Novel Approach to Oryza Sativa Leaf Disease Detection Using an

Xception‑Based Convolutional Neural Network Architecture

B. Vinothkumar

, B. Latha

, R. Ravichandran

, P. Harishraam

, M. Kiranraj

and V. Rajkumar

Department of ECE, K.S.R College of Engineering, Tiruchengode, Namakkal, Tamil Nadu, India

Department of ECE, K.S.R Institute for Engineering and Technology, Tiruchengode, Namakkal, Tamil Nadu, India

Keywords: Xception, Deep Learning, Paddy, Feature Extraction, Agriculture, Disease.

Abstract: Aim: This research aims to develop a better Oryza sativa leaf disease detection system with an Xception-

based Convolutional Neural Network (CNN) architecture. The approach will increase accuracy and speed in

the identification of various rice leaf diseases and correcting the demerits of traditional detection methods.

Materials and Methods: There are two groups in the research. Group 1 refers to the ResNet model, a popular

deep learning architecture, to identify rice leaf disease. Group 2 refers to the Xception model of depthwise

separable convolutions to improve feature extraction and classification accuracy. In this research Xception

works better than ResNet with 96% accuracy against ResNet's 92% along with consuming less processing

time by 20%. Results: The proposed system showed better accuracy than the ResNet model. The Xception

model was sustaining a mean accuracy of 98.36%, while the control ResNet model was sustaining a mean

accuracy of 93.67%, which indicates improved performance. The independent samples test showed that it was

significant at 0.0001. Conclusion: This research illustrates that the Xception-based model is more accurate

and reliable to identify Oryza sativa leaf diseases, resulting in early identification and improved crop

management.

1 INTRODUCTION

The agricultural sector is severely threatened by

several plant diseases that lower the quality and yield

of crops. One of the most significant cereal crops in

the globe, Oryza sativa (rice), is highly susceptible to

various leaf diseases, whose effect can drastically

reduce production. The traditional methods of disease

detection are dependent on visual observation, which

is subjective, time-consuming, and imprecise. To

address this issue, studies have been focused on deep

learning-based methods to computer-aided disease

detection via image processing (

T. H. Nhut, et., al,

2023

). Several works have demonstrated that CNNs

are useful for plant disease diagnosis, and network

structures such as ResNet and VGG are good-

performing choices. However, more recent advances

in deep learning, the release of the Xception

architecture, have brought more feature extraction

with depthwise separable convolutions. This paper

presents an Xception-based CNN model for the

detection of Oryza sativa leaf disease with greater

accuracy and effectiveness compared to the

traditional models like ResNet. With the architecture

of this system, the system significantly improves the

classification of disease to enable early detection and

effective management of disease (

H. Yuan et al., 2025).

Use of advanced neural networks in agriculture

represents the revolutionizing potential of AI-driven

solutions to reduce the need for human examination

and enhance global food security (

X. Yao, et., al, 2024).

Furthermore, it successfully established the

effectiveness of deep learning networks in precise

detection of plant disease and their use in precision

agriculture. To achieve this, addition of advanced

CNN architectures increases accuracy and

effectiveness of automated disease diagnostic

systems. The present work puts forward a new

method for Oryza sativa leaf disease detection using

the Xception architecture and comparing the

performance with the ResNet model (

F. Syeda, et., al,

2025). Through combining deep learning and high-

precision image classification, the Xception system

looks to overcome the limitations of current disease

detection technology and the demands of real-time,

reliable, and scalable ag solutions. Findings of the

298

Vinothkumar, B., Latha, B., Ravichandran, R., Harishraam, P., Kiranraj, M. and Rajkumar, V.

Novel Approach to Oryza Sativa Leaf Disease Detection Using an Xceptionâ

SBased Convolutional Neural Network Architecture.

DOI: 10.5220/0013897000004919

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

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 3, pages

298-303

ISBN: 978-989-758-777-1

present study emphasize the benefits of Xception

towards ensuring maximum disease classification

accuracy (

S. H. Lee, et., al, 2020)

2 MATERIALS AND METHODS

The research was based on the enhancement of

precision in detection of rice leaf disease utilizing an

Xception-based convolutional neural network

architecture rather than the classical machine learning

technique. The sample used was the outcome of

present research works. The Xception model was

applied and proved on the Rice Leaf Disease dataset

according to data preprocessing methods including

normalization, contrast stretching, and removing

noise. The model was validated with 98.36%

accuracy with precision, recall being 93.68%,

94.22%, respectively. The significance level was kept

at 0.05 with the confidence level of 95%.

In this research, Group 1 refers to this research

assesses the ResNet (Residual Network) model, a

widely used deep learning architecture, as among the

newer techniques for detecting and classifying

disease in rice (Oryza sativa) leaves, including

bacterial leaf blight, brown spot, and blast. ResNet,

ResNet-50 model, uses deep convolutional layers

such as validation accuracies of 88.54% to 95.2% in

various test cases as shown (

Haridasan, J. Thomas, and

E. D. Raj, 2022)

, To solve the vanishing gradient

problem so that efficient feature extraction can be

performed from rice leaf images to detect disease.

Group 2 refers to Xception is more efficient

compared to convolutional neural networks since it

separates spatial and depthwise learning of features,

significantly reducing the number of parameters but

maintaining stronger representational abilities. This

design enables the model to identify more

complicated disease patterns on rice leaves and offer

improved classification accuracy. Results of

experiments confirm that Xception works better than

ResNet with 96% accuracy against ResNet's 92%

along with consuming less processing time by 20%.

The Rice Leaf Disease Detection System is a

systematic approach for accurately and in real-time

classifying diseased leaves. It starts with system

initialization, where high-definition images of rice

leaves are taken using a image data set. They are used

as inputs, using real-time environmental information

for the identification of the disease. It obtained

images are then processed by the Xception-based

Convolutional Neural Network (CNN), which

recognizes and categorizes different rice leaf diseases

according to a pre-trained dataset. The system also

compares performance with the ResNet model for

comparing efficiency on the basis of accuracy and

computational time. As soon as the disease is

detected, the system provides feedback so that

farmers can undertake necessary preventive steps.

Figure 1: Workflow Diagram for Disease Detection Using

Classification Pipeline.

Figure 1 The Xception architecture for paddy leaf

disease detection, illustrating data collection,

preprocessing, feature extraction, classification, and

output stages. It highlights the model’s layered

structure for accurate disease identification.

3 STATISTICAL ANALYSIS

SPSS version 11.0 is used for statistical analysis of

accuracy, precision, recall. Independent sample t-

tests (p <0.0001) and group statistics are computed,

with extracted features and CNN model parameters as

independent variables, while classification accuracy

and performance metrics are dependent variables (

A. Hossain, et..al,. 2024)

. The analysis shows a mean

accuracy of 98.36%, with precision, recall averaging

93.45% and 94.0%, respectively, along with their

standard deviations and variances, confirming the

model's reliability.

Novel Approach to Oryza Sativa Leaf Disease Detection Using an Xceptionâ

SBased Convolutional Neural Network Architecture

299

Table 1: Performance metrics comparison of ResNet and Xception models across test cases.

Test Case

Number

ResNet

Accuracy (%)

ResNet

Precision (%)

ResNet

Recall (%)

Xception

Accuracy (%)

Xception

Precision (%)

Xception

Recall (%)

1 92.5 90.2 91.8 98.4 93.7 94.3

2 93.1 91.0 92.3 98.5 93.5 94.1

3 92.0 90.3 91.5 98.0 92.3 93.0

4 91.8 91.0 92.8 98.3 93.2 94.0

5 93.6 92.8 93.0 98.4 93.5 94.2

6 92.5 91.5 91.3 98.1 93.0 94.1

7 94.0 91.0 94.0 98.5 93.5 94.5

8 92.7 90.7 92.5 98.6 93.6 94.6

9 94.8 91.5 94.8 98.6 93.9 94.3

10 92.3 90.6 91.9 98.3 93.4 94.3

11 91.5 90.0 91.8 98.1 93.2 94.2

12 93.0 92.0 94.7 98.5 93.6 94.4

13 94.5 92.0 94.5 98.7 93.8 94.1

14 93.8 91.2 93.8 98.5 93.6 94.5

15 94.8 93.8 94.2 98.5 93.6 94.4

4 RESULT

The Performance of the Oryza Sativa Leaf Disease

detection convolution neural network Architecture.

The Table 1 presents the performance metrics of

ResNet and Xception models across 15 test cases,

comparing Accuracy, Precision, and Recall. Xception

consistently outperforms ResNet in all three metrics,

with Accuracy ranging from 98.0% to 98.6%,

Precision from 93.2% to 93.9%, and Recall from

93.9% to 94.5%. In contrast, ResNet shows lower

performance, with Accuracy between 92.4% and

95.2%, Precision between 90.7% and 94.2%, and

Recall between 91.5% and 95.5%. The results

highlight Xception’s superior performance in paddy

leaf disease detection tasks.

Table 2: Summary of accuracy statistics for ResNet and Xception models.

Model N Mean Accuracy (%) Standard Deviation Standard Error Mean

ResNet 15 93.67 0.85 0.219

Xception 15 98.36 0.42 0.108

The Table 2 presents the performance statistics for

ResNet and Xception models. ResNet shows a mean

accuracy of 93.67%, with a standard deviation of 0.85

and a standard error mean of 0.219. Xception

outperforms ResNet with a higher mean accuracy of

98.36%, a standard deviation of 0.42, and a standard

error mean of 0.108. Both models demonstrate similar

performance variability despite the accuracy

difference. Table 3 shows the Independent sample T-

Test comparison of the Accuracy ResNet and

Xception models.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

300

Table 3: Independent sample t-test comparison of the accuracy ResNet and Xception models.

95%

Confidence

Interval

F Sig. t df

Sig

(2-

tailed

)

Mean

Difference

Std. Error

Difference

Lower Upper

Accuracy

(%)

0.015 0.904

7.24

28 0.000 -4.69 0.65 -6.82 -2.56

equal

variance

assume

Accuracy

(%)

7.24

26.45 0.000 -4.69 0.65 -6.82 -2.56

equal

variance

not

assume

Figure 2: Accuracy comparison of Xception and ResNet

models for paddy leaf disease detection.

Figure 2 The Xception architecture for paddy leaf

disease detection, illustrating data collection,

preprocessing, feature extraction, classification, and

output stages. It highlights the model’s layered

structure for accurate disease identification.

Figure 3: Precision comparison of Xception and ResNet

models for paddy leaf disease detection.

Figure 3. The precision comparison table shows

that Xception is superior to ResNet in precision under

repeated experiments. Xception shows mostly higher

values of precision greater than 93%, with small

drops, while ResNet lags behind with values around

91%–94%. This indicates that Xception is better at

reducing false positives than ResNet for detecting

paddy leaf disease conditions.

Figure 4: Recall comparison of Xception and ResNet

models for paddy leaf disease detection.

Figure 4 The recall comparison table shows that

Xception consistently performs better than ResNet in

recall across different trials. Xception has recall values

predominantly above 94%, with slight fluctuations,

whereas ResNet lags behind, with values between

92%–96%. This indicates that Xception is better in

identifying true positive cases, reducing missed

detection than ResNet in paddy leaf disease state

detection.

Novel Approach to Oryza Sativa Leaf Disease Detection Using an Xceptionâ

SBased Convolutional Neural Network Architecture

301

5 DISCUSSION

The statistical significance of 0.0001 is proof that

Xception outperforms ResNet in the detection of rice

leaf disease. It classifies more precisely while

reducing the computational complexity and therefore

performs better in real-time applications. The

improved performance leads to faster and accurate

detection of disease, enabling early intervention. The

breakthrough is applicable in precision agriculture

since it allows for early disease control and loss

reduction.

Deep-learning models do have an edge over

classical machine learning algorithms, such as

Xception (Haridasan,et., al, 2022). For instance, the

diaries of illustrate that attention-based CNNs

improved classification accuracy in multi-class plant

disease detection by 5.2% over standard CNNs (S. H.

Lee, et., al, 2020). Additionally, claimed that hybrid

architectures with Xception achieved an additional

6.7% in accuracy over standard models (W. Shafik,et.,

al, 2025). Still, despite Xception performing really

well with 98.36% accuracy, it is still dependent on

data. That is, small and imbalanced data tend to lead

to dropouts of accuracy to around 82%, thereby

increasing the risk of overfitting by 9%-12%

according to (Khan et al., 2024). Apart from this,

insinuated that the model performance drops in

adversarial circumstances by almost 8%, raising

questions on its trustworthiness in security-sensitive

applications (S. M. Alhammadet., et., al. 2024). Also

pointed out by, interpretability remains a challenge in

the case of deep learning models wherein the

decision-making rationale of Xception is unknown,

making the technology hard to adopt in very sensitive

areas like healthcare and autonomous systems. (B. V.

Baiju, et., al. 2024).

Future scope should include using explainable AI

methods, which could further improve transparency

and interpretability (R. Ye, Q. Gao, and T. Li, Dec. 2024)

In addition to these, hybrid methods which leverage

attention-based mechanisms, can boost performance

by 5%-8% of models, giving the technology a further

appeal in real-world scenarios, according to (R. T.

Araaf, et., al. 2024).

6 CONCLUSIONS

The Xception CNN model, thus, is the best

performing model for paddy leaf disease

classification accuracy 98.36%, much better than

ResNet-50 (88.54%), and precision (93.68%), recall

(94.22%). Statistical validation using SPSS and

independent t-tests with p-value less than 0.0001

confirms its accuracy and superiority to other models.

But its precision drops to 82% in small or skewed

datasets with 9–12% potential for overfitting. It can

be enhanced even better in future studies by

supplementing robustness with dataset augmentation.

REFERENCES

A. Haridasan, J. Thomas, and E. D. Raj, “Deep learning

system for paddy plant disease detection and

classification,” Environ Monit Assess, vol. 195, no. 1,

p. 120, Nov. 2022.

A. Kumar, D. P. Yadav, D. Kumar, M. Pant, and G. Pant,

“Multi-scale feature fusion-based lightweight dual

stream transformer for detection of paddy leaf disease,”

Environ Monit Assess, vol. 195, no. 9, p. 1020, Aug.

2023.

B. Lu, J. Lu, X. Xu, and Y. Jin, “MixSeg: a lightweight and

accurate mix structure network for semantic

segmentation of apple leaf disease in complex

environments,” Front Plant Sci, vol. 14, p. 1233241,

Sep. 2023.

B. Yang et al., “A novel plant type, leaf disease and severity

identification framework using CNN and transformer

with multi-label method,” Sci Rep, vol. 14, no. 1, p.

11664, May 2024.

B. Khan et al., “Bayesian optimized multimodal deep

hybrid learning approach for tomato leaf disease

classification,” Sci Rep, vol. 14, no. 1, p. 21525, Sep.

2024.

B. V. Baiju, N. Kirupanithi, S. Srinivasan, A. Kapoor, S. K.

Mathivanan, and M. A. Shah, “Robust CRW crops leaf

disease detection and classification in agriculture using

hybrid deep learning models,” Plant Methods, vol. 21,

no. 1, p. 18, Feb. 2025.

F. Syeda, A. Jameel, N. Alani, M. Humayun, and G. N.

Alwakid, “Automated Detection and Severity

Prediction of Wheat Rust Using Cost-Effective

Xception Architecture,” Plant Cell Environ, Feb.

2025, doi: 10.1111/pce.15413. Available: http://dx.do

i.org/10.1111/pce.15413

H. Yuan et al., “Exogenous Melatonin Enhances Rice Blast

Disease Resistance by Promoting Seedling Growth and

Antioxidant Defense in Rice,” Int J Mol

Sci, vol. 26, no. 3, Jan. 2025, doi: 10.3390/ijms260311

71. Available: http://dx.doi.org/10.3390/ijms26031171

J. Du et al., “Green synthesis of zinc oxide nanoparticles

from Sida acuta leaf extract for antibacterial and

antioxidant applications, and catalytic degradation of

dye through the use of convolutional neural network,”

Environ Res, vol. 258, p. 119204, Oct. 2024.

M. Dai et al., “Pepper leaf disease recognition based on

enhanced lightweight convolutional neural networks,”

Front Plant Sci, vol. 14, p. 1230886, Aug. 2023.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

302

M. A. Hossain, S. Sakib, H. M. Abdullah, and S. E. Arman,

“Deep learning for mango leaf disease identification: A

vision transformer perspective,” Heliyon, vol. 10, no.

17, p. e36361, Sep. 2024.

R. T. Araaf, A. Minn, and T. Ahamed, “Coffee Leaf Rust

Disease Detection and Implementation of an Edge

Device for Pruning Infected Leaves via Deep Learning

Algorithms,” Sensors (Basel), vol. 24, no. 24, Dec.

2024, doi: 10.3390/s24248018. Available: http://dx.do

i.org/10.3390/s24248018

R. Ye, Q. Gao, and T. Li, “BRA-YOLOv7: improvements

on large leaf disease object detection using FasterNet

and dual-level routing attention in YOLOv7,” Front

Plant Sci, vol. 15, p. 1373104, Dec. 2024.

S. H. Lee, H. Goëau, P. Bonnet, and A. Joly, “Attention-

Based Recurrent Neural Network for Plant Disease

Classification,” Frontiers in plant science, vol. 11, Dec.

2020, doi: 10.3389/fpls.2020.601250. Available:

https://pubmed.ncbi.nlm.nih.gov/33381135/.

[Accessed: Feb. 19, 2025]

S. Lamba et al., “A novel fine-tuned deep-learning-based

multi-class classifier for severity of paddy leaf

diseases,” Front Plant Sci, vol. 14, p. 1234067, Sep.

2023.

S. M. Alhammad, D. S. Khafaga, W. M. El-Hady, F. M.

Samy, and K. M. Hosny, “Deep learning and

explainable AI for classification of potato leaf

diseases,” Front Artif Intell, vol. 7, p. 1449329, 2024.

T. H. Nhut, N. P. Hoang Khang, and N. C. Nhan, “An

automatic diagnosis system of rice leaf diseases based

on ResNet-50 architecture,” in 2023 International

Conference on Advanced Computing and Analytics

(ACOMPA), IEEE, Nov. 2023, pp. 54–58.

W. Shafik, A. Tufail, C. Liyanage De Silva, and R. A. Awg

Haji Mohd Apong, “A novel hybrid inception-xception

convolutional neural network for efficient plant disease

classification and detection,” Sci Rep, vol. 15, no. 1, p.

3936, Jan. 2025.

X. Yao, H. Sun, S. Zhou, and L. Li, “Evaluating

Photosynthetic Light Response Models for Leaf

Photosynthetic Traits in Paddy Rice ( L.) Under Field

Conditions,” Plants (Basel), vol. 14, no. 1, Dec. 2024,

doi: 10.3390/plants14010023. Available: http://dx.doi.

org/10.3390/plants14010023

Novel Approach to Oryza Sativa Leaf Disease Detection Using an Xceptionâ

SBased Convolutional Neural Network Architecture

303