Turkish Journal of Electrical Engineering & Computer Sciences Turk J Elec Eng & Comp Sci

(2019) 27: 1636 – 1651

http://journals.tubitak.gov.tr/elektrik/
© TÜBİTAK
Research Article doi:10.3906/elk-1809-181

Plant disease and pest detection using deep learning-based features

Muammer TÜRKOĞLU1,∗,, Davut HANBAY2 ,

1
Department of Computer Engineering, Faculty of Engineering and Architecture, Bingöl University, Bingöl, Turkey
2
Department of Computer Engineering, Faculty of Engineering, İnönü University, Malatya, Turkey

Received: 25.09.2018 • Accepted/Published Online: 25.02.2019 • Final Version: 15.05.2019

Abstract: The timely and accurate diagnosis of plant diseases plays an important role in preventing the loss of
productivity and loss or reduced quantity of agricultural products. In order to solve such problems, methods based
on machine learning can be used. In recent years, deep learning, which is especially widely used in image processing,
offers many new applications related to precision agriculture. In this study, we evaluated the performance results
using different approaches of nine powerful architectures of deep neural networks for plant disease detection. Transfer
learning and deep feature extraction methods are used, which adapt these deep learning models to the problem at
hand. The utilized pretrained deep models are considered in the presented work for feature extraction and for further
fine-tuning. The obtained features using deep feature extraction are then classified by support vector machine (SVM),
extreme learning machine (ELM), and K-nearest neighbor (KNN) methods. The experiments are carried out using
data consisting of real disease and pest images from Turkey. The accuracy, sensitivity, specificity, and F1-score are
all calculated for performance evaluation. The evaluation results show that deep feature extraction and SVM/ELM
classification produced better results than transfer learning. In addition, the fc6 layers of the AlexNet, VGG16, and
VGG19 models produced better accuracy scores when compared to the other layers.

Key words: Plant disease and pest detection, convolutional neural networks, deep learning architectures, feature
extraction, classifier methods

1. Introduction
In plants, disease indications usually occur on leaves, fruit, buds, and young branches. This situation causes
fruit to be wasted (to drop) or be damaged. In addition, these diseases lead to the formation of new infections
and the spread of the disease for reasons such as seasonal conditions. For this reason, it is very important to
determine the disease in advance and to take the necessary precautions before it spreads to other trees. As a
result, the fight against diseases and pests in plants is the single most important issue in agriculture [1–3].
There are a variety of diseases that affect plants, which can each cause economic, social, and ecological
loss. In this context, a timely and accurate diagnosis of plant diseases plays an important role in preventing
the loss of productivity and quantity of agricultural products. Detection of plant diseases is usually performed
manually. Such processes are conducted by experts such as botanists and agricultural engineers, first by visual
inspection and later in a laboratory environment. These traditional methods are often time-consuming and
complex processes [1, 3]. For these reasons, it has become important to automatically identify diseases based
on image processing and machine learning. Automatic plant disease diagnosis with visual inspection can be of
great benefit to users who have little to no knowledge of the product they are cultivating.
∗ Correspondence: [email protected]

1636
This work is licensed under a Creative Commons Attribution 4.0 International License.
 TÜRKOĞLU and HANBAY/Turk J Elec Eng & Comp Sci

There have been numerous studies in the literature with regards to the detection of plant diseases. In the
past decade, these studies have often been conducted based on the classification process by using features such
as color, shape, and texture [4–7, 9, 9, 10]. The major disadvantages of methods based on these properties are:

• Very low performance when used alone, although attempts have been made to improve performance
through a combination of several methods.

• Requires the use of segmented methods, meaning that plants have to be separated from their roots in
order to extract geometric and similar features.

• Applied using datasets that contain images that are difficult to obtain in real life.

For these reasons, such systems and methods cannot be used in real life. Recently, deep convolutional
neural networks (CNNs) have provided significant gains in areas of visual object and machine learning. One of
the biggest advantages of models and methods improved regarding these areas is that they can perform feature
extraction without applying segmented methods. Therefore, these methods and models can be easily used in
real applications.
CNNs used as basic deep learning tools have obtained significant success in several plant disease detection
studies. Amara et al. [11] used the LeNet architecture as a convolutional neural network to classify image
datasets, including being able to contrast healthy and diseased banana leaves. These were evaluated with deep
models fine-tuned by transfer learning. Mohanty et al. [12] used a dataset consisting of 54,306 images of diseased
and healthy plant leaves. They were evaluated for performance using AlexNet and GoogleNet models based on
a deep CNN to identify 14 crop species and 26 diseases. Fujita et al. [13] developed a new practical plant disease
detection system that contained seven disease types. They used classification systems based on CNNs, which
showed an average of 82.3% accuracy under a fourfold cross-validation strategy. Sladojevic et al. [14] proposed
a new approach in order to recognize 13 different types of plant diseases by using deep convolutional networks.
Fuentes et al. [15] developed a robust deep-learning-based detector for real-time usage that recognized nine
different tomato diseases and pests.
In this paper, a system based on deep CNNs is developed for the detection of plant diseases and pests as
a classification problem. In this study, we use a dataset consisting of real plant disease and pest images from
Turkey. First, we use this dataset for deep feature extraction based on deep learning architectures: AlexNet,
VGG16, VGG19, GoogleNet, ResNet50, ResNet101, InceptionV3, InceptionResNetV2, and SqueezeNet. The
deep features obtained from these deep models are classified by SVM, ELM, and KNN. Later, we fine-tune
using deep learning models based on transfer learning. We remove the last three layers of the architecture and
then add new layers to adopt the pretrained CNNs to solve the problem. Finally, we evaluate the performance
results by using transfer learning and deep feature extraction methods. The main contributions of the proposed
plant disease classification method are the following:

• Plant disease and pest classification using images is problematic for machine learning. In order to classify
these images, various methods have been proposed in the literature. Today, deep learning methods are
phenomena of machine learning, and various deep learning networks have been proposed in the literature
for supervised learning. In this article, nine of the most used deep learning networks (AlexNet, GoogleNet,
VGG16, VGG19, ResNet50, ResNet101, InceptionV3, Inception ResNetV2, and SqueezeNet) are examined
for plant diseases and pests using transfer learning and deep feature extraction, and their results are
discussed comprehensively.

1637
 TÜRKOĞLU and HANBAY/Turk J Elec Eng & Comp Sci

• This work proposes a novel classification architecture using pretrained deep learning networks and tradi-
tional classifiers. The nine deep learning networks are also utilized for feature extraction. Variable sizes
of features are extracted using these networks. The traditional classifiers, which are SVM, ELM, and
KNN, are used in the classification phase with 10-fold cross-validation. The results clearly indicate that
the proposed architecture has higher classification accuracy than transfer-learning-based networks.

• The proposed architecture has lower computational complexity than transfer-learning-based models. This
situation is proved using execution time calculation.

• In the literature, studies have generally been performed using synthetic plant disease and pest images. In
this study, we use a dataset consisting of 1965 real plant disease and pest images within eight clusters. The
proposed method was tested on this dataset and the results clearly prove that it can be used in real-world
applications.

The study is organized as follows. While the proposed method is given in detail in Section 2, the proposed
architectures are given in Section 3. The experimental work and results are given in Section 4. Finally, in Section
5, the results are discussed, as well as a presentation of the contributions of the study.

2. Materials and methods

In this section, the theoretical background and related algorithms, dataset, and recommended method are
detailed under appropriate subheadings.

2.1. Deep learning and pretrained CNN models

Deep learning is a form of machine learning technique that uses calculation models composed of multiple
processing layers in order to learn the characteristics of the data [16]. As a result of significantly high
achievements in areas such as classification and recognition using deep learning, interest in this subject has
increased. Recently, these methods have been used in many areas such as speech recognition, visual object
recognition, and object detection [17]. Although the first studies on deep learning were based on a very long
history, the main reasons for its success in recent years are the creation of large data and the generation of
faster computers with large scale memory [18].
This study evaluated performance by trying different approaches of the nine most powerful architectures
of deep neural networks for the issue of plant disease identification. These deep learning architectures are trained
on a subset of the ImageNet database. The AlexNet architecture is a deep learning algorithm consisting of a
total of 25 layers with weights that can be trained on eight layers [19]. The GoogleNet architecture, based on
a network-in-network approach, uses architecture modules that use multiple convolutions in parallel to extract
various feature points [20, 21]. The VGG network proposed by the Oxford Visual Geometry Group (VGG) is
a homogeneous architecture used to obtain better results in the ILSVRC-2014 competition [22]. The ResNet
network was developed by He et al. to train networks with even larger depth [23]. This architecture, based
on microarchitecture modules, differs from traditional consecutive network architectures such as VGGNet and
AlexNet. The Inception network was proposed by Szegedy et al. as a type of convolutional neural network
model [21]. This network consists of a large number of convolution and maximum pooling steps. In the last
stage, it has a fully connected neural network. The InceptionResNetV2 network [24] was based on the Inception-
based network structure and residual connections. InceptionResNetV2 performs nearly identically to Inception

1638
 TÜRKOĞLU and HANBAY/Turk J Elec Eng & Comp Sci

architectures, but this architecture achieved significant acceleration in training using residual connections [25].
SqueezeNet was developed by Landola et al. as a smart architecture that achieves AlexNet-level accuracy on
ImageNet with 50 times fewer parameters [26]. The characteristics of these architectures are detailed in Table 1.

Table 1. Details of the utilized architectures.

Network Depth Size Parameter (millions) Image input size
AlexNet 8 227 MB 61 227 × 227
VGG16 16 515 MB 138 224 × 224
VGG19 19 535 MB 144 224 × 224
SqueezeNet 18 4.6 MB 1.24 227 × 227
GoogleNet 22 27 MB 7 224 × 224
Inceptionv3 48 89 MB 23.9 299 × 299
InceptionResNetv2 164 209 MB 55.9 299 × 299
ResNet50 50 96 MB 25.6 224 × 224
ResNet101 101 167 MB 44.6 224 × 224

2.2. Classifiers
In this paper, traditional classifier methods of SVM, ELM, and KNN are used to construct the classification
model of deep features extracted from specific layers of pretrained deep learning models.

2.2.1. K-nearest neighbor (KNN)

The KNN classifier uses a supervised learning method. This method is widely used in fields such as machine
learning, image processing, and statistical estimation. This method is an algorithm that makes a classification
of the existing learning data when new data enter. The principle of this method is to assign new data input
to the closest cluster in a previously established sample set. The distance between these two data points is
calculated by using various distance functions. The most known functions are Euclidean distance, Minkowski
distance, and Manhattan distance. In addition, the greatest advantage of this method is that its application is
easy and understandable [27, 28].

2.2.2. Support vector machine (SVM)

SVM is a method developed by Vapnik and is based on statistical learning theory [29]. The SVM method
aims to have a linear discriminant function with the largest marginal separating the classes from each other.
The learning data closest to the hyperplane are called support vectors. SVM can classify sets of data as both
linearly distinguishable and indistinguishable [30]. This classifier has been successfully applied in order to solve
problems in many areas such as image and object recognition, voice recognition, fingerprint recognition, and
handwriting recognition [29–31].

2.2.3. Extreme learning machine (ELM)

ELM is a learning algorithm for single hidden-layer feedforward networks (SLFNs) proposed by Huang et al.
[32]. In ELM, the least squares method is used to calculate output weights while hidden layer weights are

1639
 TÜRKOĞLU and HANBAY/Turk J Elec Eng & Comp Sci

randomly generated [33]. ELM can be modeled as shown in Eq. (1);

∑
N
βi g(ωi ∗ xj + bi ) = yj , j = 1, 2, · · · , M , (1)
i=1

where [xi , yi ] is the input-output, M is the number of training samples, wi is the input weight, and bi
is the bias of the hidden layer. If the network output assumes that the actual values converge with 0 error,
Y = Hβ can be written in matrix form. Y represents the output of ELM. The H hidden layer output matrix
and β output weight are calculated by Eq. (2):
 
g(w1 .x1 + b1 ) · · · g(wN .x1 + bN )
H= ··· ··· ··· , β = β1 , β 2 , · · · , β M . (2)
g(w1 .xM + b1 ) · · · g(wN .xM + bN )

The β parameter is obtained by β = H † Y . The Moore–Penrose generalized inverse of H is shown with

H † [32–34].

2.3. Dataset
The dataset used to test the performance of the proposed method includes images of plant diseases common to
the Malatya, Bingöl, and Elazığ regions of Turkey (Figure 1). These images were obtained with a Nikon 7200d
camera. Each image in this dataset consists of 4000 × 6000 resolution and three-channel (RGB) color images.
Table 2 lists the names and numbers of plant diseases and pests in this dataset.

Table 2. Details of the utilized dataset for detection of plant diseases and pests.

Label Name Train Test Total

1 Coryneum beijerinckii (disease) 545 61 606
2 Apricot monilia laxa (disease) 76 9 85
3 Walnut leaf mite ga (pest) 62 7 69
4 Peach sphaerolecanium prunastri (pest) 310 35 345
5 Peach monilia laxa (disease) 278 31 309
6 Cherry myzus cerasi (disease) 175 19 194
7 Xanthomonas arboricola (disease) 129 14 143
8 Erwinia amylovora (pest) 193 21 214

As can be seen in Table 2, the dataset includes a total of 1965 images belonging to eight different plant
diseases. The dataset was obtained by academics in the field of plant protection in the Agricultural Faculty of
Bingöl University and İnönü University, Turkey. These images were obtained during daytime on different days.
In addition, images of the diseases were taken using a large number of different trees.

3. Proposed method
In this study, we applied deep feature extraction from various fully connected layers and transfer learning based
on pretrained deep learning architectures. The scheme of the proposed study is illustrated in Figures 2 and 3.
Transfer learning and deep feature extraction are detailed in the following subsections.

1640
 TÜRKOĞLU and HANBAY/Turk J Elec Eng & Comp Sci

Figure 1. Samples of plant diseases and pests images used for experimentations: (a) Coryneum beijerinckii, (b)
Apricot monilia laxa, (c) Walnut leaf mite ga, (d) Peach sphaerolecanium prunastri, (e) Xanthomonas arboricola, (f)
Peach monilia laxa, (g) Erwinia amylovora, (h) Cherry myzus cerasi.

3.1. Transfer learning

Transfer learning is a machine learning method that is reused as a starting point for solving a different problem
by using knowledge obtained from a model developed in the solution of a problem. The current study fine-tuned
this by using pretrained CNN models based on transfer learning. The reason for using pretrained CNN models is
that they are faster and easier than training a CNN model with randomly initialized weights [35]. In addition,
the fine-tuning process is based on transferring new layers instead of the last three layers of the pretrained
networks to our classification task, as shown in Figure 2.

Figure 2. Simple representation of transfer learning.

1641
 TÜRKOĞLU and HANBAY/Turk J Elec Eng & Comp Sci

3.2. Deep feature extraction

Deep feature extraction is based on extracting features learned from a pretrained convolutional neural net-
work. These features are used to train machine learning classifiers. In other words, this method is based on
the extraction of deep features from the fully connected layer of pretrained networks. In this study, effective
deep features were extracted from a certain layer of deep learning models, which are GoogleNet, ResNet50,
ResNet101, InceptionV3, InceptionResNetV2, and SqueezeNet and pool5-drop_7x7_s1, fc1000, fc1000, predic-
tions, predictions, and pool10, respectively. In addition, feature vectors were obtained from the these layers of
the GoogleNet, ResNet50, ResNet101, InceptionV3, InceptionResNetV2, and SqueezeNet models: 1024, 1000,
1000, 1000, 1000, and 4096, respectively. The obtained deep features are employed in the classification phase
by using the traditional classifiers of SVM, ELM, and KNN, as shown in Figure 3.

Figure 3. A simple representation for deep feature extraction and traditional classifiers.

In this study, deep features obtained from these three layers for each of the AlexNet, VGG16, and VGG19
models are used in the classification phase by using the traditional classifiers of SVM, ELM, and KNN in order
to calculate performances by using classifier methods. Then the performances of these layers are compared,
and the deep features obtained from the best layer are determined. In addition, while 4096 feature vectors are
obtained from the fc6 and fc7 layers of the AlexNet, VGG16, and VGG19 models, 1000 feature vectors are
obtained from the fc8 layer of these models.

3.3. Summary steps of proposed architecture and transfer learning

The following steps summarize the proposed deep feature extraction:
Step 1: Acquire plant images.
Step 2: Resize plant image according to deep networks using bilinear interpolation. For instance, color images
sized 224 × 224 and 227 × 227 are used in ResNet50 and AlexNet, respectively.
Step 3: Features are extracted using the fully connected layers of the deep learning models.
Step 4: Classification is performed using the deep features with the SVM, KNN, and ELM classifiers.

The following steps summarize the transfer learning:

Step 1: Acquire plant images.
Step 2: Resize plant image according to deep networks using bilinear interpolation.
Step 3: Last three layers are removed from pretrained deep models and replaced by a fully connected layer,
a softmax layer, and a classification output layer in order to adopt the pretrained CNN networks to solve the
problem.
Step 4: Classification is performed using the newly created deep model.

1642
 TÜRKOĞLU and HANBAY/Turk J Elec Eng & Comp Sci

4. Experimental studies
In this study, we evaluated the performance using different approaches of nine powerful architectures of deep
neural networks for plant disease detection. The experimental studies were implemented using the MATLAB
deep learning toolbox. All applications were run on a server with a 2xCore Intel Xeon E5 with 64 GB memory.
The server was equipped with NVIDIA Quadro M4000 and 8 GB memory. We used the SVM, ELM, and
KNN methods in order to test the classification accuracy of the deep features. The study used a one-versus-all
approach and a cubic SVM as the classifier type for the SVM classifier parameters. In addition, we used a
sigmoid function as an activation function and a hidden layer neuron number of 10,000 for the ELM classifier
parameters. Finally, we used the Euclidean distance function and 1NN-3NN value of the KNN algorithm for
the KNN classifier parameters.
In order to test the performance of the proposed method, we used a dataset containing our own plant
disease images. This dataset consists of a total of 1965 images belonging to eight different plant diseases. The
numerical split for training and testing for each plant species is given in Table 2. To calculate the performance
of the proposed methods, a 10-fold cross-validation test was applied. In addition, we used performance measures
of accuracy, sensitivity, specificity, and F1-score for classification performances of experimental studies. The
experimental results and performance comparisons are detailed under the following subheadings.

4.1. Results based on deep feature extraction with AlexNet, VGG16, and VGG19 models
In this section, we use three different fully connected layers for deep feature extraction based on pretrained
AlexNet, VGG16, and VGG19 models. For each of these models, deep features were extracted from the fc6,
fc7, and fc8 layers. Then these features were calculated for their performance using the SVM, ELM, and KNN
methods. The accuracy scores of these experimental studies are given in Table 3. In addition, these accuracy
scores were evaluated according to the calculated average accuracy score across the folds and their standard
deviations.

Table 3. Accuracy scores (%) and execution time for different training approaches based on deep feature extraction
(bold font shows the best accuracy scores for each model).
Classifier AlexNet Vgg16 Vgg19
methods fc6 fc7 fc8 fc6 fc7 fc8 fc6 fc7 fc8
SVM 95.5±1.2 94.1 ± 0.65 92.2 ± 0.92 95.0±1.4 93.7 ± 1.21 92.5 ± 0.7 94.45 ± 0.8 93.99 ± 0.74 91.76 ± 1.26
ELM 93.41 ± 0.72 92.29 ± 0.5 91.53 ± 0.96 94.84 ± 1.12 93.57 ± 0.62 92.11 ± 0.8 94.74±1.14 94.38 ± 0.81 93.01 ± 0.54
KNN 89.41 ± 1.8 89.01 ± 1.41 84.5 ± 1.02 88.55 ± 1.49 87.0 ± 1.11 84.7 ± 0.82 89.01 ± 1.52 87.94 ± 1.38 85.5 ± 0.98
Time 31 s 32 s 34 s 1 min 6 s 1 min 7 s 1 min 9 se 1 min 15 s 1 min 16 s 1 min 19 s

As shown in Table 3, the deep features obtained from the fc6, fc7, and fc8 layers for each of the AlexNet,
VGG16, and VGG19 models were tested for their performance using the SVM, ELM, and KNN methods.
According to the obtained accuracy scores, the fc6 layer was found to have the best deep features for these
models. The highest accuracy level for the AlexNet model was 95.5% using the SVM classifier. While the
highest accuracy for the VGG16 model was 95% using the SVM classifier, the highest accuracy for the VGG19
model was 94.74% using the ELM classifier. In addition, in Table 3 the execution times for each of the fully
connected layers of the AlexNet, VGG16, and VGG19 models are given. According to these results, for all fully
connected layers, AlexNet had the lowest training time among the three models, while the training times for
VGG16 and VGG19 were almost equal. In addition, the performance measures of sensitivity, specificity, and
F1-score related to these experimental studies are given in Table 4.

1643
 TÜRKOĞLU and HANBAY/Turk J Elec Eng & Comp Sci

Table 4. Performance measures (%) for different training approaches based on deep feature extraction.

Classifier Performance AlexNet VGG16 VGG19

methods measures fc6 fc7 fc8 fc6 fc7 fc8 fc6 fc7 fc8
Sensitivity 94.34 93.03 89.61 93.02 91.91 89.96 92.22 92.41 89.34
SVM
Specificity 99.35 99.22 98.78 99.31 99.12 98.88 99.19 99.13 98.79
F1-score 94.58 93.49 90.02 93.19 91.69 89.87 92.47 92.44 89.10
Sensitivity 90.75 89.44 88.20 92.12 91.51 89.85 92.67 91.87 89.98
ELM
Specificity 99.00 98.86 98.70 99.22 99.08 98.87 99.23 99.12 98.97
F1-score 91.29 90.23 89.30 92.46 91.99 90.36 93.06 91.93 90.65
Sensitivity 87.94 87.07 76.31 86.56 83.87 82.61 87.28 85.43 82.88
KNN
Specificity 98.38 98.35 97.38 98.35 98.01 97.75 98.40 98.23 97.85
F1-score 87.35 86.33 78.03 85.46 82.69 80.94 86.33 84.87 82.03

4.2. Results based on deep feature extraction with other deep learning models
In this section, we used the fully connected layer for deep feature extraction, based on pretrained CNN models of
GoogleNet, ResNet50, ResNet101, InceptionV3, InceptionResNetV2, and SqueezeNet. For each of these models,
deep features were extracted from the layers specified in Section 3. Then these features were calculated for their
performance using the SVM, ELM, and KNN methods. The accuracy scores of these experimental studies are
given in Table 5. In addition, these accuracy scores were evaluated according to the calculated average accuracy
score across the folds and their standard deviations.

Table 5. Accuracy scores (%) and execution time for other deep learning models based on deep feature extraction (bold
font shows the best accuracy scores for each model).

Classifier
GoogleNet ResNet50 Resnet101 Inceptionv3 InceptionResNetv2 SqueezeNet
methods
SVM 95.22±1.24 97.86±1.56 96.74 ± 0.64 94.96±1.22 94.76 ± 1.04 95.62±1.08
ELM 94.18 ± 0.96 97.65 ± 1.34 97.45±0.78 94.54 ± 0.87 95.20±1.16 95.10 ± 0.94
KNN 89.16 ± 2.06 90.48 ± 2.24 91.20 ± 1.41 88.65 ± 1.52 88.80 ± 2.2 87.02 ± 3.82
Time 1 min 41 s 2 min 21 s 4 min 11 s 9 min 29 s 28 min 13 s 1 min 3 s

As shown in Table 5, the deep features obtained from the specific layers of the GoogleNet, ResNet50,
ResNet101, InceptionV3, InceptionResNetV2, and SqueezeNet models were tested for their performance using
the SVM, ELM, and KNN methods. The ResNet50 model and SVM classifier produced the highest accuracy
among these other deep models with an accuracy of 97.86%. The confusion matrix of this model is shown in
Table 6. When classified with the SVM method, the highest accuracy scores were 95.22% for the GoogleNet
model, 94.96% for the InceptionV3 model, and 95.62% for the SqueezeNet model. Using the ELM classifier, the
highest accuracy scores were 97.34% for the ResNet101 model and 95.35% for the InceptionResNetV2 model.
The execution time for each of these deep models is given in Table 5. According to these results,
SqueezeNet (1 min, 3 s) showed the lowest training time among the three models, while InceptionResNetV2 (28
min, 13 s) had the highest training time. Comparatively, the GoogleNet model (1 min, 41 s) was slightly slower
than the SqueezeNet model (1 min, 3 s); the ResNet101 model (4 min, 11 s) was almost twice as slow as the

1644
 TÜRKOĞLU and HANBAY/Turk J Elec Eng & Comp Sci

Table 6. Confusion matrix based on ResNet50 model and SVM classifier (Legend: 1. Erwinia amylovora; 2.
Xanthomonas arboricola; 3. Walnut leaf mite ga; 4. Apricot monilia laxa; 5. Coryneum beijerinckii; 6. Cherry
myzus cerasi; 7. Peach sphaerolecanium prunastri; 8. Peach monilia laxa).

1 2 3 4 5 6 7 8
1 204 1 2 6
2 138 4 1
3 1 68
4 83 1 1
5 606
6 1 3 186 4
7 1 343 1
8 2 1 1 2 1 2 5 295

ResNet50 model (2 min, 21 s) in training time; and the InceptionV3 model (9 min, 29 s) was four times slower
than the ResNet50 model (2 min, 21 s). In addition, the performance measures of sensitivity, specificity, and
F1-score related to these experimental studies are given in Table 7.

Table 7. Performance measures (%) for other deep learning models based on deep feature extraction.

Classifier Performance
GoogleNet ResNet50 Resnet101 Inceptionv3 InceptionResNetv2 SqueezeNet
methods measures
Sensitivity 94.21 97.35 95.02 92.66 91.94 93.91
SVM
Specificity 99.30 99.69 99.53 99.25 99.24 99.36
F1-score 94.13 97.14 94.92 93.20 92.13 93.83
Sensitivity 92.09 96.60 96.21 91.39 92.54 93.12
ELM
Specificity 99.14 99.66 99.63 99.18 99.28 99.26
F1-score 92.57 96.70 96.26 92.31 93.17 93.51
Sensitivity 86.12 88.80 89.34 84.84 84.52 83.78
KNN
Specificity 98.40 98.59 98.71 98.25 98.33 98.10
F1-score 86.02 88.01 88.40 85.65 84.39 83.19

In this study, the utilized dataset is heterogeneous. Accordingly, the performance measures of accuracy,
sensitivity, specificity, and F1-score were calculated for each class and the obtained results using the ResNet50
model and SVM classifier are shown Figure 4.

4.3. Results based on transfer learning with deep learning models

In this section, we performed fine-tuning based on transfer learning using deep learning models from pretrained
CNN networks. For transfer learning parameters, the batch size was chosen as 10, the maximum epoch number
was set to 7, and the initial learning rate was assigned as 0.0004. In addition, training of the deep networks
was achieved with the stochastic gradient descent with momentum (SGDM) learning method. The training
procedure ended after 1250 iterations. The accuracy scores of these experimental studies are given in Table 8
and these scores were evaluated according to the calculated average accuracy score across the folds and their
standard deviations.

1645
 TÜRKOĞLU and HANBAY/Turk J Elec Eng & Comp Sci

100
100
99.82

99.67

99.68

99.89

99.77
99.75

99.77

99.42
99.42
99.63

99.27
98.84
100

98.55
98.55

97.64
97.64
97.64
99

96.91

96.87
98

96.5
96.5
96.16

95.87
95.87

95.77
97

95.46
95.46
95.32
95.32

95.1
96
95
94
Class 1 Class 2 Class 3 Class 4 Class 5 Class 6 Class 7 Class 8

Accuracy Sensitivity Specificity F1-Score

Figure 4. Performance results for each class of ResNet50 model and SVM classifier.

Table 8. Accuracy scores (%) and execution time for deep models based on transfer learning.
AlexNet VGG16 VGG19 GoogleNet ResNet50 Resnet101 Inceptionv3 InceptionResNetv2 SqueezeNet
Accuracy 93.33 ± 2.41 96.92 ± 1.26 94.87 ± 1.3 95.3 ± 1.02 95.38 ± 0.9 96.41 ± 0.86 92.31 ± 2.12 89.23 ± 3.46 91.28 ± 1.96
Time 5 min 48 s 15 min 26 s 42 min 53 s 11 min 30 s 20 min 49 s 31 min 30 s 40 min 39 s 111 min 21 s 8 min 33 s

As shown in Table 8, the VGG16 model showed the best accuracy score of 96.92% among the other
deep models, while the InceptionResNetV2 model was the worst at 89.23%. While the second best accuracy
score was produced as 96.41% for the ResNet101 model, the third best accuracy score was 95.38% using the
GoogleNet and ResNet50 models. In addition, Table 8 shows the execution time of the utilized deep models.
According to these results, AlexNet had the fastest execution time (5 min, 48 s) among the three models, while
InceptionResNetV2 had the slowest execution time (1 h, 51 min, 21 s). Comparatively, the SqueezeNet model
had a faster execution time (8 min, 33 s) than the GoogleNet (11 min, 30 s) and VGG16 (15 min, 26 s) models.
In addition, the ResNet50 model (20 min, 49 s) was almost twice as quick as the VGG19 model (42 min, 53 s) in
execution time, and one-third quicker than the ResNet101 model (31 min, 30 s). In addition, the performance
measures of sensitivity, specificity, and F1-score related to these experimental studies are given in Table 9.

Table 9. Performance measures (%) for deep models based on transfer learning.
Performance
AlexNet VGG16 VGG19 GoogleNet ResNet50 ResNet101 Inceptionv3 InceptionResNetv2 SqueezeNet
measures
Sensitivity 88.02 96.64 92.24 92.29 94.91 95.81 87.05 85.09 89.22
Specificity 99.0 99.54 99.14 95.75 99.36 99.43 98.90 98.48 98.70
F1-score 88.30 96.38 93.87 93.53 93.28 95.94 88.76 85.15 88.33

4.4. Results based on traditional methods

In this section, traditional methods of LBP, HOG, GLCM, and color features, which are widely used for plant
disease detection systems, have been applied. These methods extracted features directly from the images
without the requirement for a segmented operation. Properties obtained from these methods were calculated
for performance using the SVM, ELM, and KNN methods. The accuracy scores related to these studies are
given in Table 10.
As seen in Table 10, the best performances obtained using the SVM classifier were 70.9% for the LBP
method, 56.8% for the HOG method, 45.3% for color features, and 45.2% for the GLCM method.

1646
 TÜRKOĞLU and HANBAY/Turk J Elec Eng & Comp Sci

Table 10. Accuracy scores (%) for traditional methods (bold font shows the best accuracy scores for each method).

Classifier methods LBP HOG Color features GLCM

SVM 70.9 56.8 45.3 45.2
ELM 68.26 53.36 43.57 42.85
KNN 61.4 46.5 35.9 38.5

4.5. Discussion and comparison of simulation results

The proposed study used pretrained CNN models to obtain the best performance in the detection of plant
diseases. We evaluated the performance results of transfer learning and deep feature extraction based on the
AlexNet, VGG16, VGG19, GoogleNet, ResNet50, ResNet101, InceptionV3, InceptionResNetV2, and SqueezeNet
deep models. In addition, traditional methods of LBP, HOG, color features, and GLCM, which are widely used
in object recognition, were applied and their accuracy scores were evaluated.
First, we extracted features based on deep feature extraction by using the AlexNet, VGG16, and VGG19
models. For each of these models we separately extracted features from the three fully connected layers of
fc6, fc7, and fc8. The obtained deep features were calculated for their performance using SVM, ELM, and
KNN classifiers. According to the results, the best performance for these three models was obtained using
deep features extracted from the fc6 layer. For the AlexNet model, using the features from fc6, the highest
classification accuracy obtained was 95.5% with SVM and the lowest classification accuracy obtained was 89.41%
with KNN. For the VGG16 model, the highest classification accuracy obtained was 95.0% with SVM using the
features from fc6 and the lowest classification accuracy obtained was 88.55% with KNN. For the VGG19 model,
the highest classification accuracy obtained was 94.74% with ELM using the features from fc6 and the lowest
classification accuracy obtained was 88.01% with KNN. As a result, the features extracted from fc6 of the
AlexNet, VGG16, and VGG19 models produced better accuracy scores than the features extracted from either
fc7 or fc8. In addition, both the SVM and ELM methods achieved much better results than the KNN classifiers.
Next, we extracted features from a specific layer using the GoogleNet, ResNet50, ResNet101, InceptionV3,
InceptionResNetV2, and SqueezeNet deep learning models. The obtained deep features were calculated for their
performance using SVM, ELM, and KNN classifiers. According to the results, the GoogleNet model’s highest
classification accuracy was 95.22% with SVM and the lowest classification accuracy was 89.16% with KNN.
For the ResNet50 model, the highest classification accuracy was 97.86% with SVM and the lowest classification
accuracy was 90.48% with KNN. For the ResNet101 model, the highest classification accuracy was 97.45%
with ELM and the lowest classification accuracy was 91.2% with KNN. For the InceptionV3 model, the highest
classification accuracy was 94.96% with SVM and the lowest classification accuracy was 88.65% with KNN.
For the InceptionResNetV2 model, the highest classification accuracy was 95.2% with ELM and the lowest
classification accuracy was 88.8% with KNN. For the SqueezeNet model, the highest classification accuracy was
95.62% with ELM and the lowest classification accuracy was 87.02% with KNN. The GoogleNet, ResNet50,
InceptionV3, and SqueezeNet models obtained the best performance with the SVM classifier while the ResNet101
and InceptionResNetV2 models obtained the best performance with the ELM classifier. In addition, the worst
performance obtained for all the models was achieved using the KNN classifier.
Later, we performed fine-tuning based on transfer learning using pretrained CNN networks. According to
these performance results, while the VGG16 model produced the highest accuracy among the deep models, the
second best accuracy was achieved using the ResNet101 model. The accuracy score for the GoogleNet model

1647
 TÜRKOĞLU and HANBAY/Turk J Elec Eng & Comp Sci

was equal to that of the ResNet50 model. In addition, the VGG19 model had a better accuracy score than the
InceptionV3, InceptionResNetV2, and SqueezeNet models.
Finally, traditional methods of LBP, HOG, GLCM, and color features were used, and the obtained feature
values were calculated for their performance using SVM, ELM, and KNN classifiers. According to the results,
among these four methods, the highest classification accuracy was 70.9% using the LBP method and SVM
classifier. In addition, the best performance of these four methods was obtained using the SVM classifier and
the lowest performance was achieved with KNN. The comparisons of accuracy scores based on deep feature
extraction, transfer learning, and traditional methods are shown in Figure 5.

94.87

94.96

95.22

95.38

95.62

96.41

96.92

97.45

97.86
94.74

95.38
93.33
92.31
Accuracy Rate

95.5
95.2
91.28
89.23

95
100
90
70.9

80
56.8

70
45.3
45.2

60
50
40

Figure 5. Comparison of accuracy scores of utilized method and models.

As shown in Figure 5, the highest accuracy score was 97.86% using the ResNet50 model and SVM
classifier, while the second best accuracy score of 97.45% was produced using the ResNet101 model and SVM
classifier. The best accuracy score based on transfer learning was 96.92% with the VGG16 model. In addition,
conventional methods were found to have very poor levels of performance compared to the deep learning models.
Among the traditional methods, the highest accuracy score of 70.9% was produced for the LBP method and
SVM classifier. The accuracy scores based on deep feature extraction and transfer learning using deep models
were as follows:

• ResNet50 features and the SVM classifier had a better accuracy score than the fine-tuned ResNet50 model
based on transfer learning. ResNet101 features and the ELM classifier had a better accuracy score than
the fine-tuned ResNet101 model. SqueezeNet features and the SVM classifier had a better accuracy score
than the fine-tuned SqueezeNet model. AlexNet features and the SVM classifier had a better accuracy
score than the fine-tuned AlexNet model. InceptionResNetV2 features and the ELM classifier had a
better accuracy score than the fine-tuned InceptionResNetV2 model. InceptionV3 features and the SVM
classifier had a better accuracy score than the fine-tuned InceptionV3 model.

• The fine-tuned VGG16 model had a better accuracy score than the VGG16 features and SVM classifier.
The fine-tuned GoogleNet model had a better accuracy score than the GoogleNet features and SVM
classifier. The fine-tuned VGG19 model had a better accuracy score than the VGG19 features and SVM
classifier.

The evaluation results show that deep feature extraction and the SVM/ELM classifier obtained a higher
classification accuracy than networks based on transfer learning. In addition, each of the utilized deep models

1648
 TÜRKOĞLU and HANBAY/Turk J Elec Eng & Comp Sci

has unique features such as number of layers, number of connections, and filter types. Therefore, performance
results differed from each other. For example, the ResNet50 model has fewer layers and connections than
InceptionResNetV2, but its performance was higher.
In this study, we calculated execution times for deep feature extraction and transfer learning based on
deep learning models, as shown in Tables 5 and 8. According to the results, deep feature extraction has a much
lower training time compared to transfer learning. For example, while the AlexNet deep feature extraction was
almost ten times lower than the fine-tuned AlexNet in training time, the ResNet50-based deep feature extraction
was almost five times lower than the fine-tuned ResNet50. In addition, the testing times for transfer learning
and deep feature extraction based on deep models were mostly 10–30 s due to low test data, so there was no
need for a separate table. As a result, although the proposed architecture based on deep feature extraction
resulted in lower training times than transfer learning, it had higher accuracy scores.

5. Conclusion
This paper compared the performance results of deep feature extraction and transfer learning for the detection
of plant diseases and pests. This study used nine powerful architectures of deep neural networks for both
deep feature extraction and transfer learning. First, we extracted deep features for fully connected layers
of these deep models. The obtained deep features had their performances calculated using SVM, ELM, and
KNN classifiers. Then these deep models were fine-tuned based on plant disease and pest images. Finally, we
compared the performance results with deep learning models using traditional methods. The evaluation results
showed that the deep learning models produced better results compared to the traditional methods. The results
for deep feature extraction were better than the transfer learning achievements. In the studies based on deep
feature extraction and classifier methods, the fc6 layers of the AlexNet, VGG16, and VGG19 models produced
better accuracy scores compared to the other layers. As a result, the highest level accuracy scored was 97.86%,
obtained with the ResNet50 model and SVM classifier. In addition, the calculated execution time for deep
feature extraction and transfer learning based on deep learning models and the proposed architecture based on
deep feature extraction showed lower computational complexity than models based on transfer learning.
In future works, we are planning to use regional images and improved CNN models in order to improve
the classification performance. Additionally, we will gather images of different diseases in order to enrich the
database. The main goal for subsequent research will be developing a smart mobile device application that can
detect various plant diseases. This application, which will provide automatic plant disease diagnosis with visual
inspection, could be of great benefit to users with little to no knowledge of the plants that they are cultivating.

References

[1] Turkoglu M, Hanbay D. Apricot disease identification based on attributes obtained from deep learning algorithms.
In: IEEE 2018 International Conference on Artificial Intelligence and Data Processing; Malatya, Turkey; 2018. pp.
1-4.
[2] Güzel M. The importance of good agricultural practices (GAP) in the context of quality practices in agriculture
and a sample application. PhD, Dokuz Eylül University, İzmir, Turkey, 2012.

[3] Asma BM, Birhanlı O. Mişmiş. Malatya, Turkey: Evin Ofset, 2004 (in Turkish).

[4] Al-Hiary H, Bani-Ahmad S, Reyalat M, Braik M, Al-Rahamneh Z. Fast and accurate detection and classification
of plant diseases. International Journal of Computer Applications 2011; 17 (1): 31-38.

1649
 TÜRKOĞLU and HANBAY/Turk J Elec Eng & Comp Sci

[5] Bashir S, Sharma N. Remote area plant disease detection using image processing. IOSR Journal of Electronics and
Communication Engineering 2012; 2 (6): 31-34.

[6] Kulkarni AH, Patil A. Applying image processing technique to detect plant diseases. International Journal of
Modern Engineering Research 2012; 2 (5): 3661-3664.

[7] Arivazhagan S, Shebiah RN, Ananthi S, Varthini SV. Detection of unhealthy region of plant leaves and classification
of plant leaf diseases using texture features. Agricultural Engineering International CIGR Journal 2013; 15 (1): 211-
217.
[8] Khirade SD, Patil AB. Plant disease detection using image processing. In: IEEE International Conference on
Computing Communication Control and Automation; Pune, India; 2015. pp. 768–771.
[9] Athanikar G, Badar P. Potato leaf diseases detection and classification system. International Journal of Computer
Science and Mobile Computing 2016; 5: 76-78.
[10] Barbedo JGA, Koenigkan LV, Santos TT. Identifying multiple plant diseases using digital image processing.
Biosystems Engineering 2016; 147: 104-116.
[11] Amara J, Bouaziz B, Algergawy A. A deep learning-based approach for banana leaf diseases classification. In: BTW
Workshops; Bonn, Germany; 2017. pp. 79-88.
[12] Mohanty SP, Hughes DP, Salathé M. Using deep learning for image-based plant disease detection. Frontiers in Plant
Science 2016; 7: 1419.
[13] Fujita E, Kawasaki Y, Uga H, Kagiwada S, Iyatomi H. Basic investigation on a robust and practical plant diag-
nostic system. In: Proceedings of the 15th IEEE International Conference on Machine Learning and Applications;
Anaheim, CA, USA; 2016. pp. 989–992.
[14] Sladojevic S, Arsenovic M, Anderla A, Culibrk D, Stefanovic D. Deep neural networks based recognition of plant
diseases by leaf image classification. Computational Intelligence and Neuroscience 2016; 2016: 3289801.
[15] Fuentes A, Yoon S, Kim SC, Park DS. A robust deep-learning-based detector for real-time tomato plant diseases
and pests recognition. Sensors 2017; 17 (9): 2022.

[16] LeCun Y, Bengio Y, Hinton G. Deep learning. Nature 2015; 521 (7553): 436.

[17] Liew SS, Hani MK, Radzi SA, Bakhteri R. Gender classification: a convolutional neural network approach. Turkish
Journal of Electrical Engineering & Computer Sciences 2016; 24 (3): 1248-1264.

[18] Inik O, Ulker E. Deep learning and deep learning models used in image analysis. Gaziosmanpasa Journal of Scientific
Research 2017; 6(3): 85-104.

[19] Krizhevsk A, Sutskever I, Hinton GE. Imagenet classification with deep convolutional neural networks. In: Advances
in Neural Information Processing Systems; 2012 pp. 1097-1105.
[20] Ghazi MM, Yanikoglu B, Aptoula E. Plant identification using deep neural networks via optimization of transfer
learning parameters. Neurocomputing 2017; 235: 228-235.
[21] Szegedy C, Liu W, Jia Y, Sermanet P, Reed S et al. Going deeper with convolutions. In: Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition; Boston, MA, USA; 2015. pp. 1-9.
[22] Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition. arXiv preprint,
arXiv:1409.1556, 2014.
[23] He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. In: Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition; Las Vegas, NV, USA; 2016. pp. 770-778.
[24] Szegedy C, Loffe S, Vanhoucke V, Alemi AA. Inception-v4, inception-resnet and the impact of residual connections
on learning. In: Thirty-First AAAI Conference on Artificial Intelligence; San Francisco, CA, USA; 2017. pp. 4278-
4284.

1650
 TÜRKOĞLU and HANBAY/Turk J Elec Eng & Comp Sci

[25] Nguyen LD, Lin D, Lin Z, Cao J. Deep CNNs for microscopic image classification by exploiting transfer learning
and feature concatenation. In: 2018 IEEE International Symposium on Circuits and Systems; Florence, Italy; 2018.
pp. 1-5.
[26] Landola FN, Han S, Moskewicz MW, Ashraf K, Dally WJ, Keutzer K. Squeezenet: Alexnet-level accuracy with
50x fewer parameters and 0.5 mb model size. arXiv preprint, arXiv:1602.07360, 2016.
[27] Sengur A, Akilotu BN, Tuncer SA, Kadiroglu Z, Yavuzkilic S, Budak U, Deniz E. Optic disc determination in
retinal images with deep features. In: IEEE 26th Signal Processing and Communications Applications Conference;
İzmir, Turkey; 2018. pp. 1-4.
[28] Buttrey SE, Karo C. Using k-nearest-neighbor classification in the leaves of a tree. Computational Statistics & Data
Analysis 2002; 40: 27-37.
[29] Vapnik VN. The Nature of Statistical Learning Theory. New York, NY, USA: Springer, 1995.
[30] Ozgur A, Erdem H. Comparison of out-of-box machine learning algorithms used in intrusion detection systems.
International Journal of Informatics Technologies 2012; 5 (2): 41-48.
[31] Hajibandeh N, Faghihi F, Ranjbar H, Kazari H. Classifications of disturbances using wavelet transform and support
vector machine. Turkish Journal of Electrical Engineering & Computer Sciences 2017; 25: 832-843.
[32] Huang GB, Zhu QY, Siew CK. Extreme learning machine: a new learning scheme of feedforward neural networks.
In: Proceedings of the International Joint Conference on Neural Networks; Budapest, Hungary; 2004. pp. 985–990.
[33] Alcin OF, Sengur A, Ince MC. Forward-backward pursuit based sparse extreme learning machine. Journal of the
Faculty of Engineering and Architecture of Gazi University 2015; 30: 111-117.
[34] Ertuğrul ÖF, Tağluk ME. A fast feature selection approach based on extreme learning machine and coefficient of
variation. Turkish Journal of Electrical Engineering & Computer Sciences 2017; 25: 3409-3420.
[35] Deniz E, Şengür A, Kadiroğlu Z, Guo Y, Bajaj V et al. Transfer learning based histopathologic image classification
for breast cancer detection. Health Information Science and Systems 2018; 6 (1): 18.

1651