BACK PROPAGATION NETWORK FOR IRRIGATION SUITABILITY CLASSIFICATION OF STRESSED LANDS:
A CASE STUDY IN PAKISTAN*
Gauhar Re bin an. Senior Engineer
Abdul Fatah Shaikh. Senior Technician
M. A. Sanjrani, Principal Scientific Officer
Space & Upper Atmosphere Research Commission. Karachi, Pakistan
ISPRS COMMISSION VII
ABSTRACT
UN/FAO irrigation suitability land classification provided the working model of base relationships in an
irrigated area of Pakistan to predict the present potential of land use under intensive farming. The
mostly demonstrative study, undertaken by the national space agency (SUPARCO). aimed at exploring the
utility of SPOT XS & P coverage for aiding in the selection of test sites for field sutvey. The three-
step approach, leading to preparation of maps delineating polygonal association foi irrigation suitability
primarily consisted of a) selecting sample t.est sites through image enhancement. I>) preparation of a general
landuse map through digital classification, and processing tire agronomic, management, environmental and
socioeconomic details through a PC-based neural network for ultimate suitability selections. The results
have been encouraging in terms of 'what if assessments for qualitative uses of land against prevailing
cultural practices. Economic returns based on farm budgets for principal cash crops were evaluated using
B/C ratios in terms of prevailing market considerations.
INTRODUCTION
Allahabad Pilot Project is une of a series of pub
lic sector land reclamation programs initiated by
the Federal Government under the name of salinity
and reclamation programs (SCARPs) throughout irri
gated landscape of the country. SCARP-VI in Rahim
Yar Khan covers nearly 0.09 Mha and is buttressed
between the Cholistan Desert to the east and river
Indus on the west. There are two major inundation
canal systems running across the length of the area
after originating from the Punjnad Headworks on
river Indus. The intensively farmed region produc
es two major cash crops, i.e. cotton in the summer
season of plentiful water supply, and wheat in the
winter season. The well spread-out gravity distri
bution network of water channels is plagued by
seepage and percolation losses that in the recent
past contributed to the rise of the groundwater
level endangering the agroeconomy of the area.
Although the public sector vertical drainage has
succeeded in lowering the ' water table to safer
levels, the pumpage from saline/alkaline pockets
within the aquifer has contributed towards soil
degradation. There is no predictable pattern to
this land affliction, and within vast locales under
considerable variant farming practices, it becomes
a herculean task to keep track of unproductive
land, let alone determine the stress on productive
lands reverting to substance earnings. While .satel
lite imagery had reasonably sufficed for quantita
tive declarations on unproductive land, it failed
to afford justifications ou underlying causes for
the temporal offsets of surface behavior. However,
its versatility lends itself to much more plausible
reasoning if feature interpretability is sustained
with detailed field surveys. A step furthe1 would
be the estimation of the economic potential of the
laud in terms of deliverables that directly address
farm-level benefit-to-cost ratios. UN/FAO had
modelled just, such a scheme in 1985. based un its
"A Framework for Land Evaluation" presented in
1976, to assess the suitability of land for it-
rigated agriculture. The model matches the pre
valent characteristics with the intended land use
313
to arrive at suitability scales like SI, S3, N1, and
N2 denoting decreasing suitability for the above in-
tei actions. Prescribes to classification determin
ing factors could include crop requirements, manage
ment practices, land improvements, conservation
activities, and of course socioeconomic trends.
The national space agency of Pakistan, in extension
of its stressed land monitoring engagements, at
tempted a limited area sampling within the Allahabad
Pilot Project (fig. 1) in conformance with the
UN/FAO classification scheme for land suitability
assessment.
Inferencing for Land I'vaJ.uation
Historically, the match-up of existing land cover
types to selections prevailing on intended land use
has been done at the human operator level, on a for
mat. proposed within the UN/FAO publication No. 55.
lire tabulal descriptions allow land characteristics
to be weighed, though qualitatively, against options
for irrigated land use. The recursive exercise, more
benignly imagined as a matrixed correlation between
the existing and the desired, relies on operator-
level expertise in assigning the levels of signifi
cance to each interaction. The procedure is inhe
rently subjective because of factors 1 ike nonconfor
mance in concluding evidence across a varying land
scape, error containment at observation level, and
inadvertent remissions in recording of data elements
that tend to make the decisions bordering on the
rul e-of-t hiimh knowledge. Arrival at this decision
making can he automated towards offsetting these
hindrances, however the unconventional approaches
are less satisfactory and even impractical. Natural
resource categorizations are not always given to
discrete separations, and rather the result is »
compromise of sorts a weighted approach along pre-
establ i shed d iscr im i na t.ion .
Rolf? of Neural Networks
The growth of interest in neural networks during the
1930s has been one of the fastest for any field- in