Artificial Intelligence in Agriculture 11 (2024) 34–49

Contents lists available at ScienceDirect

Artificial Intelligence in Agriculture

journal homepage: http://www.keaipublishing.com/en/journals/artificial-
intelligence-in-agriculture/

DeepRice: A deep learning and deep feature based classification of Rice

leaf disease subtypes
P. Isaac Ritharson a, Kumudha Raimond a, X. Anitha Mary b, Jennifer Eunice Robert c,⁎, Andrew J d,⁎
a
Department of CSE Karunya Institute of Technology and Sciences, Coimbatore, Tamilnadu, India
b
Department of Robotics Engineering, Karunya Institute of Technology and Sciences, Coimbatore, India
c
Department of Mechatronics Engineering, Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, Karnataka 576104, India
d
Department of Computer Science and Engineering, Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, Karnataka 576104, India

a r t i c l e i n f o a b s t r a c t

Article history: Rice stands as a crucial staple food globally, with its enduring sustainability hinging on the prompt detection of
Received 26 August 2023 rice leaf diseases. Hence, efficiently detecting diseases when they have already occurred holds paramount impor-
Received in revised form 14 November 2023 tance for solving the cost of manual visual identification and chemical testing. In the recent past, the identification
Accepted 17 November 2023
of leaf pathologies in crops predominantly relies on manual methods using specialized equipment, which proves
Available online 23 November 2023
to be time-consuming and inefficient. This study offers a remedy by harnessing Deep Learning (DL) and transfer
Keywords:
learning techniques to accurately identify and classify rice leaf diseases. A comprehensive dataset comprising
Rice leaf disease 5932 self-generated images of rice leaves was assembled along with the benchmark datasets, categorized into
Deep learning 9 classes irrespective of the extent of disease spread across the leaves. These classes encompass diverse states in-
CNN cluding healthy leaves, mild and severe blight, mild and severe tungro, mild and severe blast, as well as mild and
Crop yield severe brown spot. Following meticulous manual labelling and dataset segmentation, which was validated by
Agriculture horticulture experts, data augmentation strategies were implemented to amplify the number of images. The
Food security datasets were subjected to evaluation using the proposed tailored Convolutional Neural Networks models.
Their performance are scrutinized in conjunction with alternative transfer learning approaches like VGG16,
Xception, ResNet50, DenseNet121, Inception ResnetV2, and Inception V3. The effectiveness of the proposed cus-
tom VGG16 model was gauged by its capacity to generalize to unseen images, yielding an exceptional accuracy of
99.94%, surpassing the benchmarks set by existing state-of-the-art models. Further, the layer wise feature extrac-
tion is also visualized as the interpretable AI.
© 2023 The Authors. Publishing services by Elsevier B.V. on behalf of KeAi Communications Co., Ltd. This is an open
access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction security for millions of people around the world (Damalas and
Koutroubas, 2016; Tudi et al., 2021) Spotting the severity of the disease
According to the Food and Agriculture Organization (FAO) (Asibi in rice plants is often determined by the extent and spread of the infec-
et al., 2019; Taw, 2022), rice is the primary food source for more than tion over the surface area of the leaves (Tudi et al., 2021; Bock et al.,
half of the world's population and damage to crops result in substantial 2022) and (Liu et al., 2008). A mild case of the disease is characterized
losses to farmers worldwide. These losses can have serious economic, by less than 10% damage to the leaf surface, while severe cases are de-
social, and environmental impacts, as they affect food security, poverty fined by more than 10% damage. It is significant because it directly im-
reduction efforts, and sustainable development. Further, International pacts the growth, development, and yield of rice crops (Shoaib et al.,
Rice Research Institute (IRRI) states that, rice diseases can deteriorate 2023) Moreover, spotting this rice disease manually seems tiring and
yields by up to 80%, leading to a decline in farm incomes, lower food does not provide early detection of diseases, which can lead to signifi-
availability, and higher prices for consumers (Shew et al., 2019). cant yield losses (Andrew et al., 2022). Accurate, real-time identification
Which further results to the use of harmful pesticides and chemicals, and categorization of crop diseases are made possible by using machine
which can have negative impacts on human health and the environ- learning and deep learning techniques, improving agricultural produc-
ment (Jones, 2021). Therefore, effective management of rice leaf dis- tivity and quality. It simultaneously helps the process become more ac-
eases is crucial for ensuring sustainable rice production and food curate overall and reduces labour costs (Aggarwal et al., 2022). Recent
advances in DL, particularly in the field of image processing, have
⁎ Corresponding authors. shown great potential in the early identification and classification of
E-mail address: [email protected] (J.E. Robert). plant diseases. Convolutional Neural networks (CNN) have been

https://doi.org/10.1016/j.aiia.2023.11.001
2589-7217/© 2023 The Authors. Publishing services by Elsevier B.V. on behalf of KeAi Communications Co., Ltd. This is an open access article under the CC BY-NC-ND license (http://
creativecommons.org/licenses/by-nc-nd/4.0/).
P.I. Ritharson, K. Raimond, X.A. Mary et al. Artificial Intelligence in Agriculture 11 (2024) 34–49

successful in identifying and classifying plant diseases based on images counts to improve the model's accuracy and generalization capacity.
of plant leaves. The use of CNNs in the early identification and classifica- The model implementation, specifications, hyperparameter tuning,
tion of rice leaf diseases can provide a reliable and quick way of detect- and data augmentation techniques of the proposed models are pre-
ing and diagnosing plant diseases, enabling timely interventions to sented in section 4. In the evaluation, we analyze the performance of
minimize yield loss (Lu et al., 2017). Compared to traditional ap- our model on diverse datasets, compare it with baseline models, and as-
proaches, DL models offer several advantages, including high accuracy, sess overfitting concerns. Results and discussions are presented under
speed, and the ability to process large amounts of data. This has signifi- section 5 wherein a detailed analysis of the model's comparative perfor-
cant implications for agriculture, as the early detection and manage- mance is done, highlighting its capability in feature extraction and spa-
ment of plant diseases are crucial for maintaining crop yield and tial hierarchy, interpretability of the Proposed custom VGG 16 model.
ensuring food security. Several studies support this fact that DL models Further, The contribution along with their broader implications in the
outperform traditional methods in identifying and classifying plant dis- field of agriculture and the future research directions are presented
eases (Javidan et al., 2023; Kumar et al., 2023). These models can effec- under section 6.
tively detect plant diseases at an early stage, providing a quick and
reliable way of detecting and diagnosing plant diseases [Andreas 2. Related works
et al.2018]. In addition to this, the precision and resilience of the DL
models can be further improved using pre-trained models and data aug- A 37% annual drop in rice yield is a consequence of rice plant dis-
mentation approaches, thereby rendering them more trustworthy for eases. Although there hasn't yet been a suitable application created
use in practical applications. Therefore, (Wang et al., 2023) in his that is capable of precisely diagnosing these rice plant diseases and
paper, explores the use of deep learning techniques, specifically CNNs, controlling those diseases, it may occur primarily because of a lack of
to detect and classify the severity of rice leaf diseases and evaluate the understanding in identifying and controlling rice plant diseases. How-
performance analysis of the transfer learning approaches. Furthermore, ever, there have only been a few studies done on the diagnosis of ill-
(Aggarwal et al., 2023a) in his work, he proposed InceptionResNet V2 nesses affecting rice plants. This section provides an insight to the
for the classification of rice plant disease. He evaluated the following recent literature summary on the existing deep learning models in
categories of rice leaf diseases namely bacterial leaf blight, blast, rice disease identification and the state-of-art which exists in classify-
Brown spot using his proposed model and achieved an accuracy of ing the rice plants disease. (Bari et al., 2021) used Faster R-CNN algo-
88% in classifying those diseases. In this article, we propose a novel ar- rithm to distinguish blast, brown spot and Hispa and identified
chitecture of VGG 16 mitigating the limitations observed in prior healthy leaves with 99.25% accuracy. The proposed models were eval-
works in plant disease detection and classification. In this proposed cus- uated using self-generated database and Kaggle with a total of 2400 im-
tomized VGG 16 architecture, we introduce critical changes, including ages. Caffe DL approach were used where feature maps of infected
the strategic incorporation of supplementary dropout layers, dense leaves are used for training purposes; An original Rice Leaf Disease
layers, and the fine-tuning of filter counts in the convolutional layers. Dataset (RLDD) was constructed from both an online database and
Through this meticulous customization, our proposed architecture their own dataset.
demonstrates significantly improved generalization capability and (Bari et al., 2021; Pandian et al., 2022) developed a ResNet19 and
heightened accuracy, effectively addressing common challenges such employed evolutionary search technique for optimising its layers. Aug-
as overfitting and elevating the model's capacity to recognize intricate mentation techniques such as scaling, cropping, flipping, padding rota-
patterns and abstract features within the images. Consequently, our tion, etc. were used to create additional data from the existing images.
proposed approach presents a substantiated solution with the existing The ResNet197 model detects various plant leaf diseases with the
state -of-art models for advancing reliable and efficient plant disease de- input image of size 224 × 224 × 3 pixels. This ResNet197 model used
tection in diverse agricultural settings. The key contribution of our work six blocks of layers and it was trained on a combined dataset consisting
is briefed below: of 154,500 images from 22 plants which included both healthy and dis-
eased leaves. While training this model in a GPU environment for up to
• We enhance VGG 16 with novel layers for better generalization in rice
1000 epochs resulted in an average classification accuracy rate of 99.58%
leaf detection. We also introduce a curated dataset of infected rice
which is better than other existing architectures or transfer learning
leaves, categorized by infection severity, aiding research in agriculture
methods.
and plant disease detection.
(Roy and Bhaduri, 2021) designed a model to address early disease
• For the subcategorization of the self-assembled repository, the data
detection and optimized both speed and accuracy in detecting apple
cleaning, labelling, and sorting is carried out using filtering, Region
diseases under complex orchard scenarios. The mean average precision
of Interest and thresholding.
(mAP) achieved was 91.2% with an F1 score of 95.9%, at a rate of 56.9 FPS
• Furthermore, to elevate the model's generalizing ability we have in-
(frames per second). Compared to existing models, the proposed model
troduced the concept data augmentation such as tilting, rotating, blur-
for multi-class plant disease detection showed significant improvement
ring as pre-processing, and making the proposed architecture more
in mAP by 9%, 0.05% and F1 Score by 7.6%. (Saberi Anari, 2022) used
reliable to other unseen data.
Model Engineering (ME) learning to classify diseased leaves. The ME
• Experimental analyses is carried on the curated datasets and the re-
learning technique employed a combination of deep transfer learning,
sults are validated with standard performance metrics. Further, we
multiple support vector machine (SVM) models as shown in Fig. 4,
performed comparative analysis with state-of-the-art literature.
and radial basis function to extract features from the images. The com-
bination of multiple techniques helped to improve the recognition of
leaf diseases, including k-NN, DT, NN, SVM-L, SVM-RBF, and ensemble
The outline of this paper is discussed under 6 sections. Under models. All the techniques were used in combination to extract features
Section 1, we provide the background and significance of plant disease from the data and classify the images. The deep transfer learning model
detection and review relevant literature to identify existing limitations. used was a modified version of a deep CNN that was used to extract fea-
Section 2 discusses the previous work, architectures used, their poten- tures from images of leaves on various fruits. In a study, (Deng et al.,
tial limits and the drawbacks. Section 3 briefs the methodology and 2021) collected 33,026 images with six varieties of rice leaf diseases
elaborates the steps involved in data collection and pre-processing, and used to train and test five sub-models. The information about the
along with a detailed overview of the VGG-16 architecture and the de- methods used, the dataset used, the camera used to capture data, the
sign of our Proposed custom VGG 16 model. We describe the incorpora- number of observations, the learning rate, the number of iterations,
tion of dropout and dense layers, as well as the fine-tuning of filter and the performance of the model are also briefed.

35
P.I. Ritharson, K. Raimond, X.A. Mary et al. Artificial Intelligence in Agriculture 11 (2024) 34–49

(Zhou et al., 2023) used a combination of FCM-KM and Faster R-CNN improvement of 3.1% producing an accuracy of 90.63%. Another ap-
fusion to diagnose rice leaf diseases. The number of observations was proach presented by (Aggarwal et al., 2023a, 2023b, 2023c) CNN's wa-
3010 with a learning rate of 0.001 and 15,000 number of iterations. tershed and graph cut segmentation algorithm and improved the
The performance of the model was 96.71% for rice blast, 97.53% for bac- classification of rice leaf blight, rice blast and spot with 94% of classifica-
terial blight, and 98.26% for blight. (Sethy et al., 2020; Sethy et al., 2020) tion accuracy but lacked scope in terms of severity-based classification.
used a Faster R-CNN model to diagnose rice false smut. The dataset used However, made a significant mark in terms of accuracy. In our study we
was from a farm field and was captured using a smartphone camera. try to overcome this problem of severity-based classification while ac-
The number of observations the number of iterations was 50, 0.001, tively sustaining the robustness of the DL model by relying on Transfer
and 5 respectively. The performance of the model was not reported. Learning (TF) and Deep-Feature Extraction methodology ensuring accu-
(Ramesh and Vydeki, 2020) used an Optimized algorithm to diagnose racy and reliability so that the proposed work can be adopted and im-
rice leaf diseases. The dataset used was from a farm field. The perfor- plemented in any similar agricultural scenario with ease. The authors
mance of the model was 98.9%, 95.78%, 92% and 94% for rice blast, bac- of (Aggarwal et al., 2023b) focused exclusively on classifying different
terial blight, sheath rot, brown spot. types of leaf diseases, without considering severity-based classification.
(Deng et al., 2021; Li et al., 2017) used a Faster-RCNN model to diag- While their work achieved notable accuracy, it lacked the dimension of
nose rice leaf diseases. The dataset used was from rice fields in Anhui, assessing disease severity. In our research, we aim to address this limi-
and Hunan Province in China and was captured using a mobile phone tation by incorporating severity-based classification. We achieve this
camera and a Sony DSC-QX10 camera. The number of observations, while maintaining the robustness of the deep learning model with
the learning rate and the number of iterations was 5320, 0.002, and Transfer Learning (TL) and Deep-Feature Extraction techniques. This ap-
50,000 respectively. (Prajapati et al., 2017) used a SVM model to diag- proach ensures both accuracy and reliability, making our proposed
nose rice leaf diseases. The dataset used was from a farm field and methodology suitable for easy adoption and implementation in various
was captured using a NIKON D90 digital SLR camera. The number of ob- agricultural scenarios with similar requirements.
servations was 120, the learning rate and the number of iterations were A detailed investigation on rice leaf disease is summarized in Table 1
not reported. The performance of the model was 93.33%, 73.33%, 83.80% below. The various categories of rice leaf diseases that prevail are
and 88.57% during training, testing, 5-fold cross-validation, and 10-fold outlined in this table along with the best classification techniques for
cross-validation respectively. (Rahman et al., 2020) used a simple CNN each category based on several evaluation metrics.
model to diagnose rice leaf diseases. Recently to bring precise classifica-
tion of the rice plant disease. (Velusamy et al., 2023) proposed a hand- 3. Methodology
crafted feature engineering in database by performing segmentation,
augmentation, and pre-processing. Followed by the hand-crafted fea- The productivity and quality of the crop will be adversely affected by
turing techniques the classification accuracy improved with an any illness that affects the rice crop. With the goal to take measures that

Table 1
Details of the different classification techniques involved in recognizing rice leaf disease.

Ref. Disease categories Feature extraction models Dataset Evaluation Metric Classification Remark
Accuracy

(Mekha and Bacterial Leaf Blight, Random forest Rice Leaf Disease – UCI Precision, Recall, Accuracy 69.4% The proposed model's
Teeyasuksaet, Brown spot, Leaf smut repository classification accuracy
2021) performance is inadequate
(Rawat et al., Bacterial Leaf Blight, ResNet 50 with added Self-generated rice leaf Precision, Recall, F1 score, 99.2% Other types of rice leaf
2023) Brown spot, Leaf smut Neural Network database, Accuracy, ROC Area diseases are not compatible
with the model.
(Thepade et al., Healthy, Hispa, Brown Thepade sorted block International Rice Accuracy 85.9% Further improvement in
2022) spot, Leaf smut truncation coding (Thepade Research Institute rice leaf disease
SBTC) and Otsu (IRRI) classification is required
thresholding
(Hossain et al., Rice blast, rice blight, CNN IRRI and Bangladesh Precision, Recall, F1 score, 97.35% Not reliable and
2020) brown spot, Tungro, rice research institute Accuracy deteriorates recognition
Seath blight. (BRRI) rice leaf data- accuracy with complex
base background.
(Haridasan Leaf blast, Sheath rot, CNN Self-generated Precision, Recall, F1-score, 88.93% Pre-processing using
et al., 2023) false smut, bacterial database and support advanced ML models may
leaf blight, brown spot improve classification
accuracy
(Kaur et al., Leaf Blight, black rot, Hybrid CNN Efficient Net B7 Plant Village dataset Precision, Recall, F1-score, 98.7% universal method for
2022) Black Measles, stable and accuracy choosing the most
appealing features is
missing
(He et al., 2022) Rice blast SVM with spectra Self-generated Classification accuracy 93% The classification accuracy
pre-processing is inadequate
(Stephen et al., Brown Spot, Hispa, ResNet34, ResNet 50, Rice leaf disease from Identification rate, accuracy, 98.54% Further model can be
2023) Leaf Blast ResNet 18 Kaggle specificity, F1 score, validated with benchmark
sensitivity, Negative databases.
predicted value.
(Win Lwin and Healthy Rice Leaf, rice AlexNet, VGG16, CNN Rice leaf disease from Classification accuracy 98.05% Further produce accurate
Htwe, 2023) Leaf Blight, Brown Kaggle classification with different
Spot and Leaf Smut rice leaf diseases.
(Gopi and Bacterial Blight, Brown DenseNet 169, Xception Benchmark datasets Classification accuracy 99.66% & 99.9 Enhance system for
Kishan Spot, Leaf smut, from mendley data and incremental learning
Kondaveeti, Tungro, False Smut Kaggle repository strategies.
2023)

36
P.I. Ritharson, K. Raimond, X.A. Mary et al. Artificial Intelligence in Agriculture 11 (2024) 34–49

Fig. 1. Workflow of our Proposed Method.

can minimize rice yield loss, improve rice quality, and increase farmer 1308 images, Blast with 1440 sample images, Bacterial Blight with
income, it is important to accurately diagnose rice leaves disease at an 1584 images, and Brown spot with1600 images. This dataset is made
early stage. Presently, farmers must diagnose and categorize diseases available in the Mendeley archive data. Apart from this, 900 images of
by hand, which takes lengthier. To get around this, automated tech- healthy leaves images were collected from the Kaggle website and UCI
niques can be used to spot plant leaf disease early on. Data accessibility Machine Learning Repository so that the proposed model possibly dis-
therefore becomes a crucial component of Convolutional Neural Net- tinguishes the healthy rice leaf plants from the different categories of
works (CNN) training since they are capable of detecting rice plant ill- rice plant diseases. Fig. 2 depicts the sample images of the dataset
ness on their own. Because CNNs are accurate at classifying and with different categories of rice leaves.
identifying images, many researchers have employed them to identify
plant diseases. Therefore, a detailed workflow of our proposed model 3.1.1. Subcategorization of dataset
is depicted in Fig. 1 wherein the evaluation is conducted with our own The images of the infected leaves were once more sub-classified into
self-generated database. groups based on the severity of the diseases and have been divided into
These pre-trained CNN's makes use of the spatial correlations be- three categories such as mild, and severe, based on the spread of infec-
tween pixels in the picture, enabling them to learn helpful features for tion above the surface of the leaf. Regions of interest (such as the spots
image classification through convolutional layers and pooling layers and marks on the leaves) were enhanced when filters were applied to
and we will be performing training and validation tests with the pre- the image data. The photos were afterward manually separated and la-
trained CNN's and proposed custom VGG 16 model. beled following a comprehensive discussion with the agricultural and
horticulture department. Table 2 shows the subcategories of the in-
3.1. Data gathering fected rice leaf plants. After segregating the image dataset and mapping
them to different class labels, the overall number of images in each class
For our evaluation we have used the dataset proposed by (Sethy was significantly decreased.
et al., 2020), in their work on SVM-based deep feature-based disease di- Consequently, data augmentation methods like tilting, rotating,
agnosis for rice leaf. The collected rice leaf dataset initially contained cropping, and blurring were used formats were employed to increase
5932 images of 4 types of diseases such as Tungro with a collection of the number of images. Sample images of nine different class labels,

Fig. 2. Depicting the sample images of four classes of diseases in rice plant (1. Blast, 2. Brown Spot, 3. Bacterial Blight, and 4. Tungro).

37
P.I. Ritharson, K. Raimond, X.A. Mary et al. Artificial Intelligence in Agriculture 11 (2024) 34–49

Table 2
Dataset description and categorization.

Sl.No Diseases Classes Sub-classes Samples No of images

1 Blight ild Blight 476

2 Severe Blight 1108

3 Blast Mild Blast 1000

4 Severe Blast 440

5 Brown Spot Mild Brown Spot 900

6 Severe Brown Spot 700

7 Tungro Mild Tungro 450

8 Severe Tungro 314

9 Healthy Leaf 900

Total Number of images 5932

which were distinctively distinguished into severely impacted and able to improve our classification. The process of data cleaning
slightly affected are shown in Fig. 3. includes,
As illustrated above, a certain threshold value was set for each Noise and Blur Removal: Identifying the noisy and blurred images
disease to identify the severity level. For example, the class Brown within the dataset.
spot was given a threshold value of 3 for the number of brown spots Duplicate Removal: Identifying and removing any duplicate entries
present on the leaf's surface; if the number of brown spots exceeded or images within the dataset to prevent bias or redundancy.
the threshold values, then the disease was considered severe as Data Augmentation: Generating additional data samples through
shown in Fig. 4. techniques like rotation, cropping, or adding noise to improve the
In the case of a disease, such as Bacterial, the entire region or surface dataset's diversity.
area affected was evaluated, and the images were classified based on Label Validation: Checking the accuracy and consistency of the labels
those surface area thresholds. Tungro was classified based on the num- assigned to each data point. Ensure that labels are correct and corre-
ber of infected leaves, as well as partially and completely-afflicted- spond to the data accurately.
leaves. After successfully collecting and categorising the disease infected Balancing Classes: In classification tasks, ensuring that each class or
images, a ML model was developed to automate the classification pro- category has enough samples to prevent class imbalance issues.
cess. The training and validation datasets each contain a total of 2826 Data Splitting: Dividing the dataset into training, validation, and test
and 180 images, respectively. As a result, 90% of the data was thus sets to assess model performance accurately and avoid data leakage.
used for training, and 10% for validation. Therefore, the process of data cleaning holds a pivotal role within the
data preparation pipeline because the dataset's quality has a direct im-
3.2. Data pre-processing and cleaning pact on the effectiveness and dependability of machine learning models
trained on it. A meticulously cleaned dataset is inclined to produce pre-
The collected dataset contained several noisy images, Blurred cise and significant outcomes in various data-driven analyses, ranging
and unwanted. They had to be cleaned so that they do not hinder from the classification of rice leaf diseases to other analytical tasks.
the performance of the model. By cleaning these images, we were Splitting the data into training and testing allows us to train the
model on a certain set of images and test the model with test data.
Hence, the entire dataset was split into two, training and testing data
containing exactly 2826 and 180 images respectively for 9 different
class labels Each class label in the training dataset contained 314 images
and testing contained 20 images. Superfluous photos like noisy images
were eliminated to prevent training from picking up on minor features
using Median filters. Further, images sent as input to the model were
reshaped. The training dataset was reshaped to size, (224 × 224) for
models like VGG16, and ResNet50 and reshaped to size (299 × 299)
for the inceptionV3 model. The image data was then converted into nu-
Fig. 3. Sample images after data augmentation. merical NumPy arrays and then normalized to the range 0 to 1. Finally,

38
P.I. Ritharson, K. Raimond, X.A. Mary et al. Artificial Intelligence in Agriculture 11 (2024) 34–49

Fig. 4. Sub-categorizations of the rice diseases based on severity.

the models were trained and tested to observe their level of generaliza- ðinput width − filterwidth þ 2∗paddingÞ
output width ¼ þ1 ð3Þ
tion given an unseen test image. stride

The max pooling layers were used to down-sample the feature maps
3.3. Transfer learning models
and reduce their dimensions, while retaining the most important infor-
mation. The max pooling operation represented by:
Transfer learning involves using VGG, ResNet, or Inception pre-
trained CNN models, as a starting point for training new image classifi- 

input height − filterheight
cation or object detection models. The pre-trained models have learned output size ¼ þ1
stride
to recognize general features in images, such as edges, lines, and tex- 

input height − filterheight
tures, which are useful for many tasks. By fine-tuning the pre-trained  þ1 ð4Þ
stride
models on a new dataset of images, the models can learn to recognize
more specific features and achieve high accuracy on the new task with
Many image-processing tasks, such as analysis and classification,
fewer training images. The models used in our experiment were
have used transfer learning.
selected based on how well pre-trained networks performed in
diverse applications at classifying data. Xception, DenseNet121,
3.3.1. Overview of various architectures of Transfer Model
InceptionResNetV2, InceptionV3, ResNet50, and VGG16 were the
VGG-16 is a relatively deep network with 16 weight layers, including
models that were chosen. Direct imports of the pre-trained models
13 convolutional layers and 3 fully connected layers. It uses small 3 × 3
from the Keras library (Keras applications) were made, and the Dense
convolutional filters, which are stacked one after another, making it
layer was changed to include labels for nine different output class
deep but computationally expensive. It follows a simple and uniform ar-
types. Accuracy, batch size = 32, standard Adam optimizer, categorical
chitecture where the convolutional layers are stacked on top of each
cross-entropy for the loss, and other parameters are utilized during
other, followed by fully connected layers. The basic architecture of
training. After 50 iterations/epochs, the model's performance was
VGG-16 is provided in Fig. 5. Xception, also known as “Extreme Incep-
assessed based on how well it generalized an input image. The detail
tion,” is even deeper than VGG-16. It has 36 convolutional layers in
specifications of the suggested pre-trained models for our experimental
total. It employs a different approach, using depth wise separable con-
analysis are shown in Table 3.
volutions. This means it applies 3 × 3 convolutions separately to each
Typically, a 2D convolution procedure between an image and a filter
channel and then combines them, reducing computational complexity.
or kernel is represented by
Xception is inspired by the Inception architecture but differs in its use
C ði, jÞ ¼ ∑∑½Iðm, nÞ ∗ K ði − m, j − nÞÞ ð1Þ of depth wise separable convolutions.
m n
DenseNet-121 has 121 layers, but unlike VGG or ResNet, it does not
have many parameters because it connects each layer to every other
where I(m, n) represents values at location (m, n) of the input picture,
layer in a densely connected fashion. It uses 3 × 3 convolutional filters.
K(i-m, j-n) is the weight of the filter at location (i-m, j-n), and location
DenseNet's architecture encourages feature reuse and reduces the
(i, j) contains values C(i, j) of the output feature map.
vanishing gradient problem, as each layer has direct access to the gradi-
The size of the convoluted image is calculated based on eq. 2 and 3.
ents from subsequent layers. ResNet-50 is relatively deep with 50


input height − filterheight þ 2 ∗ padding layers. It introduced the concept of residual learning, which uses short-
output height ¼ þ1 ð2Þ cut connections to avoid the vanishing gradient problem. It uses 7 × 7
stride

Table 3
Hyperparameter specifications for pretrained models.

Transfer Learning Model's Hyperparameter Range/Values to Tune Fine-tuning remarks

−5 −4 −3 −2
1. Xception Learning Rate [e , e , e , e ] Adjust the learning rate for training.
2. DenseNet 121 Batch Size [16, 32, 64, 128] Batch size for mini-batch gradient descent.
3. ResNet 50, Number of Epochs [10, 20, 50, 100] Number of training epochs depending on the model's convergence
4. InceptionResNet V2 Weight Decay (L2 Regularization) [1e−4, 1e−5] Regularization strength
5. Inception V3 Dropout Rate [0.2, 0.3, 0.4] Dropout for preventing overfitting.
6. VGG 16 Optimizer [Adam, RMSprop, SGD] Choice of optimization algorithm.

39
P.I. Ritharson, K. Raimond, X.A. Mary et al. Artificial Intelligence in Agriculture 11 (2024) 34–49

Fig. 5. Basic VGG-16 Architecture.

convolutions in the initial layer followed by 3 × 3 convolutions in resid- the data, fostering a hierarchical learning process. Finally, as the net-
ual blocks. ResNet's key innovation is the residual block, which allows work approaches the output layer, there is a reduction in the number
the network to skip layers, enabling the training of very deep networks. of neurons with one dense layer with 256 neurons was added followed
Inception V3 is moderately deep with 48 layers. Filter Size: It uses a by a dropout layer with 60% activation. Finally, to bring the convex
mix of filter sizes, including 1 × 1, 3 × 3, and 5 × 5 convolutions, to cap- structure of the output layers of the model, two more dense layers
ture features at different scales. Inception modules are used to with 64 and 9 neurons respectively were added. Hence, the output
parallelize different filter sizes and then concatenate their outputs. was a class label among the 9 different class labels and the class with
Inception-ResNet-V2 is even deeper than Inception V3, with 164 layers. the highest probability was the certain class. The architectures of the
Similar to Inception V3, it uses a mix of filter sizes. It combines ideas proposed custom VGG 16 model are shown in Figs. 6 and 7.
from both Inception and ResNet, incorporating residual connections In our proposed custom VGG 16 model, there were 13 convolutional
into Inception modules. Each of these architectures has its advantages layers, 5 max pooling layers, and 7 fully connected layers (can also be
and trade-offs, making them suitable for different tasks and computa- termed VGG18). The number of filters used in these layers was 64.
tional resources. The choice of architecture often depends on the spe- The next four layers were also convolutional layers, each followed by
cific requirements of the task at hand, model size and accuracy. a max pooling layer. These layers used double the number of filters as
the previous layer, 128. These layers extracted more complex features
4. Proposed approach from the image. The next four layers were also convolutional layers,
each followed by a max pooling layer. These layers used 256 filters.
In this section we discuss the design of proposed custom approach These layers extracted even more complex features from the image,
for improved generalization. Further to achieve this, the prior model allowing the model to recognize more abstract patterns. The next
was modified with additional layers to track the performance. With four layers were also CL, each followed by a max pooling layer. These
the addition of a dropout layer with 50% activation, a dense layer with layers used 512 filters. These layers extracted even more complex fea-
128 neurons and the set activation “ReLu” were added. As a result, the tures from the image, allowing the model to recognize even more ab-
number of neurons that learn the features increased significantly, stract patterns. The final three layers were fully connected layers.
which improved the model's generalization capability. Further, to These layers took the output of the last convolutional layer and flat-
bring a defined concave structure of the proposed VGG 16 custom tened them into a one-dimensional array of neurons. These neurons
model in the input layers, an architectural design that begins with a rel- were then connected to the output layer, which used a SoftMax activa-
atively smaller number of neurons in the input and early layers, facilitat- tion function to give the final probability distribution over the classes as
ing the initial processing and abstraction of fundamental features from given in eq. (5).
the input data. Subsequently, as it progresses through the network, it
traverses layers with progressively fewer neurons, effectively reducing softmaxðxÞ ¼ exp ðxÞ=sumð exp ðxÞÞ ð5Þ
the dimensionality and complexity of the representations.
After reaching a minimum point in terms of the number of neurons, The proposed custom model used very small filters (3 × 3) and mul-
the network then undergoes expansion, typically occurring within the tiple layers of convolution and pooling to increase the depth of the net-
middle layers, where the number of neurons increases. This expansion work and to extract more abstract features from the image. This
allows the network to capture more intricate and abstract features in contrasted with architectures that used larger filters or fewer layers.

Fig. 6. Custom-CNN Model built on top of functional VGG-16 Model.

40
P.I. Ritharson, K. Raimond, X.A. Mary et al. Artificial Intelligence in Agriculture 11 (2024) 34–49

Fig. 7. Proposed custom VGG 16 model Structure Diagram/Architecture.

The spatial hierarchy of patterns in the Custom Model was achieved validate the model, as shown in Fig. 13, to see how well it could recog-
using multiple CL. It is responsible for extracting features from the nize and categorize the images and, in turn, evaluate how well it could
input image at different scales and positions shown in Fig. 20. The generalize in terms of performance metrics. The confusion table shows
max pooling layers were then used to down-sample the feature maps, how well a classification model performs when put to the test on
which helped to reduce the dimensionality of the data and increase fresh data with known actual labels. Metrics The table's entries are
the spatial invariance of the features. In the early layers of the custom displayed below.
model, convolutional filters were applied to the input image to extract
▪ True Positive (TP): defines how many instances of positivity the
low-level features such as edges, corners, and textures.
model accurately identified as positive.
These features were then passed through the max pooling layer. This
▪ True Negative (TN): defines how many negative samples were prop-
allowed the network to extract features at a coarser scale and increased
erly categorized as negative by the model.
the model's ability to recognize patterns that were present in different
▪ False Positive (FP): defines how many negative cases the model mis-
parts of the image. As the architecture progressed through deeper
takenly categorized as good.
layers, convolutional filters were applied to increasingly abstract feature
▪ False Negative (FN): defines the quantity of positive samples that
maps, allowing the network to extract more complex patterns such as
the model mistook for negative.
parts of objects or even entire objects. The max pooling layers continued
to down-sample the feature maps, allowing the network to extract fea-
tures at an even coarser scale. The final feature maps were then fed into
Precision is the proportion of TP predictions (i.e., correctly predicted
the fully connected layers, which were used to classify the image into
positive instances) out of all positive predictions made by the model. It
one of the predefined classes. The spatial hierarchy of patterns in cus-
measures how accurate the model's positive predictions are, as shown
tom model allowed the network to recognize patterns at different scales
in eq. (6)
and positions, which was crucial for accurate image classification.
Table 4 displays the specifications involved in training the proposed TP
Precision ¼ ð6Þ
custom model. ðTP þ FP Þ

5. Experimental analysis Recall is the proportion of TP predictions out of all actual positive in-
stances in the data. Refer eq. (7)
5.1. Performance evaluation metrics
TP
Recall ¼ ð7Þ
After the model had successfully been trained for 50 iterations using ðTP þ FNÞ
images containing categories of various diseases, it was finally time to
F1 score represents the harmonic mean of precision and recall and is
used to balance the trade-off between precision and recall. Refer eq. (8)

Table 4 2 ∗ precision ∗ recall

F1 Score ¼ ð8Þ
Hyperparameter Specifications of different pre-trained networks. precision þ recall
Model Name Training Time Model Size No of Parameters
in seconds in megabytes Accuracy is the percentage of correct outputs of all the outputs made
Xception 10,000 90.3 21,764,657
by the model as given in eq. (9)
DenseNet121 15,000 90.3 7,489,097
ðTP þ TN Þ
InceptionResnetV2 12,500 213.2 54,682,345 Accuracy ¼ ð9Þ
Resnet 50 11,000 100.8 24,490,889 ðTP þ TN þ FP þ FNÞ
VGG 16 14,000 58.8 14,940,489
InceptionV3 5000 89.3 22,263,593 Precision indicates how many of the examples that the model antic-
Our Custom-Model 14,500 59.5 14,991,826
ipated to have a disease occur. In this situation, the disease in question is

41
P.I. Ritharson, K. Raimond, X.A. Mary et al. Artificial Intelligence in Agriculture 11 (2024) 34–49

classified as a rice disease. Recall lets us know how many cases truly had further would require more time, on top of more experimentation, sub-
the sickness that the model predicted would. stantial deployment of resources, and risk of overfitting. It was crucial to
select a model that required less training time while yet producing accu-
5.2. Experimental setup rate results. After comparing the models, we found that Xception per-
formed well, requiring just 10,000 s for training as opposed to
Initially the proposed model was implemented using google collab 14,000 s for VGG16. However, the size of the model had to be consid-
with GPU machines and after successfully training the model for 50 ered before the model could be deployed. For which VGG16 only has
epochs with images containing categories of different diseases, testing the size of 58.8 mb whereas Xception had a size of 90.3 which is
of the model was performed as shown in Fig. 8. to find out how well heavy and also the accuracy i.e., the generalization of the VGG model
the model can identify and classify the images, hence measure its ability was quite better when compared to the Xception model for our
to generalize in terms of performance metrics. application.

5.3. Results of pretrained models 5.4. Visualization of feature extraction using proposed custom model

The confusion matrix results for all 9 classes of rice leaf dataset is VGG-16 also had comparatively less parameters to tune and was
provided in Fig. 9. Here we have a 9 × 9 matrix for different classes of true for easy optimization of the model. Thus, it was an optimal choice
rice leaf disease. In addition to accuracy and the AUC-ROC curve, metrics to go for the VGG-16 group for this project. Further, from the confusion
used to assess the performance of ML models include recall and preci- matrices as shown in Fig. 10, the VGG-16 model performed better than
sion. The confusion matrix enables us to quantify these metrics. The cor- the other transfer learning models for this application. Although it cor-
responding performance metrics of all the pretrained models are rectly identified 178 images, it misclassified 2 images of the severe
discussed in Table 5. brown spot and the moderate brown spot. Hence, an insight had been
From the confusion matrix of the various transfer learning models, it gained to tweak the layers of the VGG-16 model and to observe the per-
is observed that the classification error is minimum in VGG 16 model formance of the customized CNN model. The spatial hierarchy of the
when compared to DenseNet 121, ResNet 50, InceptionResNet V2, In- proposed custom model is depicted in Fig. 11.
ception V3, and Xception models. Therefore, from this performance ob- The layer wise visualization of the proposed model's feature extrac-
servation, the VGG −16 model opted and we have altered the VGG 16 tion and model training is shown in Fig. 12 (a), 12 (b), and 12 (c). Fig. 12
layers by integrating dense and dropout layers along with filter struc- (a) visualizes the features that are learned by the proposed model for
tures. The Table 5. Shows the classification performance of all the testing “Severe Brown spot” disease. Fig. 12 (b) visualizes the features that
models. are learned by the proposed model for “Severe Blast’” disease. Fig. 12
In Fig. 10. We observe the ROC graph plots, showing the perfor- (c) visualizes the class severe Bacterial Blast in Rice leaves.
mance measure for the classification of different categories with respect
to various threshold settings. Through this graph the distinctive classifi- 5.5. Performance evaluation with state of the art
cation between different classes of rice plant disease is observed.
Fig. 10. depicts the various classification properties of the different 5.5.1. Comparison of VGG-16 and proposed custom VGG − 16 model
transfer learning approaches. The x- axis in this graph represents the The structural comparison between VGG16 and the proposed cus-
False positive rate (FPR) and the y-axis represents the True positive tom VGG 16 model reveals several similarities and differences. Both
rate (TPR). The micro average and macro average of all the classes models utilize convolutional layers with 3 × 3 filters and max pooling
have been evaluated in terms of the true positive and false positive layers for feature extraction. However, the custom model has a larger
rates. From this graph, we were able to observe the time taken by number of convolutional layers (13 compared to 16 in VGG16) and a
each model to train for 50 iterations (epochs) and saw that the models different filter Configuration, starting with 64 filters and doubling the
having different training times, which was also a key deciding factor for number with each set of four layers. In contrast, VGG16 uses a fixed
the choice of models. Since the length of training directly affects the number of 64 filters throughout. Additionally, the custom model has 7
time required to generate results for a test picture, training a model fully connected layers compared to 3 in VGG16. The custom model's

Fig. 8. Test inputs (Healthy image) to validate the generalization of the model.

42
P.I. Ritharson, K. Raimond, X.A. Mary et al. Artificial Intelligence in Agriculture 11 (2024) 34–49

Fig. 9. Confusion Matrices of (a) VGG16, (b) Xception, (c) ResNet50, (d) DenseNet121, (e) InceptionResnetV2, and (f) Inception V3.

43
P.I. Ritharson, K. Raimond, X.A. Mary et al. Artificial Intelligence in Agriculture 11 (2024) 34–49
Table 5
Performance Measure Analysis of all the transfer learning approach.

Model Name Precision Recall F1_score Accuracy

Xception 0.977 0.977 0.977 97.7%

DenseNet121 0.938 0.938 0.938 93.8%
InceptionResnetV2 0.933 0.933 0.933 93.3%
Resnet 50 0.811 0.811 0.811 81.1%
InceptionV3 0.955 0.955 0.955 95.5%
VGG 16 0.988 0.988 0.988 98.8%

Fig. 10. Shows the Area Under the Curve Plot TPR vs FPR of (a) ResNet-50, (b) Inception ResNetV2, (c)InceptionV3, (d)DenseNet121, (e) Xception, and (f) VGG16.
P.I. Ritharson, K. Raimond, X.A. Mary et al. Artificial Intelligence in Agriculture 11 (2024) 34–49

Fig. 11. Feature Extraction using Custom CNN (Spatial hierarchy of patterns)-the stages involved in feature learning.

use of multiple convolutional layers and max pooling layers allows for a of the model is maintained so can take effective measures based on
spatial hierarchy of feature extraction, aiding in recognizing patterns at the accurate detection of the diseases. The generalization capacity of
different scales and positions for accurate image classification. various transfer learning models are shown in (Table 7).
Besides, in comparison to previous works by other authors, our CNN The performance evaluation of the proposed Custom-VGG Model is
model demonstrates superior performance in accuracy, model size, and evaluated from Fig. 15. There the confusion matrix generated by the
parameters. Details are shown in Table 6. Our model achieves 99.7% ac- proposed model has lowest classification error rate.
curacy using a dataset with 5932 images with only 50 epochs of train-
ing, outperforming Faster-RCNN with 98.85% accuracy on 2400 images 6. Conclusion
(not recorded epochs), ResNet197 with 99.58% accuracy on 154,500 im-
ages trained for 1000 epochs, YOLOv4 with 95.9% accuracy on 56 images The outcomes of this investigation show the potential of Deep Learn-
(not recorded epochs), and ensemble models with 99.1% accuracy on ing, and more specifically CNNs, in the early detection and categoriza-
33,026 images trained for 100 epochs. Furthermore, our model has a tion of illnesses affecting rice leaves. The proposed custom VGG 16
significantly smaller model size of 14,991,826 parameters compared model showed high performance in identification and categorization
to ResNet197 with 58,450,784 parameters and YOLOv4 with over of 9 different disease class labels of rice leaf with improved accuracy
60,000,000 parameters, making it more efficient and less computation- of 99.94% and maximum precision and recall scores, providing a quick
ally demanding. and reliable way of detecting and diagnosing plant diseases. Six transfer
Additionally, when the number of epochs grew throughout the models were selected based on criteria such as model size, parameter
training phase of both the VGG-16 and Custom-CNN models, a few size (with a major focus on real-time deployment), and subsequently
changes were observed. There were a few surges in the validation accu- retrained. Among these models, VGG16 exhibited the most outstanding
racy of the Custom-CNN Model during the training phase, for example, performance in terms of precision, recall, and accuracy. Following this,
in 18th iteration the validation accuracy abruptly dropped to 0.81 we conducted further parameter fine-tuning, which involved the addi-
(which is 81%) and the loss had increased, but over time the accuracy tion of extra filters, dense layers, and dropouts, optimized through
reverted back to its place, achieving 0.9889 accuracy value in the next trial and error. As a result of these efforts, the proposed CNN demon-
iteration, and ultimately achieving an accuracy of 0.997 (shown in strated an exceptional accuracy rate. This high-performing model
Fig. 13(a), (b)). In contrast, the VGG-16 model's training phase was con- holds promise for implementation in various agricultural scenarios,
stant, no sudden increases in accuracy were noticed, and the accuracy where it can be utilized to identify crop diseases based on their severity
continuously increased to offer a final average accuracy. Nonetheless, levels. The study also involved the analysis of confusion matrices and
when trained for 100 iterations, the proposed custom VGG 16 and ROC plots for all six selected models, illustrating key metrics such as
Transfer learning models was found to have no significant improve- true positives (TP), true negatives (TN), false positives (FP), false nega-
ments in the performance during evaluation, rather only increasing tives (FN), and the relationship between true positive rate (TPR) and
the risk of model instability due to over training. Hence strongly believe false positive rate (FPR). Moreover, the model's robustness and reliabil-
the learning has converged at the 50th iteration and training the model ity were ensured by using data augmentation techniques and pre-
further would contribute to the overfitting of the model. trained models. The contribution of this research to agriculture is signif-
Our proposed custom VGG 16 model showed superior performance icant, as it can help farmers detect plant diseases at an early stage,
in the generalization of transfer learning for individual class labels of allowing them to take necessary measures and avoid crop losses.
rice leaf diseases, as evidenced by the results in the table. Compared to In future, the proposed custom VGG 16 architecture can be applied
the Transfer Model, our custom model exhibited higher accuracy in to other crop diseases, contributing to the development of more effec-
identifying the severity of rice leaf diseases. This indicates that our tive and efficient disease diagnosis tools in agriculture. The study also
model was more effective in accurately classifying the severity of differ- highlighted the negative impacts of rice leaf diseases on crop yields,
ent diseases affecting rice leaves. The results highlight the robustness food security, and poverty, emphasizing the urgency to find effective so-
and efficacy of our custom model in disease classification, showcasing lutions. The traditional methods of disease detection have been found to
its potential as a reliable tool for diagnosing and managing rice leaf dis- be time-consuming, unreliable, and often require the expertise of
eases in agricultural settings as shown in Fig. 14. The symptoms of infec- trained personnel, making them inaccessible to small-scale farmers.
tion can appear in different areas of the plant, and leaves are often used The proposed model's high accuracy and efficiency make it a promising
to diagnose plant diseases. More effectively, the models can be deployed tool for disease diagnosis, ultimately benefiting farmers and promoting
on platforms like android to be utilized by farmers and agriculturalists sustainable crop production. This study's findings, therefore, contribute
to identify the disease and severity. Some potential crops that can re- significantly to the emerging field of precision agriculture, where DL-
quire such an application could range from Tomato crops, Grape, based approaches are increasingly being used to improve crop yield
Apple, Blueberry to Cherry. Although we must ensure, the robustness and quality.

45
P.I. Ritharson, K. Raimond, X.A. Mary et al. Artificial Intelligence in Agriculture 11 (2024) 34–49

Fig. 12. (a) Visualization of Custom-VGG 16 Model Trying to learn the class ‘Severe Brown spot’. (b). Visualization of Custom-VGG-16 Model learning the class ‘Severe Blast’.
(c) Visualization of Custom-VGG 16 Model learning the class “Severe Bacterial Blight”.

Table 6
Comparision of the Proposed custom VGG 16 model Results with State-of-the-art Research.

Methods Accuracy No.of Images Epochs Parameters

Faster- RCNN Bifta Sama Bari et al., 2021 98.85% 2400 Not Recorded Not Recorded
ResNet197 model (Pandian et al., 2022) 99.58% 154,500 1000 58,450,784
YOLOv4 model (Roy et al., 2021; Roy and Bhaduri, 2022) 95.9% 56 Not Recorded Over 60,000,000
Ensemble models (Saberi Anari, 2022) 99.1% 33,026 100 30,000,000
Proposed custom VGG 16 model 99.7% 5932 50 14,991,826
P.I. Ritharson, K. Raimond, X.A. Mary et al. Artificial Intelligence in Agriculture 11 (2024) 34–49

Fig. 13. (a) Showing the Training, validation loss for Custom-CNN Model (b)Showing Training and validation accuracy for Custom-CNN Model (c) Showing the Training, validation loss of
VGG 16, and (d) Training and validation accuracy for VGG-16.

Fig. 14. Performance evaluation of the proposed custom VGG 16 model and pretrained network.

47
P.I. Ritharson, K. Raimond, X.A. Mary et al. Artificial Intelligence in Agriculture 11 (2024) 34–49

Table 7
Generalization of Transfer Model for individual class labels.

Disease Type Xception denseNet Inception V3 InceptionResnetV2 Resnet 50 VGG 16 Custom-Model

Mild Blight 1.00 0.95 0.95 1.00 0.95 1.00 1.00

Severe Blight 0.95 1.00 1.00 1.00 0.95 0.95 1.00
Mild Blast 0.95 1.00 1.00 1.00 0.85 0.95 1.00
Severe Blast 1.00 1.00 1.00 1.00 0.90 1.00 1.00
Mild Brown Spot 0.95 0.50 0.90 0.50 0.05 1.00 1.00
Severe Brown Spot 1.00 1.00 1.00 0.95 0.90 1.00 1.00
Mild Tungro 1.00 1.00 1.00 0.95 0.85 1.00 1.00
Severe Tungro 1.00 1.00 0.95 1.00 0.95 0.95 0.95
Healthy Leaf 1.00 1.00 1.00 0.85 0.90 1.00 1.00

Fig. 15. Confusion Matrix of Proposed Custom-VGG Model (20 unseen images/sub-class).

CRediT authorship contribution statement Aggarwal, M., Khullar, V., Goyal, N., 2023a. Exploring classification of Rice leaf diseases
using machine learning and deep learning. Proceedings of 2023 3rd international
conference on innovative practices in technology and management, ICIPTM 2023.
P. Isaac Ritharson: Conceptualization, Methodology, Writing – https://doi.org/10.1109/ICIPTM57143.2023.10117854.
original draft. Kumudha Raimond: Conceptualization, Methodology, Aggarwal, M., Khullar, V., Goyal, N., Alammari, A., Albahar, M.A., Singh, A., 2023b. Light-
weight federated learning for Rice leaf disease classification using non independent
Writing – original draft, Supervision. X. Anitha Mary: Data and identically distributed images. Sustainability 15 (16), 12149. https://doi.org/10.
curation, Writing – original draft, Supervision. R. Jennifer Eunice: 3390/SU151612149.
Data curation, Writing – original draft, Visualization, Investigation, Aggarwal, M., Khullar, V., Goyal, N., Singh, A., Tolba, A., Thompson, E.B., Kumar, S., 2023c.
Pre-trained deep neural network-based features selection supported machine learn-
Writing – review & editing. J. Andrew: Conceptualization, Methodology,
ing for Rice leaf disease classification. Agriculture 13 (5), 936. https://doi.org/10.
Visualization, Investigation, Writing – review & editing. 3390/agriculture13050936.
Andrew, J., Eunice, J., Popescu, D.E., Chowdary, M.K., Hemanth, J., 2022. Deep learning-
based leaf disease detection in crops using images for agricultural applications.
Declaration of Competing Interest
Agronomy 12 (10), 2395. https://doi.org/10.3390/AGRONOMY12102395.
Asibi, A.E., Chai, Q., Coulter, J.A., 2019. Rice blast: A disease with implications for global
The authors declare that they have no known competing financial food security. Agronomy 9 (8), 451.
interests or personal relationships that could have appeared to influ- Bari, B.S., Islam, M.N., Rashid, M., Hasan, M.J., Razman, M.A.M., Musa, R.M., Nasir, A.F.A.,
Majeed, A.P.P.A., 2021. A real-time approach of diagnosing rice leaf disease using
ence the work reported in this paper. deep learning-based faster R-CNN framework. PeerJ Computer Science 7, e432.
https://doi.org/10.7717/PEERJ-CS.432/SUPP-1.
References Bock, C. H., Chiang, K. S., & Del Ponte, E. M. (2022). Plant disease severity estimated visu-
ally: a century of research, best practices, and opportunities for improving methods
Aggarwal, M., Khullar, V., Goyal, N., 2022. Contemporary and futuristic intelligent Tech- and practices to maximize accuracy. Tropical Plant Pathology, 47(1), 25–42.
nologies for Rice Leaf Disease Detection. 2022 10th international conference on reli- https://doi.org/10.1007/S40858-021-00439-Z/TABLES/3.
ability, Infocom technologies and optimization (trends and future directions), ICRITO Damalas, C.A., Koutroubas, S.D., 2016. Farmers’ exposure to pesticides: toxicity types and
2022. https://doi.org/10.1109/ICRITO56286.2022.9965113. ways of prevention. Toxics 4 (1), 1.

48
P.I. Ritharson, K. Raimond, X.A. Mary et al. Artificial Intelligence in Agriculture 11 (2024) 34–49

Deng, R., Tao, M., Xing, H., Yang, X., Liu, C., Liao, K., Qi, L., 2021. Automatic diagnosis of Rice Rawat, P., Pandey, A., Panaiyappan, K., 2023. Rice leaf diseases classification using deep
diseases using deep learning. Front. Plant Sci. 12, 701038. https://doi.org/10.3389/ learning techniques. Proceedings of the 1st IEEE International Conference on Net-
FPLS.2021.701038/BIBTEX. working and Communications 2023. ICNWC. https://doi.org/10.1109/ICNWC57852.
Gopi, S.C., Kishan Kondaveeti, H., 2023. Transfer learning for Rice leaf disease detection. 2023.10127315.
Proceedings of the 3rd International Conference on Artificial Intelligence and Smart Roy, A.M., Bhaduri, J., 2021. A deep learning enabled multi-class plant disease detection
Energy, ICAIS 2023, pp. 509–515. https://doi.org/10.1109/ICAIS56108.2023. model based on computer vision. AI 2021 2 (3), 413–428. https://doi.org/10.3390/
10073711. AI2030026.
Haridasan, A., Thomas, J., Raj, E.D., 2023. Deep learning system for paddy plant disease de- Roy, A.M., Bhaduri, J., 2022. Real-time growth stage detection model for high degree of oc-
tection and classification. Environ. Monit. Assess. 195 (1). https://doi.org/10.1007/ cultation using DenseNet-fused YOLOv4. Computers and Electronics in Agriculture
s10661-022-10656-x. 193, 106694.
He, Y., Zhang, W., Ma, Y., Li, J., Ma, B., 2022. The classification of Rice blast resistant seed Saberi Anari, M., 2022. A hybrid model for leaf diseases classification based on the mod-
based on Ranman spectroscopy and SVM. Molecules 27 (13). https://doi.org/10. ified deep transfer learning and ensemble approach for agricultural AIoT-based mon-
3390/molecules27134091. itoring. Comput. Intell. Neurosci. 2022. https://doi.org/10.1155/2022/6504616.
Hossain, S.M.M., Tanjil, M.M.M., Ali, M.A., Bin Islam, M.Z., Islam, M.S., Mobassirin, S., Saberi Anari, M., 2022. A hybrid model for leaf diseases classification based on the mod-
Sarker, I.H., Islam, S.M.R., 2020. Rice Leaf Diseases Recognition Using Convolutional ified deep transfer learning and ensemble approach for agricultural aiot-based mon-
Neural Networks. Lecture Notes in Computer Science (Including Subseries Lecture itoring. Computational Intelligence and Neuroscience 2022.
Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 12447 LNAI, Sethy, P.K., Barpanda, N.K., Rath, A.K., Behera, S.K., 2020. Deep feature based rice leaf dis-
pp. 299–314. https://doi.org/10.1007/978-3-030-65390-3_23. ease identification using support vector machine. Comput. Electron. Agric. 175,
Javidan, S.M., Banakar, A., Vakilian, K.A., Ampatzidis, Y., 2023. Diagnosis of grape leaf dis- 105527. https://doi.org/10.1016/J.COMPAG.2020.105527.
eases using automatic K-means clustering and machine learning. Smart Agricultural Shew, A.M., Durand-Morat, A., Putman, B., Nalley, L.L., Ghosh, A., 2019. Rice intensification
Technology 3. https://doi.org/10.1016/j.atech.2022.100081. in Bangladesh improves economic and environmental welfare. Environ. Sci. Policy 95,
Jones, R.A.C., 2021. Global plant virus disease pandemics and epidemics. Plants (Basel, 46–57. https://doi.org/10.1016/J.ENVSCI.2019.02.004.
Switzerland) 10 (2), 1–41. https://doi.org/10.3390/PLANTS10020233. Shoaib, M., Shah, B., EI-Sappagh, S., Ali, A., Ullah, A., Alenezi, F., Gechev, T., Hussain, T., Ali,
Kaur, P., Harnal, S., Tiwari, R., Upadhyay, S., Bhatia, S., Mashat, A., Alabdali, A.M., 2022. Rec- F., 2023. An advanced deep learning models-based plant disease detection: A review
ognition of leaf disease using hybrid convolutional neural network by applying fea- of recent research. Front. Plant Sci. 14, 1158933. https://doi.org/10.3389/FPLS.2023.
ture reduction. Sensors 22 (2). https://doi.org/10.3390/s22020575. 1158933/BIBTEX.
Kumar, V.S., Jaganathan, M., Viswanathan, A., Umamaheswari, M., Vignesh, J., 2023. Rice Stephen, A., Punitha, A., Chandrasekar, A., 2023. Designing self attention-based ResNet ar-
leaf disease detection based on bidirectional feature attention pyramid network chitecture for rice leaf disease classification. Neural Comput. & Applic. 35 (9),
with YOLO v5 model. Environmental Research Communications 5 (6), 065014. 6737–6751. https://doi.org/10.1007/s00521-022-07793-2.
https://doi.org/10.1088/2515-7620/ACDECE. Taw, Nay Pyi, 2022. In: Taw, Nay Pyi (Ed.), Handbook on the integrated crop management
Li, X., Fang, J., an, & Li, H., 2017. Exponential adaptive synchronization of stochastic of rice and paddy for farmer field schools in central dry zone of Myanmar Climate-
memristive chaotic recurrent neural networks with time-varying delays. Friendly Agribusiness Value Chains Sector (CFAVC) project Global Agriculture and
Neurocomputing 267, 396–405. https://doi.org/10.1016/j.neucom.2017.06.049. Food Security Programme (GAFSP) Food and Agriculture Organization of the United
LIU, Z.Y., HUANG, J.F. and TAO, R.X., 2008. Characterizing and estimating fungal disease se- Nations Nay Pyi Taw. Food and agricultural organization of the United States.
verity of rice brown spot with hyperspectral reflectance data. Rice Science, 15(3), https://doi.org/10.4060/cc1163en.
pp.232-242. Thepade, S.D., Abin, D., Chauhan, K.H., 2022. Improved Rice leaf disease detection using
Lu, Y., Yi, S., Zeng, N., Liu, Y., Zhang, Y., 2017. Identification of rice diseases using deep fusion of Otsu thresholding and Thepade SBTC features. Proceedings - 2022 Interna-
convolutional neural networks. Neurocomputing 267, 378–384. https://doi.org/10. tional Conference on Recent Trends in Microelectronics, Automation, Computing and
1016/J.NEUCOM.2017.06.023. Communications Systems, ICMACC 2022, pp. 444–449 https://doi.org/10.1109/
Mekha, P., Teeyasuksaet, N., 2021. Image classification of Rice leaf diseases using random ICMACC54824.2022.10093433.
Forest algorithm. 2021 Joint 6th International Conference on Digital Arts, Media and Tudi, M., Ruan, H.D., Wang, L., Lyu, J., Sadler, R., Connell, D., Chu, C., Phung, D.T., 2021. Ag-
Technology with 4th ECTI Northern Section Conference on Electrical, Electronics, riculture development, pesticide application and its impact on the environment. Int.
Computer and Telecommunication Engineering, ECTI DAMT and NCON 2021, J. Environ. Res. Public Health 18 (3), 1–24. https://doi.org/10.3390/IJERPH18031112.
pp. 165–169. https://doi.org/10.1109/ECTIDAMTNCON51128.2021.9425696. Velusamy, D., Mounika, A., Ragulraj, D., Vigneshwaran, K., 2023. Automatic disease classi-
Pandian, J., Kanchanadevi, K., Rajalakshmi, N.R., Arulkumaran, G., 2022. An improved deep fication of Paddy leaf images using deep learning algorithms. ICSPC 2023 - 4th Inter-
residual convolutional neural network for plant leaf disease detection. Comput. Intell. national Conference on Signal Processing and Communication, pp. 325–329 https://
Neurosci. 2022. https://doi.org/10.1155/2022/5102290. doi.org/10.1109/ICSPC57692.2023.10125872.
Pandian, A., Kanchanadevi, K., Rajalakshmi, N.R., Arulkumaran, G., 2022. An improved Wang, Y., Bian, J., Sun, X., Ruan, D., Gu, Z., 2023. Sensitivity-dependent dynamic searching
deep residual convolutional neural network for plant leaf disease detection. Compu- approach coupling multi-intelligent surrogates in homotopy mechanism for ground-
tational Intelligence and Neuroscience 2022. water DNAPL-source inversion. J. Contam. Hydrol. 255. https://doi.org/10.1016/j.
Prajapati, H.B., Shah, J.P., Dabhi, V.K., 2017. Detection and classification of rice plant dis- jconhyd.2023.104151.
eases. Intelligent Decision Technologies 11 (3), 357–373. https://doi.org/10.3233/ Win Lwin, L.Y., Htwe, A.N., 2023. Image classification for Rice leaf disease using AlexNet
IDT-170301. model. IEEE International Conference on Control and Automation, ICCA, 2023-
Rahman, C.R., Arko, P.S., Ali, M.E., Iqbal Khan, M.A., Apon, S.H., Nowrin, F., Wasif, A., 2020. February, pp. 124–129 https://doi.org/10.1109/ICCA51723.2023.10181847.
Identification and recognition of rice diseases and pests using convolutional neural Zhou, C., Zhong, Y., Zhou, S., Song, J., Xiang, W., 2023. Rice leaf disease identification by
networks. Biosyst. Eng. 194, 112–120. https://doi.org/10.1016/J.BIOSYSTEMSENG. residual-distilled transformer. Eng. Appl. Artif. Intel. 121, 106020. https://doi.org/10.
2020.03.020. 1016/J.ENGAPPAI.2023.106020.
Ramesh, S., Vydeki, D., 2020. Recognition and classification of paddy leaf diseases using
optimized deep neural network with Jaya algorithm. Information Processing in Agri-
culture 7 (2), 249–260. https://doi.org/10.1016/J.INPA.2019.09.002.

49