Multi-Disease Detection and Classification in Paddy Using Deep

Convolutional Neural Networks

Sornalakshmi K

, Valadri Pardhavan Reddy and Ravipati Deepthi

Department of Data Science and Business Systems,

Faculty of Engineering and Technology,

SRM Institute of Science and Technology, Kattankulathur Campus, 600023, India

Keywords: Convolutional Neural Networks, Paddy Leaf Disease, Multi Label Classification, Computer Vision

Abstract: The world is expecting an exponential growth in food production in the recent future. Rice, a staple food for

a large part of the world's population, faces the threat of various diseases that can seriously affect the crop.

The proposed solution uses advanced deep learning algorithms on images of paddy leaves to predict leaf

diseases. Using data containing high-resolution images of healthy and diseased leaves, convolutional neural

network (CNN) model was implemented to accurately identify the disease. Preprocessing is used to improve

the quality of the image and remove features that hinder accurate classification. The system has been shown

to be useful in diagnosing many types of foliar diseases, providing good results for early disease detection

and good agronomic management. The Resnet-50, efficient net B3 architectures of Convolutional Neural

Networks (CNNs), a specialized deep learning architecture, has been trained on diverse datasets containing

images of healthy and diseased rice leaves for the diseases bacterial leaf blight, Hispa and brown spot. Once

trained, these models can accurately classify diseases with up to 90% accuracy thereby supporting timely

interventions, ultimately preventing extensive crop losses and fostering sustainable practices. In addition to

this, deep learning's image recognition capabilities is also used in sorting and grading rice leaves based on

various parameters such as size, color, and ripeness. A user interface using Streamlit is developed for

uploading test images and the system would identify the diseases.

1 INTRODUCTION

Rice is the first staple food crop in South India. The

demand for rice is increasing due to the population

growth. It is a crop chosen by many small scale and

marginal farmers. It is cultivated in diverse climatic

and soil bases, across south India. In contrast to the

demand for rice, the production of paddy is facing

many challenges. One of the major challenges faced

in paddy cultivation are the bacterial diseases like

foot rot, grain rot, sheath brown rot, fungal diseases

like blast, brown spot, narrow brown leaf spot, sheath

blight, false smut and viral diseases like Rice Tungro,

Rice Grassy Stunt, Rice Yellow Dwarf (TamilNadu

Agricultural University, 2024). These diseases when

identified at early stages can be prevented from

spreading and affecting other crops. Crop

phenotyping is gaining popularity in precision

agriculture recently. In phenotyping different classes

https://orcid.org/0000-0002-3579-3384

of images such as RGB, multi-spectral and remote

sensing are used for disease identification and

different vegetative indices calculation. Such early

interventions increases productivity, sustainability,

and quality of the paddy crops.

The traditional machine learning models had

several challenges like manual image feature

extraction, multiple phases for feature engineering,

sub set identification and ranking before

classification. Deep learning, a subset of artificial

intelligence, utilizes powerful deep neural networks

to enhance various aspects of image classification.

Deep learning's capacity to analyze and understand

complex data, particularly images and time-series

data, has demonstrated proficiency in image analysis

and recognition tasks. However, challenges remain in

implementing this technology, including the lack of

comprehensive datasets and the interpretability of

deep learning models. Additionally, concerns about

K., S., Reddy, V. P. and Deepthi, R.

Multi-Disease Detection and Classiﬁcation in Paddy Using Deep Convolutional Neural Networks.

DOI: 10.5220/0012881000004519

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Emerging Innovations for Sustainable Agriculture (ICEISA 2024), pages 32-39

ISBN: 978-989-758-714-6

scalability, adaptability, and computational resource

limitations are essential to address.

In this paper, we deploy a Convolution Neural

Network (CNN) ResNet 50 architecture and Visual

Transformer Model (ViT) that classifies 3 paddy

diseases namely Brown Spot , Hispa and LeafBlast.

CNN is a type of deep neural network, employing

supervised learning, which automatically extracts

features from the training images and classifies the

images. CNNs, a specialized deep learning

architecture, can be trained on diverse datasets

containing images of healthy and diseased paddy

leaves. Once trained, these models can accurately

diagnose potential issues and support timely

interventions, ultimately preventing extensive crop

losses and fostering sustainable practices.

Furthermore, deep learning's image recognition

capabilities can aid in sorting and grading Rice leaves

based on various parameters such as size, color, and

ripeness.

2 RELATED WORK

In (Batchuluun et al., 2022) the authors use thermal

images for image classification using CNN and

explainable AI on open paddy and self-collected rose

leaves data sets. The proposed CNN 16 architecture

used Class Activation Map XAI layer followed by a

GAN discriminator in the custom architecture. The

authors (Singh et al., 2022)propose a custom

developed CNN architecture which is evaluated by

two optimizers Adam and Stochastic Gradient

Descent with Momentum. The dataset was custom

collected in the fields of Orissa and later augmented.

The work compares the performance of both the

optimizers with different kernel sizes. Most of the

erroneous classification occurred in healthy images

because of the background noise and the Adam

optimizer performed better. Eight diseases that are

more predominant in Bangladesh are identified (Ahad

et al., 2023). The best performing CNN architectures

were used in the classification of the eight diseases.

Six original individual architectures, three transfer

learning architectures and one ensemble architecture

were analyzed. The architectures using transfer

learning performed better than individual models.

The proposed ensemble model DEX provided

consistent accuracy across diseases. The VGG 19

model is used for detecting brown spot diseases in

paddy leaves(Dogra et al., 2023). The model employs

transfer learning which provides improvement in

accuracy.

The other type of work integrates sequencing

models like Bi-GRU (Bi directional gated recurrent

unit) with the CNN architectures(Lu et al., 2023). The

outputs of the two models are concatenated and

passed to the classification layer. The original block

attention module in the inception layer also was

improvised to use convolution block attention

mechanism to generate the feature map. The work in

a hierarchical model for detecting Rice Blast from

UAV images of different types of rice with

noise(Shaodan et al., 2023). The initial phase had a

Swin transformer for the fine-grained recognition of

features. The following phases used trinomial tree

structures to capture detailed local information,

finally predicting the output label. The authors

developed a mobile application to detect multiple rice

diseases and nutritional deficiencies (Nayak et al.,

2023). The work combined open data sets and field

collected data set for training the model. The images

were resized, and the background was removed, along

with using stochastic depth cut optimization made the

models to converge faster in training phase. The

comparison of models on mobile and cloud-based

platforms are also discussed. The real time detection

using phone image capture on field without the

requirement of internet connection made the mobile

app more usable by farmers on field. The work in

(Stephen et al., 2023) have used a fine tuned ResNet

architecture that enhances the feature extraction

capability of the model. Some authors extract the

features using different convolutional layers and then

the features are classified using Machine Learning

Algorithms (Aggarwal et al., 2023). The authors in

(Simhadri et al., 2023) compared the performance of

15 CNN architectures with transfer learning included,

and observed that the Inception v3 model performed

better. The work in (Abasi et al., 2023) proposed a

customized CNN architecture tailored for paddy leaf

diseases. This work employed using transfer learning

with EfficientNet and Inception v3 architectures.

When a single plant is suffering from multiple

diseases, the work proposed in (Yang et al.,

2023)delivers a solution to find the affected area and

classify diseases. The authors of (Bouacida et al.,

2024) develop a small inception architecture to test

how a model trained on one crop classifies the

diseases on other crops. The work in (Trinh et al.,

2024) uses both CNN and YOLO v8 as a multi step

identification and classification of paddy leaf

diseases. The summary of the different paddy leaf

disease classification is given in Table 1.

Multi-Disease Detection and Classiﬁcation in Paddy Using Deep Convolutional Neural Networks

Table 1: Comparison of computer vision based leaf disease classification

Reference Data Set Algorithm

Used

Disease

Classifi

GPU Accura

cy (%)

Image

Type

Image

Count

Augmentati

Pixel

Size

(Batchulu

un et al.,

2022)

Custom

(Rose)

and Open

(Paddy)

PlantXDAI

(Custom)

Blast,

Bacteri

al leaf

blight,

Hispa,

Leaf

fodder,

Leaf

Spot

Yes

(NVIDI

A Titan

90.04 Therm

636

augment

ed to

3576

Yes 220x220

(Singh et

al., 2022)

Custom Custom

CNN

Bactrei

blight,

Blast,

Brown

spot,

Tungro

No 99 RGB 7332

augment

ed to

35190

Yes 256x256

(Ahad et

al., 2023)

Banglade

sh Rice

Research

institute,

Custom

Inception

V3,

DenseNet,

Mobile

Net,

ResNet,

SeresNet,

EfficientNe

t, Xception

Bactrei

blight,

Blast,

Brown

spot,

Tungro,

Leaf

Scald,

Leaf

Smut,

Hispa,

Shath

Blight

Yes.

Google

Colab

Tesla

97.62

(DEX)

RGB 1800

augment

ed to

85752

Yes 132x132

(Lu et al.,

2023)

Custom CNN+BiG

Rice

blast,

Sheath

blight,

Brown

Spot,

Leaf

Blight

Yes

RTX30

98.21 RGB 2414

augment

ed to

6000

Yes 224x224

(Shaodan

et al.,

2023)

Custom Custom Rice

Blast

Yes

Three

Tesla

V100

92.5 UAV 1702 No 5742x36

ICEISA 2024 - International Conference on ‘Emerging Innovations for Sustainable Agriculture: Leveraging the potential of Digital

Innovations by the Farmers, Agri-tech Startups and Agribusiness Enterprises in Agricu

(Nayak et

al., 2023)

PlantVilla

ge +

Custom

DenseNet,

ResNet 50,

Xception,

MobileNet

diseases

Yes,

NVidia

P100,

Mali

G52

MC2 in

mobile

98 RGB 2259 No 300x300

(Stephen

et al.,

2023)

Kaggle Improvised

ResNet

Hispa,

Brown

Spot,

Leaf

Blast

Yes

NVIDI

P5000

98 RGB 3355 Yes 256x256

(Aggarwa

l et al.,

2023)

Custom Multiple

Convolutio

n layers for

feature

extraction,

Bacteri

al leaf

blight,

Brown

Spot,

Blast

No 94 RGB 551 None 224x224

3 PROPOSED METHODOLOGY

3.1 Data Set

The data set we used was from Kaggle(Kaggle, n.d.)

which had a total of 3355 images with 779 for Leaf

blast, 565 for Hispa, 523 for Brown spot, 1488 for

Healthy images. We converted the images to

448x448, The experiments were performed on a

system with Intel i5 processor, 64 bit OS and 16GB

RAM.

3.1.1 Leaf Blast (Magnaporthe oryzae)

Leaf blast is a fungal disease occurring at all stages of

growth. It occurs in regions of frequent rainfalls or

cool temperature. Green or grey or white lesions start

to appear on the leaves as shown in Figure 1. The

lesions are broad in the central parts of the leaf and

pointed at the end. The lesions in due course of

growth cycles might enlarge and kill the leaves.

3.1.2 Brown Spot (Helminthosporium

oryzae):

Brown spot disease are result of infected seeds, After

blast it is the second most occurring disease in paddy.

It affects the quality and quantity of rice produced.

This mainly occurs between seedling to milky stage.

The disease starts as small brown spots and later

spreads a big dark brown oval spot with a yellow halo

as shown in Figure 2.

Figure 1: Leaf Blast.

Figure 2: Brown Spot.

Multi-Disease Detection and Classiﬁcation in Paddy Using Deep Convolutional Neural Networks

3.1.3 Hispa (Dicladispa armigera, Olivier)

Hispa is a damage done due to the Rice Hispa insect.

The insect grows a lot in monsoon and in pre

monsoon seasons. Proximal weed will cause this

insect to grow fast. The insect scrapes the upper part

of the leaf leaving only the lower epidermis part of

the leaf as shown in Figure 3.

Figure 3: Hispa.

Hispa effect can cause up to 20% loss in production.

Care must be takes to identify the infestation at an

early stage. Effort and experience help identify the

streaks and lower epidermis in parallel leaves along

with identification of accurate feeding marks left by

the insect.

3.2 Disease Classification Process

In our work, we have chosen two fungal diseases and

one insect-based damage. All three of these have to

be identified at an early stage and should be classified

correctly. Few symptoms between fungal diseases is

common and high precision analysis is required to

classify them from images. Similarly, Rice Hispa also

requires more accurate model for classification.

We apply two different Convolutional Neural

Network (CNN) architectures to classify the paddy

leaf diseases. CNNs are a type of deep learning

networks with different architectures which are

extremely good in learning the features from images

automatically. A CNN network has many layers

which are useful in applying features extracted to the

tasks like image classification.

1. The convolutional layer is the cornerstone of

CNNs, serving as the fundamental building block. By

sliding specialized filters (known as kernels) across

the input image, it carries out convolutions to capture

important local features such as edges, color depth,

curvatures, textures, and patterns. These resulting

feature maps represent the various spatial locations of

the extracted features in the images.

2. At the conclusion of each convolutional

operation, an activation function is employed to add

a non-linear element, thereby boosting the model's

ability to comprehend intricate connections within the

dataset. This strengthens the model's capability to

learn intricate relationships within the data.

3. Pooling layers are crucial components that

reduce the spatial dimensions of the output feature

maps from the convolution layer and alleviate the

computational burden, while preserving the most

relevant information. Among the popular techniques

commonly used in pooling layers are max-pooling

and average-pooling, which extract the highest or

average values from within specific regions of the

feature map.

4. Fully Connected Layer: This is one of the final

layers in a CNN that takes the output of the previous

layers which is a flattened matrix, representing the

extracted features. This is present just before the

output layer that uses this feature matrix to classify

the output class.

5. Flattening: It prepares the feature maps for

traditional neural network processing by converting

them into a one-dimensional vector before passing

them onto the fully connected layers. This allows for

a smooth and streamlined flow of data.

6. Dropping Out for Better Results: When it

comes to preventing overfitting, one effective

approach is using dropout. This handy technique

involves randomly omitting neurons during training,

with a specified chance of setting their output to zero.

By forcing the network to rely on alternate options for

making predictions, dropout can greatly enhance

overall accuracy.

7. Normalization:

Batch normalization is applied between layers in

CNN in order to standardize data instead of only on

the raw input image. It can help using larger learning

rates and faster convergence possibilities.

8. Output Layer: The last layer that converts the

activation functions and classifies producing the

classification probabilities of the different classes.

All the above layers with different configuration

comprise of various architectures of CNN. The layers

should be able to learn the features efficiently without

over fitting or missing any vital information.

By training on diverse datasets, these models can

identify subtle variations and defects that might be

imperceptible to the human eyes.

ICEISA 2024 - International Conference on ‘Emerging Innovations for Sustainable Agriculture: Leveraging the potential of Digital

Innovations by the Farmers, Agri-tech Startups and Agribusiness Enterprises in Agricu

We apply ResNet50 and ViT architecture to the

dataset and compare the performance of the models.

The general process flow of ResNet 50 architecture

and the ViT models are represented in Figure 4 and

Figure 5 respectively.

Figure 4: ResNet 50 process flow.

The ResNet 50 version of CNN is 50 layers deep.

This architecture learns residual connections that

helps to map input to output.

The Vision Transformers uses attention

mechanism to differentially weigh different parts of

the input data. The ViT model converts the images to

patches without overlaps, converted into vectors and

then processed using transformer architectures. This

is the process used in text by Large Language Models

and is now significantly producing results for image

classification tasks.

Figure 5: Vision Transformers process flow.

4 RESULTS AND DISCUSSION

4.1 ResNet 50 CNN Model

The training and validation loss, accuracy for epochs

of 10, 20 and 50 are given in Figure 6a,6b and 6c.

Figure 6a: Training and Validation Loss, Accuracy for 10

epochs.

Figure 6b: Training and Validation Loss, Accuracy for 20

epochs.

Figure 6c: Training and Validation Loss, Accuracy for 50

epochs.

Similarly, the confusion matrix for 10, 20 and 50

epochs are presented in Figure 7a, 7b and 7c

respectively.

Multi-Disease Detection and Classiﬁcation in Paddy Using Deep Convolutional Neural Networks

Figure 7a: Confusion Matrix for 10 epochs.

Figure 7b: Confusion Matrix for 20 epochs.

We developed a user interface using Streamlit

where an image can be uploaded and the

classification result can be viewed. The below

Figures 8a and 8b are the sample images for Brown

Spot and Healthy classes.

Figure 7c: Confusion Matrix for 50 epochs.

Figure 8a: Brown spot classification.

Figure 8b: Healthy leaf classification.

4.2 Vision Transformer Model

The training and validation accuracy of the Vision

Transformer model is given in Figure 9 for 15 epochs.

ICEISA 2024 - International Conference on ‘Emerging Innovations for Sustainable Agriculture: Leveraging the potential of Digital

Innovations by the Farmers, Agri-tech Startups and Agribusiness Enterprises in Agricu

Figure 9: Training and Validation Loss, Accuracy for 15

epochs.

5 CONCLUSIONS

Paddy leaf disease identification and classification at

earlier stages could be very useful to farmers in

treating the crops. In this work we have taken two

fungal diseases and an insect base disease which have

similar patterns when viewed with normal eyes. The

image classifiers we used are ResNet 50 architecture

of CNN and the Vision Transformer model. We got

close to 90% accuracy in ResNet50, whereas we got

upto 45% accuracy with 15 epochs in the ViT model.

The ViT model is more promising in classifying

generic images and could be improved for better

accuracy in our task of paddy leaf disease

classification by providing better runtime resources

and more images in class.

REFERENCES

Abasi, A. K., Makhadmeh, S. N., Alomari, O. A., Tubishat,

M., & Mohammed, H. J. (2023). Enhancing Rice Leaf

Disease Classification: A Customized Convolutional

Neural Network Approach. Sustainability

(Switzerland), 15(20). doi: 10.3390/su152015039

Aggarwal, M., Khullar, V., Goyal, N., Singh, A., Tolba, A.,

Thompson, E. B., & Kumar, S. (2023). Pre-Trained

Deep Neural Network-Based Features Selection

Supported Machine Learning for Rice Leaf Disease

Classification. Agriculture (Switzerland), 13(5). doi:

10.3390/agriculture13050936

Ahad, M. T., Li, Y., Song, B., & Bhuiyan, T. (2023).

Comparison of CNN-based deep learning architectures

for rice diseases classification. Artificial Intelligence in

Agriculture, 9, 22–35. doi:

https://doi.org/10.1016/j.aiia.2023.07.001

Batchuluun, G., Nam, S. H., & Park, K. R. (2022). Deep

learning-based plant classification and crop disease

classification by thermal camera. Journal of King Saud

University - Computer and Information Sciences,

34(10), 10474–10486. doi:

10.1016/j.jksuci.2022.11.003

Bouacida, I., Farou, B., Djakhdjakha, L., Seridi, H., &

Kurulay, M. (2024). Innovative deep learning approach

for cross-crop plant disease detection: A generalized

method for identifying unhealthy leaves. Information

Processing in Agriculture. doi:

10.1016/j.inpa.2024.03.002

Dogra, R., Rani, S., Singh, A., Albahar, M. A., Barrera, A.

E., & Alkhayyat, A. (2023). Deep learning model for

detection of brown spot rice leaf disease with smart

agriculture. Computers and Electrical Engineering,

109, 108659. doi:

https://doi.org/10.1016/j.compeleceng.2023.108659

Kaggle. (n.d.). Rice Leaves Data Set. Retrieved from

https://www.kaggle.com/datasets/shayanriyaz/riceleafs

Lu, Y., Wu, X., Liu, P., Li, H., & Liu, W. (2023). Rice

disease identification method based on improved CNN-

BiGRU. Artificial Intelligence in Agriculture, 9, 100–

109. doi: 10.1016/j.aiia.2023.08.005

Nayak, A., Chakraborty, S., & Swain, D. K. (2023).

Application of smartphone-image processing and

transfer learning for rice disease and nutrient deficiency

detection. Smart Agricultural Technology, 4. doi:

10.1016/j.atech.2023.100195

Shaodan, L., Yue, Y., Jiayi, L., Xiaobin, L., Jie, M.,

Haiyong, W., Zuxin, C., & Dapeng, Y. (2023).

Application of UAV-Based Imaging and Deep

Learning in Assessment of Rice Blast Resistance. Rice

Science, 30(6), 652–660. doi:

10.1016/j.rsci.2023.06.005

Simhadri, C. G., & Kondaveeti, H. K. (2023). Automatic

Recognition of Rice Leaf Diseases Using Transfer

Learning. Agronomy, 13(4). doi:

10.3390/agronomy13040961

Singh, S. P., Pritamdas, K., Devi, K. J., & Devi, S. D.

(2022). Custom Convolutional Neural Network for

Detection and Classification of Rice Plant Diseases.

Procedia Computer Science, 218, 2026–2040. doi:

10.1016/j.procs.2023.01.179

Multi-Disease Detection and Classiﬁcation in Paddy Using Deep Convolutional Neural Networks