply at peak demand periods were
included in the assessments foi the evaluation of
B/C ratios.
V
Network Implementation
The class determining fact ms were arranged in the
neural network so as to allow ranged selections ac
ross physical and chemical characteristics of the
soil and the quality of irrigation water such that
they become input not only for the economic returns
but also introduced a measure of suggestibility
through land development costs in the use of amend
ments (as is typical of saline/alkali and non-saline
alkali soils).
The network was configured to 100 agronomic and ma
nagement characteristics that were to serve as the
primer for ultimate reflections on equity considera
tions. The original variable selections were divid
ed into 4 subgroups, namely water quality, physical
attributes ol the soil, and saturation extract (1:5)
cationic and anionic concentrations. Twenty-five
sample case summaries were prepared from amongst the
combines of subgroups for irrigation suitability
categor izations SI, S2, S3, N1. ami N2 (in the order
of most suitable to permanently not suitable). The
intermediate node settings were defaulted to tire
number of input nodes for model simplification and
to minimize on the computing time. The error cri
teria for weighted convergence was set to 0.005.