computers

Article
Improved Optimization Algorithm in LSTM to Predict
Crop Yield
Usharani Bhimavarapu 1, * , Gopi Battineni 2 and Nalini Chintalapudi 2

1 Department of Computer Science and Engineering, Koneru Lakshmaiah Education Foundation,

Vaddeswaram 522502, Andhra Pradesh, India
2 Medical informatics Centre, School of Medicinal and Health Science Products, University of Camerino,
62032 Macerata, Italy
* Correspondence: ushafdp1122@[Link]

Abstract: Agriculture is the main occupation across the world with a dependency on rainfall. Weather
changes play a crucial role in crop yield and were used to predict the yield rate by considering
precipitation, wind, temperature, and solar radiation. Accurate early crop yield prediction helps
market pricing, planning labor, transport, and harvest organization. The main aim of this study is to
predict crop yield accurately. The incorporation of deep learning models along with crop statistics can
predict yield rates accurately. We proposed an improved optimizer function (IOF) to get an accurate
prediction and implemented the proposed IOF with the long short-term memory (LSTM) model.
Manual data was collected between 1901 and 2000 from local agricultural departments for training,
and from 2001 to 2020 from government websites of Andhra Pradesh (India) for testing purposes.
The proposed model is compared with eight standard methods of learning, and outcomes revealed
that the training error is small with the proposed IOF as it handles the underfitting and overfitting
issues. The performance metrics used to compare the loss after implementing the proposed IOF
were r, RMSE, and MAE, and the achieved results are r of 0.48, RMSE of 2.19, and MAE of 25.4. The
evaluation was performed between the predicted crop yield and the actual yield and was measured
in RMSE (kg/ha). The results show that the proposed IOF in LSTM has the advantage of crop yield
prediction with accurate prediction. The reduction of RMSE for the proposed model indicates that
Citation: Bhimavarapu, U.; the proposed IOFLSTM can outperform the CNN, RNN, and LSTM in crop yield prediction.
Battineni, G.; Chintalapudi, N.
Improved Optimization Algorithm in Keywords: crop yield; deep learning; LSTM; model performance; optimizer
LSTM to Predict Crop Yield.
Computers 2023, 12, 10. https://
[Link]/10.3390/computers12010010

Academic Editors: Phivos Mylonas, 1. Introduction

Katia Lida Kermanidis and Agricultural yield mainly depends on rainfall, which is a vital factor in the growth and
Manolis Maragoudakis production of food crops. Reports highlighted that there is a direct relationship between
Received: 2 December 2022
crop growth and rain patterns. The actual rainfall recorded for the year 2020 by the
Revised: 30 December 2022 southwest monsoon is 704.8 mm, and by the northeast monsoon is 370.3 mm [1]. In India,
Accepted: 31 December 2022 the land usage for agriculture is about 200.2 million hectares, which means a big source
Published: 3 January 2023 of livelihood [2]. The prediction of crop yield plays a vital role in the management of
controlling the crop parameters and guarantees the maximum yield.
Prediction of yield is one of the most significant and well-known themes in accuracy
farming as it characterizes yield planning and assessment, coordinating harvest supply
Copyright: © 2023 by the authors. with the request, and crop management [3]. Smart farming makes it possible to conduct
Licensee MDPI, Basel, Switzerland. agriculture more effectively and efficiently through high-precision algorithms. In particular,
This article is an open access article Machine Learning (ML) drives the mechanism of machines which are able to learn without
distributed under the terms and the involvement of manpower [4]. ML emerges jointly with high-performance computing
conditions of the Creative Commons
and big data knowledge, creating novel opportunities to understand, unravel, and quantify
Attribution (CC BY) license (https://
data-intensive procedures in agricultural operations. State-of-the-art models have gone a
[Link]/licenses/by/
long way past basic predictions dependent on past data and consolidate computer vision
4.0/).

Computers 2023, 12, 10. [Link] [Link]

Computers 2023, 12, 10 2 of 19

advances to give information in an urgent and far-reaching multidimensional examination

of crops, climate, and financial conditions to take advantage of the yield for the population.
Deep learning is a subpart of ML knowledge that works on artificial neural networks
(ANN) with representation learning. As deep learning was successfully implemented in
various domains such as computer vision, medicine, chatbots, etc. [5,6], it recently entered
agriculture and was majorly used for crop classification [7], predicting and estimating
crop yield [8]. In particular, ANN models are largely utilized to predict crop yields by
considering various factors such as water content, solar radiation, CO2 fixation, and
transpiration [9–11]. It is reported that the Recurrent Neural Network (RNN) can predict
crop yield with high accuracy [12]. Similarly, combining convolutional neural networks
(CNN) and long short-term memory (LSTM) models can estimate the yield for country-level
soybean crops based on remote sensing data [13].
Yang et al. [14] implemented the region-based CNNs to the forest of the strawberry
yield and detected and counted the number of flowers, mature strawberries, and immature
strawberries. Helena et al used the 3D CNN and RNN models, which are largely incorpo-
rated to predict crop yield, and concluded that the proposed 3DCNN model outperforms
some machine learning algorithms [15,16]. In addition, the LSTM model forecasts corn
yield by considering weather data and soil data parameters to achieve better performance
in yield prediction [17,18]. However, all these models do not guarantee the low process
time and the reduction of loss that are largely attributed to the prediction of crop yields.
This research forecasts future crop fields by considering the weather parameters
and rainfall time series. This study aimed to propose a new and improved optimization
function (IOF) to train the LSTM model to reduce the training and testing loss. The
proposed IOFLSTM model can enhance most of the advanced LSTM models on the public
crop yield and available production datasets. In addition, the performance evaluation of
the developed optimizer IOF was performed by comparing the proposed IOF with the
standard training of the seven gradient descent optimizers, which can be elaborated on in
further sections. The scarcity of research on reducing overfitting issues among crop yield
datasets has motivated us to develop the IOFLSTM model by handling the bias variance.
The novel contributions of this study are:
• Propose a new optimized IOF function to train the LSTM model to reduce the training
and testing loss;
• LSTM uses the new optimization function to calculate the model error;
• The proposed IOFLSTM model was superior to most of the state-of-the-art LSTM
models on the public crop yield and available production datasets;
• Evaluate the performance of the proposed optimizer IOF by comparing the proposed
IOF with the standard Stohastic Gradient Descent (SGD), Momentum, AdaGrad, Root
Mean Square prop, and Adam by training using the Root Mean Square (RMSE), Mean
Square Error (MSE) functions. The results show that the proposed IOF prevents the
model from overfitting by handling the bias-variance.

2. Materials and Methods

2.1. Dataset Description
We collected the rainfall data from 1901 to 2020 and the crop yield data of paddy, maize,
ragi, sugarcane, and cotton from the Indian government website [19]. Additionally, the
metrological data from the Indian metrological department website was incorporated [20].
From 1901 to 2000 as the training data manually collected the total crop yield yearly
production from the local agricultural departments and from 2001 to 2020 as the testing data
from the Andhra Pradesh government websites and the Indian government agricultural
websites [21]. The crop yield dataset consists of 25,515 rows and the number of features
in the dataset is 26. Table 1 tabulates the features used in this research. The maximum
temperature varies from 30.1 ◦ C to 45.4 ◦ C, the minimum temperature varies from 19 ◦ C
to 32.5 ◦ C, the wind speed varies from 12 km/h to 21 km/h, and the relative humidity
Computers 2023, 12, 10 3 of 19

varies from 52 to 79, and the solar radiation varies from 8.1 to 10.9. The rainfall varies from
0.2 mm to 7 mm.

Table 1. Dataset features used in this study.

N Feature Unit of Measurement

1 Min. temperature ◦C

2 Max. temperature ◦C

3 Avg. temperature ◦C

4 Total rainfall mm
5 Humidity %
6 Solar radiance w/m2
7 Southwest monsoon rainfall mm
8 Northeast monsoon rainfall mm
9 Production kg/hectare
10 Pest standing crop Paddy, wheat, jowar, bajra . . .
11 District pest affected Krishna, Guntur, . . .
12 Pest affected area Hectares
13 Area treated Hectares
14 Seasonal crop yield Tones in millions
15 Total pulses %
16 Total food grains %
17 Total oilseeds %
18 Total cropped area Area in lakh hectares
19 Target sown area Area in lakh hectares
20 District Krishna, Guntur, . . .
21 Seasonal sown area Area in lakh hectares
22 Seasonal area under production Area in lakh hectares
23 Total yield Tones in millions
24 Cropping season Kharif or rabi
25 Crop Paddy, wheat, jowar, bajra . . .
26 Total crop yield kg/hectare

Figure 1 shows the major crops that were cultivated in India and their yields in the
year 2000. Rice is one of the major crops cultivated in India. State-based statistics were
collected from the agricultural departments and from the Indian government agricultural
websites [21]. The yielded data were reported in units of kg per hectare.

2.2. Predictive Modeling

Based on the differences in rainfall, we used predictive modeling to forecast the Kharif
crop yield for the year 2021. The feature selection was performed on the attributes by
combining the rainfall data with the crop yield data. As the Kharif crop mainly depends
on the information on rainfall, this study only used data from June to September as the
training data to do model training. Therefore, weather data is helping to provide input to
train the model and crop yield data as predicted outcomes.
Computers 2023, 11, x FOR PEER REVIEW 4 of 18
Computers 2023, 12, 10 4 of 19

Figure1.
Figure CropYield
[Link] yield data.
data.

2.3. Padding and Optimization

2.2. Predictive Modeling
The imported data may consist of unfilled data. To complement the incomplete
Based on
measures, thethe
”−differences
1” paddingintechnique
rainfall, we used
was predictive
used. modeling
Optimizers to forecast
minimize the Kha‐
the error by
rif crop yield for the year 2021. The feature selection was performed
updating the weight parameters and reducing the training time. An optimizer is an on the attributes by
combining the rainfall data with the crop yield data. As the Kharif crop
algorithm or a function that changes the attributes of the deep learning models such mainly depends
on
as the information
learning rate and onweights
rainfall,tothis studythe
reduce only used
error data
[22]. from June
Gradient to September
descent as the
is a first-order
training data to
optimization do model
algorithm, training.
and Therefore,
it is dependent on weather data is
the first-order helpingoftothe
derivate provide input to
error function
train
of thethedeep
model and crop
learning yield Optimizers
model. data as predicted
find aoutcomes.
way to alter the weights to minimize
the error.
2.3. Padding and Optimization
2.4. Existing Models with
The imported data LSTM Model of unfilled data. To complement the incomplete
may consist
The other
measures, deeppadding
the ”−1” learning technique
models used wasto used.
compare with the proposed
Optimizers minimize model are by
the error CNN,up‐
RNN, and GRU.
dating the weight parameters and reducing the training time. An optimizer is an algo‐
rithm or a function that changes the attributes of the deep learning models such as learn‐
2.4.1. Convolution Neural Network (CNN)
ing rate and weights to reduce the error [22]. Gradient descent is a first‐order optimization
CNN and
algorithm, comprises three layers;
it is dependent namely,
on the convolution,
first‐order pooling,
derivate and fully
of the error connected
function of theones,
deep
and this model automatically detects distinctly. The term convolution
learning model. Optimizers find a way to alter the weights to minimize the error. represents the
mathematical function of convolution, i.e., a linear operation where two functions are
multiplied
2.4. Existing to get the
Models withthird
LSTMfunction.
Model The pooling layer reduces computational costs by
reducing the size of the convolved feature map. The fully connected layer predicts the class
basedTheonother deepfeatures
extracted learningand
models used to compare
the convolution with
process the proposed model are CNN,
output.
RNN, and GRU.
2.4.2. Recurrent Neural Network (RNN)
2.4.1. RNN
Convolution
saves theNeural
outputNetwork (CNN)
of a specific layer and feeds it back to the input of another
CNN the
to predict comprises
result ofthree layers;
the layer. namely,
It can memorizeconvolution, pooling,
the previous inputs and
duefully
to itsconnected
internal
ones,
memory. and this
RNNmodel
has a automatically
vanishing gradient detects
and distinctly.
exploding The term convolution
gradient issues. represents the
mathematical function of convolution, i.e., a linear operation where two functions are
2.4.3. GatedtoRecurrent
multiplied Unit function.
get the third (GRU) The pooling layer reduces computational costs by
reducingGRUtheuses gates
size to control
of the the flow
convolved of information.
feature It maintains
map. The fully two layer
connected gates predicts
called reset
the
and update
class based onwhere the reset
extracted gate and
features contains the information
the convolution to remember,
process output. and the update
gate determines future data based on past information.
2.4.2. Recurrent Neural Network (RNN)
2.4.4. Long Short-Term Memory (LSTM)
RNN saves the output of a specific layer and feeds it back to the input of another to
LSTM is a special kind of RNN and can learn long-time dependencies [23]. The
predict the result of the layer. It can memorize the previous inputs due to its internal
standard LSTM consists of three gates, and these gates are responsible for regulating
memory. RNN has a vanishing gradient and exploding gradient issues.
the information and passing that information to the next unit. The forget value either
forgets everything or does not forget the information based on the values of the forget gate
2.4.3. Gated Recurrent Unit (GRU)
(i.e., the forget gate forgets everything if the value is zero, and nothing if the value is one).
The GRUinputuses
gategates to control
controls the new theinformation
flow of information.
to add theIt maintains two gates
next cell state, and itcalled
worksreset
in
and
twoupdate where
parts. The firstthe
partreset gate
of the contains
input gate isthe
theinformation
sigmoid layer,to remember,
which controlsand the
theoutput
update
gate determines future data based on past information.
 rything or does not forget the information based on the values of the forget gate (i.e
forget gate forgets everything if the value is zero, and nothing if the value is one)
input gate controls the new information to add the next cell state, and it works in
Computers 2023, 12, 10
parts. The first part of the input gate is the sigmoid layer, which controls5the of 19
output
stored in the cell state. The input gate’s second part is the Tanh layer, and it crea
vector of new feature values stored in the cell state. The output gates output the up
cell state
value information.
stored Through
in the cell state. thegate’s
The input gates’ structure,
second the
part is the statistics
Tanh execute
layer, and it createsselectivel
a
arevector
handedof new feature values
through to updatestoredand
in the cell state.
hold The outputstatistics
the historical gates output
andtheupdate
updatedthe cell
cell state information. Through the gates’ structure, the statistics execute selectively and
LSTM considers the previous historical values, analyzes the present unknown patter
are handed through to update and hold the historical statistics and update the cell state.
adjusting itself according
LSTM considers the previoustohistorical
the complete patterns,
values, analyzes theand makes
present future
unknown forecasts
patterns by ahead
functionality
adjusting itselfofaccording
LSTM istopresented
the completeinpatterns,
Figureand 2. makes future forecasts ahead. The
functionality of LSTM is presented in Figure 2.

Figure 2. The LSTM functionality representation.

Figure 2. The LSTM functionality representation.
An LSTM cell, ht−1 , is the previous memory output, and ct is the current memory
output. LSTM cell
An LSTM is explained
cell, ht−1, is theas:previous memory output, and ct is the current memor
• LSTM
put. It calculates
cell is the current memory
explained as: (cgt ), the weight matrix (wtCg ), and the bias is
the (bscg ).
 It calculates the current memory (cgt), the weight matrix (wtCg), and the bias
(bscg). cgt = Tahn(wtCg × [hdcg−1 ,xCg ] + bscg ) (1)

• The input gate manages thecg t = Tanh(wt

update Cg × [hd
of the current cg‐1,x input
memory ] + bsdata
cg) to the value of
the memory cell, the weight matrix (wtig ), and the bias (bsig ) and the sigmoid function.
 The
Theinput gateismanages
input gate calculated the
as: update of the current memory input data to the va
the memory cell, the weight matrix (wtig), and the bias (bsig) and the sigmoid fun
The input gate is calculated
igt = σ(wtas:
ig × [hdig−1 ,xig ] + bsig ) (2)

• igt = σ(wt
The forget gate controls the update × [hdig‐1memory
of theigprevious ,x ] + bs ig) to the value of the
data
memory cell, the weight matrix (wtf ), and the bias (bsfg ) and is the sigmoid function.
 The
Theforget gateis controls
forget gate calculatedthe
as: update of the previous memory data to the value
memory cell, the weight matrix (wtf), and the bias (bsfg) and is the sigmoid fun
The forget gate is calculated fg × [hdfgpedicle ] + bsfg )
fgt = wtf wtas: (3)

• lct−1 is the last LSTM cell value,

fgand the current memory cell can be calculated as:
t = wtfwtfg × [hdfgpedicle] + bsfg)

 lct−1 is the last LSTM cell value, and

cut = fit × lcthe current memory cell can be calculated
t−1 + cgt (4) a

2.5. Proposed Approach cut = fit × lct−1 + cg

This research proposes an enhanced optimized function, i.e., an Improved Optimiza-
tion Function (IOF), to reduce the error in crop yield prediction. The flowchart presentation
of the present study is presented in Figure 2.
In this study, time series forecasting helps to forecast future crop yield. LSTM considers
the previous historical values, analyzes unfamiliar patterns by regulating itself according
to the complete patterns, and makes further predictions. This study proposes an enhanced
optimization function, i.e., an Improved Optimization Function (IOF), to analyze the
 tion Function (IOF), to reduce the error in crop yield prediction. The flowc
tion of the present study is presented in Figure 2.
In this study, time series forecasting helps to forecast future crop yiel
siders the previous historical values, analyzes unfamiliar patterns by regul
Computers 2023, 12, 10
cording to the complete patterns, and makes further predictions.6 ofThis 19
stud
enhanced optimization function, i.e., an Improved Optimization Function
lyze the prediction.
prediction. Theinflow
The flow followed followed
this study is shownininthis study
Figure is shown
3. Xn and ininput
On are the Figure 3. X
the the
and input and
output the output
patterns, patterns,
and the observed andisthe
output the observed
actual [Link] is the actual va

Figure
Figure 3. Flow chartchart
3. Flow for thefor
IOF-LSTM.
the IOF‐LSTM.
In the proposed modeling, the IOF function works after error calculation, and only the
In the proposed
backpropagation modeling,
will start that theweights
can update the IOF function works
of the matrices after
and the cell error
states ofcalcula
the
the backpropagation will start that can update the weights of thethematrice
LSTM cell. The IOF function reduces the error and converges fast depending upon
backpropagation used. The algorithmic representation of IOF is given below.
states
An of the LSTM
improved cell. The
optimization IOFalgorithm
function function reduces
is given the error
in Algorithm and
1. The converges f
algorithm
upon until
repeats the backpropagation
θt is not convergent. Inused.
step 3, gThe algorithmic
t is the representation
current gradient and m is the decayingof IOF is g
mean over the previous updates. Step 3 allows the model to move
An improved optimization function algorithm is given in Algorithm 1. faster along dimensions
where the update is consistently smaller and smaller along turbulent dimensions where the
repeats until θt is not convergent. In step 3, gt is the current gradient and m i
update is significantly oscillating. β1 mt−1 and β2 vt−1 in step 4 and 5 do not depend on the
mean gradient
current over the butprevious
gives higherupdates.
quality byStep 3 allows
updating the model
the parameters to move
with the momentum faster alon
t t
step
wherebeforethecomputing
update the gradient. In steps
is consistently 6 and 7,and
smaller the denominator
smaller along 1 − βturbulent
1 , 1 − β2 dime
corrects the bias, which helps to continue changing direction even when the learning rate
the update is significantly oscillating. β1mt−1 and β2vt−1 in step 4 and 5 do n
the current gradient but gives higher quality by updating the parameters
mentum step before computing the gradient. In steps 6 and 7, the denomina
β corrects the bias, which helps to continue changing direction even whe
rate has annealed significantly toward the end of the training.
Computers 2023, 12, 10 7 of 19

has annealed significantly toward the end of the training.

Algorithm 1: IOF
α: step Size
η: Learning rate
β1 , β2 ∈ [0,1]: Exponential Decay rate to the moment estimation
1. while θt is not joined, repeat
2. t < −t + 1
3. gt ← 5θ ft (θt−1 )
4. mt ← log(β1 mt−1 +(1 – β1 )gt )
5. vt ← log(β2 vt−1 +(1 – β2 )g2t )
6. mt ← mt /(1 – βt1 )
7. mt ← vt /(1 –βt2 ) √
8. θt ← θt−1 – αmt /( ϑt + ε)
end
return θt

The updated algorithm of IOFLSTM is presented in Algorithm 2. In step 1, the LSTM

model is trained using the crop yield data. In step 3(a), consider I the input, i.e., fed to the
embedding layer to produce the output f(x). While coupling LSTM and IOF, the output
f(x) from the LSTM model and the ground truth y calculate the IOF for each instance. In
step 3(b), the IOF algorithm is implemented. The error is then passed through a nonlinear
function to create the weight used to adjust the loss. The forward propagation of the LSTM
cells is where the preprocessed input has taken and passed through the hidden states,
and then the output has been achieved. Then, the error between the predicted and the
actual values are calculated. The output value converges between zero and one. Then,
the predicted values are computed with the specified number of epochs until the RMSE
value converges. After completing the training procedure, it is tested on the test data, and
the accuracy is calculated in step 6. The obtained time ahead values of the IOFLSTM are
validated using the statistical evaluation parameters to analyze the forecast.
The hyperparameters applied during the training process were 11 hidden layers and
50 neurons in each layer. The proposed model was trained and tested with the same
hyperparameters of epoch size 1000, learning rate 1 × 10−2 , batch size 64, and dense layers
4 with weight decay constants of 0.0001. We used the maxout activation function for all
neurons in the model.

2.6. Performance Metrics

In regression models, the mathematical equations help to understand the relationship
between two variables. Generally, these models are largely useful in the estimation and
prediction of outputs. The models consider the different aspects of inputs (independent
variables) and predict output values (dependent variables). The difference between the
estimated model outputs versus actual values is called residuals, and these residuals play
a vital role in judging model efficiency. The collection of residual quantity is opposite to
model performance. From a statistical point of view, different parameters help to inspect the
residuals in judging the model’s accuracy. These performance metrics are: mean absolute
percentage error (MAPE), mean absolute error (MAE), root mean square error (RMSE), and
Adjusted R squared.
Mean Absolute Percentage Error (MAPE): It is the sum of the individual absolute errors
divided by everyone, and the average of those individual absolute errors. It determines
how far the proposed model prediction deviates from its corresponding output and is
defined as:
1 n yi − ŷi
m i∑
MAPE = (5)
=1 yi
Computers 2023, 12, 10 8 of 19

where yi is the actual value and ŷi is the predicted value and m is the total number
of observations

Algorithm 2: Updated IOFLSTM

Dataset S = { xm m
p , yp }, 1 ≤ m ≤ N
Input: rainfall historical data, crop yield historical data
Output: reduced loss value and reduced processing time of the f(x) data
1. Initialisation: gx hg fx hf xi hi xo ho g i f o
 w, w , w , w , w ,w , w , w , w , b , b , b , b , b
m gx m hg
2. ap = tanh wp xp + wp hp−1 + bg
 
im = σ w xi xm + whi hm + b
t p p−1 i
 p p 
fm
t = σ w fx xm + whf hm + b
p p p p−1 f
 
om xo m ho m
t = σ wp xp + wp hp−1 + +mt + νt + bo
m
Tm m
p−1 ⊗ ft + it ⊗ gm
 p
m
ht = ot ⊗ tanh Tm
m
p
m
3. om


t = σ wp ∗ h t
+b
m
4. δm m m


p = σ ot − yP
⊗ σ wp ∗ ht + b

0 m
δm m m
o,p = σ ot − yP ⊗ σ wp ∗ ht + b    
m
δf,p = wM σ om m m
⊗ tanh0 Tm 0 xo m ho m




t − yP ⊗ σ wp ∗ ht + b p ⊗ σ wp xp + wp ht−1 + bo ⊗
   
Tm xo m ho m
p−1 ⊗ σ wp xp + wp ht−1 + bo    
δm M σ om − ym ⊗ σ w ∗ hm + b ⊗ tanh0 Tm 0 xi m hi m




i,p = w t P p t p ⊗ σ wt xp + wp ht−1 + bi ⊗
   
gx hg ho m
tanh wp xm xo m
p + wp ht−1 + bg ⊗ σ wp xp + wp ht−1 + bo
   
M σ om − ym ⊗ σ w ∗ hm + b 0 m 0 gx m hg m
δm




g,p = w t P p t ⊗ tanh T p ⊗ tanh wp x p + w p ht−1 + bg ⊗
   
xi m hi m x0 m h0 m
σ wp xp + wp ht−1 + bi ⊗ σ wp xp + wp ht−1 + bo
N
m M
5. ∂LN
∂W = ∑ δm
p ( ht )
p=1
N
∂LN
∂b = ∑ δm p
p=1
N  M
m
xM = ∑ δp
∂LN
xm
p
∂W
p=1
N
m
M
∂LN
= ∑ δm p ht−1
∂WhM p=1
N
m
∂bp = ∑ δp
∂LN
p=1
6. Update weights and biases.

Mean Absolute Error (MAE): It is the absolute arithmetical mean of the absolute differ-
ence between the actual values and the predicted values. It represents the average of the
importance of the errors and is defined as:

1 n
m i∑
MAE = |yi − ŷi | (6)
=1

where yi is the actual value and ŷi is the predicted value and m is the total number
of observations.
Root Mean Square Error (RMSE): It is the standard deviation of the predicted errors. It
determines how well the information is concentrated on the best fit line and is defined as:
s
1 n
m i∑
RMSE = (yi − ŷi )2 (7)
=1
Computers 2023, 11, x FOR PEER REVIEW 9 of 18

Computers 2023, 12, 10 1 9 of 19

RMSE y y (7)
m

where yi is the actual value and ŷyi isis the

where the predicted
predicted value
value and m is the total number
number ofof
observations
observations
R Square
R (R2):2 ):This
Square(R This calculates
calculates thethe variance
variance of proposed
of the the proposed model
model over over the vari‐
the total total
variance. The higher the score, the smaller the difference between the observed
ance. The higher the score, the smaller the difference between the observed and the pre‐ and the
predicted values, which are defined
dicted values, which are defined as: as:

∑∑(yy − ŷ)y
2
RR22
= 1 1− ∑ yi (8)
(8)
y2
∑(yi − y)
where y is the actual value and y is the predicted value, y is the mean value of y .
where yi is the actual
Optimizers play value
a veryand
vitalŷi role
is the
inpredicted
efficientlyvalue, y is the mean
and effectively valueaofmodel
training yi . and
Optimizers play a very vital role in efficiently and effectively training
producing accurate results. The optimal parameter values minimize the loss, which affectsa model and
producing accurate results. The optimal parameter values minimize the loss, which affects
the overall training time. The proposed optimization algorithm navigates the possible set
the overall training time. The proposed optimization algorithm navigates the possible set
of weights to make good predictions that reduce the loss and improve the accuracy of the
of weights to make good predictions that reduce the loss and improve the accuracy of the
model. If the model loss is near zero then the model prediction is perfect, i.e., the lower
model. If the model loss is near zero then the model prediction is perfect, i.e., the lower
the loss the better the proposed model. So, the main goal of this paper is to minimize the
the loss the better the proposed model. So, the main goal of this paper is to minimize the
loss concerning the training model parameter through different optimization parameters.
loss concerning the training model parameter through different optimization parameters.
These experiments implemented the IOFLSTM using the Keras in Python on the Intel Core
These experiments implemented the IOFLSTM using the Keras in Python on the Intel Core
i5 3.4 GHz processor.
i5 3.4 GHz processor.
3.
3. Results
Results
This
This section presents the
section presents the comparison
comparisonof
ofdifferent
differentperformance
performancemetrics
metricsbetween
between dif‐
differ-
ferent deep learning algorithms.
ent deep learning algorithms.

3.1.
3.1. Prediction
Prediction of
of the
the Temperature, Yield
Temperature, Monsoon Rainfall, and Crop Yield
The
The purpose
purposeofofthis study
this studyis to
is predict
to predictrainfall, temperature,
rainfall, temperature,and crop
and yields for mon‐
crop yields for
soon
[Link].
During the monsoon
During the monsoonseason, high high
season, rainfall and high
rainfall temperatures
and high may con‐
temperatures may
tribute to high
contribute crop
to high yields.
crop Based
yields. Basedonon thethe
trends
trends inin
rainfall
rainfalland
andtemperature,
temperature,we wepredicted
predicted
crop yields.
crop yields. The
The proposed
proposed model
model was
was trained
trained from
from 1901
1901 to
to 1999
1999 and
and tested
tested on monsoon
monsoon
rainfall and
rainfall and temperatures
temperatures from
from 2000
2000 toto 2020.
2020. Finally,
Finally, the model was testedtested on
on rainfall
rainfall and
and
temperatures for
temperatures forthe
theyear
year2021.
[Link]
Figure 4 shows
4 shows thethe predicted
predicted temperature
temperature andand rainfall
rainfall for
for 2021.
2021. Figure
Figure 5 summarizes
5 summarizes the yield
the crop crop prediction
yield prediction
resultsresults of the IOFLSTM
of the IOFLSTM model inmodel
var‐
in various
ious districtsdistricts of Andhra
of Andhra [Link].
The crop The cropofyield
yield of a million
a million tons pertons per hectare
hectare per yearper
is
year is directly proportional to the annual rainfall in Andhra Pradesh. Figure
directly proportional to the annual rainfall in Andhra Pradesh. Figure 6 shows the predic‐ 6 shows the
prediction
tion of cropofproduction
crop production
for thefor the2021.
year year 2021.

Figure 4. Cont.
Computers 2023,12,
Computers2023, 11,10
x FOR PEER REVIEW 10 of 19
10 18

(a)

(b)
[Link]
Figure Temperatureand
andrainfall
rainfallprediction
predictionfor
forthe
theyear
year2021.
2021.(a)
(a)Temperature
Temperatureand
and(b)
(b)Rainfall.
Rainfall.

3.2. Performance Comparison

Comparing the proposed IOF with the default optimizer Adam using different deep
learning models were presented in Table 2. The performance comparison of major crops in
AP with different deep learning models (LSTM, RNN, ANN, and CNN) and our proposed
IOFLSTM using the statistical parameters are presented in Table 2. In an LSTM model,
the RMSE values range from 0.021 to 3.03, the ‘r’ values range from 0.29 to 0.45, and the
MAE values range from 0.69 to 3.54 during the model training. Experimental outcomes
provide evidence that the IOFLSTM outperforms others and produced accurate prediction
numbers. The loss with IOF is less than the default optimizer ADAM. Table 3 tabulates the
comparison of different optimizers using the LSTM model. A comparison of the proposed
model with the RNN, LSTM, ANN, and CNN models is tabulated in Table 4.
 Computers2023,
Computers 2023,12,
11,10
x FOR PEER REVIEW 1111ofof18
19

Figure 5. Crop yield prediction

Figure
Figure 5. Crop
5.
(million
Crop yield
yield prediction
tons pertons
prediction (million
hectare)
(million tons per forformajor
per hectare)
hectare) for major crops
major crops
crops for
for districts
for districts
districts in
in Andh
in Andhra
Andhra
Pradesh from 2017 to [Link]
Pradesh
Pradesh from 2017
2017 to
to 2020.
2020.

Figure 6. Prediction of crop yield for the year 2021 in Andhra Pradesh.

3.2. Performance Comparison

Comparing the proposed IOF with the default optimizer Adam using different deep
learning models were presented in Table 2. The performance comparison of major crops
in AP with different deep learning models (LSTM, RNN, ANN, and CNN) and our pro‐
Predictionusing
Figure 6. Figure
posed
Prediction of crop yieldoffor
[Link] cropthe
yield
the for the year
statistical
year 2021 2021 in Andhra
parameters
in Andhra Pradesh. in Table 2. In an LSTM
are presented
Pradesh.

3.2. Performance Comparison

Comparing the proposed IOF with the default optimizer Adam using different dee
Computers 2023, 12, 10 12 of 19

Table 2. Comparison of various deep learning models for crop yield prediction (million tons per
hectare) from 2010 to 2020.

CNN+ LSTM+ CNNLSTM+ GRU+

Year IOFLSTM Observed
Adam Adam Adam Adam
2010 215.13 215.34 217.17 216.18 218.06 218.11
2011 246.44 245.56 244.33 245.43 244.44 244.49
2012 259.22 255.67 258.75 258.25 259.20 259.29
2013 257.18 258.67 256.86 256.16 256.56 257.12
2014 162.41 163.44 164.95 166.45 164.05 165.05
2015 250.54 252.76 252.16 251.56 252.07 252.02
2016 253.64 255.78 253.75 253.65 251.22 251.54
2017 272.15 274.54 275.05 274.15 274.67 275.11
2018 281.58 283.87 286.04 286.54 284.51 285.01
2019 282.59 283.22 284.16 284.56 285.31 285.21
2020 291.29 292.32 295.03 294.23 295.17 295.67

Table 3. Comparison of various optimizers for crop yield prediction (million tons per hectare) from
2010 to 2020.

LSTM+ LSTM+ LSTM+ LSTM+ LSTM+

Year IOFLSTM Observed
SGD AdaGrad RMSP Adam Momentum
2010 214.45 216.54 212.85 215.34 214.05 218.06 218.11
2011 244.85 247.84 247.42 245.56 246.91 244.44 244.49
2012 256.29 252.89 254.26 255.67 257.29 259.20 259.29
2013 257.20 254.25 256.18 258.67 259.92 256.56 257.12
2014 166.84 167.68 164.29 163.44 165.08 164.05 165.05
2015 250.30 259.29 254.90 252.76 253.20 252.07 252.02
2016 252.20 257.52 250.04 255.78 250.23 251.22 251.54
2017 273.29 272.19 272.82 274.54 273.91 274.67 275.11
2018 282.19 289.39 281.02 283.87 284.61 284.51 285.01
2019 282.84 280.27 283.19 283.22 282.09 285.31 285.21
2020 290.73 297.85 291.94 292.32 293.75 295.17 295.67

Table 4. Comparison of performance statistics for major crops in Andhra Pradesh.

Crop Metrics RNN LSTM GRU CNN IOFLSTM

MAE 0.917 0.878 0.893 1.005 0.802
Paddy RMSE 1.243 1.211 1.221 1.868 1.145
r 0.983 0.972 0.984 0.964 0.992
MAE 1.320 1.297 1.319 1.457 0.951
Red gram RMSE 3.036 3.039 3.044 3.545 2.879
r 0.983 0.983 0.983 0.977 0.993
Computers 2023, 12, 10 13 of 19

Table 4. Cont.

Crop Metrics RNN LSTM GRU CNN IOFLSTM

MAE 1.531 1.522 1.526 1.549 0.912
Sugarcane RMSE 2.035 2.045 2.032 2.065 1.948
r 0.981 0.981 0.981 0.981 0.995
MAE 1.669 1.737 1.662 1.678 0.954
Cereals RMSE 2.248 2.332 2.242 2.273 2.045
r 0.987 0.986 0.987 0.986 0.995
MAE 0.053 0.052 0.053 0.054 0.051
Pulses RMSE 0.074 0.074 0.074 0.077 0.069
r 0.978 0.978 0.978 0.976 0.995
MAE 1.198 1.157 1.899 1.937 0.936
Groundnut RMSE 2.530 2.480 2.532 2.570 2.315
r 0.990 0.990 0.990 0.989 0.997
MAE 1.228 1.211 1.226 1.341 0.994
Chilli RMSE 1.604 1.584 1.604 2.255 1.513
r 0.983 0.983 0.983 0.967 0.993

The comparisons between district (region) wise performance metrics for the proposed
IOFLSTM with similar state-of-the-art frameworks such as LSTM and GRU are presented
in Table 5. The performance of the IOFLSTM is better for all the metrics and gives accurate
predictions compared to the standard LSTM and GRU.

Table 5. Performance statistics of 13 states of AP using IOFLSTM.

LSTM GRU IOFLSTM

District
MAE RMSE r MAE RMSE r MAE RMSE r
Anantapur 1.96 5.23 0.90 1.89 6.78 0.92 1.82 3.11 0.93
Chittoor 1.82 4.24 0.91 1.76 3.25 0.90 1.53 1.08 0.95
East
1.24 2.24 0.99 1.89 2.78 0.99 1.03 2.43 0.98
Godavari
Guntur 1.46 4.25 0.99 1.31 3.67 0.90 1.27 2.22 0.97
Kadapa 1.56 1.46 0.92 1.43 1.01 0.91 1.31 2.31 0.96
Krishna 1.37 2.42 0.91 1.32 2.31 0.92 1.13 1.26 0.95
Kurnool 1.56 4.25 0.94 1.49 3.93 0.93 1.35 2.63 0.94
Nellore 1.82 5.61 0.91 1.76 5.32 0.91 1.49 2.42 0.94
Prakasam 1.49 5.63 0.90 1.63 4.63 0.91 1.42 2.45 0.96
Srikakulam 1.40 5.92 0.91 1.31 5.72 0.90 1.05 2.15 0.98
Visakhapatnam 1.37 5.21 0.92 1.43 4.96 0.91 1.18 2.73 0.91
Vijayanagaram 1.93 8.35 0.94 1.21 7.92 0.93 1.34 3.32 0.99
West
1.37 6.25 0.92 1.95 5.84 0.795 1.16 4.28 0.99
Godavari
 Computers 2023, 12, 10 14 of 19

The time series comparison of the proposed model with the different state-of-the-art
PEER REVIEW 14 of it18
techniques for crop yield prediction are presented in Figure 7. From these figures, is
clear that the IOFLSTM model was the closest to the observed rainfall data (in blue). The
outperformance of IOFLSTM methods in the prediction of crop yield is acquired.

Figure 7. Comparison of the [Link]

Figure Comparisonstate‐of‐the‐art deep learning
of the proposed state-of-the-art models
deep learning for for
models crop
cropyield predic‐
yield prediction
(million tons per hectare)
tion (million tons per hectare) from 2015 to 2021. from 2015 to 2021.

Comparison of the proposed models with the different statistical measures is presented
Comparison of intheTable
proposed models with
6. The performance the different
comparison statistical
of major crops in AP withmeasures is pre‐
different measures
(Multivariate Logistic Regression (MLR), Support Vector Regression (SVR), Analysis of
sented in Table 6. The performance comparison of major crops in AP with different
varainace (ANOVA), Partial Leasr Squares Regression (PLSR)) and our proposed IOFLSTM
measures (Multivariate
usingLogistic Regression
the statistical parameters (MLR), Support
are presented in TableVector
6. In an Regression
MLR model, the (SVR),
RMSE
Analysis of varainace (ANOVA), Partial Leasr Squares Regression (PLSR)) and our pro‐
values range from 0.071 to 2.321, ‘r’ values range from 0.989 to 0.994, and MAE values
posed IOFLSTM using range from
the 0.054 to 2.321
statistical during the model
parameters are training.
presentedThe PLSR exhibits
in Table [Link]
Inworst perfor-
an MLR
mance, with RMSE values ranging from 0.089 to 3.321, ‘r’ values ranging from 0.964 to
model, the RMSE values range
0.977, MAE from
values 0.071from
ranging to 2.321,
0.074 to ‘r’ values
2.684, range
and, MASE from
values 0.989from
ranging to 0.836
0.994,
to
and MAE values range from
1.245 0.054
during to 2.321
the model during
training. the modeloutcomes
Experimental training. The PLSR
provide evidence exhibits
that the
IOFLSTM
the worst performance, with outperforms
RMSE values othersranging
and produced
fromaccurate
0.089 prediction numbers.
to 3.321,.‘r’ values ranging
To validate the performance of the proposed optimizer, the dataset is decomposed into
from 0.964 to 0.977, MAE
trainingvalues ranging
and testing [Link] 0.074 dataset
The training to 2.684,
helpsand, MASE
to carry out thevalues ranging
model training and
from 0.836 to 1.245 during the model training. Experimental outcomes provide evidence
the model test is performed by testing the dataset. The accuracy and loss of both training
and testing datasets
that the IOFLSTM outperforms othersforand
different optimizing
produced algorithms
accurate can be visualized
prediction in Figure 8. The
numbers.
model outcomes of crop yield prediction implemented by IOFLSTM (in brown) produced
the best accuracy with the minimum loss. The outperformance of IOFLSTM methods in the
Table 6. Comparison ofprediction
various ofstatistical
crop yieldmeasures
is [Link] crop yield prediction (million tons per hec‐
tare) from 2010 to 2020.

Metrics MLR ANOVA SVR PLSR IOFLSTM

MAE 0.907 0.956 0.914 1.005 0.802
RMSE 1.149 1.225 1.151 1.869 1.145
r 0.989 0.971 0.987 0.964 0.992
MASE 0.682 0.977 0.755 1.245 0.215
MAE 2.223 2.257 2.367 2.684 0.951
RMSE 2.890 3.134 2.896 3.321 2.879
r 0.990 0.981 0.983 0.973 0.993
MASE 0.894 1.153 0.986 1.198 0.455
MAE 1.528 1.548 1.530 1.569 0.912
Computers 2023, 12, 10 15 of 19

Table 6. Comparison of various statistical measures for crop yield prediction (million tons per hectare)
from 2010 to 2020.

Crop Metrics MLR ANOVA SVR PLSR IOFLSTM

MAE 0.907 0.956 0.914 1.005 0.802
RMSE 1.149 1.225 1.151 1.869 1.145
Paddy
r 0.989 0.971 0.987 0.964 0.992
MASE 0.682 0.977 0.755 1.245 0.215
MAE 2.223 2.257 2.367 2.684 0.951
RMSE 2.890 3.134 2.896 3.321 2.879
Red gram
r 0.990 0.981 0.983 0.973 0.993
MASE 0.894 1.153 0.986 1.198 0.455
MAE 1.528 1.548 1.530 1.569 0.912
RMSE 1.951 2.138 1.959 2.283 1.948
Sugarcane
r 0.991 0.985 0.989 0.971 0.995
MASE 0.377 0.797 0.576 0.927 0.235
MAE 1.659 1.678 1.662 1.689 0.954
RMSE 2.248 2.332 2.242 2.373 2.045
Cereals
r 0.992 0.981 0.989 0.970 0.995
MASE 0.424 0.736 0.515 0.836 0.314
MAE 0.054 0.072 0.056 0.074 0.051
RMSE 0.071 0.083 0.074 0.089 0.069
Pulses
r 0.993 0.981 0.990 0.974 0.995
MASE 0.368 0.637 0.396 0.945 0.205
MAE 1.841 1.863 1.845 1.945 0.936
RMSE 2.321 2.526 2.327 2.627 2.315
Groundnut
r 0.994 0.984 0.990 0.979 0.997
MASE 0.473 0.737 0.516 0.978 0.344
MAE 1.208 1.236 1.210 1.348 0.994
RMSE 1.527 1.670 1.531 1.738 1.513
Chilli
r 0.990 0.983 0.988 0.977 0.993
MASE 0.357 0.583 0.389 0.847 0.248
 into training and testing portions. The training dataset helps to carry out the model train‐
ing and the model test is performed by testing the dataset. The accuracy and loss of both
training and testing datasets for different optimizing algorithms can be visualized in Fig‐
ure 8. The model outcomes of crop yield prediction implemented by IOFLSTM (in brown)
produced the best accuracy with the minimum loss. The outperformance of IOFLSTM
Computers 2023, 12, 10 16 of 19
methods in the prediction of crop yield is acquired.

Figure 8. Comparison
Figure 8. Comparisonofofthe
theproposed withthe
proposed with thestate-of-the-art
state‐of‐the‐art optimizers.
optimizers.

4. Discussion
4. Discussion
The present study proposed an improved optimizer on LSTM to predict the crop yield
The present study proposed an improved optimizer on LSTM to predict the crop
that depends on the monsoon rainfall in Andhra Pradesh. Using the essential features,
yield that depends
predicted on crops
the Kharif the monsoon rainfall
in Andhra Pradeshin Andhra Pradesh.
like paddy, Usingred
sugar cane, thegram,
essential
and fea‐
tures,
maize. This study used rainfall parameters only to predict crop yield. We exhibited thegram,
predicted the Kharif crops in Andhra Pradesh like paddy, sugar cane, red
IOFLSTM algorithm for country-level crop yield prediction based on the historical rainfall
data collected from the Andhra Pradesh government websites.
Some studies implemented both CNN and RNN to forecast crop yield by considering
the climate [24]. This was achieved by implementing the Gaussian process component into
CNN and LSTM and the Caffe-based deep learning regression model to forecast crop yield
at the country level [25,26]. Other data-centric deep learning modeling techniques such as
the CNN model implementation from RGB data can effectively forecast crop yield [27,28].
The novelty of the proposed algorithm by coupling LSTM with IOF optimization has
produced a tremendous performance compared with other existing optimizers. For instance,
the authors of [29] have implemented the simple LSTM model to predict the tomato crop
and ficus Benjamina stem growth by considering the environmental parameters. With
the implementation of CNN, RNN models forecast the crop yield at the country level by
considering the spatial parameters and the phenological characteristics [30].
Computers 2023, 12, 10 17 of 19

Our study is in line with the study by Japneet Kaur et al. [31] that used the state-level
dataset of the Indian major crop paddy, wheat, sugarcane, and cotton from 2004 to 2013.
We considered various climatic conditions from the seven states and analyzed the impact
of climate change on food security and Indian agriculture. In deep learning, optimization
is a process that iteratively reduces error and improves accuracy. Overfitting is a significant
issue in deep neural networks, and there are various optimization techniques to improve
the performance. So, including optimization modeling reduces the loss and processing
time by fitting the function appropriately on the training data and avoiding overfitting.
This research proposed an improved optimization function to predict crop yield so that it
reduces forecasting errors by tuning the hyperparameters in the deep learning models [32]
and finding the optimal values for the hyperparameters [33].
Furthermore, to extend the proposed model, we would like to train our model not
only with a greater number of training data but also to increase the diversity concerning
the types of crops, thereby enabling our research to perform better forecasts among all
seasons. This will make our model more generic, enabling the automated task to work at
its best in different farming regions. Future work will be on fuzzy-based deep learning
on crop yield [34] and will improve the classifier as proposed in [35]. This scope of future
work will focus on the advancement of tools that can be used in the field where there are
major challenges, which in turn leads us to anticipate the necessity for better segmentation
algorithms. In addition, a detailed evaluation of the features extracted from different pre-
trained models of the deep learning platforms is needed, as well as gauging the excellence
of these features. Additionally, the model architecture will be further advanced toward
upgrading the performance of the prediction model.

5. Conclusions
The proposed work can be beneficial to evade difficulties and get the best yield from
the rainfall and climatic parameters. The performance metrics with IOFLSTM achieved r of
0.48, RMSE of 2.19, and MAE of 25.4, with the best performance in the crop yield forecast.
We compared the proposed optimizer IOF with eight standard methods of learning. The
comparison reveals that the training error is small in IOF and handles the underfitting
and overfitting issues. IOF gradually decreases the error to the minimum values and
updates the parameters, and IOF effectively performs best for higher learning rates on a
high dimensional dataset and converges to the global minimum. IOF works efficiently
for the high dimensional dataset and converges for the specified number of epochs with
high learning rates with the slightest error. We compared the proposed optimizer IOF with
the standard optimizers, proving that it is suitable for predicting crop yield, and it avoids
overfitting by efficiently managing the bias-variance tradeoffs.

Author Contributions: Conceptualization, U.B. and G.B.; methodology, U.B.; software, N.C.; vali-
dation, U.B. and G.B.; formal analysis, U.B.; investigation, U.B.; resources, G.B.; data curation, U.B.;
writing—original draft preparation, U.B.; writing—review and editing, U.B. and G.B.; visualization,
U.B. All authors have read and agreed to the published version of the manuscript.
Funding: This research received no external funding.
Data Availability Statement: Not applicable.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Wang, E.; Martre, P.; Zhao, Z.; Ewert, F.; Maiorano, A.; Rötter, R.P.; Kimball, B.; Ottman, M.; Wall, G.; White, J.; et al. The
uncertainty of crop yield projections is reduced by improved temperature response functions. Nat. Plants 2017, 3, 17102.
[CrossRef] [PubMed]
2. Annual Report 2020–21, Department of Agriculture, Cooperation and Farmers Welfare, Ministry of Agriculture and Farmers
Welfare, Government of India. Available online: [Link] (accessed on 23 May 2021).
3. Sharifi, A. Yield prediction with machine learning algorithms and satellite images. J. Sci. Food Agric. 2020, 101, 891–896. [CrossRef]
[PubMed]
Computers 2023, 12, 10 18 of 19

4. Obsie, E.Y.; Qu, H.; Drummond, F. Wild blueberry yield prediction using a combination of computer simulation and machine
learning algorithms. Comput. Electron. Agric. 2020, 178, 105778. [CrossRef]
5. Wu, R.; Yan, S.; Shan, Y.; Dang, Q.; Sun, G. Deep image: Scaling up image recognition. arXiv 2015, arXiv:1501.02876.
6. Miotto, R.; Li, L.; Dudley, J.T. Deep learning to predict patient future diseases from the electronic health records. In European
Conference on Information Retrieval; Springer: Cham, Switzerland, 2016; pp. 768–774.
7. Cai, Y.; Guan, K.; Peng, J.; Wang, S.; Seifert, C.; Wardlow, B.; Li, Z. A high-performance and in-season classification system of
field-level crop types using time-series Landsat data and a machine learning approach. Remote Sens. Environ. 2018, 210, 35–47.
[CrossRef]
8. Khaki, S.; Wang, L. Crop Yield Prediction Using Deep Neural Networks. Front. Plant Sci. 2019, 10, 621. [CrossRef] [PubMed]
9. Ehret, D.L.; Hill, B.D.; Helmer, T.; Edwards, D.R. Neural network modeling of greenhouse tomato yield, growth and water use
from automated crop monitoring data. Comput. Electron. Agric. 2011, 79, 82–89. [CrossRef]
10. Gholipoor, M.; Nadali, F. Fruit yield prediction of pepper using artificial neural network. Sci. Hortic. 2019, 250, 249–253. [CrossRef]
11. Salazar, R.; López, I.; Rojano, A.; Schmidt, U.; Dannehl, D. Tomato yield prediction in a semi-closed greenhouse. In Proceedings
of the XXIX International Horticultural Congress on Horticulture: Sustaining Lives, Livelihoods and Landscapes (IHC2014),
Brisbane, Australia, 17 August 2014; pp. 263–270, 1107.
12. Elavarasan, D.; Vincent, P.M.D. Crop Yield Prediction Using Deep Reinforcement Learning Model for Sustainable Agrarian
Applications. IEEE Access 2020, 8, 86886–86901. [CrossRef]
13. Sun, J.; Di, L.; Sun, Z.; Shen, Y.; Lai, Z. County-Level Soybean Yield Prediction Using Deep CNN-LSTM Model. Sensors 2019,
19, 4363. [CrossRef] [PubMed]
14. Chen, Y.; Lee, W.S.; Gan, H.; Peres, N.; Fraisse, C.; Zhang, Y.; He, Y. Strawberry Yield Prediction Based on a Deep Neural Network
Using High-Resolution Aerial Orthoimages. Remote Sens. 2019, 11, 1584. [CrossRef]
15. Russello, H.; Wenling, S. Convolutional Neural Networks for Crop Yield Prediction Using Satellite Images; IBM Center for Advanced
Studies: Armonk, NY, USA, 2018.
16. Kulkarni, S.; Mandal, S.N.; Sharma, G.S.; Mundada, M.R.; Meeradevi. Predictive Analysis to Improve Crop Yield using a
Neural Network Model. In Proceedings of the 2018 International Conference on Advances in Computing, Communications and
Informatics (ICACCI), Bangalore, India, 19–22 September 2018; pp. 74–79.
17. Jiang, Z.; Liu, C.; Hendricks, N.P.; Ganapathysubramanian, B.; Hayes, D.J.; Sarkar, S. Predicting county level corn yields using
deep long short term memory models. arXiv 2018, arXiv:1805.12044.
18. Haider, S.A.; Naqvi, S.R.; Akram, T.; Umar, G.A.; Shahzad, A.; Sial, M.R.; Khaliq, S.; Kamran, M. LSTM Neural Network Based
Forecasting Model for Wheat Production in Pakistan. Agronomy 2019, 9, 72. [CrossRef]
19. Indian Metrological Department. Available online: [Link] (accessed on 7 May 2021).
20. Central Pollution Control Board. Available online: [Link] (accessed on 7 May 2021).
21. India Crop Production—State Wise. Available online: [Link] (accessed on 10 May 2021).
22. Zhang, Z. Improved Adam Optimizer for Deep Neural Networks. In Proceedings of the 2018 IEEE/ACM 26th International
Symposium on Quality of Service (IWQoS), Banff, AB, Canada, 4–6 June 2018; pp. 1–2. [CrossRef]
23. Hochreiter, S.; Schmidhuber, J. Long short-term memory. Neural Comput. 1997, 9, 1735–1780. [CrossRef] [PubMed]
24. Ma, J.W.; Nguyen, C.H.; Lee, K.; Heo, J. Regional-scale rice-yield estimation using stacked auto-encoder with climatic and MODIS
data: A case study of South Korea. Int. J. Remote Sens. 2019, 40, 51–71. [CrossRef]
25. You, J.; Li, X.; Low, M.; Lobell, D.; Ermon, S. Deep gaussian process for crop yield prediction based on remote sensing data. In
Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence, San Francisco, CA, USA, 4–9 February 2017.
26. Kuwata, K.; Shibasaki, R. Estimating crop yields with deep learning and remotely sensed data. In Proceedings of the 2015 IEEE
International Geoscience and Remote Sensing Symposium (IGARSS), Milan, Italy, 26–31 July 2015; pp. 858–861.
27. Nevavuori, P.; Narra, N.; Lipping, T. Crop yield prediction with deep convolutional neural networks. Comput. Electron. Agric.
2019, 163, 104859. [CrossRef]
28. Yang, Q.; Shi, L.; Han, J.; Zha, Y.; Zhu, P. Deep convolutional neural networks for rice grain yield estimation at the ripening stage
using UAV-based remotely sensed images. Field Crop. Res. 2019, 235, 142–153. [CrossRef]
29. Alhnaity, B.; Pearson, S.; Leontidis, G.; Kollias, S. Using deep learning to predict plant growth and yield in greenhouse
environments. In Proceedings of the International Symposium on Advanced Technologies and Management for Innovative
Greenhouses: GreenSys2019, Angers, France, 16–20 June 2019; pp. 425–432, 1296.
30. Shi, X.; Chen, Z.; Wang, H.; Yeung, D.Y.; Wong, W.K.; Woo, W.C. Convolutional LSTM network: A machine learning approach for
precipitation nowcasting. In Proceedings of the Advances in Neural Information Processing Systems, Montreal, QC, Canada,
7–12 December 2015; p. 28.
31. Kar, S.; Das, N. Climate Change, Agricultural Production, and Poverty in India. In Poverty Reduction Policies and Practices in
Developing Asia. Economic Studies in Inequality, Social Exclusion and Well-Being; Heshmati, A., Maasoumi, E., Wan, G., Eds.; Springer:
Singapore, 2015. [CrossRef]
32. Yang, J.; Yang, G. Modified Convolutional Neural Network Based on Dropout and the Stochastic Gradient Descent Optimizer.
Algorithms 2018, 11, 28. [CrossRef]
33. Bello, I.; Zoph, B.; Vasudevan, V.; Le, Q.V. Neural optimizer search with reinforcement learning. In Proceedings of the International
Conference on Machine Learning, PMLR, Sydney, Australia, 6–11 August 2017; pp. 459–468.
Computers 2023, 12, 10 19 of 19

34. Elavarasan, D.; Durai Raj Vincent, P.M. Fuzzy deep learning-based crop yield prediction model for sustainable agronomical
frameworks. Neural Comput. Appl. 2021, 33, 13205–13224. [CrossRef]
35. Elavarasan, D.; Vincent, P.M.D.R. A reinforced random forest model for enhanced crop yield prediction by integrating agrarian
parameters. J. Ambient. Intell. Humaniz. Comput. 2021, 12, 10009–10022. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.