Available via license: CC BY 4.0
Content may be subject to copyright.
September, 2022 Int J Agric & Biol Eng Open Access at https://www.ijabe.org Vol. 15 No. 5 187
Maize leaf disease identification using deep transfer convolutional neural
networks
Zheng Ma1, Yue Wang1, Tengsheng Zhang1, Hongguang Wang2, Yingjiang Jia1, Rui Gao3*,
Zhongbin Su1*
(1. Institute of Electrical and Information, Northeast Agricultural University, Harbin 150030, China;
2. Agricultural Products and Veterinary Drug Feed Technical Identification Station, Department of Agriculture and Rural Affairs of
Heilongjiang Province, Harbin 150090, China; 3. Key Laboratory of Northeast Smart Agricultural Technology, Ministry of Agriculture
and Rural Affairs, Heilongjiang Province, Harbin 150030, China)
Abstract: Gray leaf spot, common rust, and northern leaf blight are three common maize leaf diseases that cause great
economic losses to the worldwide maize industry. Timely and accurate disease identification can reduce economic losses,
pesticide usage, and ensure maize yield and food security. Deep learning methods, represented by convolutional neural
networks (CNNs), provide accurate, effective, and automatic diagnosis on server platforms when enormous training data is
available. Restricted by dataset scale and application scenarios, CNNs are difficult to identify small-scale data sets on mobile
terminals, while the lightweight networks, designed for the mobile terminal, achieve a better balance between efficiency and
accuracy. This paper proposes a two-staged deep-transfer learning method to identify maize leaf diseases in the field.
During the deep learning period, 8 deep and 4 lightweight CNN models were trained and compared on the Plant Village dataset,
and ResNet and MobileNet achieved test accuracy of 99.48% and 98.69% respectively, which were then migrated onto the field
maize leave disease dataset collected on mobile phones. By using layer-freezing and fine-tuning strategies on ResNet and
MobileNet, fine-tuned MobileNet achieved the best accuracy of 99.11%. Results confirmed that disease identification
performance from lightweight CNNs was not inferior to that of deep CNNs and transfer learning training efficiency was higher
when lacking training samples. Besides, the smaller gaps between source and target domains, the better the identification
performance for transfer learning. This study provides an application example for maize disease identification in the field
using deep-transfer learning and provides a theoretical basis for intelligent maize leaf disease identification from images
captured with mobile devices.
Keywords: maize leaf disease, deep learning, transfer learning, convolutional neural networks
DOI: 10.25165/j.ijabe.20221505.6658
Citation: Ma Z, Wang Y, Zhang T S, Wang H G, Jia Y J, Gao R, et al. Maize leaf disease identification using deep transfer
convolutional neural networks. Int J Agric & Biol Eng, 2022; 15(5): 187–195.
1 Introduction
Maize is an important food crop and industrial raw material
globally, and ensuring maize yield stability is of great importance
to food security, agricultural development, and the national
economy. Over ten kinds of common maize diseases directly
affect maize yield and quality, including in the leaves, ears, and
roots. Although gray leaf spots, common rust, and northern leaf
blight in the leaves can severely reduce maize yield, timely
identification and disposal lead to minimum harm caused by the
Received date: 2021-04-06 Accepted date: 2022-05-13
Biographies: Zheng Ma, PhD candidate, research interest: deep learning for
agricultural applications, agricultural remote sensing data processing, Email:
zavier_ma@outlook.com; Yue Wang, Under Postgraduate, research interest:
machine vision for agricultural applications and agricultural intelligent
equipment, Email:605714643@qq.com; Tengsheng Zhang, Master, research
interest: hyperspectral data processing and deep learning, Email:
153843193@qq.com; Hongguang Wang, Bachelor, Researcher, research
interest: maize and rice phenotypes, Email: 570229092@qq.com; Yinjiang Jia,
PhD, Associate Professor, research interest: hyperspectral data processing in
agriculture, Email: jiayinjiang@126.com.
*Corresponding author: Rui Gao, PhD, Lecturer, research interest:
hyperspectral data processing and crop phenotype inversion, Email:
415730327@qq.com; Zhongbin Su, PhD, Professor, research interest: smart
agriculture and big data in agriculture. Institute of Electrical and Information,
Northeast Agricultural University, Harbin 150030, China. Tel:
+86-13303609163, Email: suzb001@163.com.
disease. Traditional identification requires agricultural or forestry
experts to diagnose in the field or from a distance, which is quite
subjective, time-consuming, laborious, and inefficient. Therefore,
realizing an intelligent, rapid, and accurate automatic identification
method is of great significance.
Deep learning methods can be applied in hyperspectral
images[1] and RGB images. Identifying crop phenotypic diseases
with deep learning methods has become a strong research focus in
precision agriculture[2], especially by using Convolution Neural
Networks (CNN). CNN has diverse structures and offers
outstanding capabilities as a consequence of gradual optimization,
contributing to being the prevailing disease identification classifier
for large- or small-scale tasks. Inchoate researchers considered
CNNs as feature extractors, followed by machine learning
classifiers (mostly SVM)[3,4]. After CNNs were gradually used for
classification directly, small dataset size problem came into view.
In practice, low disease incidence and high cost of acquisition
result in only a few training data collected, which limits the
application of deep learning methods in identification[5].
Therefore, most studies in plant disease identification and detection
are based on the prevailing public dataset Plant Village[6], which
contains 38 categories by species and disease, adding up to 54 303
images. Jaiswal et al.[7] sampled 5 diseases of every species in
Plant Village to carry out their research, which focused on the
hyperparameters’ effect on GoogLeNet model performance.
188 September, 2022 Int J Agric & Biol Eng Open Access at https://www.ijabe.org Vol. 15 No. 5
Sravan et al[8] utilized 20 639 images from Plant Village on
ResNet50 model, which achieved 99.26% classification accuracy,
and Agarwal et al[9] selected 10 kinds of tomato diseases with
image preprocessing and brightness enhancement, reaching an
accuracy of 98.7% on the proposed simplified CNN model.
Besides, it is common to reassemble dataset with Plant Village and
real field images on specific species, with data amplification,
synthesis and generation followed. Liu et al.[10] used 4023 field
images and 3646 Plant Village images to generate 107 366 grape
leaf images for training the proposed DICNN model, with an
overall accuracy of 97.22%. Furthermore, investigated
approaches like structure modification, module enhancement and
data preprocessing (augmentation, segmentation or background
removal) have been validated to enhance the performances[11-14],
which can be seen as the solution to dataset size problem.
Although deep learning method mitigates manual
misjudgments, reliance on expert experience, and reduces
workforce and material resource requirements, the performances of
CNN models are based on intensive image computing processing,
which indicates a large amount of manual labeling work. Besides,
recent researches[15-17] showed that CNN methods lack robustness
to environmental conditions. Training datasets are mostly
collected under controlled conditions like laboratory background or
public datasets, which causes problems with generalization,
adaptability, and anti-interference capability[18]. Another
application issue of CNN in disease identification is the limited
amount of existing data for specific species, bringing about
unstable training processes and overfitting models. Therefore,
transfer learning methods are introduced to solve the above
challenges. Transfer learning leverages knowledge from the
source domain to offer solutions to the target domain, and stronger
similarity between source and target domains improves accuracy,
indicating better transferability[19]. The heavy workload of
manual annotation is reduced greatly on account of source domain
training, denoting high resource utilization. And the overfitting
problem caused by limited existing data is solved simultaneously,
since a fully trained CNN model on source dataset provides good
feature extraction capacity which satisfies requirements on smaller
targeted datasets. Xu et al.[20] adopted transfer learning to solve
overfitting as well as a replacement of fully-connected layer into
global pooling layer, which achieved 93.28% accuracy on a maize
disease identification dataset, superior to four previous states of the
art models.
Pre-trained CNN models based on the ImageNet dataset are
commonly employed in transfer learning, making full use of their
feature extracting or fine-tuning in follow-up processing on another
public or local disease dataset. Chen et al.[21] selected a CNN
model pre-trained on ImageNet with initialized weights to achieve
at least 91.83% validation accuracy on a public dataset. Average
accuracy reached 92.00% on rice plant images with complex
backgrounds. Yin et al.[22] extracted deep features for pepper
disease and insect pest dataset using 8 CNN models pre-trained on
ImageNet for identification, achieving 85.6% and 93.62% accuracy
for disease and insect pest identification, respectively, using
ResNet depth features. However, still unresolved are the lacking
robustness to environmental conditions, which is more of dataset
problem, and the less similarity between domains, indicating a
more similar source dataset. Besides, high CNN model accuracy
depends on computing power supported by high-performance
hardware, whereas the general trend is toward lightweight and
mobile agricultural equipment. Several recent studies have
imported lightweight CNNs, such as MobileNet[23,24] and
EfficientNet[25], for crop disease identification. Concessions must
be made by accuracy against the network scale.
Therefore, a more feasible deep-transfer learning method is
proposed in this paper. By training and comparing deep and
lightweight CNNs on the Plant Village public dataset, optimal
pre-trained models are transferred onto a maize leaf diseases
dataset collected in a real field and optimized by fine-tuning, which
solves the challenges mentioned above: manual annotation, model
robustness, small dataset, domain similarity, and mobile simulation.
In summary, this study provides an application example for maize
disease identification under complex (field) background with deep
transfer learning, which provides a theoretical basis for intelligent
in-field identification using mobile terminal devices.
2 Materials and methods
2.1 Dataset
This experiment was divided into two parts: training on the
Plant Village public dataset and transfer learning on the local maize
leaf diseases dataset using pre-trained models in part one. The
Plant Village dataset contains 39 classes which are composed of 38
kinds of diseases and 1 background. The 38 diseases may occur
in 14 different crops and all categories contain 61 486 images in
sum. Each category is stored in an independent folder,
representing the label. Images were augmented by image flipping,
gamma correction, noise injection, PCA color enhancement,
rotation, and scaling. Figure 1 shows some sample images of
Plant Village.
a. Apple leaf with cedar rust
b. Corn leaf with gray leaf spot
c. Grape leaf with black rot
d. Peach leaf with bacterial spot
e. Pepper leaf with bacterial spot
f. Potato leaf of health
g. Strawberry leaf with scorch
h. Tomato leaf with yellow leaf curl virus
Figure 1 Augmented Plant Village dataset sample images
September, 2022 Ma Z, et al. Maize leaf disease identification using deep transfer convolutional neural networks Vol. 15 No. 5 189
The local maize disease dataset was collected from a maize test
field in Zhaodong City, Heilongjiang Province, China. Image
capture was achieved using iPhone 7 Plus rear camera with
3024×4032 pixel resolution, and shooting times included morning,
noon, and afternoon. The camera supported 2× optical zoom, 10×
digital zoom at most, and optical image stabilization. Since a
couple of leaves with several diseases may appear in one
photograph, a 300×300 pixels clipping frame was predefined for
separating the principal part of each leaf. Pictures saved as JPG
formats were manually clipped into 300×300 pixels images, and
actual complex backgrounds (from the field) were retained
wherever possible. The dataset contained maize leaf health status
and other three diseases of gray leaf spot, common rust, and
northern leaf blight, comprising 4 categories and 1189 images in
total, while approximately 10% of images were acquired from
search engines to enlarge the dataset. Table 1 lists the local maize
dataset details and Figure 2 shows sample images.
Table 1 Local maize dataset details
No.
Name
Train
Test
Total
1
Gray leaf spot
233
58
291
2
Common rust
331
83
414
3
Northern leaf blight
229
57
286
4
Health
158
40
198
Total
4 categories
951
238
1189
a. Gray leaf spot
b. Common rust
c. Northern leaf blight
d. Health
Figure 2 Local maize dataset sample images
Each dataset was split training and test sets (ratio 4:1
respectively)[26]. Training set image resolution was adopted as the
network entry size (224×224 pixels) or (299×299 pixels) in the
following entrance of CNNs. Training images were enhanced
online through numerical normalization, rotation (20° or 40°
randomly), translation (horizontal or vertical), scaling, flipping (up
and down, left and right), and cross-cutting, which can make the
model more generalizable and robust.
2.2 Convolutional neural networks
Convolutional neural networks are pipeline multi-processing
layer network models, comprising multiple convolutions (C),
pooling (P), and fully connected (FC) layers generally. Deeper
CNN architectures tend to extract better features, reduce loss levels,
and improve fit, which also require more training data and
computing resources. 12 different CNN models were trained on
the Plant Village dataset and their performances for disease
identification were compared. VGG16 and VGG19[27], ResNet[28],
InceptionV3[29], InceptionResNetV2[30], DenseNet121,
DenseNet169, and DenseNet201[31] were divided into a set of deep
CNNs; whereas Xception[32], MobileNet[33], MobileNetV2[34], and
ShuffleNet[35] were divided into a set of lightweight CNNs.
2.2.1 Dropout
Complex feedforward neural networks can cause overfitting
when trained on small datasets. Dropout[36] helps prevent
overfitting by reducing joint feature detector actions, improving
overall CNN performance. In this study, dropout parameter = 0.5,
i.e., each training batch ignores 50% of the feature detectors (we set
50% hidden layer node value=0), reducing interdependence
between feature detectors (hidden layer nodes) to ensure local
feature independence and enhance generalization.
2.2.2 Hyperparameters
The experiment was accomplished in 2 sections: deep learning
on Plant Village and transfer learning on maize dataset, thus
hyperparameters were introduced respectively.
In both sections, “Callback” functions were applied to enhance
the efficiency of training, namely “Early Stopping” and “Learning
Rate Scheduler”. The training began with a relatively big learning
rate (LR) at first, and a metric (the accuracy or the loss value) was
monitored at every step of every epoch. If the training proceeded
rationally and smoothly, the LR would keep its value. Otherwise,
a decay rate would minish the LR, and change the pace of training
to reach convergence efficiently. However, the mechanism would
not run endlessly. If the training hadn’t improved over a patience
number of epochs even if the LR reached its least value, the
training would be faced with a risk of overfitting and fluctuation.
Then the training would be stopped and the best model weights
would be saved.
In the first section, all CNNs trained on the Plant Village
dataset shared the same hyperparameters: Learning rate (LR)=
0.001 with 0.000001 threshold. LR decay was set within 5 epoch
patience, then LR would become half of the original when
cross-entropy loss reduced slowly. Early stopping was
implemented, and the epoch where loss converged to the minimum
was recorded (Epoch convergent). Maximum epochs = 100
epochs. Preliminary trials determined maximum batch size for a
single training = 32. Therefore, one epoch was completed after
1538 steps if all 49 193 images in the training set were included.
The training process in every batch is not used as a reference, but
the recording of accuracy and loss value after each round of epoch.
We used the RMSprop optimizer.
In the second section, LR would begin at a smaller value, that
is 0.0001 with 0.000001 threshold. LR decay was set within 3
epoch patience, other parameters remained unchanged. Therefore,
one epoch was completed after 30 steps if all 951 images in the
training set were included.
2.3 Transfer learning
A two-staged model-based transfer learning approach was
applied in this study. Conventional transfer learning trains the
CNN in the source domain and fine-tunes it in the target domain,
solving overfitting and instability due to insufficient training data.
In this paper, a closer source domain to the target domain was
applied to obtain better performance. Challenges lay in that the
source domain dataset was collected under lab conditions while the
target domain dataset was shot in the field, indicating an impact on
performance. Besides, in order not to destroy the ability of the
previous layers of the network to extract features, pre-trained
models needed to be fine-tuned temperately. And the target
domain was a small-scale dataset thus overfitting could easily
190 September, 2022 Int J Agric & Biol Eng Open Access at https://www.ijabe.org Vol. 15 No. 5
occur. Therefore, the transfer learning employed two training
strategies: layer-freezing (LF), freezing all layers’ parameters and
initializing the classifier layer randomly; or fine-tuning (FT),
defrosting part of the intermediate layer close to the classifier layer
gradually and initializing them all. Moreover, training set images
in the self-built maize disease dataset were enhanced online. The
current applied both strategies to verify identification efficiency
and accuracy. Figure 3 shows the proposed process.
Figure 3 Leaf disease identification process
2.4 Hardware and software
The platform employed for this study was a deep learning
workstation equipped with Intel® Core™ i9-9900K 16 CPU @
3.6 GHz processor, RTX 3090 24G memory graphics card, 32 GB
RAM, 2.5 TB HDD, graphics card driver version 455.45.01,
CUDA version 11.1, and CUDNN version 8005. The operating
system was 64 bit Ubuntu 20.04 LTS, and programming was
implemented using Python 3.7.9 and Keras 2.4.3 under the
TensorFlow-GPU framework.
3 Results and discussion
3.1 Pre-trained model performances
Table 3 shows pre-trained model performances. ResNet and
InceptionResNetV2 test accuracy was superior to other models,
achieving 99.48% and 99.31% respectively; whereas InceptionV3,
MobileNet, and MobileNetV2 achieved 98.82%, 98.69%, and
98.00% respectively, and the remaining rest networks ranged from
96% to 98%.
Table 3 Baseline CNN model performances
Network
Epoch
(Convergent)
Training
Test
Loss
Accuracy
Loss
Accuracy
VGG16
33
0.1666
95.79%
0.1767
96.76%
VGG19
41
0.0798
97.79%
0.1147
97.04%
ResNet
47
0.0076
99.82%
0.0183
99.48%
DenseNet121
27
0.0628
98.27%
0.1253
97.47%
DenseNet169
23
0.0662
98.37%
0.1391
97.67%
DenseNet201
34
0.0486
98.59%
0.1024
97.99%
InceptionV3
18
0.0515
98.42%
0.1524
98.82%
InceptionResNetV2
24
0.0151
99.63%
0.2096
99.31%
Xception
10
0.0452
98.90%
0.1863
96.81%
MobileNet
32
0.0178
99.44%
0.0493
98.69%
MobileNetV2
36
0.0297
99.04%
0.0814
98.04%
ShuffleNet
24
0.0352
98.82%
0.1473
96.08%
ResNet, MobileNet, and MobileNetV2 losses were also
smaller, achieving 0.0183, 0.0492, and 0.0814, respectively,
whereas the other considered other networks were between 0.1 and
0.21. All model test accuracies exceeded 96%, indicating the
excellent CNN performance on a sufficient training dataset.
Ferentinos et al.[37] trained CNN models on a large self-constructed
dataset, including 25 plants, 58 categories, and 87 848 images in
total, achieving the highest accuracy for classic CNN = 99.53%.
Therefore, sufficient training data is critical for optimal
performance. More complex networks can theoretically extract
features better[38] but require more training data, hence DenseNet
achieved the best accuracy in contrast to the present study[37].
In this study, the Test accuracy for the proposed deep CNN
ranged from 96.76% to 99.48% and convergence occurred within
30 epochs on average; whereas test accuracy for lightweight
CNNs=96.08% to 98.69% and convergence<25 iterations. This
small test accuracy difference will ensure the lightweight network
is easier to train.
Table 4 shows the model scales and related parameters.
Under the premise that the accuracy is higher than 98% or the loss
value is not greater than 0.1, MobileNet and MobileNetV2
employed relatively small parameter values=28.9 and 34.4,
respectively, whereas ResNet, InceptionV3, and
InceptionResNetV2 employed 12.5, 88.9, and 104.7, respectively.
ResNet and MobileNet network models were relatively small, 144
and 208 MB, respectively, outperforming all other models.
Figure 4 shows training accuracy and loss changes for ResNet and
MobileNet during training.
Table 4 ResNet and MobileNet training accuracy and loss
Network
No. parameters
(million)
Model size/MB
Input size (pixel)
VGG16
27.6
210
224×224
VGG19
32.8
251
224×224
ResNet
12.5
144
224×224
DenseNet121
49.5
223
224×224
DenseNet169
54.4
367
224×224
DenseNet201
66.5
438
224×224
InceptionV3
88.9
679
299×299
InceptionResNetV2
104.7
400
299×299
Xception
125.7
959
299×299
MobileNet
28.9
208
224×224
MobileNetV2
34.4
262
224×224
ShuffleNet
1.9
17.3
224×224
September, 2022 Ma Z, et al. Maize leaf disease identification using deep transfer convolutional neural networks Vol. 15 No. 5 191
Figure 4 ResNet and MobileNet training processes
Consequently, ResNet and MobileNet have relatively small
parameters and model sizes while providing high identification
performances, and hence are more suitable to be mounted on
mobile agricultural equipment. Therefore, after removing the
bottom FC layer, these pre-trained models were retained for
transfer learning in the next stage.
3.2 Feature extraction
Transfer learning requires a new FC layer for pre-trained
models. Figures 5 and 6 show MobileNet and ResNet network
structure diagrams for transfer learning training, respectively, with
the numbers at the bottom in the shape format of the output for
each layer. All convolutional layers included batch normalization
(BN) and ReLU activation layers.
MobileNet and ResNet features were extracted before the
second stage of transfer learning. Feature maps from the
intermediate layer output can be obtained by importing three
different disease images from the local maize disease dataset into
the network. Figure 7 shows feature samples for gray leaf spots,
northern leaf blight, and common rust in MobileNet. The number
of channels selected = 10, 11, 12, 18, 19, and 20.
Figure 5 MobileNet internal layer features (a) original image; feature images from (b) convolution, (c) BN,
and (d) ReLU6 layers; feature images from (e) depthwise convolution, (f) BN, and (g) ReLU6 layers
Figure 6 shows feature graphs from the 23rd channel within
block 1. Comparing the original image with the feature image,
the 23rd feature image can not only resist influence from complex
backgrounds but can also classify lesion areas more accurately.
Ahmad et al.[39] concluded performance on laboratory data was
superior to field data regardless of network type, and complex
backgrounds can reduce identification accuracy. Chen et al.[40]
confirmed these conclusions for maize disease identification.
Therefore, removing background interference will have significant
impact on identification accuracy.
192 September, 2022 Int J Agric & Biol Eng Open Access at https://www.ijabe.org Vol. 15 No. 5
Figure 7 shows feature maps from the same three images as
above in ResNet. The feature maps show sensitivity towards
northern leaf blight and common rust, and the highlighted area
indicates the lesion area. Most blade texture differences are also
noted. However, edge outlines and considerable background remains
in the feature map, which will impact disease identification.
Figure 6 Features from Channel No.23 MobileNet Block 1
a. Features in Block0_cnv1 layer b. Features in Block0_res_module layer
Figure 7 ResNet internal layer feature maps
3.3 Transfer learning model performances
The pre-trained ResNet and MobileNet weights were loaded
and performed transfer learning training on the local maize disease
dataset. Four new models were derived depending on the
different training strategies, layer-freezing (LF) or fine-tuning (FT):
R-LF, R-FT, M-LF, and M-FT. Figure 8 compares accuracy and
September, 2022 Ma Z, et al. Maize leaf disease identification using deep transfer convolutional neural networks Vol. 15 No. 5 193
loss for these four models, and Table 5 lists training and test set
performances, including early accuracy and loss after the first
training epoch.
Figure 8 Training process for ResNet and MobileNet models with
layer-freezing (LF) and fine-tuning (FT) training strategies: R-LF,
R-FT, M-LF, and M-FT
Table 5 ResNet and MobileNet transfer learning performance
Strategy
Epoch
Early
Training
Test
Loss
Accuracy
Loss
Accuracy
Loss
Accuracy
R-LF
28
1.0976
81.91%
0.0904
98.74%
0.1266
97.82%
R-FT
41
1.0197
84.61%
0.0722
99.23%
0.1368
98.77%
M-LF
24
1.2255
79.18%
0.0823
98.30%
0.1987
96.92%
M-FT
30
0.9486
81.38%
0.0918
99.05%
0.109
99.11%
All transfer models achieved test accuracy ≥96%.
Fine-tuning strategy models performed better than LF models with
0.95% and 2.19% improvement for ResNet and MobileNet,
respectively. Fine-tuning models also reduced losses by 0.0098
and 0.0893 compared with LF models. M-LF achieved the best
test accuracy=99.11%. Thus, transfer learning provided good
identification performance. The general outcome is consistent
with Too et al.[41], who migrated pre-trained ImageNet models onto
the Plant Village dataset, achieving accuracy >90% for all models
except VGG16, whereas DenseNet and ResNet performed better
and converged more easily. In contrast, the present study used the
Plant Village dataset as the source rather than target domain.
Plant Village is smaller capacity than ImageNet (1.28 million
images tagged training sets, 1 thousand categories), hence
deep-CNN fails to show significantly better performance, and Plant
Village was more similar to the local maize disease dataset, hence
improving accuracy.
On the other hand, LF effectively reduced training epochs at
the expense of accuracy, reducing overall hardware resource
requirements compared with FT. Transfer learning accuracy
≥79% and loss ≤1.3 after the first epoch, with FT achieving better
initial accuracy and loss than LF. Thus, transfer learning training
required less initial learning rate and patience with greater decay
rate, hence improving training efficiency compared with training
from scratch.
Figure 9 shows confusion matrices for the four model’s disease
identification performances. Gray leaf spot and northern leaf
blight slightly misidentify each other, and common rust may be
misidentified as gray spot, but gray spot does not tend to be
misclassified as rust. ResNet and MobileNet were the superior
pre-trained models on Plant Village, and also achieved outstanding
performance on the local maize leaf disease dataset collected in the
field. MobileNet achieved the best performance after FT (M-FT),
with test accuracy = 99.11%, 2.19% higher than the original model.
In contrast, R-FT test accuracy = 98.77%, 0.95% improvement.
a. M-LF
b. M-FT
c. R-LF
d. R-FT
Note: Label 0 to 3 refers to gray leaf spot, common rust, northern leaf blight and healthy leaves respectively; Test set predictions made by
models are in the 4×4 grid and larger values are assigned darker colors.
Figure 9 Confusion matrices for R-LF, R-FT, M-LF, and M-FT models
Table 6 lists the comparison with current state-of-the-art
methods of recent researches on plant disease identification, where
P.V. is short for “Plant Village” dataset and TL is short for
“Transfer Learning”.
Along with the shrinking scale of dataset, more training
strategies and innovative modifications to models are applied, and
194 September, 2022 Int J Agric & Biol Eng Open Access at https://www.ijabe.org Vol. 15 No. 5
transfer learning becomes a common practice. Large scale dataset
with conventional deep CNNs promises an extra high accuracy,
which indicates sufficient samples are of the essence when
networks have the capacity to extract deep features. Among the
studies listed above, this paper proves the effectiveness of transfer
learning from Plant Village with its high accuracy and supplies the
solution to the dataset size problem. Moreover, fine-tuned
MobileNet is guaranteed to be deployed on mobile terminal devices
in real field scenario tasks, which is expected to be a direction of
future development.
Table 6 Comparison with current state-of-the-art methods of diseases identification
Research
Dataset
Strategy
Model
Accuracy
a[37] (2018)
Expanded P.V. (87848)
Training from scratch
VGG
99.53%
b[41] (2018)
P.V.
TL (ImageNet)
DenseNets
99.75%
c[42] (2019)
Local (46135)
Background removal; Dataset expansion
GoogLeNet
94%-96%
d[21] (2020)
Maize (3852, P.V.);
Local (rice&maize, 500 &466 each)
TL (ImageNet); Augmentation
INC-VGGN
92.00%
e[43] (2020)
Local (maize, 466)
TL (Plant Village); Attention mechanism
Mobile-DANet
95.86%
f [44] (2021)
Local (coconut, 1564)
TL (ImageNet); Segmentation
InceptionResNetV2, MobileNet
81.48%, 82.10%
g
Local (maize, 1192)
TL (Plant Village); Augmentation
MobileNet
99.11%
In this section, transfer learning was an efficient method and
provided high-precision outcomes. MobileNet was more robust to
interference from complex backgrounds than ResNet, which may
explain why MobileNet achieved better identification performance
on the local dataset.
4 Conclusions
Deep-transfer learning method was validly effective especially
when the dataset was on small scale, and transfer learning
improved initial model performance and training efficiency,
illustrated by fine-tuned MobileNet achieving the best performance.
This study provides a theoretical foundation for mobile collection
terminal maize disease identification with deep-transfer learning
method, which establishes the foundation for further practical
development of models and enrichment of data set.
Acknowledgements
This work was financially supported by the Science and
Technology Innovation 2030-"New Generation of Artificial
Intelligence" Major Project (Grant No. 2021ZD0110904), the
Central Government to Support the Reform and Development Fund
of Heilongjiang Local Universities (Grant No. 2020GSP15) and
Key R&D plan of Heilongjiang Province (Grant No.
GZ20210103).
[References]
[1] Gao Z, Shao Y, Xuan G, Wang Y, Liu Y, Han X. Real-time hyperspectral
imaging for the in-field estimation of strawberry ripeness with deep
learning. Artificial Intelligence in Agriculture, 2020; 4: 31–38.
[2] Gao Z, Luo Z, Zhang W, Lyu Z, Xu Y. Deep learning application in plant
stress imaging: a review. AgriEngineering, 2020; 2(3): 430–446.
[3] Yalcin H, Razavi S. Plant classification using convolutional neural
networks. In: 2016 Fifth International Conference on
Agro-Geoinformatics (Agro-Geoinformatics), IEEE, 2016; pp.1–5. doi:
10.1109/agro-geoinformatics.2016.7577698
[4] Fuentes A, Lee J, Lee Y, Yoon S, Park D S. Anomaly detection of plant
diseases and insects using convolutional neural networks. 2017.
Elsevier In: Conference ISEM 2017 - The International Society for
Ecological Modelling Global Conference, Jeju, South Korea, 2017.
[5] Liu J, Wang X. Plant diseases and pests detection based on deep learning:
a review. Plant Methods, 2021; 17(1): 1–18.
[6] Hughes D, Salathé M. An open access repository of images on plant
health to enable the development of mobile disease diagnostics. arXiv:
Computer Science, 2015; arXiv: 151.08060. doi: 10.48550/
arXiv.1511.08060.
[7] Jaiswal A, Pathak S, Rathore Y K, Janghel R R. Detection of disease
from leaf of vegetables and fruits using deep learning technique.
Advances in Biomedical Engineering and Technology, Springer, 2021;
pp.199–206. doi: 10.1007/978-981-15-6329-4_18.
[8] Sravan V, Swaraj K, Meenakshi K, Kora P. A deep learning based crop
disease classification using transfer learning. Materials Today:
Proceedings, 2021; In Press. doi: 10.1016/j.matpr.2020.10.846.
[9] Agarwal M, Gupta S K, Biswas K. Development of efficient CNN model
for tomato crop disease identification. Sustainable Computing:
Informatics and Systems, 2020; 28: 100407. doi: 10.1016/
j.suscom.2020.100407
[10] Liu B, Ding Z, Tian L, He D, Li S, Wang H. Grape leaf disease
identification using improved deep convolutional neural networks.
Frontiers in Plant Science, 2020; 11: 1082. doi: 10.3389/fpls.2020.01082
[11] Anagnostis A, Asiminari G, Papageorgiou E, Bochtis D. A convolutional
neural networks based method for anthracnose infected walnut tree leaves
identification. Applied Sciences, 2020; 10(2): 469. doi:
10.3390/app10020469.
[12] Fang T, Chen P, Zhang J, Wang B. Crop leaf disease grade identification
based on an improved convolutional neural network. Journal of
Electronic Imaging, 2020; 29(1): 013004. doi: https://doi.org/10.1117/
1.jei.29.1.013004.
[13] Waheed A, Goyal M, Gupta D, Khanna A, Hassanien A E, Pandey H M.
An optimized dense convolutional neural network model for disease
recognition and classification in corn leaf. Computers and Electronics in
Agriculture, 2020; 175: 105456. doi: 10.1016/j.compag.2020.105456.
[14] Wu Y. Identification of maize leaf diseases based on convolutional neural
network. Journal of Physics: Conference Series, IOP Publishing, 2021;
1748: 032004. doi: 10.1088/1742-6596/1748/3/032004.
[15] Boulent J, Foucher S, Théau J, St-Charles P-L. Convolutional neural
networks for the automatic identification of plant diseases. Frontiers in
plant science, 2019; 10: 941. doi: 10.3389/fpls.2019.00941.
[16] Nigam S, Jain R. Plant disease identification using deep learning: A
review. Indian Journal of Agricultural Sciences, 2020; 90(2): 249–257.
[17] Pardede H F, Suryawati E, Krisnandi D, Yuwana R S, Zilvan V. Machine
learning based plant diseases detection: A review. In: 2020 International
Conference on Radar, Antenna, Microwave, Electronics, and
Telecommunications (ICRAMET), IEEE, 2020; pp.212–217. doi:
10.1109/icramet51080.2020.9298619.
[18] Barbedo J G A. Impact of dataset size and variety on the effectiveness of
deep learning and transfer learning for plant disease classification.
Computers and Electronics in Agriculture, 2018; 153: 46–53.
[19] Afifi A, Alhumam A, Abdelwahab A. Convolutional neural network for
automatic identification of plant diseases with limited data. Plants, 2021;
10(1): 28. doi: 10.3390/plants10010028.
[20] Xu Y, Zhao B, Zhai Y, Chen Q, Zhou Y. Maize diseases identification
method based on multi-scale convolutional global pooling neural network.
IEEE Access, 2021; 9: 27959–27970.
[21] Chen J, Chen J, Zhang D, Sun Y, Nanehkaran Y A. Using deep transfer
learning for image-based plant disease identification. Computers and
Electronics in Agriculture, 2020; 173: 105393. doi: 10.1016/
j.compag.2020.105393.
[22] Yin H, Gu Y H, Park C-J, Park J-H, Yoo S J. Transfer learning-based
search model for hot pepper diseases and pests. Agriculture, 2020; 10(10):
439. doi: 10.3390/agriculture10100439.
[23] Chen J, Zhang D, Suzauddola M, Nanehkaran Y A, Sun Y. Identification
September, 2022 Ma Z, et al. Maize leaf disease identification using deep transfer convolutional neural networks Vol. 15 No. 5 195
of plant disease images via a squeeze‐and‐excitation MobileNet model
and twice transfer learning. IET Image Processing, 2021; 15(5):
1115–1127.
[24] Chen J, Zhang D, Nanehkaran Y. Identifying plant diseases using deep
transfer learning and enhanced lightweight network. Multimedia Tools
and Applications, 2020; 79(41): 31497–31515.
[25] Atila Ü, Uçar M, Akyol K, Uçar E. Plant leaf disease classification using
efficientnet deep learning model. Ecological Informatics, 2021; 61:
101182. doi: 10.1016/j.ecoinf.2020.101182.
[26] Mohanty S P, Hughes D P, Salathé M. Using deep learning for
image-based plant disease detection. Frontiers in Plant Science, 2016; 7:
1419. doi: 10.3389/fpls.2016.01419.
[27] Simonyan K, Zisserman A. Very deep convolutional networks for
large-scale image recognition. arXiv: Computer Science, 2014; arXiv:
1409.1556. doi: 10.48550/arXiv.1409.1556.
[28] He K, Zhang X, Ren S, Sun J. Deep residual learning for image
recognition. Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, Las Vegas: IEEE, 2016; 770–778. doi:
10.1109/cvpr.2016.90.
[29] Szegedy C, Vanhoucke V, Ioffe S, Shlens J, Wojna Z. Rethinking the
inception architecture for computer vision. In: Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition, Las Vegas: IEEE,
2016; 2818–2826. doi: 10.1109/cvpr.2016.308.
[30] Szegedy C, Ioffe S, Vanhoucke V, Alemi A. Inception-v4,
inception-resnet and the impact of residual connections on learning.
Proceedings of the AAAI Conference on Artificial Intelligence; 2017;
pp.4278–4284.
[31] Huang G, Liu Z, Van Der Maaten L, Weinberger K Q. Densely connected
convolutional networks. Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition (CVPR); 2017; pp.2261–2269.
doi: 10.1109/CVPR.2017.243.
[32] Chollet F. Xception: Deep learning with depthwise separable
convolutions. Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, Honolulu, HI, USA, 2017; pp.1800–1807. doi:
10.1109/cvpr.2017.195.
[33] Howard A G, Zhu M, Chen B, Kalenichenko D, Wang W, Weyand T, et al.
Mobilenets: Efficient convolutional neural networks for mobile vision
applications, arXiv: Computer Science, 2017; arXiv: 1704.04861. doi:
10.48550/arXiv.1704.04861.
[34] Sandler M, Howard A, Zhu M, Zhmoginov A, Chen L-C. Mobilenetv2:
Inverted residuals and linear bottlenecks. In: Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition, Salt Lake City:
IEEE, 2018; pp.4510–4520. doi: 10.1109/cvpr.2018.00474.
[35] Zhang X, Zhou X, Lin M, Sun J. Shufflenet: An extremely efficient
convolutional neural network for mobile devices. Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition, 2018. doi:
10.1109/cvpr.2018.00716.
[36] Hinton G E, Srivastava N, Krizhevsky A, Sutskever I, Salakhutdinov R R.
Improving neural networks by preventing co-adaptation of feature detectors.
arXiv: Computer Science, 2012; arXiv: 1207.0580. doi:
10.48550/arXiv.1207.0580.
[37] Ferentinos K P. Deep learning models for plant disease detection and
diagnosis. Computers and Electronics in Agriculture, 2018; 145:
311–318.
[38] Sethy P K, Barpanda N K, Rath A K, Behera S K. Deep feature based rice
leaf disease identification using support vector machine. Computers and
Electronics in Agriculture, 2020; 175: 105527. doi: 10.1016/
j.compag.2020.105527.
[39] Ahmad I, Hamid M, Yousaf S, Shah S T, Ahmad M O. Optimizing
pretrained convolutional neural networks for tomato leaf disease detection.
Complexity, 2020; 2020: 8812019. doi: 10.1155/2020/8812019.
[40] Chen J, Zhang D, Zeb A, Nanehkaran Y A. Identification of rice plant
diseases using lightweight attention networks. Expert Systems with
Applications, 2021; 169: 114514. doi: 10.1016/j.eswa.2020.114514.
[41] Too E C, Yujian L, Njuki S, Yingchun L. A comparative study of
fine-tuning deep learning models for plant disease identification.
Computers and Electronics in Agriculture, 2019; 161: 272–279.
[42] Barbedo J G A. Plant disease identification from individual lesions and
spots using deep learning. Biosystems Engineering, 2019; 180: 96–107.
[43] Chen J, Wang W, Zhang D, Zeb A, Nanehkaran Y A. Attention embedded
lightweight network for maize disease recognition. Plant Pathology, 2021;
70(3): 630–642.
[44] Singh P, Verma A, Alex J S R. Disease and pest infection detection in
coconut tree through deep learning techniques. Computers and
Electronics in Agriculture, 2021; 182: 105986. doi: 10.1016/
j.compag.2021.105986.
