A Parameter based ANFIS Model for crop yield prediction
Aditya Shastry1 , Sanjay H A2 and Madhura Hegde3
Nitte Meenakshi Institute of Technology
Bangalore, India
e-mail: 1shastryaditya@gmail.com, 2sanju.smg@gmail.com,3madhuhegede@gmail.com
Abstract — Crop yield prediction is important as it can
support decision makers in agriculture sector. It also assists in
identifying the relevance of attributes which significantly affect
the crop yield. Wheat is one of the widely grown crops around
the world. Its accurate prediction can solve various problems
related to wheat farming. This work analyses how yield of a
particular crop is determined by few attributes. In this paper,
the models of Fuzzy logic (FL), Adaptive Neuro Fuzzy
Inference System (ANFIS) and Multiple Linear Regression
(MLR) are used for predicting the yield of wheat by
considering biomass, extractable soil water (esw), Radiation
and rain as input parameters. The outcome of the prediction
models will assist agriculture agencies in providing farmers
with valuable information as to which factors contribute to
high wheat yield. We compare all these models based on
RMSE values. Results show that the ANFIS model performs
better than MLR and FL models with a lower RMSE value.
Keywords- Fuzzy Logic; Multiple Linear Regression; Adaptive
Neuro Fuzzy Inference System; yield; parameters
I. INTRODUCTION
Prediction of any natural event requires information
regarding its time of occurrence and nature, based on logical
analysis. Manual analysis generates inconsistencies due to
some factors like fatigue, contradiction of personal
perceptions, etc. Soft Computing like fuzzy logic, neural
computing, etc. can be applied to a wide variety of real
world applications as they can handle uncertainties better
than traditional methods. It resembles how human reasons
which are quite different from how regular methods based
on sentential logic, predicate logic operate.
Predictive models take historic information and are able
to predict the future values with less expense and more
quickly. They can provide support for human decisions,
making them more efficient or in some cases; they can be
used to automate entire decision-making processes.
The motivation behind the proposed work is to build and
customize such a predictive model which can be used for
predicting the crop yields by providing different attributes
on which crop yield is dependent.
In this paper, we have applied FL, MLR and ANFIS
models for predicting the yield of wheat crop obtained from
Agriculture Production Simulator (APSim) version 7.5 [1].
The data set comprises of various weather, crop, soil and
historic yield information. This dataset is preprocessed by
removing outliers, redundant, inconsistent and missing
values. We have conducted experiments on the wheat data
set using FL, MLR and ANFIS models. In our work four
978-1-4799-8047-5/15/$31.00 2015
c IEEE
input parameters which are biomass, esw, radiation and
amount of rainfall during sowing are considered for wheat
yield prediction. The data set consists of 50 records from
1960 to 2009 [1]. Based on the experiments we found that
the ANFIS model performs better than FL and MLR models
with a lower RMSE value.
The rest of this paper is organized as follows: Section II
discusses some related work. Section III gives the details of
the data source along with the description of the parameters
used for wheat yield prediction. Sections IV, V and VI
discuss the applications of FL, MLR and ANFIS models for
wheat yield prediction. Section VII gives the details about
the experiments conducted and results obtained along with
the comparison of results followed by the conclusion and
future work.
II. RELATED WORK
The previous works for predicting the yields of various
crops like rice, wheat using different prediction models are
surveyed in this section.
V.R Thakare et.al [2] proposed a fuzzy system that
predict maximum yield from crops. This system is able to
predict the crop name which can provide maximum yield. It
is also able to find the particular soil type and climatic
condition which is more suitable to the crop. 15 Soil
parameters and 22 crops are considered here.
Rajeshwari G Joshi et.al [3] proposed a Decision Support
System (DSS) for predicting crop suitability. They have
summarized fuzzy related aspects which can be incorporated
in an online farmer assisting system. Various factors for
crop selection have been considered by the authors while
proposing the DSS. They observe that such a DSS would
minimize losses due to fewer yields.
O.K. Chaudhari et.al [4] proposed ANFIS Based Model
for maximizing the profit of rice using Multi Objective
Linear Programming Problem by optimization method.
Labour wages, Machinery Cost, Fertilizer_manure_Cost and
Seed_Cost were considered as input variables and the profit
through produced yield is the output for maximizing the
profit.
V Mohan Vizhakar et.al [5] proposed a Expert Fuzzy
Model which could predict Avalanche. The authors have
made an attempt to develop a simpler and better technique
for Avalanche prediction using algorithm based on Fuzzy
Logic.
253
 S.P.Srinivasan et.al [6] proposed a new ANFIS approach Table I depicts a sample of wheat data set used for our
which is able to efficiently predict yield of Crop in the experiments. The following input parameters were
supply chain of Jatropa. considered in our work for predicting the yield of wheat:
Agus Buono et.al [7] implemented a Fuzzy Inference • Biomass: is the accumulated energy in plants
System which was able to increase the resilience of Rice measured in kg/ha
Yield. The authors have considered two data sets: SOI data
• esw (mm): is the extractable soil water measured in
set from 1877 to 2011 and rainy season dataset from
Indramaya District. They have observed a correlation value mm
of 0.68 between rainy onset and its predicted value. • Rain(mm): Amount of rainfall since sowing
Jesus Soto et.al [8] proposed how Type 2 and 1 Fuzzy measured in mm
Integrators can be optimized using Genetic Algorithms. • Radiation (MJ/m2): Amount of solar radiation
Their goal was to develop an ensemble approach for ANFIS measured in MJ/m2
models which could make the prediction error as minimal as
The wheat yield is the output that will be predicted by the
possible.
MLR, FL and ANFIS prediction models. It is measured in
Kefaya Qaddom et.al [9] proposed Adaptive Neuro-
kg/ha. Historic wheat yield information was considered
Fuzzy Modeling for Crop Yield Prediction. The authors
from 1960 to 2009.
have predicted tomato yield using neural networks by
considering various environmental variables. The TABLE I. Sample Wheat Data Set [1]
forecasting accuracy is calculated by averaging it over all Biomass esw Radiation Rain Wheat Yield
the years for which prediction is considered. 5253.3 105.3 28.3 33 2006.1
Maria Rossana et.al [10] proposed a Prediction Model 6905.3 113.5 25.5 106 2908.5
Framework for Crop Yield Prediction. This study 5911.9 94.2 23.8 39 2135.4
investigates the development of a crop prediction model 5163.9 113.2 30.9 34 1819.2
framework that can pre-process and fuse input raw data 5739.7 98.5 25.8 8 2087.2
5822.7 93.5 24.1 35 2572.1
from multiple sources, provide an accurate prediction of
5163.3 98.4 27.3 32 1952.5
crop yield, identify significant variables that affect crop
yield, and identify useful prediction policies using four IV. FUZZY LOGIC FOR YIELD PREDICTION
climate-related variables, 13 agronomic variables, and
For fuzzy logic implementation in predicting wheat yield
variables on weather disturbance. Comparison experiments
the Fuzzy Inference System editor in Matlab was used.
are performed to determine which data mining technique is
The particular Fuzzy Component is developed in
used for each framework component.
MATLAB. The wheat data set is collected as described in
Most of the related works have not considered critical
Section III. In order to predict the wheat yield we
parameters like crop biomass, rainfall, extractable soil water
considered input parameters like esw, biomass, rain and
esw and solar radiation for wheat yield prediction. In this
radiation [1]. The membership function is designed
work, we have considered these critical parameters and
carefully so that accurate output is obtained. Finally thirteen
found that crop yield of wheat can be predicted more
rules are generated for the fuzzy inference system. Figure 1
accurately by considering them. Further, we have combined
shows the methodology used for predicting the wheat yield
the aspects of fuzzy logic and artificial neural networks
using Fuzzy system.
using ANFIS and shown that it can be successfully applied
to predict the yield of wheat.
Crop Data Collection
III. DATA SOURCE AND DESCRIPTION
This section describes the source of data from which Selection of Input parameters
wheat data set was obtained to perform our experiments. It
also describes the input parameters considered for wheat
yield prediction. Membership function design
The wheat data set of 50 records collected from a period
of 16-03-1960 to 17-03-2009 was obtained from
Agricultural Production System Simulator (APSim) version List of Rules in Rule Base
7.5. APSim collected the data as part of ACIAR (Australian
Centre for International Agricultural Research) project
“Expanding Rabi Season Cropping in Southern Bangladesh” Implementation and testing of Fuzzy System
in 2012 from North Joynags, Bhola region, Southern
Fig.1. Methodology for Fuzzy system
Bangladesh with AP Soil number: 673. The Soil Texture
was Silt with Latitude and Longitude of 22.585 and 90.666
respectively [1].

254 2015 IEEE International Advance Computing Conference (IACC)

 We have made use of the Mamdani fuzzy inference
system to construct four input variables which are esw,
biomass, rain and radiation.
Figure 2 illustrates the structure of fuzzy system used for
wheat yield prediction. One output variable is used which is
the wheat yield. The input variables use membership
functions such as small, medium and large. The output
variables membership functions such as small, medium,
large are used. The inputs are assigned with the specific
condition. f(u) in Fig.2 represents the first order polynomial
which is the output of the if-then rule. In our case, it
specifies whether the yield is large or small.
Biomass
esw
Sugeno f(u)
Rain
Radiation Yield
Fig. 2. Fuzzy system structure for wheat yield prediction
Rule editor in Matlab was used for creating the fuzzy
rules. It was invoked using rule edit (‘a) to modify the rules
of FIS structure stored in a file named a.fis. We set the
specific rules in the rule editor using AND, OR and NOT
operations. We were able to build 13 rules in total for
predicting the wheat yield. Sample two rules generated by
the fuzzy system for predicting the yield of wheat is shown
below:
• Rule-1: if (esw is large) and (biomass is small) and
(rain is small) and (radiation is small) then (yield
is small)
• Rule-2: if (esw is medium) and (biomass is large)
and (rain is medium) and (radiation is medium)
then (yield is large)
V. MULTIPLE LINEAR REGRESSION
Multiple Linear regression is implemented using Matlab
R2013a. Equation 1 is applied on the wheat data set.
y = -3.0493+0.0006x1+0.0021x2+0.0137x3–0.0003x4+Ķ (1)
Where:
• y is the predicted yield of wheat crop,
• x1 is the biomass (kg/ha),
• x2 is esw (mm),
• x3 is Rain (mm),
• x4 is Solar Radiation (MJ/m2)
• Ķ (Error variance) = 2.6899e+004.
Here, the training data is modeled using linear predictor
functions and the unknown coefficients are estimated. Table
II shows the steps for implementing MLR on the wheat data
set using Matlab R2013a.
2015 IEEE International Advance Computing Conference (IACC)
TABLE II. Steps for wheat yield prediction using MLR
Input: Experimental data set of weather data, crop data and soil data
Output: Predicted wheat yield for the experimental dataset.
Method: Step1: Preprocess the data.
Step 2: Load the data.
Step 3: Fill ‘1’s in the first column of data set and
store it in x[ ].
Step 4: Store yield column in Y [ ].
Step 5: Use Regress function to apply MLR on the
given dataset.
Step 6: Display the STATS [1] which gives RMSE
Statistics of data
VI. ANFIS MODEL
Here we have applied the ANFIS model for predicting
the wheat yield by configuring its parameters. ANFIS uses
hybrid training algorithm which is a combination of back-
propagation and least mean square methods. Its main
purpose is to minimize the error due to approximation.
When fuzzy inference file (FIS) is trained, number of
epochs are used which are used to minimize the error. More
the number of epochs lesser will be the error. In ANFIS,
epochs are used to adjust the weights in order to minimize
the error.
Figure 3 shows the ANFIS methodology adopted for
predicting wheat yield. The input data set consisting of 50
records of esw, biomass, rain, radiation and historic wheat
yield which is loaded into the ANFIS. The ANFIS is then
adjusted by setting the variables and is trained. The ANFIS
model is trained using training data set and is validated for
the test set. The training and validating processes are
repeated until the RMSE value reduces. This result in
various ANFIS models being generated with different
configuration parameters. Among these models, the ANFIS
model with a relatively least RMSE value is chosen as the
best model. This model is able to predict the yield of wheat
more accurately.
Dataset
ANFIS development
ANFIS variable setting
Training
No
Validating?
Yes
Results
Fig.3. ANFIS Methodology
255
 TABLE III show the implementation process of the
customized ANFIS model.
TABLE III. Steps for wheat yield prediction using ANFIS
Input: Experimental data set of weather data, crop data and soil data
Output: Predicted wheat yield for the experimental dataset.
Method:
Step1: Wheat crop data set of 50 records with esw, biomass, rain and
radiation as attributes is collected.
Step2: The data set is split into training, validation and test sets with
input-output pairs. Here the input is the dataset with
parameters esw, biomass, rain and radiation while output is
the predicted wheat yield.
Step3: The initial first order TSK fuzzy inference system consisting
of first order polynomial as the output of rule is generated.
This is done using Subtractive Clustering which is useful
when the numbers of clusters are not known. Here, the first
cluster centre is selected from the data set where more data
points are closer. Again calculation is done until clusters
become stable.
Step3-1: The number of clusters and data points in each
cluster is estimated
Step3-2: The Initial parameters of the rules and the rules from
the rule base are mined
Step4: In this step the configuration parameters of the ANFIS model
are adjusted continuously until the error goes below a certain
threshold value or the numbers of epochs reach their
maximum limit.
Step5: The performance measure of the ANFIS model which in this
case is the RMSE value is computed.
Step6: The cluster size is set again along with the step size and the
ANFIS model is executed again.
Step7: The ANFIS model having the lowest RMSE value for the test
set is considered as the best model for predicting the yield of
wheat.
Instead of choosing the membership function
parameters in ANFIS model randomly, they are chosen so
that they match the input/output data. This facilitates the
model to take into account the variations present in the data
values. The hybrid combination of fuzzy logic and neural
networks allows the fuzzy modeling procedure to determine
the membership function parameters which can be better
suited for the input/output data. The FIS is constructed by
the ANFIS based on the input/output data set of wheat. The
function parameters of FIS are adjusted using a hybrid
combination of Back Propagation (BP) and Least Squares
methods or only BP so that the resulting ANFIS model is
able to model the dataset properly.
Implementation is done using M file editor in
matlab for ANFIS modeling. When the Matlab Script file is
run an RMSE value equal to 3.3282 is obtained. Figure 4
depicts the ANFIS structure for this modeling where
inputmf is the input membership function and outputmf is
the output membership function. The inputmf maps the
inputs esw, biomass, rain and radiation to the rules. The
outputmf maps the rules to the output which is the wheat
yield in our case.
256 2015 IEEE International Advance Computing Conference (IACC)
inputmf rules outputmf
inputs
biomass
output
esw
. .
rain
. . (wheat
yield)
. .
radiation
Fig.4. ANFIS structure for wheat yield prediction
VII. EXPERIMENTAL SETUP AND RESULTS
This section explains the experiments carried out on
wheat data set using FL, MLR and ANFIS models. It also
compares the results obtained from them.
The MLR, FL and ANFIS models were implemented
using Matlab R2013a on windows 7 operating system ,Intel
® Pentium ® CPU P6200 @2.13 GHz processor with 4GB
RAM and 500 GB Hard Disk.
A. Prediction Results Using Fuzzy Logic on Wheat Crop:
We conducted experiments using FL editor of Matlab.
Thirteen rules were generated for the given dataset. An
RMSE value of 6.4251 was found when fuzzy logic model
was applied on the wheat data set.
B. Prediction Results Using MLR on Wheat Crop:
We conducted the experiment using regress function in
Matlab. The RMSE VALUE was found to be 9.252 when
applied on the wheat data set taking crop biomass, esw,
Radiation and Rain as input parameters. The output
parameter was the predicted wheat yield.
When the MLR was applied on the wheat data set the
coefficients (ß0 = -3.0493, ß1 = 0.0006, ß2 = 0.0021, ß3 =
0.0137, ß4 = -0.0003) and error variance (Ķ = 2.6899e+004)
were derived.
C. Prediction Results using ANFIS on Wheat Crop
We conducted experiments using ANFIS editor of
Matlab. 81 fuzzy rules were generated by the ANFIS model
using the input parameters esw, biomass, rainfall and
radiation. The output parameter was the predicted wheat
yield. The RMSE value for the ANFIS model was found to
be 3.3282.
D. Comparison of Fuzzy Logic, MLR and ANFIS
prediction models
We compared Fuzzy crop prediction model, ANFIS crop
prediction model and MLR crop prediction model based on
RMSE values. The experimental results of MLR, FL and
ANFIS models are shown in Table IV.
TABLE IV. Experimentation results of prediction models
Prediction Models RMSE Values
Multiple Linear Regression (MLR) Model 9.252
Fuzzy Logic (FL) Model 6.4251
Adaptive Neuro Fuzzy Inference System (ANFIS)
Model 3.3282
Figure 5 shows the comparison of the prediction models
based on their RMSE values.
Fig.5. Comparison of prediction models
Among all 3 models we observe that the ANFIS model
gives better accuracy than MLR and FL models with a
lower RMSE value.
VIII. CONCLUSION
In this paper we have compared crop yield prediction
using three prediction models which are Fuzzy Logic,
ANFIS and MLR models. APSim was used for obtaining
the wheat data set. MLR, FL and ANFIS models were
applied on this data set. Prediction results of the models
were compared based on the RMSE values which were
found to be 9.252, 6.4251 and 3.3282 for MLR, FL and
ANFIS models respectively. Based on this we conclude that
ANFIS was able to predict the yield of wheat more
accurately then the MLR and FL models. We observe that
this was due to the fact that the ANFIS model was able to
capture the nonlinear relationship between the data variables
more effectively than the other two models.
In future we plan to extend this work for different types
of crops by considering different parameters. Also, we plan
to develop better or improved prediction models for this
purpose.
2015 IEEE International Advance Computing Conference (IACC)
REFERENCES
[1] N. Dalgliesh, M. Saifuzzaman, I. Khan, A.B.S Hossain, H.
Rawson, P. Carberry, P. Poulton, Q.K. Alam and J. Kabir,
“Expanding the area for Rabi-season cropping in southern
Bangladesh”. Australian Centre for International Agricultural
Research (ACIAR), Australia, Tech. Report. FR2012-01, Jan.
2012.
[2] V.R.Thakare and H.M. Baradkar. “Fuzzy System for
Maximum Yield from Crops”, International Journal of
Applied Information Systems, pp. 4-9 , March 2013.
[3] Rajeshwar G Joshi, Parag Bhalchandra and S.D.Khmaitkar.
“Predicting Suitability of Crop by Developing Fuzzy Decision
Support System”, International Journal of Emerging
Technology and Advanced Engineering, vol. 3, pp. 10-16 , Jan
2013.
[4] O.K. Chaudhari and P.G. Khot. “ANFIS based Model based
in Decision Making to Optimize the Profit in Farm
Cultivation”, International Journal of Engineering Science
and Technology, vol. 4, pp. 442-448, Feb. 2012.
[5] Mohan Vizhakat. “Expert Fuzzy Model for Avalanche
Prediction”, Defence Science Journal, vol. 53, No. 4, pp. 443-
451, October 2003.
[6] S.P.Srinivasan and P.Malliga. “A New Approach of Adaptive
Nuero Fuzzy Inference System Modelling For Crop
Prediction in the Supply Chain of Jatropa”, in Proc.17th IEEE-
IE&EM, Oct 2010, pp. 1249-1253.
[7] Agus Buono and Mushthofa. “An Implementation of Fuzzy
Inference System for Onset Prediction Based on Southern
Oscillation Index for Increasing the Resilience of Rice
Production Against Climate Variability”, in Proc. IEEE-
ICACSIS, Dec. 2012, pp. 281-286.
[8] Jesus Soto and Patricia Melin. “Optimization of Interval
Type-2 and Type-l Fuzzy Integrators in Ensembles of ANFIS
Models with Genetic Algorithms”, in Proc. IEEE-World
Congress on NaBIC, Aug. 2013, pp 41-46.
[9] Kefaya Qaddom and Evor Hines. “Adaptive Neuro-
FuzzyModeling For Crop Yield Prediction”, in Proc. 10th
ACM-WSEAS, 2011, pp 199-204.
[10] Maria Rossana C. de Leon, Eugene Rex L, Jalao (2013) A
Prediction Model Framework for Crop Yield Prediction. In
Proc. 14th Asia Pacific Industrial Engineering and
Management Systems Conference. pp 185.
257