Measurement 171 (2021) 108796

Contents lists available at ScienceDirect

Measurement
journal homepage: www.elsevier.com/locate/measurement

Leaf disease segmentation and classification of Jatropha Curcas L. and

Pongamia Pinnata L. biofuel plants using computer vision based approaches
Siddharth Singh Chouhan a, Uday Pratap Singh b, *, Utkarsh Sharma c, Sanjeev Jain d
a
School of Computer Science and Engineering-Specialization, Jain University, Bangalore 562112, Karnatka, India
b
School of Mathematics, Shri Mata Vaishno Devi University, Katra 182320, Jammu and Kashmir, India
c
Department of Computer Science and Engineering, JUET, Guna 473226, Madhya Pradesh, India
d
Department of Computer Science and Engineering, IIITDM, Jabalpur 482005, Madhya Pradesh, India

A R T I C L E I N F O A B S T R A C T

Keywords: Several efforts have been made in finding alternate sources of energy. The production of bio-fuel from the ex­
Computer vision tracts of plants like Jatropha Curcas L. and Pongamia Pinnata L. is most favored among all. But, due to certain
Image classification biotic factors, the growth of these plants get affected, therefore reducing the overall production. To formulate the
Image segmentation
demand and automate the disease diagnosis system a Computer vision methodology is proposed in this work. For
Machine learning
Neural network
disease region segmentation, a Hybrid Neural Network incorporated with Superpixel clustering is proposed.
Simple linear iterative clustering Color, shape, and texture features are evaluated using different algorithms. Finally, seven different Machine
Learning techniques were used to classify the images among three categories. Segmentation results with average
Specificity = 0.9534, 0.9795, Sensitivity = 0.9637, 0.9805 and average Classification accuracy = 0.9857 ±
0.0285, 0.9095 ± 0.0688 and 0.9607 ± 0.0256 for both the plants when evaluated separately proved the su­
premacy of the proposed work.

1. Introduction achieving high-level understanding therefore automating the human

visual learning process [5–7]. And, also because of high computational
With an imperative search for alternate resources nowadays, the power, low-cost solutions, ease of availability of devices, and increasing
exploitation of plants like Jatropha Curcas L. (JCL) and Pongamia Pinnata interest, computer vision and image processing methodologies have
L. (PPL) for generating energy (biofuel) is becoming very prevalent. This been used in several fields to achieve artificial intelligence [7–10].
is because of their eco-friendly nature and economical importance [1,2]. Another important aspect of these approaches lies with their fast and
Moreover, these plants are also well-known for their affluent medicinal accurate solutions, thus overpowering the deficiencies present in the
values. Several studies also exemplify their importance in controlling manual intervention. In recent times, these methods have also shown a
soil erosion, developing a degraded land, and acting as carbon seques­ wide development in agricultural engineering for plant phenotyping,
ters. But because of various biotic and abiotic factors, the normal growth fruit disease detection, and classification, estimation of the crop, weed
of these plants gets affected. The leaves, flowers, and fruits of these detection, leaves disease detection, plant growth, and yield prediction,
plants get afflicted in most of the cases, but it may also result in a grading and sorting of plants, soil condition, water management, and so
complete plant loss number of times. These agents likewise reduce on [6–14].
plants environmental and economical values as well [3,4]. Therefore, Therefore, a computer vision and image processing based method for
inorder to sustain the supply chain demand and to automate the disease plant leaf disease segmentation and classification is proposed in this
(biotic factor) detection and treatment process an intelligent system is work. Two important well-known plants which are the rich source of
required. renewable biodiesel production named as Jatropha Curcas L. and Pon­
Images play a decisive role in the number of applications. They have gamia Pinnata L. were selected for this purpose. To begin with, the
been adopted in fields like medical, defense, agriculture for pattern healthy and diseased plant leaf images were self-acquired for both the
recognition, classification, object detection, and so on. Images also act as plants separately. Then these images were the preprocessed for main­
a backbone for computer vision methodology where these are used for taining uniformity and attaining better results. Further, with the help of

* Corresponding author.
E-mail address: [email protected] (U.P. Singh).

https://doi.org/10.1016/j.measurement.2020.108796
Received 15 September 2020; Received in revised form 17 November 2020; Accepted 25 November 2020
Available online 1 December 2020
0263-2241/© 2020 Elsevier Ltd. All rights reserved.
S.S. Chouhan et al. Measurement 171 (2021) 108796

Simple Linear Iterative Clustering (SLIC) algorithm the pixels with (continued )
similar attributes were initially clustered, and then using the Adaptive Algorithm 1. Basic steps for leaf disease segmentation and classification
Linear Neuron (ADALINE) network the disease region is segmented from 3. SLIC algorithm for grouping of pixels into a similar type of cluster and pixel labeling using
the plant leaves images. Color, shape, and texture features were the level of equivalence
extracted using different algorithms. Finally, seven different Machine 4. Disease segmentation using Adaptive Linear Neural network
Learning (ML) classifiers had been used for classification. A three-way 5. Feature extraction using L*a*b*, SIFT, and LBP algorithms
6. Ternary classification of leaf images using seven machine learning classifiers
classification methodology was adopted that first classifies the images
7. Results evaluation and comparison using different metrics
among two discrete plant leaf classes, then among healthy and diseased
for Jatropha Curcas L. plant and at last healthy and diseased for Pongamia
The block diagram for the proposed method has been depicted in Fig. 1.
Pinnata L. plant respectively.
The rest of the paper has been structured as follows: Section 2 pre­
2.2. Data acquisition
sents the various materials and methods used in this work, including
basic methodology or the algorithm used for segmentation and classi­
For the proposed work, Jatropha Curcas L. and Pongamia Pinnata L.
fication, data and image acquisition filed and setup, Section 3 introduces
plant leaf images were acquired from the beautiful campus of Shri Mata
the proposed hybrid neural network, feature extraction algorithms, and
Vaishno Devi, University (32◦ 56′ 30.6′′ N 74◦ 57′ 14.4′′ E), Katra, 182320,
different machine learning algorithms. The experimental evaluation and
Jammu and Kashmir, India [18]. The site location has been shown in
results for disease segmentation and classification have been discussed
Fig. 2. These plants are well known for their higher ecological, medical,
in Section 4. The summary and limitations of the proposed work is
and economical importance. Another interesting purpose for selecting
discussed in Section 5. Finally, Section 6 concludes the manuscript and
these plants is of their richness in bio-fuel (energy) production. This
with possible future improvements.
dataset consists of leaf images separated among healthy and diseased
classes. A total number of 133 healthy and 124 diseased leaf images for
2. Materials and methods
Jatropha Curcas L. and a total number of 322 healthy and 276 diseased
leaf images for Pongamia Pinnata L. were acquired for segmentation and
2.1. Basic methodology for leaf diseases segmentation and classification
classification. All the captured images are of uniform size. Both plant
images belong to the P6 and P7 class of our data repository available
The different steps used for the disease segmentation and classifi­
online at Mendeley Data (https://data.mendeley.com/datasets/
cation of Jatropha Curcas L. and Pongamia Pinnata L. plant leaf images
hb74ynkjcn/4).
has been given by Algorithm 1. Beginning with the process the leaf
images were first acquired and then preprocessed to make them uniform
2.3. Image acquisition setup
and delimit the complications of images. Then the pixel of images is
divided and grouped into several blocks of similar pixels by using the
Fig. 3 shows the image acquisition setup. All the plant’s leaf images
SLIC algorithm [15–17]. In the next step, the initially segmented
were captured with the help of a setup that was installed in a room. This
diseased images were processed by our proposed hybrid neural network
was done to reduce the effect of unwanted information like background
for extracting the region of interest. Also, the labeling of pixels among
details, variable atmospheric conditions, the taxonomy of plants, etc.
different classes has been defined by the level of equivalence. The
Also, to maintain the uniformity and to eliminate the surrounding illu­
different features like color, texture, and shapes were extracted using
mination conditions four white color led lamps were used. The leaf
different algorithms. After feature extraction, a three-way classification
images were first cleaned for the removal of dust. Nikon D5300 digital
strategy has been adopted for the classification between two plant leaf,
camera was used for maintaining the high quality of images. All the
classification among healthy and diseased classes for Jatropha Curcas L.,
devices were wirelessly connected. Then the image acquisition process
and classification among healthy and diseased classes for Pongamia
has been carried out by maintaining the classes of the images differen­
Pinnata L. plant images separately. Finally, the results for segmentation
tiating between the plants, healthy, and diseased leaves. The camera was
and classification were evaluated and compared with different methods
fixed at one position and angle. This was done to maintain the unifor­
by using different metrics.
mity in the database [18].
Algorithm 1. Basic steps for leaf disease segmentation and classification

1. Acquire the healthy and diseased leaves of Jatropha Curcas L. and Pongamia Pinnata L. 3. Proposed hybrid model
2. Image prepossessing like resizing, contrast enhancement, etc.
(continued on next column) In this section, the various steps followed for leaf disease

Fig. 1. Block diagram for the proposed approach.

2
S.S. Chouhan et al. Measurement 171 (2021) 108796

The height and the width was fixed to 10th times proportional to orig­
inal image measurements by using the resize function. This method
upholds the quality of the images. Further, for better analysis, the
contrast of all the images had been enhanced by using the Histogram of
Gradients (HoG) approach [19].

3.1. Initial segmentation using simple linear iterative clustering (SLIC)

Through features like accuracy, effectiveness, and simplicity SLIC

algorithm is the most widely used methodology for superpixel clus­
tering. In this approach, first, all the given RGB images were converted
to LAB color space. This is because, SLIC method adopts a 5-dimensional
compromise of three color channels as L*a*b* and two as pixel co­
ordinates of an image given by x, y [15,20,21]. Using the formula S =
N/k, the grid interval has been calculated, where N corresponds to the
total no. of given pixels in an image, k is the superpixels centers that
were opted at regular grid intervals given by S. Then around the cluster
center, the searching area for pixels in every region has been calculated
by 2S × 2S [15,20,21]. The similarity metrics are obtained by calcu­
lating the distance measures from every pixel to the cluster center. Based
upon this distance measure, every pixel is then associated with its new
nearest center label. The distance measure is computed by
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
(m)2 ̅
Dik = d2lab + d2xy (1)
S

where,
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
dlab = (li − lk )2 + (ai − ak )2 + (bi − bk )2 (2)

and
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
dxy = (xi − xk )2 + (yi − yk )2 (3)

m is the weight factor within the range [1,20]. (li , ai , bi ) is the color
value for pixel i at position (xi , yi ) for the acquired image.
Since, the segmented regions maintain the similarity of attributes
and continuity within every patch, pixels in a cluster are then recognized
by its mean value. This mean value also acts as a mean filter, which in
turn also reduces the sensitivity against noise [17,20,21]. Fig. 4 shows a
sample superpixel clustered image.

3.2. Adaptive Linear Neuron (ADALINE) for image segmentation

A single-layer neural network with adaptive learning was first

introduced by Bernard Widrow and Marcian E. Hoff in 1960 is
commonly known as Adaptive Linear Neuron (ADALINE). It could be
found in nearly every real-time noise filtering and its training is done
using the Widrow-Hoff rule or delta learning rule. The advantages of the
delta learning rule over the perceptron network are its adaptability i.e. if
the difference between target and the actual output of the network is not
small then changes the weight of the network accordingly [22]. Some
nature-inspired algorithm is used for intilialization of weight in ADA­
LINE, input–output relation of this hybrid neural network is linear and
Fig. 2. Shows (a) Jatropha Curcas L. and (b) Pongamia Pinnata L. sites from
the delta learning rule is based on the gradient-descent approach, which
where the leaves are collected.
updates the weights between the connection to minimize the difference
between calculated output and target [22,23].
segmentation and classification have been discussed. Image enhance­
For the training of the ADALINE we consider following notations, let
ment is the most vital step in computer vision methodology. Images,
us assume that input and output patterns of the input layer are I = (II1 ,
when captured in real-time environments, differ in their characteristics
II2 , ..., IIM ) and I = (IO1 ,IO2 ,...,IOM ), where IIi , IOi input and output at the
because of varied environmental conditions. To eliminate this redun­
ith unit at the input layer respectively. Similarly, H = (HI1 , HI2 , ..., HIN )
dancy, in the proposed approach all the images were captured in closed
surroundings. But due to the varied texture, color, and shape of a leaf, and H = (HO1 ,HO2 ,...,HON ), where HIj , HOj input and output at the jth unit
normalization is essential for better understanding and perception of an at the hidden layer respectively also let O = (OI1 , OI2 , ..., OIK ) and O =
image. So to start with, all the acquired images used in this work was (OO1 ,OO2 ,...,OOK ), where OIk , OOk input and output at the kth unit at the
resized to 600 × 400 size using the skimage library of python language. output layer respectively and targets is denoted as T = (TO1 ,TO2 ,...,TOK ),
where, 1 ≤ i ≤ M, 1 ≤ j ≤ N, and 1 ≤ k ≤ K. Also denotes vij is the

3
S.S. Chouhan et al. Measurement 171 (2021) 108796

Fig. 3. Image acquisition setup [18].

∂Ek
= − (TOk − OOk ) (8)
∂OOk

∂OOk
=1 (9)
∂OIk

∂OIk
= HOj (10)
∂wjk
Using Eqs. (8), (9) and (10), Eq. (7) yield

∂E 1∑P
∂
= (T − OOk )2
∂wjk P k=1 ∂wjk Ok
Fig. 4. Initial segmentation using SLIC (The superpixel blocks).
∂Ek
= − (TOk − OOk )HOj (11)
connecting weight between ith unit of the input layer to the jth unit of the ∂wjk
hidden layer, and wjk is the connecting weight between jth hidden unit to Step 2: Update the connecting weight between input to hidden layer.
the kth output unit as shown in Fig. 5. The delta learning rule for weight
adjusting is given by: Δwjk = α(TOk − OOk )IIi . For the training set vij (updated) = vij (previous) + Δvij (12)
(I, T)with cardinalityP, let us consider the cost function given as in Eq. { }
(4) [22,23]: where Δvij = − η ∂∂vEij and
avg
1∑ P
E= (TOk − OOk )2 (4) {
∂E
}
1∑ P
∂Ek
P k=1 =
∂vij avg P k=1 ∂vij
1∑ P
E= (Ek )2 , where Ek = TOk − OOk ∂Ek ∂Ek ∂OOk ∂OIk ∂HOj ∂HIj
P k=1 = (usingchainruleatkth neuron) (13)
∂vij ∂OOk ∂OIk ∂HOj ∂HIj ∂vij
To minimize the cost function, the incremental mode of training is
performed in two steps: ∂Ek
where = − (TOk − OOk ) (14)
Step 1: Update the connecting weight between the hidden and the ∂OOk
output layer.
∂OOk
=1 (15)
wjk (updated) = wjk (previous) + Δwjk (5) ∂OIk

∂E ∂OIk
where Δwjk = − η (6) = wjk (16)
∂wjk ∂HOj
Evaluate the gradient: ∂HOj
= a1 HOj (1 − HOj ); (activationfunctionislog − sigmoidal) (17)
∂Ek ∂Ek ∂OOk ∂OIk ∂HIj
= (usingchainruleatkth neuron) (7)
∂wjk ∂OOk ∂OIk ∂wjk
∂HIj
= IIj (18)
where ∂vij

4
S.S. Chouhan et al. Measurement 171 (2021) 108796

Fig. 5. Adaptive Linear Neuron (ADALINE).

Using Eqs. (14)–(18), Eq. (13) yield accuracy was evaluated by using the two segmentation evaluation
metrics known as specificity and sensitivity [33–36] given by the
∂ Ek
= − a1 (TOk − OOk )HOj (1 − HOj )wjk IIj (19) following equation
∂vjk
TrueNegative
Specificity = (20)
TrueNegative + FalsePositive
3.3. Feature extraction
TruePositive
Three different algorithms had been used for feature extraction. Sensitivity = (21)
TruePositive + FalseNegative
L*a*b* color model [24] is used for color (using a* channel), Scale-
Invariant Feature Transform (SIFT) [25,26] for shape, and Local Bi­ where True Positive = no. of pixels exactly classified, False Positive =
nary Pattern (LBP) [27] has been used for texture features extraction. no. of pixels incorrectly classified, True Negative = no. of pixels exactly
misclassified, and False Negative = no. of pixels incorrectly
3.4. Image classification misclassified.
Segmentation results are given by Figs. 6 and 7 for Jatropha Curcas L.
Seven different commonly known machine learning classification (Leaf spot) and Pongamia Pinnata L. (Colletotrichum leaf blight).
algorithms had been used for plant leaf image classification. These al­ This section presents the various segmentation results for different
gorithms are Logistic Regression (LR) [14], Linear Discriminant Analysis approaches and their comparison with our proposed hybrid ADALINE
(LDA) [14], K-Nearest Neighbours (KNN) [28], Classification and network. Segmentation performance has been validated using Speci­
Regression Trees (CART) [29], Random Forest (RF) [14,16], Naïve ficity and Sensitivity evaluation parameters and results are given by
Bayes (NB) [29], and Support Vector Machine (SVM) [17,30–32] Table 1 and Fig. 8 for Jatropha Curcas L. and Table 2 and Fig. 9 for
respectively. Pongamia Pinnata L respectively. The average Specificity and Sensitivity
in case of Jatropha Curcas L. for randomly chosen eight images from the
4. Experimental evaluation and results database has been evaluated to be 0.8333 and 0.8343 for k-means al­
gorithm [15], 0.8745 and 0.8753 for Multiple Linear Regression Model
4.1. Segmentation results [36], and 0.9495 and 0.9551 respectively for proposed hybrid ADALINE
network. This shows the effectiveness of the proposed hybrid network.
Segmentation of diseases from the leaf images had been accom­ Also to further validate the performance of the network the results were
plished using the proposed superpixel clustering-based hybrid neural evaluated on the complete dataset for Jatropha Curcas L. (124 plant
network. A total number of 124 and 162 leaf images from the database leaves images). In this case, also the proposed network with the average
have been selected for this purpose. The experiments were performed Specificity = 0.9534 and Sensitivity = 0.9637 shows the higher per­
working on a system with 4 GB RAM, 1 TB HDD, and I3 processor in the formance when compared with that of 0.8419 and 0.8496 for the k-
python language environment. The results for the segmentation means algorithm, 0.8798 and 0.8825 for Multiple Linear Regression

5
S.S. Chouhan et al. Measurement 171 (2021) 108796

Fig. 6. Jatropha Curcas L. plant disease segmentation results.

Fig. 7. Pongamia Pinnata L. plant disease segmentation results.

Model respectively. Average specificity and sensitivity with 0.9795 and 0.9805, 0.8906 and
Similarly, for Pongamia Pinnata L. plant leaves first the average 0.8934, 0.8712, and 0.8798 were found to be for the proposed hybrid
specificity and sensitivity have been evaluated on randomly chosen ADALINE network, Multiple Linear Regression Model, and k-means al­
eight images and then on a complete dataset that consists of about 162 gorithm for the complete database. In both cases, the proposed method
plant leaf images. Average specificity and sensitivity for the proposed proves its supremacy over the other state of the art methods. In both
hybrid ADALINE network were found to be 0.9722 and 0.9749, for cases, it has been observed that segmentation accuracy increases with
Multiple Linear Regression Model is 0.8832 and 0.8856, and for the k- the increase in the sample size. Superpixel clustering and level of
means algorithm is 0.8633 and 0.8709 for eight images respectively. equivalence methods defined for the labeling of pixels among healthy

6
S.S. Chouhan et al. Measurement 171 (2021) 108796

Table 1
Jatropha Curcas L. segmentation results.
Leaf sample k-means algorithm Multiple linear regression model Proposed hybrid ADALINE network

Specificity Sensitivity Specificity Sensitivity Specificity Sensitivity

Leaf 1 0.8152 0.8264 0.8756 0.8642 0.9534 0.9547

Leaf 2 0.8215 0.8289 0.8878 0.8810 0.9529 0.9480
Leaf 3 0.8119 0.8210 0.8239 0.8369 0.9315 0.9488
Leaf 4 0.8235 0.8196 0.8315 0.8477 0.9446 0.9585
Leaf 5 0.8491 0.8347 0.8914 0.8962 0.9617 0.9722
Leaf 6 0.8396 0.8412 0.8889 0.8783 0.9540 0.9619
Leaf 7 0.8439 0.8477 0.8953 0.8991 0.9427 0.9539
Leaf 8 0.8620 0.8548 0.9018 0.8987 0.9549 0.9426
Average for 8 samples 0.8333 0.8343 0.8745 0.8753 0.9495 0.9551
Average for 124 samples 0.8419 0.8496 0.8798 0.8825 0.9534 0.9637

Fig. 8. Segmentation results comparison for Jatropha Curcas L.

Table 2
Pongamia Pinnata L. segmentation results.
Leaf sample k-means algorithm Multiple linear regression model Proposed hybrid ADALINE network

Specificity Sensitivity Specificity Sensitivity Specificity Sensitivity

Leaf 1 0.8652 0.8569 0.8856 0.8569 0.9814 0.9896

Leaf 2 0.8536 0.8610 0.8781 0.8798 0.9638 0.9669
Leaf 3 0.8515 0.8597 0.8624 0.8790 0.9834 0.9875
Leaf 4 0.8568 0.8639 0.8710 0.8839 0.9914 0.9825
Leaf 5 0.8614 0.8777 0.8897 0.8877 0.9616 0.9732
Leaf 6 0.8623 0.8812 0.8795 0.8862 0.9533 0.9678
Leaf 7 0.8711 0.8793 0.8970 0.8993 0.9606 0.9525
Leaf 8 0.8841 0.8876 0.9025 0.9116 0.9819 0.9792
Average for 8 samples 0.8633 0.8709 0.8832 0.8856 0.9722 0.9749
Average for 162 samples 0.8712 0.8798 0.8906 0.8934 0.9795 0.9805

Fig. 9. Segmentation results comparison for Pongamia Pinnata L.

and diseased improves the segmentation performance of the proposed Jatropha Curcas L. leaf image or Pongamia Pinnata L. leaf image
method. (including disease and healthy images of both the plants), (ii) to classify
among healthy and disease leaf images of Jatropha Curcas L. plant and,
(iii) to classify the leaf images among healthy and diseases leaf images of
4.2. Classification results
Pongamia Pinnata L. plant. Results for the different cases have been
evaluated on all the seven classifiers and the classification accuracy [35]
The classification results have been validated for three cases
has been calculated by using the given equation
including (i) to classify the given image among two plant classes i.e.

7
S.S. Chouhan et al. Measurement 171 (2021) 108796

Truepositive + TrueNegative
Accuracy =
Truepositive + TrueNegative + FalsePositive + FalseNegative
(22)

4.2.1. To classify among two plant leaf images

In the first phase, the priority of the system is to classify the given
images among two different plant classes i.e. among Jatropha Curcas L.
and Pongamia Pinnata L. All the acquired leaf images was moved into a
single class that contains healthy and diseased leaf images of Jatropha
Curcas L. and Pongamia Pinnata L. Then, the results were evaluated using
seven different classifiers that were used in the proposed work. LR
classifier attains the best result among all the classifiers with an accuracy
of 0.9857 ± 0.0285 (Accuracy ± S.D.). The classification accuracy for
different classifiers has been given in Table 3 and the box plot in Fig. 10
show the variabilities in the accuracy of the results among different
methods.

4.2.2. Classification of Jatropha Curcas L. Leaves among healthy and

disease
Fig. 10. Jatropha Curcas L. and Pongamia Pinnata L. leaves classification.
The foremost part is to identify whether the given leaf image is
healthy or diseased. So, in this phase, the different classifier has been
used to classify the leaf images among healthy and diseased leaf image Table 4
classes of Jatropha Curcas L. plant. The classification accuracy with Jatropha Curcas L. leaf classification results.
standard deviation has been tabularized in Table 4 and the box plot
Methods Accuracy ± S.D.
graph has been depicted for comparison of the variabilities among the
different models in Fig. 11. In this case, the RF classifier attains higher LR 0.8733 ± 0.0882
LDA 0.7276 ± 0.1280
classification accuracy with 0.9095 ± 0.0688. KNN 0.7676 ± 0.0901
CART 0.8161 ± 0.0971
4.2.3. Classification of Pongamia Pinnata L. Leaves among healthy and RF 0.9095 ± 0.0688
disease NB 0.7047 ± 0.1572
SVM 0.4023 ± 0.1040
The last phase of our ternary classification model is again classifi­
cation of given leaf images among the healthy and diseased class, but
this time the results were evaluated for Pongamia Pinnata L. plant im­
ages. The results for different classifiers are given in Table 5. In this case,
LR and RF with an accuracy of 0.9607 ± 0.0256 attain equivalent results
for classification. The box plot for the different classifiers is given in
Fig. 12.

5. Discussion

Because of various inconsistent features like shape, size, texture,

color, etc. of plant leaves and their diseases, the task of segmentation
and classification is very difficult to perform. Also the disadvantages of
background details, temperature, complex plant structure, etc. affect the
performance of methods. Many researchers and academicians had pro­
posed different methods to solve this problem. The images of leaves can
be captured in the close environment so to limit the various environ­
mental effects. The traditional approaches like k-means algorithm when
used for segmentation purposes were not capable enough to provide
accurate results. So, improvement in results can be observed when the
images are initially segmented. In the proposed work, this was achieved Fig. 11. JCL leaves healthy and diseased classification result.
with the help of the SLIC algorithm. As a result higher segmentation
performance for the proposed ADALINE hybrid network had been
observed. But the performance of the network can further be improved Table 5
by optimizing the connection weights between the input and hidden Pongamia Pinnata L. leaf classification results.
Methods Accuracy ± S.D.
Table 3 LR 0.9607 ± 0.0256
Leaf classification. LDA 0.7983 ± 0.1306
KNN 0.9385 ± 0.0388
Methods Accuracy ± S.D.
CART 0.8993 ± 0.0486
LR 0.9857 ± 0.0285 RF 0.9607 ± 0.0256
LDA 0.9228 ± 0.0489 NB 0.8316 ± 0.0615
KNN 0.9504 ± 0.0455 SVM 0.5405 ± 0.2054
CART 0.9019 ± 0.0986
RF 0.9790 ± 0.0320
NB 0.7385 ± 0.0894
SVM 0.4023 ± 0.1040

8
S.S. Chouhan et al. Measurement 171 (2021) 108796

Acknowledgement

This work was supported by grant No. 8-68/FDC/RPS (POLICY-1/

2019/20) from the All India Council of Technical Education (AICTE),
India.

References

[1] E. Venturino, et al., A model for the control of the mosaic virus disease in Jatropha
curcas plantations, Energ. Ecol. Environ. 1 (6) (2016) 360–369, https://doi.org/
10.1007/s40974-016-0033-8.
[2] D. Harreh, et al., An experimental investigation of karanja biodiesel production in
Sarawak, Malaysia, J. Eng. (2018), https://doi.org/10.1155/2018/4174205.
[3] P. More, et al., Geminiviruses: Molecular biodiversity and global distribution in
Jatropha, Physiol. Mol. Plant Pathol., vol. 108, doi: 10.1016/j.pmpp.2019.101439.
[4] L.M.R. Al Muqarrabun, et al., Medicinal uses, phytochemistry and pharmacology of
Pongamia pinnata (L.) Pierre: A review, J. Ethnopharmacol. 150 (2013) 395–420.
[5] D.I. Patricio, R. Rieder, Computer vision and artificial intelligence in precision
agriculture for grain crops: A systematic review, Comput. Electron. Agric. 153
(2018) 69–81, https://doi.org/10.1016/j.compag. 2018.08.001.
[6] S.S. Chouhan, et al., Applications of computer vision in plant pathology: a survey,
Fig. 12. PPL leaves healthy and diseased classification result.
Arch. Comput. Methods Eng. 27 (2020) 611–632, https://doi.org/10.1007/
s11831-019-09324-0.
layer of the network. This can be done using some nature-inspired al­ [7] A. Cruza, et al., Detection of grapevine yellows symptoms in Vitis vinifera L. with
artificial intelligence, Comput. Electron. Agric. 157 (2019) 63–76, https://doi.org/
gorithm. Improving the learning rate of the network will eventually 10.1016/j.compag.2018.12.028.
increase the segmentation accuracy. Also, for the classification results, [8] J. Waldchen, P. Mader, Plant species identification using computer vision
most of the machine learning approaches achieve higher performance, techniques: a systematic literature review, Arch. Comput. Methods Eng. 25 (2018)
507–543, https://doi.org/10.1007/s11831-016-9206-z.
but SVM suffers to attain better results. The accuracy of SVM can further [9] Yogesh, et al., Computer vision based analysis and detection of defects in fruits
be improved by parameter tuning and adopting different activation causes due to nutrients deficiency, Cluster Computing (2019), https://doi.org/
functions. Increasing the sample size may further increase the accuracy 10.1007/s10586-019-03029-6.
[10] L. Gao, X. Lin, Fully automatic segmentation method for medicinal plant leaf
of the algorithms. For future advancement, a convolutional neural images in complex background, Comput. Electron. Agric. 164 (2019), https://doi.
network can be used for validating the performance of the proposed org/10.1016/j.compag.2019.104924.
model. [11] S.K. Behera, et al., Identification, classification & grading of fruits using machine
learning & computer intelligence: a review, J. Ambient Intell. Hum. Comput.
(2020), https://doi.org/10.1007/s12652-020-01865-8.
6. Conclusion and future scope [12] S.S. Chouhan, et al., Bacterial Foraging Optimization Based Radial Basis Function
Neural Network (BRBFNN) for Identification and Classification of Plant Leaf
Diseases: An Automatic Approach Towards Plant Pathology, IEEE Access 6 (2018)
The proposed work is focused upon the leaf disease segmentation 8852–8863, https://doi.org/10.1109/ACCESS.2018.2800685.
and classification for Jatropha Curcas L. and Pongamia Pinnata L. plants [13] S.K. Muruganandham, et al., Study on leaf segmentation using K-means and K-
respectively. First, the disease region from the images has been medoid clustering algorithm for identification of disease, Indian J. Public Health
Res. Dev. 9 (2018) 289–293, https://doi.org/10.5958/0976-5506.2018.00456.4.
segmented using a proposed superpixel clustering-based hybrid neural [14] G. Dhingra, et al., A novel computer vision based neutrosophic approach for leaf
network. Based on the features like color, texture, and shape the images disease identification and classification, Measurement 135 (2019) 782–794,
were classified among different classes. The average segmentation re­ https://doi.org/10.1016/j.measurement.2018.12.027.
[15] S. Zhang, et al., Plant diseased leaf segmentation and recognition by fusion of
sults for the proposed approach were found to be Specificity = 0.9534 superpixel, K-means and PHOG, Optik 157 (2018) 866–872, https://doi.org/
and Sensitivity = 0.9637 for the Jatropha Curcas L. plant leaf images and 10.1016/j.ijleo.2017.11.190.
Specificity = 0.9795 and Sensitivity = 0.9805 for the Pongamia Pinnata [16] P. Wang, et al., An maize leaf segmentation algorithm based on image repairing
technology, Comput. Electron. Agric. 172 (2020), https://doi.org/10.1016/j.
L. plant leaf images show higher segmentation performance when compag.2020.105349.
compared with other methods. LR with classification accuracy = 0.9857 [17] S. Zhang, et al., Fusion of superpixel, expectation maximization and PHOG for
± 0.0285 achieves higher classification accuracy in the first case, RF recognizing cucumber diseases, Comput. Electron. Agric. 140 (2017) 338–347,
https://doi.org/10.1016/j.compag. 2017.06.016.
with classification accuracy = 0.9095 ± 0.0688 achieves higher classi­
[18] S.S. Chouhan, et al., A data repository of leaf images: practice towards plant
fication accuracy in the second case, and LR, RF with classification ac­ conservation with plant pathology, 2019 4th International Conference on
curacy = 0.9607 ± 0.0256 achieves superior performance when Information Systems and Computer Networks (ISCON), Mathura, India, 2019, pp.
evaluated with other approaches. Using some nature-inspired algorithm 700-707, doi: 10.1109/ISCON47742.2019.9036158.
[19] U.P. Singh, et al., Multilayer Convolution Neural Network for the Classification of
to optimize the connection weights of the network, thus improving the Mango Leaves Infected by Anthracnose Disease, IEEE Access 7 (2019)
learning rate to increase the segmentation performance and adopting a 43721–43729, https://doi.org/10.1109/ACCESS.2019.2907383.
deep neural network for classification can be a part of the future work. [20] M. Ye, et al., Crop feature extraction from images with probabilistic superpixel
Markov random field, Comput. Electron. Agric. 114 (2015) 247–260, https://doi.
org/10.1016/j.compag.2015.04.010.
CRediT authorship contribution statement [21] Y. Sun, et al., SLIC_SVM based leaf diseases saliency map extraction of tea plant,
Comput. Electron. Agric. 157 (2019) 102–109, https://doi.org/10.1016/j.
compag.2018.12.042.
Siddharth Singh Chouhan: Conceptualization, Writing - original [22] S. Kumar, Neural Networks: A Classroom Approach, Tata McGraw-Hill Education,
draft, Data curation. Uday Pratap Singh: Methodology, Validation, ISBN: 0070482926, 2004.
[23] R.V. Rao, et al., Teaching–learning-based optimization: A novel method for
Writing - review & editing. Utkarsh Sharma: Software, Visualization. constrained mechanical design optimization problems, Comput. Aided Des. 43
Sanjeev Jain: Supervision. (2011) 303–315, https://doi.org/10.1016/j.cad.2010.12.015.
[24] G. Garcia-Mateos et al., “Study and comparison of color models for automatic
image analysis in irrigation management applications,” Agricultural Water
Management, vol. 151, pp. 158-166, 2015, doi: 10.1016 /j.agwat.2014.08.010.
Declaration of Competing Interest [25] N. Charfi, et al., Bimodal biometric system for hand shape and palmprint
recognition based on SIFT sparse representation, Multimed Tools Appl 76 (2017)
The authors declare that they have no known competing financial 20457–20482, https://doi.org/10.1007/s11042-016-3987-9.
[26] S.S. Chouhan, et al., Web facilitated Anthracnose disease segmentation from the
interests or personal relationships that could have appeared to influence leaf of Mango tree using Radial basis function (RBF) neural network, Wireless Pers.
the work reported in this paper. Commun. (2020), https://doi.org/10.1007/s11277-020-07279-1.

9
S.S. Chouhan et al. Measurement 171 (2021) 108796

[27] H. Ali, et al., Symptom based automated detection of citrus diseases using color [32] E. Omrani, et al., Potential of radial basis function-based support vector regression
histogram and textural descriptors, Comput. Electron. Agric. 138 (2017) 92–104, for apple disease detection, Measurement (2014), https://doi.org/10.1016/j.
https://doi.org/10.1016/j.compag. 2017.04.008. measurement.2014.05.033.
[28] S. Prasad, et al., Multi-resolution mobile vision system for plant leaf disease [33] S.S. Chouhan, et al., Image segmentation using fuzzy competitive learning based
diagnosis, SIViP 10 (2016) 379–388, https://doi.org/10.1007/s11760-015-0751-y. counter propagation network, Multimedia Tools and Applications 78 (2019)
[29] J. Han, M. Kamber, Data mining: concepts and techniques, Morgan Kaufmann 35263–35287, https://doi.org/10.1007/s11042-019-08094-y.
Publishers, Third Edition, ISBN: 978-0-12-381479-1, 2012. [34] S.S. Chouhan, et al., Soft computing approaches for image segmentation: a survey,
[30] X. Qiao, et al., Underwater sea cucumber identification based on principal Multimedia Tools and Applications 77 (2018) 28483, https://doi.org/10.1007/
component analysis and support vector machine, Measurement 133 (2019) s11042-018-6005-6.
444–455, https://doi.org/10.1016/j.measurement.2018.10.039. [35] S.S. Chouhan, et al., Image segmentation using computational intelligence
[31] S. Sabzi, et al., Using video processing to classify potato plant and three types of techniques: review, Arch. Comput. Methods Eng. 26 (2019) 533–596, https://doi.
weed using hybrid of artificial neural network and partincle swarm algorithm, org/10.1007/s11831-018-9257-4.
Measurement 126 (2018) 22–36, https://doi.org/10.1016/j. [36] I.P. Sutawinaya, et al., Comparison of adaline and multiple linear regression
measurement.2018.05.037. methods for rainfall forecasting, 2nd International Joint Conference on Science and
Technology (IJCST), J. Phys.: Conf. Series 953, 2017, doi: 10.1088/1742-6596/
953/1/012046.

10