Professional Documents
Culture Documents
luis.manuel.callejas@itesm.mx
ABSTRACT.
This paper shows how an information system based on a fuzzy neural network
carry out a land suitability assessment. The process done by the information
system is base on an fuzzy logic algorithm in order to determine which can be
the most optimum crop or tree to grown in a given plot of land.
The information system is capable to assess around to 1711 different types of
crops or trees, using information about to climate (rainfall intensity,
temperature, sun radiation and humidity regime), Soil characteristics such as
depth, moisture, slope and other important soil characteristics.
The computer system was tested in field situations and it showed to be more
accurate than the current land evaluation methods and systems.
Introduction
The approach proposed in this research differs from the current approaches of land
suitability assessment [3] by (1) translating the evaluation of land characteristics into a
fuzzy suitability class avoiding crisp suitability classes, (2) developing and applying a
fuzzy neural network in order to estimate crop yield productions, and (3) developing
digital knowledge bases with the climatic and soil requirements for 1711 different
types of crops and trees.
The FAO Framework for Land Evaluation is based on a matching exercise between
specific land utilization type (LUT) requirements and the natural conditions of the land
(Figure 1).
Suitability Classes
Depending on the objectives of the evaluation, the suitability of an LMU evaluated for
a given LUT can be described by four levels of detail (Figure):
Suitability orders: All land is divided into two suitability orders, according to whether
the land is suitable or not for a given LUT:
‘S’ = suitable
‘N’ = not suitable.
Suitability classes: These are divisions of suitability orders that indicate the degree of
suitability. These can be thought of as modifiers of suitable vs. not suitable
‘S1’ = suitable
‘S2’ = moderately suitable
‘S3’ = marginally suitable
‘N1’ temporarily unsuitable
‘N2’ = unsuitable
The limits between S1 and S2, S2 and S3/N1 are arbitrary or based on single-factor
yield reductions.
Suitability subclasses: These are divisions of suitability classes that indicate not only
the degree of suitability (as in the suitability class) but also the nature of the limitations
that make the land less than completely suitable. Only suitability class S1 has no
subclasses. E.g. ‘S3e’: marginally suitable (‘S3’) because of erosion hazard (‘e’).
S3e-2
S management unit
order: 'suitable' 2
e
3
subclass
class: 'marginal'
'erosion hazard'
Huajun et al. (1991) [4] proved that their approach to land suitability assessment based
on fuzzy set theory is a potentially fruitful option for improving land evaluation results.
The authors showed that there is evidence that the fuzzy set approach yields more
accurate results in comparison to parametric and limitation methods.
But, when this research tested the Huajun's algorithm some errors and missed
specifications were found:
• Only one membership function is defined for all kind of land characteristics to
be evaluated.
• Only the suitability class S (suitable) is evaluated, and other three suitability
classes (S1 suitable, S2 moderately suitable, and S3 marginally suitable) are
not used.(see figure 4)
• Huajun’s research only applies to corn production, and his member function is
very generic in order to be applied to a wide number of different types of
crops.
1; x = γ
Where: γ= the highest value of the land characteristic, there is the point where
membership is 1; β is an intermediate value and it can be found by β=(α+γ)/2, here
the membership value is 0.5; α is the lowest value of the membership function. x is the
value assessed on a given requirement
Table 1: Land Characteristic (LC) Types according to their effect on crop performance
Land Characteristic (LC) Types Crop performance pattern charts
according to their effect on crop
performance
Type 1: “Positive Effect”
These type of LC have a crop response
pattern, which corresponds to an increase 150
(not necessarily monotonic nor linear) to a 100
ar abi ca cof f ee
maximum of crop performance with banana
50
increases in the values of the LC followed, in bar l ey
crop performance
120
cassave
tailed” curve behaviour of crop performance 100
80 cotton
LGP (days)
performance as the values of the LC
continue to increase.
Type 3: “Negative Effect”
These LC have typically a negative effect on
Crop Performance
150
arabica coffee
crop performance, which decreases with an 100 banana
increase in the values of the LC until cassave
50
reaching a minimum. cocoa
0
coconut
0 2 4 6 8
Dry season (months)
Fig. 4: Relationship between crisp suitability classes, Huajun suitability class (S) and Fuzzy suitability
classes
E
N S3 S2 S1
E
R
G 1
E
D
P
I S
H
S
R
E
B 0.5
M
E
M 0 S1
0
S2
0 25 50 75 100
PERFORMANCE (%)
S3
In fact, the values variation on most land characteristics, (i.e. requirements) do not
follow the shape of the membership function defined by Huajun. Thus the impact of
using such fuzzy function generically to any land characteristic could lead to large
interpretation errors.
Moreover, it was found that the relationship between each land characteristic and crop
performance is not generic either. Thus it is not the same for each land characteristic.
These relationships only changes on some particular crops.
Land characteristics on crop performance was determined after plotting the data
contained in the knowledge bases on land characteristics and their relationship with
crop performance.
The investigation of the nature of the patterns that were found graphically by plotting
Land Characteristics versus crop performance, lead to interesting results, which are
shown in Table 1.
After adopting some aspects of Huajun’s approach and using the knowledge bases
developed to determine and classify the LC into the three performance pattern types.
In this research the membership function to the class “suitable” is broken into
membership functions for suitability classes (S), and suitability classes S1, S2, and S3,
and a FMF for each of these suitability subclasses. In contrast, Huajun defined the
suitability class S as the whole universe of suitability possibilities of crop performance.
Since, Super class “S” contain the three basic suitability classes (Figure 4), each one of
these suitability classes should have a different LC pattern type and its own FMF.
The suitability classes (S1, S2, and S3) have their corresponding FMF, which are
inferred from the generic membership function defined by Huajun.
Following a set of rules starting from the generic function, the rest of the FMF, one for
each suitability subclass, can be generated. Figure 4 illustrates in graphical form the
nature of the functions generated. The generic suitability function (S) of Huajun is
represented stretching from a point of maximum performance γ (100%) and FMF = 1,
to a point of minimum performance α with a FMF= 0.
Where: H is the γ value calculated from a suitability class, m and M are the minimum and maximum
values in the interval of the crisp class. Crisp classes can be retrieved form the knowledge bases.
f1 ( x; γ 2, β , γ ) = (3)
1 − 2[(x − γ ) /(γ − γ 2)]2
; x ∈[β , γ )
1; x ∈[γ , ∞)
FMF for Class S2 0; x ∈ (−∞, γ 3)orx ∈ (γ , ∞)
2[( x − γ 3) /(γ − γ 3)] ; x ∈ [γ 31 , β 1 )
2
f 2 ( x; γ 2, γ 3, β 1 , β 2 , γ ) = (4)
1 − 2[( x − γ 2) /(γ − γ 3)] ; x ∈ [ β 1 , γ 2)
2
1; x = γ 3
When the Fuzzy Membership Functions (FMF) for S1, S2 and S3 are combined with the
three different types of land characteristics (Type 1, Type 2 and Type 3) allowed for
deriving nine new membership functions. These membership functions (one for each
type of land characteristics) are resulting from the combinations of membership
functions and LC types.
0; x ∈ (−∞, γ 2 ) (6)
2[(x − γ 2 ) /(γ − γ 2 )] ; x ∈[γ 2 , β )
2
f1 ( x; γ 2 , β , γ ) =
1− 2[(x − γ ) /(γ − γ 2 )] ; x ∈[β , γ )
2
1; x ∈[γ , ∞)
2[( x − γ ) /(γ − γ )] 2 ; x ∈ ( β , γ ]
f 2 ( x; γ , γ 2 , γ 3 β 1 , β 2 ) = 2 2
1 − 2[( x − γ 2 ) /(γ 2 − γ 3 )] ; x ∈ [ β 1 , γ 2 )
2
1 − 2[( x − γ ) /(γ − γ )] 2 ; x ∈ (γ , β ]
2 2 2 2
1; x = γ 2
1 − 2[( x − γ 3 ) /(0 − γ 3 )] ; x ∈ [ β 1 , γ 3 )
2
1; x = γ 3
f 4 ( x; γ 2 , γ 2' , β1 , β 2 , γ ) =
1 − 2[(x − γ ) /(γ − γ 2 )] ; x ∈[ β1 , γ )
2
1; x = γ
1; x = γ 2 ∨ x = γ 2 '
1 − 2[( x − γ 2 ) /(γ 2 − γ )] ; x ∈ (γ 2 , β 2 ]
2
0; x = α1 (11)
2[(x − α1 ) /(γ 2 − α1 )] ; x ∈(α1 , β 2 )
2
2[(x − α1 ) /(γ 1 − α1 )] ; x ∈(β1 ,α1 )
2
f 6 (x;α1 , β1 , β 2 , γ 1 , γ 2 ) =
1 − 2[(x − γ 1 ) /(γ 1 − α1 )] ; x ∈(γ 1 , β1 ]
2
f7 (x;α1, β1,γ ) =
1 − 2[(x − γ ) /(γ − α1)] ; x ∈[β1,γ )
2
0; x ∈(α , ∞)
1
1; x =γ
f9 (x;α1, β1,γ ) =
1−2[(x −γ ) /(γ −α1)] ; x∈[β1,γ )
2
1; x∈[γ,∞)
The definition of the nine membership functions, as introduced above, shows how the
effect of the range of values of land characteristics on crop performance, can be
accounted for through Fuzzy Set Theory. It also shows how land characteristics can be
reclassified and grouped in patterns or types. These two findings are important
contributions from this research to the field of land suitability assessment and
automated land evaluation systems in natural resources management.
It must be noted that for each land characteristic the critical values ( i i i) of the
nine membership functions derived above, are usually difficult to determine by
traditional methods of analysis. An important feature of these functions is that such
critical values ( i i i) are selected here based on expert knowledge and experience
about the ranges and critical values for the different suitability classes. As it has been
shown in these results, the nine (9) different membership functions are directly related
to the type of land characteristic and are the direct result too of investigations into
global knowledge bases. Particularly, on how a most generic membership function
described the behaviour of the three different land characteristics types (types 1, 2 and
3) on the membership functions to three different land suitability classes (S1, S2, S3).
The findings about the Fuzzy Membership Functions, the typification of land
characteristics (LC) and the procedures for the computation of fuzzy suitability classes
described in the sequence of steps above were all incorporated into the architecture of
an automated fuzzy neural network system for land suitability assessment. The
development of such a system is the central contribution of this research.
Land suitability assessment using a fuzzy neural network
Neural network characteristics have been found useful in a variety of applications, but
its application in agriculture has been limited, the few existing applications fall into
five groups:
This project based on the goodness of the neural networks and in order to apply the
algorithm defined, attempted to design of a fuzzy neural network, which can carry out
land suitability assessment using the algorithm described in this work. And applying it
to evaluate land utilization types on rain fed agriculture. The neural fuzzy network
(FNN) proposed is a network with a typical feed-forward multi-layer perceptron (see
figure 5).
The weights are calculated by the back propagation algorithm, which is commonly
used among the current types of ANN systems available.
A back propagation ANN learns from a set of input values to a set of output values and
its name comes from the fact that the error of hidden layer or units are derived from
propagating backward the errors associated with the output units since the target values
for the hidden units are not given.
The ANN calculated the weights using the data from Sys (1983), and “trained” it with a
cycle of 10000 epoch or calculations for each crop. Later, the ANN was tested with
some standard values from Sys research in order to determine the range of expected
errors from using the weights.
The accuracy of predictions of the neural network was calculated using a cost function.
The cost function calculates the deviation in predictions, which results from the
difference between the network output and the desired behaviour.
The development of a system based on the approach proposed in this research was
undertaken.
The computer system developed was named as “LANSAS” from the acronym of
LANd Suitability Assessment System. LANSAS is a computer system based on a
fuzzy artificial neural network approach and based on the framework defined by [13],
and it has the capability to carry out land suitability assessments using a recursive and
parallel method allowing users to access three different knowledge bases.
These knowledge bases provide information regarding to LUT requirements for a wide
variety of crops (mostly for rain fed agriculture). Additionally, LANSAS is designed as
to use geographic information from standard GIS software (raster and vector). The
capability of LANSAS to import-export data from standard relational database
management systems, are also part of the design.
Results
The new paradigm was tested using climatic and soil data from a watershed in centre of
Mexico, the test was based on an evaluation of estimated crop yield production made
by each of the current land suitability assessment methods such as parametric and
limitation method and the current computer approach called ALES [14] and the new
approach (LANSAS).
The results shown new paradigm is more accurate than current methods (see table 2).
The weights calculated define an average value estimated for each land characteristic in
order to use those in future evaluations as universal knowledge.
Table 3: Impact weights of land characteristics for rain-fed agriculture
Agronomically speaking, the calculated weights indicate that soil depth, air
temperature, the soil organic matter, air humidity and the length of the growing season
are the determining factors, in that order, for most crops under rain-fed agriculture in
the Texcoco River watershed.
References
[1] [FAO, 1983]. FAO; Guidelines: Land evaluation for rain fed agriculture; FAO Soils
bulletin 52, FAO, Rome, 1983.
[2] [Fu, 1994] Fu, LiMin; neural networks in computer intelligence; McGraw Hill, 1994.
[3] [Sys, 1985] Sys, Ir C.; Land Evaluation; State University of Ghent; Ghent, 1985.
[4] [Huajun, 1991] Huajun, Tang; Debaveye, J.; Da, Ruan; Van Ranst, E.; Land suitability
classification based on fuzzy set theory; Pedologie XLI-3 pp 277 - 290, 1991.
[5] [Liao et al., 1992] Liao, K., M.R. Paulsen, J.F. Reid, B. Ni, and E. Bonifacio; Corn Kernel
shape identification by machine vision using neural network classifier; ASAE paper No. 927017,
St. Joseph, Michigan, 1992.
[6] [Zhuang et al., 1992] Zhuang, X., A. Heztroni, and C.J. Precetti; Neuro-segmentation of
colour images in quality evaluation of pork images; ASAE paper No. 92-3595, St. Joseph,
Michigan, 1992.
[7] [Thai and Shewfelt, 1991] Thai, C.N., and R.L. Shewfelt; Modelling sensory colour quality
of tomato and peach: Neural networks and statistical regression; Transactions of the ASAE
34(3): 950-954, St. Joseph, Michigan, 1991.
[8] [Ozer and Engel, 1994] Ozer, N., B.A. Engel; An adaptative technique for mutisensor
fusion and classification of fruit; ASEA paper No. 946025, St. Joseph, Michigan, 1994.
[9] [Panigrahi, 1998] Panigrahi, Suranjan; Neuro-Fuzzy Systems: applications and potential in
biology and agriculture; Dept. of agricultural and biosystems engineering, North Dakota State
University, USA, 1998.
[10] [Easson and Barr, 1996] Easson, Gregory L.; Barr, David J.; Integration of GIS and
artificial neural networks for natural resource applications;
www.esri.com/userconf/proc96/TO150/PAP126/P126.html
[11] [Decatur, 1992] Decatur, S.E.; Application of neural networks to terrain classification;
Hughes Company, P.O. box 3310, Fulenton, California, 1992.
[12] [Callejas, 2002] Callejas Saenz, Luis Manuel; A computer system for land suitability
assessment based on fuzzy neural network; Trent University, Canada; 2002.
[13] [Callejas, 1999] Callejas Saenz, Luis Manuel; Ponce Hernandez, Raul; An approach to
develop a spatial decision support system for natural resources; Memorias del 2do. Encuentro
Internacional de Computo, Sociedad Mexicana de Ciencias Computacionales, México, 1999.
[Rossiter, 1986] Rossiter, D; ALES, an automated land evaluation system. Cornell University,
Ithaca, N.Y. 1986