Survey on Application of Intelligent Agriculture Based on Deep Belief
Network
Yuanyuan Zhang
1*
, Shengwei Shi
2
1
College of Computer and Information Engineering, Beijing Agricultural University, Beijing100096, China
2
College of Biological Resources and Environment, Beijing Agricultural University, Beijing100096, China
Keywords: DBN, Deep Learning, Intelligent Agricultural.
Abstract: With the improvement of modern technology, agricultural development continues to be precise, refined and
intelligent. The traditional model has low recognition accuracy. Deep belief network (DBN) has strong
feature extraction and learning ability. Based on the principle of transmission DBN network, this paper
gives the application research progress of the model in intelligent agriculture from the perspective of DBN
improvement direction, and finally discusses the challenges and new directions of DBN in the application
field of intelligent agriculture.
1 INTRODUCTION
With the rapid development of Internet of things,
big data, artificial intelligence and other
technologies, agricultural development is moving
towards precision, refinement and intelligence (Guo
2019). These agriculture data have the
characteristics of large volume, high authenticity,
fast generation speed and many data types (Guo et
al. 2019). The traditional methods include linear
polarization, wavelet filtering and machine learning
models have poor generalization ability, low
recognition accuracy and instability
(LV, Li, et al.
2019).
In order to analyze the growth status of various
crops and animals and monitor the agricultural
growth environment, deep learning theory is widely
used. Deep belief network (DBN) is a typical deep
learning model, which constructs a nonlinear deep-
seated network structure through self-learning, so as
to extract the high-level features of data and
accurately fit complex functions (Li et al. 2020). It
has been widely used in intelligent agriculture fields
such as crop classification, pest prediction, weed
identification, breeding monitoring and crop yield
prediction.
2 DBN THEORY
DBN is formed by superposition of multiple neurons,
and its constituent element is restricted Boltzmann
machines (RBMs), which belongs to a probability
generation model (Liu, Wang, et al. 2018). By
training the weights between its neurons, the whole
neural network generates training data according to
the maximum probability. As shown in Figure 1, it
is a three-tier DBN model.
Figure 1: The structure of DBN.
Zhang, Y. and Shi, S.
Survey on Application of Intelligent Agriculture Based on Deep Belief Network.
DOI: 10.5220/0011757800003607
In Proceedings of the 1st International Conference on Public Management, Digital Economy and Internet Technology (ICPDI 2022), pages 739-744
ISBN: 978-989-758-620-0
Copyright
c
2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)
739
2.1 RBM Structure and Working
Principle
RBM is composed of one layer of dominant neurons
and one layer of recessive neurons, and the two
layers of neurons are fully connected in two
directions, so it is also called neural perceptron. The
structure is shown in Figure 2.
Figure 2: The structure of RBM.
An RBM contains m visual units and n hidden
units. For a given set of states (v, h), the energy
possessed by RBM as a system is defined as (HAN
etc. 2017):
𝐸
𝑣, ℎ
|
𝜃
∑
𝑎
𝑣
∑
𝑏
ℎ
∑∑
𝑣
𝑊
ℎ
(1)
Where θ𝑊
, 𝑎
, 𝑏
is the parameter of RBM,
𝑊
is the connection weight between the visible
unit i and the hidden unit j, 𝑎
represents the paranoia
of the visible unit i, and j represents the bias of the
hidden unit. When the parameters θ are determined,
the joint probability distribution of (v, h) can be
obtained based on formula:
P
v, h
|
θ
𝑣, ℎ
𝜃
(2)
Z
θ
∑
𝑒
𝑣, ℎ
𝜃
,
(3)
Z
θ
is the normalization factor.
Because RBM has a special structure of
connection between layers and no connection within
layers, when the visible cell state is given, the
activation states of each hidden cell are
conditionally independent. Therefore, the activation
probability of the hidden unit is:
Pℎ
1v, θ σ𝑏
∑
𝑣
𝑊
(4)
σ
x
Sigmoid activation function.
The maximum likelihood estimation method is
used to maximize the above formula to obtain the
RBM parameters θ, and then the contrast divergence
algorithm is used to obtain the RBM parameters θ
𝑊
, 𝑎
, 𝑏
. The update rules are as follows:
∆𝑤
𝜀
〈
𝑣
ℎ
〉
ℎ
𝑣
〈
𝑣
ℎ
〉
(5)
∆𝑏
𝜀
〈
𝑣
〉
ℎ
𝑣
〈
𝑣
〉
(6)
∆𝑎
𝜀
〈
ℎ
〉
ℎ
𝑣
〈
ℎ
〉
(7)
Where, ε is the learning rate,
〈
∙
〉
|
represents
the limit of partial derivative function under P(h|v)
distribution,
〈
∙
〉
represents the limit of the
partial derivative function under the distribution of
the reconstructed model.
3 APPLICATION OF DBN IN
INTELLIGENT AGRICULTURE
The application of DBN in intelligent agriculture
generally includes three processes: application data
acquisition, feature extraction, recognition and
prediction. Data acquisition: collect influencing
factors and other parameters according to the actual
needs of agriculture; Feature extraction mainly
completes the mapping from data space to feature
space by DBN unsupervised learning; Recognition
and prediction mainly realizes the transformation
from feature space to decision space through DBN
supervised learning. At present, the main research
directions focus on the direct application of DBN
and the improvement of network topology, learning
process and network parameters.
3.1 Direct Application
In order to improve the efficiency of maize breeding,
Article (Yu et al. 2019) is proposed a haploid
identification method based on DBN multi variety
mixed maize. Through the comparative analysis of
SVM, BPR and DBN, the recognition rate of DBN
method increased most significantly and the
recognition accuracy improved fastest; Aiming at
the problems of strong subjectivity and insufficient
accuracy of the existing benchmark land price
evaluation model, Article (Wang et al. 2018) is
proposed an agricultural land benchmark land price
evaluation method based on the idea of deep earning.
Through repeated experiments, the number of
visible and hidden layers and the number of neurons
in each layer were determined, and compared with
SVM and BP, The accuracy of the model is 6.76%
higher than that of the other two models when the
time consumption of the model is basically the same;
Article (Li et al. 2018) is carried out disease early
warning by monitoring pig cough sound, proposed a
method to identify pig cough sound based on deep
confidence network, and constructed a 5-layer pig
cough sound recognition model, with a total
ICPDI 2022 - International Conference on Public Management, Digital Economy and Internet Technology
740
recognition rate of 93.21%; Article (Wang et al.
2018) is proposed a depth belief network model
with two hidden layers to improve the accuracy of
wheat aphid prediction; Article (Guo et al. 2019) is
used Gaussian filtering to preprocess the number of
images, and designed a DBN containing three
hidden layer restricted Boltzmann machines for rice
sheath blight recognition, with an average
recognition accuracy of 94.05%. The specific
progress analysis is shown in Table 1.
3.2 Topology
The network structure of DBN determines the
change of network performance, and there is no
clear system for the construction of DBN model. On
the one hand, the current research is mainly based n
the standard structure optimization of DBN.
Coefficient DBN has been designed in the field of
intelligent agriculture, which is connected with other
neural networks such as convolution and self-
encoder; On the other hand, the combination with
other methods is reflected in the combination of
preprocessing method and feature extraction method
to design a new optimization model (sohu 2017).
Article (Liu 2018) is proposed a forest land change
detection method based on sparse DBN model,
which is, adding regular items to RBM to complete
the automatic classification of image spots, which
Table 1: Progress in direct application of DBN.
author Propose
metho
d
data sources input output Accuracy Insufficient
Yu
Yunhua
DBN
Collection of data from a
suburban experimental base
in Beijing
Haploid and
diploid grains
of 10 varieties
Corn seed
species
+6.92%
Strong dependence on
sample quality
Wang
Hua
DBN
Puning farmland
benchmark land price
evaluation standard s
y
ste
m
19 index
evaluation
factors
Predicted
land price
+6.76%
Subjective setting of hidden
layers and neurons
Li Xuan
DBN
School owned boutique pig
farm
The sound of
10 Landrace
Pigs Cough
recognition
-
Lack of comparative analysis
with other algorithms
Xiu Mei
Wang
DBN
Yantai plant protection
station and related data sets
in Shandong Province
mpact data
related to
wheat aphid
occurrence
Occurrence
degree of
wheat aphids
8.1%
Aphid data samples are few,
and the hidden layer is 2
Guo Dan
DBN
Collect the images of Rice
Sheath Blight in northern
cold area
Preprocessed
image data
Occurrence
degree of rice
sheath blight
3.65%
Manual experience value
setting of model parameters
Table 2: Progress in topology structure of DBN.
author Propose method data sources input output Accuracy Insufficient
Liu Run
dong
Sparse DBN
Correlation
speckle image
Forest land cover
index
Forest land and non
forest land
identification
-
practicability
needs to be
further verifie
d
Du Jiaxin
DBN+CNN
Multi place video
image data set
Forest smoke and
smoke-free image
Smoke and smokeless
identification
+10.03%
Convolution
kernel is set
manuall
y
Zhu
Zhihui
PCA+DBN
Jingzhou yukou
poultry industry
Co., Lt
d
180 pieces of
Jingfen No. 1
male and female
classification
+20%
The training time
of DBN model is
lon
g
Zhou
Xiangyu
PCA +
wavelet+DBN
Pond 2-6 of
balidian
ex
p
erimental base
16 pond
environmental
data
Ammonia nitrogen
content
MAPE
-0.013 9
High calculation
time cost
Ying Yi
Chen
EMD+GRU+
DBN
Microclimate data
of a greenhouse in
Nanjing
3 kinds of
temperature
related monitoring
information
Temperature in the next
1H
+11.1%
Less feature
selection
Lu Wei EEMD+PCA+D
BN
Fluorescence
spectra of rice seed
soakin
g
solution
Lianjing 7 and
Wuyunjing
germination percentage 10%
optimization
cannot be
g
uarantee
d
Survey on Application of Intelligent Agriculture Based on Deep Belief Network
741
can effectively identify forest land and non-forest
land information in high-resolution remote sensing;
A deep confidence convolution network forest fire
image recognition model is proposed in article (Du
2020), which makes feature learning, feature
extraction and classifier reconstruction form a
DBN_ CNN model improves the probability of
forest fire smoke recognition; The machine vision
acquisition system is constructed in (Zhu et al.
2018). For the full information features of chicken
egg images, the principal component analysis and
deep belief network model are used to realize the
male and female recognition of three kinds of
chicken embryos, and the accuracy is as high as
83.33%;There are many influencing factors of
ammonia nitrogen content in pond culture
environment and the correlation is complex, (Chen
et al. 2019) through principal component analysis
PCA, the main factors affecting the change of
ammonia nitrogen content are selected as model
input, the noise is eliminated by wavelet threshold
method, and then the prediction is realized by DBN
network. Compared with the traditional model, the
average percentage error is significantly reduced;
The greenhouse is complex and changeable, the
representation ability of environmental factors is
low, and the learning time is long, which brings
inconvenience to the temperature prediction of
agricultural greenhouse (Zhou et al. 2019) A
greenhouse prediction method based on improved
depth belief network combined with empirical mode
decomposition and gated cycle unit is proposed; The
spectral data are decomposed into EEMD by
ensemble empirical mode in (Lu 2018), and the
dimensionality is reduced by PCA through principal
component analysis. The prediction model of
germination rate based on DBN is established to
predict the germination rate of rice seeds
nondestructively, which has high accuracy. It’s
shown in table 2.
3.3 Learning Process
The learning process of DBN includes two stages:
forward unsupervised training and reverse
supervised fine tuning of RBM. Optimizing the
number and structure of RBM stack or improving
the reverse fine tuning algorithm are the two main
research hotspots. Article (Zhou et al. 2019)
introduces glia to improve the depth belief network.
Each neuron in the hidden layer will be connected
with a glia to predict the greenhouse temperature,
and verified its effectiveness in the complex time
series environment of the greenhouse; Article
(Zhang et al. 2017) added a priori information of
winter jujube diseases and pests into RBM based on
the improved deep belief network, and the
prediction accuracy was improved by more than 20
percentage points through experimental comparison
(Xu et al. 2017) proposed a DBN-LSSVR model
based on deep belief network fusion least squares
support vector regression machine to predict the
dissolved oxygen content, which has high prediction
Table 3: Progress in train learning of DBN.
author Propose method data sources input output Accuracy Insufficient
Zhou
Xiang
yu
Glial RBM Pond 2-6 of balidian
experimental base
16 pond
environmental
data
Ammonia
nitrogen
content
MAPE
-0.013 9
High calculation
time cost
Zhang
Shan
wen
Prior
information
RBM
Winter jujube planting
base in Dali County
Soil,
meteorological
and pest
information
Classification of
5 diseases and
insect pests
+21% Manual
experience setting
of model
p
arameters
Xu Long
Qin
DBN+LSSVR A pond in Haiou Island
shrimp culture base,
Panyu District,
Guan
g
zhou
6 breeding
environmental
factors
Predicted value
of dissolved
oxygen
MAPE
-8.84%
Parameter
adaptive learning
can be improved
Yu an
Hong
chun
ARIMA+DBN Construction data of
aquatic product
traceability and safety
early warning platfor
m
Historical data of
hydrolyzed
oxygen
Future forecast MAPE
-7.2%
Time series data
can be
preprocessed
Li
Jialang
Cross entropy +
DBN
Yam ani shi
Proposed dataset
Drug protein Protein
sequence
similarity
+5% Practical
application
scenarios can be
considere
d
ICPDI 2022 - International Conference on Public Management, Digital Economy and Internet Technology
742
Table 4: Progress in network parameter of DBN.
author Propose method data sources input output Accuracy Insufficient
Deng
Xiang
wu
DBN hidden
layer node
determination
method
Paddy field of agricultural
experimental base of
Jiangmen Institute of
Agricultural Sciences
Rice field weed
image data
Weed category - Lack of
comparative
analysis with
other algorithms
Xian
Feng
Wan
g
Adaptive
learning rate
DBN
Data of 10 cotton planting
demonstration bases in
Dali Count
y
, Weinan Cit
y
12 kinds of
environmental
information data
Classification of
5 diseases and
insect
p
ests
+2.8% The calculation
efficiency needs
to be im
p
rove
d
Min
Min
Wu
PSW / GW
optimize DBN
initial wei
g
ht
University experimental
greenhouse
Lettuce sample
image
Lead content +11.1% High
computational
com
p
lexit
y
Pang
Qihua
Adaptive step
size DBN
Breiman experiment case Data of 12
environmental
factors
Fruit category +4.2% Time series data
can be
p
reprocesse
d
accuracy and generalization ability. Article (yuan et
al. 2017) the first mock exam and the deep belief
network model are analyzed. An ARIMA-DBN
model for predicting the water quality of
aquaculture is established. The mean square error is
significantly lower than that of the single model.
Article (Li et al. 2018) is proposed a drug protein
interaction prediction algorithm based on DBN,
which outputs the interaction probability in the
reverse fine-tuning stage, uses cross entropy as the
loss function and soft-max as the output. It’s shown
in table 3.
3.4 Network Parameters
The network parameters of DBN mainly include the
number of hidden layers, the number of neurons in
each layer and the value of learning rate. At present,
the determination of these parameters mainly
depends on a priori knowledge or repeated
experiments. Its disadvantages are large subjectivity
and high time complexity. Therefore, it is also the
research direction of DBN optimization. Article
(Deng et al. 2018) in rice seedling stage as the
research object, and optimized the number of DBN
nodes in double hidden layer by selecting three node
combination modes of constant value type,
appreciation type and value reduction
type. Through comparative analysis, they have a
higher recognition rate; (Wang et al. 2018) is
proposed a DBN training model with adaptive
learning rate in order to overcome the slow
convergence speed of DBN in the prediction of
cotton diseases and pests; Article (Wu et al. 2020) is
proposed to build a lead content prediction model of
lettuce leaves based on improved DBN algorithm. In
order to avoid falling into local optimization in
DBN training, PSO algorithm and GWO algorithm
are used to optimize the initial weight and paranoia
of DBN network, and the results have higher
stability. Article (Pang et al. 2019) is analyzed the
planting adaptability of tropical fruit trees and
proposed an ALS-DBN recognition model. The
adaptive step size algorithm and momentum term
are introduced, and the recognition accuracy is as
high as 97.72%. It’s shown in table 4.
4 CONCLUSION
The application of DBN in intelligent agriculture is
still in its infancy (Pang et al. 2020). This paper
introduces the network structure and training
process of DBN, and expounds its application in
crop classification, pest prediction, weed
identification, breeding monitoring and crop yield
prediction from the perspective of DBN algorithm
improvement and optimization. Future research can
be improved from the following aspects: ①improve
the topology of DBN and reduce the computational
complexity; ②Feature extraction models suitable
for different application scenarios; ③Expand the
application of DBN in more fields of intelligent
agriculture. DBN provides new ideas for the
development of smart agriculture and ecological
agriculture in the future.
ACKNOWLEDGMENTS
This work was financially supported by young
teachers' Scientific Research Fund Project of Beijing
University of Agriculture (17ZK007).
Survey on Application of Intelligent Agriculture Based on Deep Belief Network
743
REFERENCES
Application of deep machine learning combined with big
data in medical image analysis. July 2017
https://www.sohu.com/a/154356777_ one hundred
thousand six hundred and sixty-three.
Chen Yingyi et al. Study on ammonia nitrogen prediction
model of pond aquaculture water based on improved
depth belief network [J]. Journal of agricultural
engineering, 2019.35 (7): 96-101.
Du Jiaxin. Based on DBN_ CNN network forest fire risk
image recognition [D], Taiyuan University of
technology, June 2020.
Deng Xiangwu et al. Weed identification at seedling stage
of paddy field based on multi feature fusion and deep
belief network [J], Journal of agricultural engineering,
2018,34 (14): 165-168.
Guo Xiao. Research on Intelligent Agricultural Decision
System Based on machine learning algorithm [D].
Xi'an University of Electronic Science and technology.
2019.
Guo Xiangyun et al. Application and Prospect of deep
learning in field planting [J]. Journal of China
Agricultural University, 2019, 24 (1): 119-129.
Guo Dan et al. Research on identification method of rice
sheath blight based on deep belief network [J],
research on agricultural mechanization. 2019.12:42-46.
HAN Dongying etc. A new fault diagnosis method based
on deep belief network and support vector machine
with Teager-Kaiser energy operator for hearings[J].
Advances in Mechanieal Engineering,2017,9(12):1-11.
LV Shingling, Li Denghui, et al. Application and research
status of deep learning in agriculture in China [J].
Computer engineering and application, 2019, 55 (20):
24-26.
Li Shaobo et al. Overview of mechanical equipment fault
diagnosis research based on deep confidence network
[J]. Modern manufacturing engineering, 2020.10:156-
158.
Liu Fangyuan, Wang Shuihua, et al. Review of deep
confidence network model and application [J].
Computer engineering and application, 2018.54 (1):
11-14.
Li Xuan et al. Pig cough recognition based on deep belief
network [J]. Journal of agricultural machinery, 2018,
34 (21): 180-184.
Liu Rundong. Forest land change detection method based
on sparse DBN model in remote sensing images.
China, cn109635836a [P], November 9, 2018.
Lu Wei. Study on detection method of rice seed
germination rate based on fluorescence spectroscopy
and deep belief network [J], spectroscopy and spectral
analysis, 2018.38 (4): 1303-1010.
Li Jialang et al. Study on prediction of drug targeted
protein action based on DBN [D], South China
Agricultural University, 2018, 6.
Pang Qihua et al. Analysis on planting suitability of
tropical fruit trees based on als-dbn [D], Guangxi
University, and March 2019.
Pang Haitong et al. Overview of pest identification
technology based on deep learning [J], agricultural
engineering, 2020, 10 (10): 19-24.
Wang Hua et al. Agricultural land benchmark land price
evaluation model based on deep belief network [J],
Journal of agricultural engineering, 2018, 34 (21):
606-610.
Wang Xiumei et al. Prediction and early warning of wheat
aphids based on deep learning [J], Jiangsu agricultural
science, 2018,46 (5): 180-184.
Wang Xianfeng et al. Prediction method of cotton diseases
and insect pests based on performance improvement
depth belief network [J], Zhejiang Agricultural Journal,
2018,30 (10): 1790-1797.
Wu Minmin et al. Research on nondestructive detection of
lead content in lettuce leaves based on hyperspectral
image technology and DBN [D], Jiangsu University,
2020.5.
Xu Yi, Li Beibei, Song Wei. Research on improved deep
confidence network classification algorithm [J].
Computer science and exploration, 2019, 13 (4): 596-
607.
Xu LongQin et al. Prediction of dissolved oxygen in
Litopenaeus vannamei culture based on deep belief
network and least squares support vector regression
[J], Journal of agricultural engineering, 2017,30 (4):
1-4.
yuan Hongchun et al. Simulation Research on abnormal
optimization prediction of aquaculture water quality
[J], computer simulation, 2017,34 (12): 447-453.
Yu Yunhua et al. Study on qualitative identification
method of multi variety Maize Haploid based on deep
belief network [J], spectroscopy and spectral analysis,
2019.39 (3): 906-909.
Zhu Zhihui et al. Early chicken embryo male and female
recognition based on egg image blood line features
and deep belief network [J]. 2018.34 (6): 197-201.
Zhou Xiangyu et al. Temperature prediction method of
agricultural greenhouse based on improved depth
belief network [J]. 2019, 39 (4): 1053-1058.
Zhang Shanwen et al. Prediction model of diseases and
insect pests of greenhouse winter jujube based on
improved deep belief network [J], Journal of
agricultural engineering, 2017,33 (19): 202-205.
ICPDI 2022 - International Conference on Public Management, Digital Economy and Internet Technology
744
