INDIAN JOURNAL OF SCIENCE AND TECHNOLOGY

RESEARCH ARTICLE

Enhancing Plant Disease Detection

using Advanced Deep Learning Models
Kamaldeep Kaur1 ∗ , Khusboo Bansal2

OPEN ACCESS 1 Research Scholar, Desh Bhagat University, Mandi Gobindgarh, Punjab, India
2 Associate Professor, Department of Engineering and Applied Science, Desh Bhagat
Received: 24-02-2024 University, Mandi Gobindgarh, Punjab, India
Accepted: 01-04-2024
Published: 19-04-2024
Abstract
Objective: The aim of this paper is to analyze the efficacy of employing multiple
Citation: Kaur K, Bansal K (2024) advanced convolutional neural networks (CNNs) for the purpose of enhancing
Enhancing Plant Disease Detection
using Advanced Deep Learning
the accuracy in detecting and classifying various plant diseases. Methods:
Models. Indian Journal of Science The research involves the analysis of 7623 training images as well as 1906
and Technology 17(17): 1755-1766. validation images of different plant diseases and employed advanced deep
https://doi.org/
10.17485/IJST/v17i17.536
learning models like DenseNet169, Xception, InceptionV3, MobileNetV2, and
∗ ResNet50V2 to classify them. At first the RGB images are converted to Grayscale
Corresponding author.
and later to enhance their quality, few techniques have been used such as Otsu
[email protected] thresholding, noise removal, distance transform, and watershed techniques.
Funding: None Subsequently, contour features are extracted by calculating morphological
Competing Interests: None values to obtain the necessary region that correspond to diseased areas in
Copyright: © 2024 Kaur & Bansal.
plant images. Findings: On evaluating the performance of the applied models
This is an open access article on the basis of various metrics, MobileNetV2 and ResNetV2 achieved the
distributed under the terms of the highest validation accuracy scores of 99.42% each, with their respective loss
Creative Commons Attribution
License, which permits unrestricted
values of 0.19 and 0.49. In terms of recall, precision, and F1 score, all models,
use, distribution, and reproduction except MobileNetV2 and InceptionV3, attained optimal scores of 0.99 each.
in any medium, provided the Novelty: The novelty of this paper resides in the incorporation of multiple
original author and source are
credited. image segmentation techniques with fine tuning the parameters of advanced
Convolutional Neural Network (CNN) models on the basis of various factors
Published By Indian Society for
Education and Environment (iSee) such as the number of images, size, channel, classes etc to generate the optimal
results.
ISSN
Print: 0974-6846 Keywords: Agriculture; Plant diseases; Artificial Intelligence; Advanced CNN
Electronic: 0974-5645 models; Watershed Technique

1 Introduction
Plant diseases have always been a threat to the global food security despite having
significant steps in the technology of agricultural sector. These diseases are responsible
of disrupting supply chains, jeopardizing crop yields, as well as escalate the growth of
malnutrition and hunger worldwide. There have been various traditional techniques to
combat such diseases, but they rely on human inspection, call for excessive labor, and
consume lot of time (1,2) . In addition to this, it is a known fact that agriculture systems
have become complex at a large scale which requires the urgency towards the reliable

https://www.indjst.org/ 1755
Kaur & Bansal / Indian Journal of Science and Technology 2024;17(17):1755–1766

as well as efficient detection method. Hence, in light of these pressing concerns, there is a critical need to harness emerging
technologies and innovative approaches to bolster plant disease management efforts (3,4) . Among these, the application of AI
based learning techniques holds promise for revolutionizing disease detection in agriculture. By automating the process of
analyzing plant images and identifying disease symptoms, AI-based systems offer the potential to enhance diagnostic accuracy,
streamline decision-making processes, and ultimately mitigate the socioeconomic impacts of plant diseases on a global scale (5,6) .
From the past years, a number of research studies have put forth various machine learning and deep learning methodologies
which aimed at the detection of various disease-based plants. Deep Learning models like MobileNet architectures and
Convolutional Neural Networks were used to accurately and swiftly identify plant illnesses by Khalid and Karan (2024) (7) . They
employed GradCAM and eXplainable Artificial Intelligence (XAI) to visualize plant disease signs. That clarified model decision-
making. CNN model had 89% accuracy, 96% recall, F1 score, and precision. In contrast, MobileNet has 96% accuracy. Precision,
recall, and F1-score were lower at 90%, 89%, and 89%, respectively. According to Nawaz et al. (2024) (8) proposed CoffeeNet
as a new Deep Learning (DL) model. This methodology was created to address coffee plant leaf disorder identification issues.
To investigate disease-related sample characteristics, they used an upgraded CenterNet technique with a ResNet-50 model
and spatial-channel attention strategy. The study focused on the Arabica coffee leaf repository and its practical and complex
environmental constraints. CoffeeNet produced excellent classification accuracy of 98.54% and 0.97 as mAP. Khanna et al.
(2024) (9) employed Deep Convolutional Neural Network image analysis to identify and classify plant diseases in leaf photos
at different resolutions. PlaNet was compared against 18 popular CNN models, including a top-5 ensemble model. PlaNet
performed well throughout all testing experiments, with an average-wise accuracy of 97.95%, AUC of 0.9752, F1-score of 0.9686,
sensitivity of 0.9707, precision of 0.9576, and specificity of 0.9456. Haridasan et al. (2023) (10) proposed a computer vision-based
approach to reliably detect and classify rice plant illnesses. They used image processing and AI learning to combat fake smut,
bacterial leaf blight, rice blast, brown leaf spot, and sheath rot in Indian rice fields. First, image pre-processing and segmentation
identify the affected area, then convolutional neural networks and support vector machine classifiers classify diseases. Deep
learning had 0.9145 validation accuracy. Javidan et al. (2023) (11) employed SVM to diagnose and classify grape leaf diseases
as black rot, black measles, and leaf blight. K Means clustering technique is used to automatically separate illness symptoms
from healthy leaf portions. HSV, l*a*b, and RGB color models extract features. Principal component analysis for dimension
reduction gave SVM 98.97% accuracy. The suggested technique classifies grape leaf diseases more accurately and faster than
CNN and GoogleNet. Guo et al. (2020) (12) proposed a deep learning-based mathematical model for plant disease diagnosis. The
goal was to improve accuracy, variety, and training efficiency. This method utilizes the region proposal network (RPN) to locate
leaves in complex situations. Using RPN results, the Chan-Vese (CV) algorithm segmented images and extracted symptom
information. The model outperformed conventional methods in bacterial plaque, black rot, and rust disease evaluations with
an accuracy rate of 83.57%. Chen et al. (2020) (13) investigated plant leaf disease identification using deep convolutional neural
networks and transfer learning. Pre-trained models like VGGNet on ImageNet and Inception were used. Pre-trained ImageNet
networks were used to initialize weights. Their method performed well and had an average accuracy of 92.00% for rice plant
picture class prediction even in difficult backdrop settings. Chohan et al. (2020) (14) developed a deep learning-based ”plant
disease detector” to detect plant illnesses using leaf pictures. The method used the PlantVillage dataset for training, augmented
the dataset to boost sample size, and used a CNN with several convolution and pooling layers. After training, the model was
extensively evaluated and obtained 98.3% accuracy in studies utilizing 15% of the PlantVillage dataset, which contained healthy
and diseased plant photos. Ahmed and Yadav (2023) (15) detected plant diseases using linear regression, random forest-nearest
neighbors, neural networks, SVM, and Naive Bayes. The ensemble plant disease model outperformed other suggested and
developed plant disease detection methods. Panigrahi et al. (2020) (16) examined maize plant disease detection using supervised
machine learning. These categorization algorithms were compared to get the best disease prediction model. Other algorithms
were less accurate than the Random Forest algorithm (79.23%). The models were trained to help farmers prevent new plant
illnesses by detecting and classifying them early. In 2020, Jasim et al. (17) developed a deep learning strategy to detect plant leaf
diseases using convolutional neural networks. The study examined tomatoes, peppers, and potatoes using Plant Village dataset
photographs. The CNN classified 15 groups, including 12 disease classes (bacteria, fungi, etc.) and three healthy leaf classes.
They found 98.29% training accuracy and 98.029% testing accuracy across the dataset with their proposed approach. Atila et
al. (2021) (18) suggested classifying plant leaf disease with EfficientNet deep learning. The original and enhanced PlantVillage
datasets containing 55,448 and 61,486 photos were utilized for training. Transfer learning made all layers trainable in the
EfficientNet architecture and other models. EfficientNetB5 had 99.1% accuracy and 98.42% precision in the test dataset, while
EfficientNetB4 had 99.39%.
Albeit, several research have studied the application of deep learning and machine learning approaches for plant disease
detection and classification, but there is still room for improvement.

https://www.indjst.org/ 1756
Kaur & Bansal / Indian Journal of Science and Technology 2024;17(17):1755–1766

Inadequate Performance: Various learning models have been applied by the researchers, but they fail in achieving the
satisfactory scores of accuracy. Despite considerable efforts, many studies have stated the suboptimal performance of their
applied models which showed a significant gap in achieving reliable as well as robust classification outcomes.
Model Comparison and Benchmarking: In the realm of detecting diseases in plants, various models have been used by the
researchers, but they neither measure the performance of the models for the different classes of the plant diseases nor with the
existing techniques.
Limited Model Diversity: Researchers used limited models for the identification as well as classification of plant disease. They
mostly focused on the small subset of the AI learning models and restricts in exploring the usage of more advanced models in
order to generate to optimal results.
By focusing these gaps, this paper focuses on the development of an advanced deep learning model that uses various
classifiers to generate the optimal results as well as enhance the accuracy of plant disease detection and classification. Apart
from this, the paper also addresses the merits and limits of these technologies and their impact on the area of plant pathology.

1.1 Contribution of the work

The following are the contributions made in designing an AI-based system for identifying and classifying diseases in plants:
• This research focuses on using various machine learning and deep learning techniques to accurately detect and classify
plant diseases.
• In order to improve the dataset’s quality, several operations have been employed. These include converting the images to
grayscale and applying Otsu thresholding, noise removal, distance transform, and watershed techniques.
• To classify diseases effectively, contour features have been computed using morphological values to identify significant
characteristics. The purpose of this phase is to accurately define the specific areas in plant images that correspond to
diseased regions.
• Within the domain of convolutional neural networks (CNNs), various advanced architectures have been used, such as
DenseNet169, Xception, InceptionV3, MobileNetV2, and ResNet50V2. The assessment of these models is carried out
using evaluative metrics such as precision, accuracy, recall, loss, as well as F1 score.

1.2 Structure of the manuscript

The manuscript is structured such that Section 1 acts as an Introduction which defines the role of traditional approaches and
AI techniques in identifying plant diseases. Section 2 outlines the approach for constructing the deep learning system, which
involves several phases. Section 3 then evaluates the system based on numerous metrics and in the end Section 4 provides a
comprehensive summary of the entire work.

2 Methodology
The section includes the phases (Figure 1) that have been taken into account during the design of the deep learning model for
identifying and classifying different plant diseases.

Fig 1. Proposed system to detect and classify plant diseases

https://www.indjst.org/ 1757
Kaur & Bansal / Indian Journal of Science and Technology 2024;17(17):1755–1766

2.1 Depiction of the dataset

The dataset included in the article was sourced from the New Plant Diseases dataset, which encompasses a collection of leaf
images from 38 distinct plant categories. This study focuses on four categories of plant diseases (Figure 2): Corn (Maize)
Northern Leaf Blight, Grape Esca (Black Measles), Common Rust (Corn), and Grape Black Rot. The diseases are represented
as RGB photographs, with a total of 9529 images, each measuring 224x224 pixels (19) .

Fig 2. Original sample of plant diseases

2.2 Data Preprocessing

In this process, we convert a color image from the RGB color space to a grayscale image, as shown in Equation (i) (Figure 3).
Each pixel in the grayscale image is given a luminance value that represents its perceived intensity (20) .

Gray scale value = 0.299 x Red + 0.587 x Green + 0.114 x Blue (i)

Fig 3. RGB to grayscale conversion of plant disease images

2.3 Segmentation
In this research work, we have used various techniques to enhance the quality of image such as Otsu thresholding, Noise removal,
distance transformation, and watershed segmentation.
Otsu’s thresholding is applied to convert the image into black and white format (binary) and to identify a threshold which
splits the pixels into two classes i.e. background and foreground (21) . In the context of plant disease identification, Otsu’s
thresholding segments the images of plant leaves or tissues and efficiently distinguishes healthy and unhealthy regions in the
images. The mathematical Equation (ii) for Otsu’s thresholding is as follows:

σb2 (t) = ωb (t) . ω f (t) .[µb (t) . µ f (t)]2 (ii)

Here, σb2 (t) is the inter class variance for a given threshold t, ωb (t) and ω f (t)are the probabilities of the background and
foreground classes, respectively, and µb (t) and µ f (t) are the mean pixel intensities of the background and foreground classes,
respectively.

https://www.indjst.org/ 1758
Kaur & Bansal / Indian Journal of Science and Technology 2024;17(17):1755–1766

Fig 4. Enhanced the visualization of plant disease images through Otsu thresholding techniques

In the context of plant disease detection, noise removal is a crucial preprocessing step to enhance the accuracy of subsequent
analyses. Image noise, such as random variations in pixel values, can adversely affect the performance of algorithms by
introducing spurious information or obscuring meaningful features. Therefore, effective noise removal is essential for improving
the overall reliability of plant disease detection systems (21) . One common method for noise removal is image smoothing or
filtering. Hence, in this work, technique such as Gaussian smoothing has been employed to reduce high-frequency noise while
preserving the essential structures of the image (Figure 5). This filter helps in creating a cleaner and more uniform representation
of the plant tissues or leaves.
The mathematical equation for a 2D Gaussian filter is given by:

x 2 + y2
1 − (iii)
G (x, y) = e 2σ 2
2πσ 2
Here, G (x, y)is the value of the Gaussian function at coordinates (x,y), π refers to the mathematical constant which represents
the ratio of the circumference of a circle to its diameter, and σ refers to the standard deviation of the Gaussian distribution.

Fig 5. Removing noise in plant disease images

After removing noise, distance transform technique is applied to analyze and quantify the severity or extent of diseases
affecting plant structures such as leaves. The distance transform computes the distance of each pixel in an image to the nearest
boundary or edge (21) . In the context of plant images, the boundary could represent the outline of the plant structure or the
contour of a diseased region (Figure 6). The Euclidean distance transform is a commonly used method, and its mathematical
expression for a 2D image can be represented as follows:
Let I (x, y) be the binary input image where I (x, y) = 1 for foreground (object) pixels and I (x, y) = 0 for background pixels.
The distance transform D(x,y) is given by :
√ 2 2
D (x, y) = ∑i, j (i − x) + ( j − y) (iv)

Here, the sum is over all foreground pixels (i,j)in the binary image I(x,y). The distance D(x,y) represents the Euclidean distance
from the pixel(x,y) to the nearest foreground pixel.

https://www.indjst.org/ 1759
Kaur & Bansal / Indian Journal of Science and Technology 2024;17(17):1755–1766

Fig 6. Distance transformation on plant disease images

The last technique used is a watershed segmentation which is useful in separately identifying the affected diseased areas
in plants (Figure 7). It is able to accurately differentiate overlapping as well as adjacent regions which makes it more useful in
instances where boundary delineation is crucial for accurately identifying and analyzing plant diseases (21) .
The general equation for the watershed transformation can be expressed as follows:
Let I(x,y) be the input image, and |∇I | be the gradient magnitude of the image. The watershed function W(x,y) is computed
as the negative of the gradient magnitude:

W (x, y) = −|∇I| (v)

Fig 7. Watershed technique on plant disease images

2.4 Feature Extraction

Various parameters such as area, perimeter, width, height, aspect ratio, etc, as shown in Table 1 are used to provide contour
information to characterize as well as analyze diseased regions in plants for identification and interpretation (21) .

Table 1. Characteristic values of images

Perimeter Grape Esca (Black Corn (maize) Common Corn (maize) Northern Grape Black rot
Measles) rust Leaf Blight
Area 0.0 2.0 0.0 0.0
Perimeter 0.0 5.65 0.0 0.0
Epsilon 0.0 0.565 0.0 0.0
Width 1 3 1 1
Height 1 3 1 1
Equivalent Diameter 1.0 1 1.0 1.0
Extent 0.0 0.22 0.0 0.0
Aspect Ratio 0.0 1.59 0.0 0.0
Max Value 151.0 115.0 134.0 131.0
Continued on next page

https://www.indjst.org/ 1760
Kaur & Bansal / Indian Journal of Science and Technology 2024;17(17):1755–1766

Table 1 continued
Min Value 151.0 155.0 134.0 131.0
Min Value Loc (231,255) (239,249) (43,255) (66,255)
Max value loc (231,255) (238,249) (43,255) (66,255)
Mean color (151.0) (133.8) 134.0 (131.0)
Extreme Leftmost (231,255) (238,249) (43,255) (66,255)
point
Extreme rightmost (231,255) (240,249) (43,255) (66,255)
point
Extreme topmost point (231,255) (239,248) (43,255) (66,255)
Extreme bottommost (231,255) (239,250) (43,255) (66,255)
point

2.5 Applied Classifiers

InceptionV3 uses inception modules with multiple filter sizes, batch normalization, factorized convolution, as well as auxiliary
classifiers for capturing complex features and hierarchical representations. Likewise, the variant of DenseNet architecture,
DenseNet169, has densely connected blocks which receives input from the previous layers, allows the reusing of plant disease
image features, as well as mitigates the vanishing gradient problem (22) . Another classifier is Xception which is a derivative of
Inception architecture. It is also useful in capturing the intricate features of images as it contains depth wise as well as point
wise separable convolutions. The design of the network reduces the computational complexity and allows discerning fine details
present in the image. A variant of ResNet architecture, ResNet50V2 has also showcased its effect in detecting diseases in plants.
It contains residual blocks in stacked form that facilitates to preserve the information of the image via skip connections. At
the end, a lightweight architecture, MobileNetV2, is used which has inverted residuals, linear bottlenecks as well as depth wise
separable convolutions in its architecture. The network also contains shortcut and skip connections that enable the smooth flow
of gradients during training and facilitate the learning of both local and global features (23) . The layered architectures of all the
applied models are presented in Table 2.

Table 2. Parameters of the applied classifiers

Layer Output Shape Param#
Inception_V3 (None,2048) 21802784
Flatten_9 (None,2048) 0
Dense_45 (None,1024) 2098176
Dense_46 (None,512) 131328
Dropout_9 (None, 256) 0
Dense_48 (None, 128) 32896
Dense_49 (None,8) 1032
Total Parameters 24591016
Trainable Parameters 24556584
Non-Trainable Parameters 34432
DenseNet169 (None,1664) 12642880
Flatten_1 (None,1664) 0
Dense_5 (None,1024) 1704968
Dense_6 (None,512) 524800
Dropout_1 (None, 256) 0
Dense_8 (None, 128) 32896
Dense_9 (None,8) 1032
Total Parameters 15037896
Trainable Parameters 14879496
Non-Trainable Parameters 158400
Xception (None,2048) 20861480
Flatten_2 (None,2048) 0
Dense_10 (None,1024) 2098176
Dense_11 (None,512) 524800
Dropout_2 (None,256) 0
Dense_13 (None,128) 32896
Dense_14 (None,8) 1032
Continued on next page

https://www.indjst.org/ 1761
Kaur & Bansal / Indian Journal of Science and Technology 2024;17(17):1755–1766

Table 2 continued
Total Parameters 23649712
Trainable Parameters 23595184
Non-Trainable Parameters 54528
MobileNetV2 (None,1280) 2257984
Flatten_3 (None,1280) 0
Dense_15 (None, 1024) 1311744
Dense_16 (None,512) 524800
Dense_17 (None, 256) 131328
Drouput_3 (None, 256) 0
Dense_18 (None,128) 32896
Dense_19 (None,8) 1032
Total Parameters 4259784
Trainable Parameters 4225672
Non-Trainable Parameters 34112
ResNet50V2 (None,2048) 23564800
Flatten_4 (None,2048) 0
Dense_20 (None,1024) 2098176
Dense_21 (None,512) 524800
Dense_22 (None,256) 131328
Dropout_4 (None,256) 0
Dense_23 (None,128) 32896
Dense_24 (None,8) 1032
Total Parameters 26353032
Trainable Parameters 26307592
Non-Trainable Parameters 45440

2.6 Performance Metrics

When it comes to detecting plant diseases, there is several important performance metrics that help us evaluate how well a
model is working. Accuracy is a fundamental metric which represents the proportion of correctly classified instances among
the total samples. Loss, however, is a way to measure how much the predicted values differ from the actual values. Lower values
indicate that the model is performing better (24) .

True Negative + True Positive

Accuracy = (vi)
True Negative + True Positive + False Positive + False Negative

(Actual Value − Predicted Value)2

Loss = (vii)
Number o f observations

Precision measures the accuracy of positive predictions by measuring the proportion of true positives out of the total of true
positives and false positives. Recall, or sensitivity, measures the model’s capacity to correctly identify all relevant instances by
computing the ratio of true positives to the sum of true positives and false negatives. F1 score represents the harmonic mean
of precision and recall, providing a balanced assessment that is especially valuable when there is an imbalance between disease
and non-disease instances (24) .
True Positive
Precision = (viii)
True Positive + False Positive

True Positive
Recall = (ix)
True Positive + False Negative

Precision ∗ Recall
F1 score = 2 (x)
Recall + Precision

https://www.indjst.org/ 1762
Kaur & Bansal / Indian Journal of Science and Technology 2024;17(17):1755–1766

3 Results and Discussion

To predict and classify plant diseases, the performance metrics of five different architectures were assessed, namely
DenseNet169, Xception, MobileNet V2, ResNet50 V2, and Inception V3 on the basis of training and validation accuracy and
loss (Table 3).

Table 3. Analysis of the applied models for plant disease dataset

Training Validation
Models
Accuracy Loss Accuracy Loss
DenseNet 169 99.11 0.30 99.24 0.001
Xception 99.60 0.32 99.26 0.34
MobileNet V2 99.14 0.34 99.42 0.19
ResNet50 V2 99.49 0.59 99.42 0.49
Inception V3 99.60 0.59 98.49 0.46

The DenseNet169 model exhibited a commendable performance with a training accuracy of 99.11% and a negligible training
loss of 0.30, while on the validation set, it achieved an impressive accuracy of 99.24% with an extremely low loss of 0.001.
Similarly, the Xception model demonstrated a high level of accuracy in both training (99.60%) and validation (99.26%) sets,
accompanied by relatively low losses (0.32 and 0.34, respectively). MobileNet V2 and ResNet50 V2 also performed well, with
the training accuracy above 99% i.e. 99.14% and 99.49% respectively and relatively low losses. But these models computed the
best accuracy at the validation phase with 99.42% each. On the other hand, the Inception V3 model showed a slightly lower
validation accuracy of 98.49% compared to its training accuracy of 99.60%, suggesting a potential issue with generalization.

Table 4. Analyzing the execution of the models

Models Precision Recall F1 Score
DenseNet169 0.99 0.99 0.99
Xception 0.99 0.99 0.99
MobileNet V2 0.97 0.97 0.97
ResNet50 V2 0.99 0.99 0.99
Inception V3 0.98 0.99 0.98

Likewise, as shown in Table 4, DenseNet169, ResNet50V2, and Xception indicated strong classification performance by
obtaining the highest values of precision, recall, and F1 score of 0.99 followed by InceptionV3 with the best recall value of
0.99 but slightly lower value of precision and F1 score with 0.98. On the contrary, MobileNet V2 showed lower performance as
compared to the other models in terms of precision, recall, and F1 score with 0.97.
As shown in Figure 8, for the classes ”Grape___Esca_(Black_Measles)”, ”Corn_(maize)___Common_rust,”and
”Grape___Black_rot,” DenseNet169 achieved perfect scores of 1.00 in precision, recall, and F1and showcase its
ability to accurately identify instances of these diseases in grape and corn plants. On the contrary, for the class
”Corn_(maize)___Northern_Leaf_Blight,” the DenseNet 169 model exhibited little low precision (0.97), recall (0.99), and
F1 score (0.98), which highlightes its effectiveness in distinguishing and classifying instances of Northern Leaf Blight in maize
plants. Likewise, for ”Grape___Esca_(Black_Measles),” Xception model achieved perfect precision (1.00), which indicate accu-
rate positive predictions, along with high recall (0.98) and F1 score (0.99). In the case of ”Corn_(maize)___Common_rust,”
the model shows flawless precision, recall, and F1 scores with 1.00. But regarding ”Corn_(maize)___Northern_Leaf_Blight,”
the model obtained lower value of precision (0.96), recall (0.97), and an overall high F1 score (0.99). For ”Grape___Black_rot,”
the Xception model demonstrates accurate classification with a balanced 0.97 as precision, 1.00 as recall, and 0.98 as
F1 score. In the ”Grape___Esca_(Black_Measles)” class, MobileNetV2 shows high recall (1.00) but slightly lower preci-
sion (0.94) and results in a balanced F1 score of 0.97. For ”Corn_(maize)___Common_rust,” the model achieves flaw-
less precision, recall, and F1 scores which indicates its precise identification of instances of common rust. In the case of
”Corn_(maize)___Northern_Leaf_Blight,” the model displays perfect precision but a lower recall i.e., 0.85 which leads to a
slightly lower F1 score of 0.92. For ”Grape___Black_rot,” the model demonstrates perfect precision (1.00) but a somewhat lower
recall of 0.93 but results in a balanced F1 score of 0.96. The ResNet50V2 model exhibits robust and consistent performance
in classifying various plant diseases. For the ”Grape___Esca_(Black_Measles)” and “Grape___Black_rot” class, the model
obtained 1.00 as precision, F1 score, and recall value respectively which thereby showcases its accuracy in identifying instances

https://www.indjst.org/ 1763
Kaur & Bansal / Indian Journal of Science and Technology 2024;17(17):1755–1766

Fig 8. Performance evaluation of classifiers for different class of plant disease

of this disease in grape plants. In the ”Corn_(maize)___Common_rust” class, the model maintains flawless precision, which
indicates accurate positive predictions, but with a slightly lower recall of 0.99, resulting in a high F1 score of 0.99. Similarly,
for ”Corn_(maize)___Northern_Leaf_Blight,” the model demonstrates a precise classification with a high precision of 0.99,
though with a slightly lower recall of 0.96, thereby results in a commendable F1 score of 0.97. Similarly, the InceptionV3 model
demonstrates strong and precise performance in classifying specific plant diseases. For the ”Grape___Esca_(Black_Measles)”
and ”Grape___Black_rot,” class, the model obtains perfect precision, recall, and F1 score which indicates accurate identifica-
tion and classification of instances of this disease in grape plants. In the ”Corn_(maize)___Common_rust” class, the model
maintains a high precision of 0.99, capturing the majority of true positives with perfect recall and resulting in a commendable
F1 score of 0.99. For ”Corn_(maize)___Northern_Leaf_Blight,” the model exhibits perfect precision and a high recall of 0.99,
though the F1 score is slightly lower at 0.94.
In addition to this, the performance of the applied models has been also compared with the existing ones in Table 5 through
rigorous evaluation and analysis.

Table 5. Comparative Analysis

Ref Dataset Techniques Accuracy Remarks
(4) Plant disease GradCAM+MobileNetV2 96% High processing time
(5) Coffee plant data CoffeeNet 98.54% Limited dataset
(6) Images of healthy and dis- PlaNet 97.95% Lack of generalizability
eased plants
Continued on next page

https://www.indjst.org/ 1764
Kaur & Bansal / Indian Journal of Science and Technology 2024;17(17):1755–1766

Table 5 continued
(7) Real time data CNN+SVM 91.45% Not reliable
(8) Plant Village SVM 98.97% Class imbalance issue
(9) Leaf dataset RPN+CV 83.57% Low detection accuracy
(15) Plant Village Dataset EfficientNetB5 99.1% Fine tuning of model is required
New Plant Diseases ResNet50V2 99.42% Improved accuracy and generalizability
Our Study
Dataset MobileNetV2 99.42% by using multiple plant diseases

4 Conclusion
The paper presents that advanced CNN models has yielded significant quantitative observations and findings in plant disease
detection. Through rigorous experimentation and evaluation, we observed that the ResNet50V2 and MobileNetV2 models
achieved impressive validation accuracy scores of 99.42%. These high accuracy rates demonstrate the effectiveness of using the
advanced deep learning architectures in accurately identifying and classifying various plant diseases. Furthermore, our research
highlighted the importance of preprocessing plant images through operations such as converting to grayscale, applying Otsu
thresholding, noise removal, distance transform, and watershed techniques along with the contour feature extraction. These
preprocessing steps played a crucial role in improving the quality of the dataset and enhancing the performance of the applied
deep learning models. Hence, by incorporating multiple image segmentation techniques and fine-tuning the parameters of
advanced CNN models, we were able to optimize the results and achieve optimal scores for recall, precision, and F1 score. This
novel approach not only enhanced the efficiency of disease detection but also provided valuable insights into the potential
of AI-based systems in revolutionizing plant disease management efforts. However, there are also certain limitations and
challenges like the risk of overfitting as well as use of small data. Addressing these issues through robust validation techniques,
regularization methods, and data augmentation strategies in future will be crucial in ensuring the reliability and effectiveness of
AI-based plant disease detection systems in practical agricultural settings. Apart from this, Explainable AI techniques should
be also integrated to enhance the interpretability of model predictions and develop real-time plant disease monitoring systems
for proactive management. By addressing these limitations and leveraging the future scope of research, the potential for AI
technologies to transform plant health management and agricultural sustainability remains promising.

References
1) Kanna GP, Kumar SJKJ, Kumar YJ, Changela A, Woźniak M, Shafi J, et al. Advanced deep learning techniques for early disease prediction in cauliflower
plants. Scientific Reports. 2023;13(1):1–21. Available from: https://doi.org/10.1038/s41598-023-45403-w.
2) Panchal AV, Patel SC, Bagyalakshmi K, Kumar P, Khan IR, Soni M. Image-based Plant Diseases Detection using Deep Learning. Materials Today:
Proceedings. 2023;80(Part 3):3500–3506. Available from: https://dx.doi.org/10.1016/j.matpr.2021.07.281.
3) Liu J, Wang X. Plant diseases and pests detection based on deep learning: a review. Plant Methods. 2021;17(1):1–18. Available from: https://dx.doi.org/
10.1186/s13007-021-00722-9.
4) Kumar Y, Singh R, Moudgil MR, Kamini. A Systematic Review of Different Categories of Plant Disease Detection Using Deep Learning-Based Approaches.
Archives of Computational Methods in Engineering. 2023;30(8):4757–4779. Available from: https://dx.doi.org/10.1007/s11831-023-09958-1.
5) Naralasetti V, Bodapati JD. Enhancing Plant Leaf Disease Prediction Through Advanced Deep Feature Representations: A Transfer Learning Approach.
Journal of The Institution of Engineers (India): Series B. 2024;10:1–4. Available from: https://dx.doi.org/10.1007/s40031-023-00966-0.
6) Shoaib M, Shah B, Ei-Sappagh S, Ali A, Ullah A, Alenezi FA, et al. An advanced deep learning models-based plant disease detection: A review of recent
research. Frontiers in Plant Science. 2023;14:1–22. Available from: https://doi.org/10.3389/fpls.2023.1158933.
7) Khalid MM, Karan O. Deep Learning for Plant Disease Detection. International Journal of Mathematics, Statistics, and Computer Science. 2023;2:75–84.
Available from: https://dx.doi.org/10.59543/ijmscs.v2i.8343.
8) Nawaz M, Nazir T, Javed A, Amin ST, Jeribi F, Tahir A. CoffeeNet: A deep learning approach for coffee plant leaves diseases recognition. Expert Systems
with Applications. 2024;237(Part A):121481. Available from: https://dx.doi.org/10.1016/j.eswa.2023.121481.
9) Khanna M, Singh LK, Thawkar S, Goyal M. PlaNet: a robust deep convolutional neural network model for plant leaves disease recognition. Multimedia
Tools and Applications. 2024;83(2):4465–4517. Available from: https://dx.doi.org/10.1007/s11042-023-15809-9.
10) Haridasan A, Thomas J, Raj ED. Deep learning system for paddy plant disease detection and classification. Environmental Monitoring and Assessment.
2023;195. Available from: https://dx.doi.org/10.1007/s10661-022-10656-x.
11) Javidan SM, Banakar A, Vakilian KA, Ampatzidis Y. Diagnosis of grape leaf diseases using automatic K-means clustering and machine learning. Smart
Agricultural Technology. 2023;3:1–14. Available from: https://dx.doi.org/10.1016/j.atech.2022.100081.
12) Guo Y, Zhang J, Yin C, Hu X, Zou Y, Xue Z, et al. Plant Disease Identification Based on Deep Learning Algorithm in Smart Farming. Discrete Dynamics
in Nature and Society. 2020;2020:1–11. Available from: https://dx.doi.org/10.1155/2020/2479172.
13) Chen J, Chen J, Zhang D, Sun Y, Nanehkaran YA. Using deep transfer learning for image-based plant disease identification. Computers and Electronics in
Agriculture. 2020;173:105393. Available from: https://dx.doi.org/10.1016/j.compag.2020.105393.
14) Chohan M, Khan A, Chohan R, Katpar SH, Mahar MS, , et al. Plant Disease Detection using Deep Learning. International Journal of Recent Technology
and Engineering (IJRTE). 2020;9(1):909–914. Available from: https://dx.doi.org/10.35940/ijrte.a2139.059120.
15) Ahmed I, Yadav PK. Plant disease detection using machine learning approaches. Expert Systems. 2023;40(5). Available from: https://dx.doi.org/10.1111/
exsy.13136.

https://www.indjst.org/ 1765
Kaur & Bansal / Indian Journal of Science and Technology 2024;17(17):1755–1766

16) Panigrahi KP, Das H, Sahoo AK, Moharana SC. Maize Leaf Disease Detection and Classification Using Machine Learning Algorithms. In: Progress in
Computing, Analytics and Networking;vol. 1119 of Advances in Intelligent Systems and Computing. Singapore. Springer. 2020;p. 659–669. Available from:
https://doi.org/10.1007/978-981-15-2414-1_66.
17) Jasim MA, Al-Tuwaijari JM. Plant Leaf Diseases Detection and Classification Using Image Processing and Deep Learning Techniques. In: 2020
International Conference on Computer Science and Software Engineering (CSASE). IEEE. 2020;p. 259–265. Available from: https://doi.org/10.1109/
CSASE48920.2020.9142097.
18) Atila Ü, Uçar M, Akyol K, Uçar E. Plant leaf disease classification using EfficientNet deep learning model. Ecological Informatics. 2021;61:101182. Available
from: https://dx.doi.org/10.1016/j.ecoinf.2020.101182.
19) Bhattarai S. New Plant Diseases Dataset. . Available from: https://www.kaggle.com/datasets/vipoooool/new-plant-diseases-dataset.
20) Bansal K, Batla RK, Kumar Y, Shafi J. Artificial Intelligence Techniques in Health Informatics for Oral Cancer Detection. In: Connected e-Health;vol.
1021 of Studies in Computational Intelligence. Springer, Cham. 2022;p. 255–279. Available from: https://doi.org/10.1007/978-3-030-97929-4_11.
21) Koul A, Bawa RK, Kumar Y. An Analysis of Deep Transfer Learning-Based Approaches for Prediction and Prognosis of Multiple Respiratory Diseases
Using Pulmonary Images. Archives of Computational Methods in Engineering. 2024;31(2):1023–1049. Available from: https://dx.doi.org/10.1007/s11831-
023-10006-1.
22) Hammou DR, Boubaker M. Tomato Plant Disease Detection and Classification Using Convolutional Neural Network Architectures Technologies. In:
Networking, Intelligent Systems and Security;vol. 237 of Smart Innovation, Systems and Technologies. Singapore. Springer. 2022;p. 33–44. Available from:
https://doi.org/10.1007/978-981-16-3637-0_3.
23) De Silva M, Brown D. Plant disease detection using multispectral imaging. In: International Advanced Computing Conference: IACC 2022;vol. 1781 of
Communications in Computer and Information Science. Springer, Cham. 2022;p. 290–308. Available from: https://doi.org/10.1007/978-3-031-35641-4_
24.
24) Koul A, Bawa RK, Kumar Y. Enhancing the detection of airway disease by applying deep learning and explainable artificial intelligence. Multimedia Tools
and Applications. 2024;p. 1–33. Available from: https://dx.doi.org/10.1007/s11042-024-18381-y.

https://www.indjst.org/ 1766