A Ensemble Deep Learning Approach to Classify Common

Rice Leaf Diseases

MD. Istiak Mondol, Md. Tanvir Hossain Khan, Kasfia Afrose Binti, Sazzadur Rahman Anik,
Raiyan Rahman*
United International University
United City, Madani Ave, Dhaka 1212
{mmondol201367,mkhan201389,kbinti201375,sanik201020}@bscse.uiu.ac.bd
[email protected]*

ABSTRACT of the global population. However, its production is often threat-

Bangladesh ranks as the third-largest producer of rice globally and ened by various diseases, leading to substantial yield losses. Timely
reached about 39.1 million tonnes in 2023. On the other hand, the and accurate diagnosis of these diseases is crucial for implement-
average per capita rice consumption in Bangladesh was the second ing effective control measures and ensuring food security. With
highest in Asia from 2016 to 2019. Bangladesh relies significantly advancements in technology, particularly in the field of computer
on rice to support its economy and fulfill its food requirements. It’s vision and deep learning, automated disease diagnosis systems have
very crucial to timely identify any diseases in rice plants to ensure emerged as promising tools for crop disease management.
their healthy and optimal growth and ensure proper treatment. In recent years, convolutional neural networks (CNNs) have
Since manual detection of disease is time-consuming and costly, an shown remarkable success in image recognition tasks, including
automated system can be effective in getting rid of manual disease the classification of plant diseases. By leveraging large datasets of
detection. This paper is about rice leaf disease detection using annotated images, CNNs can learn intricate patterns and features
machine learning techniques. The most common five rice plant indicative of specific diseases, thereby enabling accurate disease
diseases of Bangladesh namely bacterial leaf blight, brown spot, leaf diagnosis. Moreover, ensemble learning techniques, which combine
blast, leaf scald, and narrow brown spot diseases, and a ’healthy’ multiple models to improve predictive performance, have further
class for Unaffected leaves are detected in this work. Images of enhanced the efficacy of CNN-based disease classification systems.
affected leaves with a white background were used as the input. Inspired by the success of CNNs in various domains, particularly
After necessary pre-processing, the dataset was trained on CNN- in plant disease diagnosis, this research paper presents an investi-
based pre-trained deep learning models DenseNet121, InceptionV3, gation into the classification of rice leaf diseases using CNN-based
and ResNet50V2. ensemble learning. The proposed approach aims to leverage the
collective intelligence of multiple CNN models to achieve superior
CCS CONCEPTS accuracy and robustness in disease identification.
Drawing on the methodology established in previous studies
• Computer systems organization → Embedded systems; Re-
and adapted to the context of rice leaf diseases, this paper explores
dundancy; Robotics; • Artificial intelligence → Computer vision.
the utilization of pre-trained CNN models, such as DenseNet121,
KEYWORDS InceptionV3, and ResNet50V2R, in conjunction with ensemble learn-
ing techniques. By assembling these models and aggregating their
CNN, ANN, Deep Learning, Rice Leafs, Disease Detection predictions through a majority voting scheme, the research endeav-
ACM Reference Format: ors to enhance the reliability and generalization capability of the
MD. Istiak Mondol, Md. Tanvir Hossain Khan, Kasfia Afrose Binti, Sazzadur classification system.
Rahman Anik,, Raiyan Rahman*. 2018. A Ensemble Deep Learning Ap- The contributions of this research lie in its application of ad-
proach to Classify Common Rice Leaf Diseases. In ICCA 2024: International vanced deep learning techniques to address the pressing challenge
Conference on Computing Advancements, 05-06th September 2024, Dhaka,
of rice disease diagnosis. Through comprehensive experimentation
Bangladesh. ACM, New York, NY, USA, 5 pages. https://doi.org/10.1145/
1122445.1122456
and analysis, this paper aims to demonstrate the effectiveness of
CNN-based ensemble learning in accurately identifying and classify-
1 INTRODUCTION ing common rice leaf diseases. Furthermore, by presenting detailed
insights into the methodology and experimental results, this paper
Rice (Oryza sativa) is one of the most important staple crops world-
seeks to provide valuable guidance for future research directions
wide, serving as a primary food source for a significant portion
in the field of automated crop disease diagnosis.
Permission to make digital or hard copies of all or part of this work for personal or The structure of this paper is organized as follows: Section II
classroom use is granted without fee provided that copies are not made or distributed provides an overview of related works in the domain of plant disease
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than ACM classification. Section III delineates the methodology employed
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, in this study, including dataset preparation, model selection, and
to post on servers or to redistribute to lists, requires prior specific permission and/or a
fee. Request permissions from [email protected].
ensemble learning techniques. Subsequently, Section IV presents
ICCA 2024, 05-06th September 2024, Dhaka, Bangladesh the results of the investigation, offering a thorough analysis of the
© 2018 Association for Computing Machinery. experimental findings. Finally, Section V concludes the paper with
https://doi.org/10.1145/1122445.1122456
a summary of key findings, implications, and suggestions for future 3 METHODOLOGY
research endeavors in the field of rice disease classification.
3.1 Convolutional Neural Network
CNNs are neural networks designed specifically for tasks involving
2 LITERATURE REVIEW images, patterns, and spatial hierarchies. They excel at recognizing
Kawcher Ahmed and others [2] aimed to predict rice leaf diseases edges and complex patterns, making them essential for computer
using a Decision tree with 10-fold cross-validation using a logis- vision tasks like picture classification [9] In our paper, we use CNN
tic regression algorithm that will automate the system to detect with DenseNet121, InceptionV3, and ResNet50V2.
diseases quickly and easily which will help the farmers to take
proper steps at an early stage. Their dataset contains the three 3.2 Experimental Design
most common rice leaf disease classes namely leaf smut, bacterial In Fig 1, our initial step involved collecting datasets, where 81% of
leaf blight, and brown spot. They have achieved an accuracy of the data was allocated for training purposes, while the remaining
over 97% when applied to the test dataset. Shreya Ghosal and oth- 19% was set aside for testing. Subsequently, we executed data pre-
ers [3] Deep learning models, particularly Convolutional Neural processing procedures to enhance the quality and relevance of our
Networks (CNNs), have been employed to accurately identify rice dataset. Following this, we employed three distinct algorithms with
leaf diseases, achieving an impressive accuracy of 92.46%. Transfer CNN (Convolutional Neural Networks) and meticulously compared
Learning, specifically fine-tuning a pre-trained VGG-16 model, has their performance and outcomes.
notably enhanced the performance of the deep learning model when
trained on a small dataset of rice leaf images. The proposed model
achieved a test accuracy of 92.4% after training on 1509 images
and testing on 647 images. A comparison with a CNN model with-
out Transfer Learning demonstrated that the model with Transfer
Learning surpassed in accuracy, indicating the efficacy of Transfer
Learning in this domain. Reinald Adrian DL. Pugoy and others
conducted digital image analysis, which accurately quantifies the
color parameters of turfgrass, assisting in disease detection in rice
leaves. The study employs image segmentation, outlier detection,
and clustering to acquire the major colors of abnormal regions in
diseased rice leaves. Likelihood values derived from this analysis
can indicate the probability of a rice leaf being infected with a
particular disease, thereby aiding in disease identification. Prabira Figure 1: Experimental Design
Kumar Sethy and others [5] proposed a study that evaluated 13
CNN models in transfer learning and deep feature plus SVM ap-
proach. AlexNet was found to be better among all CNN models. 3.3 Dataset Description
The best classification model was chosen using statistical analysis,
In Fig 2, we use a rice life disease picture dataset. This dataset
with ResNet50 plus SVM performing the best. Data availability: The
contains six classes: bacterial leaf blight, brown spot, leaf blast,
diseased rice leaf images used in the study are available upon re-
healthy, leaf scald, and narrow brown spot. Each class contains 670
quest. Dataset: 5932 diseased rice leaf images were used, including
images for training and 152 images for validation.
bacterial blight, blast, brown spot, and tungro varieties. S. Phadikar
and others [6] developed an automated system to classify rice leaf
diseases based on morphological changes using Bayes and SVM
Classifier. The system initially categorizes uninfected and diseased
leaves based on histogram peaks, then utilizes Bayes’ classifier and
SVM for disease classification. Bayes’ classifier assigns patterns
to classes based on probabilities, while SVM achieves high per-
formance with maximum-margin hyperplane and kernel function.
Preprocessing steps involve enhancing image quality using mean
filtering techniques. SethuMadhavi Rallapalli and others [8] pro-
pose an improved CNN architecture named M-Net for plant disease
identification and classification. The M-Net model is designed to
effectively work on plant disease identification and classification
using datasets from the Rice Disease Image Dataset and Rice Leaf
Diseases Dataset. The model achieved 71% accuracy, outperform-
ing benchmarked methods. The model utilizes deep learning-based
CNN variants like AlexNet, customized specifically for plant disease Figure 2: Dataset of medicinal plants
identification, showing efficacy over other deep learning models
for plant disease-based applications.
2
3.4 Work Flow and pinpointing any possible shortcomings is easier when the ma-
First of all, the images are collected from kaggle rice leaf disease trix compares the model’s projected labels with the exemplification
detection, and undergo preprocessing steps to prepare them for ma- labels in a tabular format. [10]
chine learning analysis. Here we augmented the images to produce
some more images so that our model can learn better. Machine 4.1 Model Specific Result Analysis
learning algorithms are used for image classification, CNN. Three In Table II, we see that the average accuracy is 92.67 based on f1-
well-established models, namely InceptionV3, ResNet50V2, and score. CNN with DenseNet121 and InceptionV3, we see the highest
DenseNet121, are utilized for the CNN architecture. A majority vot- accuracy. The precision, Recall, and f1-score of both are the same.
ing mechanism is applied to combine the predictions. The training In Figure 3, DenseNet121 with CNN shows the height accuracy on
process involves using a dataset of 822 images distributed across the leaf scald class and the lowest accuracy on the healthy class. In
six classes, with a total dataset size of 4932 images. Figure 4, InceptionV3 with CNN shows the height accuracy on both
bacterial leaf blight and leaf scald class. and the lowest accuracy
3.4.1 Image Preprocessing: Initially, we enhanced the image quality in the healthy class. In Figure 5, ResNet50V2 with CNN shows
by removing noise and resizing the images while ensuring pixel the height accuracy on the leaf scald class and the second height
values were normalized, and finally used augmentation to increase accuracy on bacterial leaf blight class and the lowest accuracy is
the variety in the dataset. on the healthy class.
3.4.2 Machine Learning Algorithm Selection: Our selection leaned
towards CNN, a deep learning approach tailored for image recogni- Table 1: DenseNet121, InceptionV3, and ResNet50V2 base
tion tasks. models with CNN layers on top

3.4.3 Model Training: We fed the data into the models, fine-tuning DenseNet121
their parameters until they exhibited accurate image classification. Layer Output Shape Param#
densenet121(Functional) (None, 16, 16, 1024) 7,037,504
3.4.4 Model Assessment: To gauge performance, we evaluated the Flatten_1(Flatten) (None, 262,144) 0
models using a separate dataset, measuring metrics like accuracy, Dense_3(Dense) (None, 50) 13,107,250
Dense_4(Dense) (None, 20) 1,020
precision, and recall. Dense_5(Dense) (None, 6) 126
Total params: 20,145,900
3.4.5 Consensus Building: Utilizing predictions from each model, Trainable params: 13,108,396
Non-trainable params: 7,037,504
we ranked and collectively decided on the most frequent prediction
as the final outcome. Inception v3
Layer Output Shape Param#
3.5 Model Architecture Inception_v3(Functional) (None, 14, 14, 2048) 21,802,784
flatten(Flatten) (None, 401,408) 0
In Table I, the erecting and training of a CNN using the DenseNet121, dense(Dense) (None, 50) 20,070,450
InceptionV3, and ResNet50V2 architectures as base models for im- dense_1(Dense) (None, 20) 1,020
dense_2(Dense) (None, 6) 126
age classification, employing transfer learning. In terms of this, the
Total params: 41,874,380
pre-trained models are loaded with their layers frozen to preserve Trainable params: 26,215,596
learned features from the ImageNet dataset. CNN layers are added Non-trainable params: 21,802,784

on top, with a final softmax output for multi-class classification.

ResNet50 v2
The models are compiled with the RMSprop improver and categori-
Layer Output Shape Param#
cal cross-entropy loss, suitable for multi-class tasks, and trained for
resnet50v2(Functional) (None, 16, 16, 2048) 23,564,800
twenty epochs with a batch size of 8 for train and 32 for validation. Flatten_2(Flatten) (None, 524,288) 0
This transfer learning strategy harnesses the knowledge embedded Dense_6(Dense) (None, 50) 26,214,450
Dense_7(Dense) (None, 20) 1,020
in pre-trained models, offering the potential for effective classifica- Dense_8(Dense) (None, 6) 126
tion with limited training epochs, making it valuable for various Total params: 49,780,396
image recognition tasks. This approach to transfer learning taps Trainable params: 26,215,596
Non-trainable params: 23,564,800
into the wealth of knowledge already embedded within pre-trained
models, providing a pathway to achieve accurate classification even
with a relatively small number of training epochs. It’s a powerful
technique that holds significant promise for a wide range of image
Table 2: Classification Report of Models
recognition tasks.
Model Precision Recall f1-score Accuracy
4 RESULT ANALYSIS DenseNet121 0.93 0.92 0.92 –
A confusion matrix is a fundamental tool in statistics and machine InceptionV3 0.94 0.93 0.93 –
learning to assess the effectiveness of classification models. It gives a ResNet50V2 0.94 0.93 0.93 –
graphical representation of how well a model categorizes instances
into various classes or groups. Analyzing the model’s performance
3
 4.2 Ensemble Result Analysis
In Table III, The provided classification detail summarizes the per-
formance of a machine learning model on a dataset with six classes.
The model achieved an impressive accuracy of 96%, specifying that
it accurately predicted the class labels for most of the 96 instances
out of 100 instances. With minor dissimilarities, it gives an idea of
high precision, recall, and F1-score for most classes, exhibiting its
solid predictive capabilities. Classes bacterial leaf blight, leaf scald,
and brown spot achieved perfect scores across accuracy, recall,
and F1-score, specifying the model’s capability to correctly classify
these classes with high confidence. The model’s performance is
excellent, as indicated by the high weighted average F1 score of 96,
suggesting its effectiveness in multi-class classification tasks.
Table 3: Ensemble Classification Report

Figure 3: DenseNet121 with CNN precision recall f1-score

bacterial leaf blight 99 100 100
brown spot 89 97 192
healthy 97 94 95
leaf scald 100 100 100
leaf blast 95 88 91
narrow brown spot 100 100 100
accuracy - - 96
macro avg 97 96 96
weighted avg 97 96 96

Figure 4: InceptionV3 with CNN

Figure 6: Ensemble classification matrix

5 CONCLUSION AND FUTURE WORK

In this research, we provide a unique CNN-based automated sys-
tem for the recognition and classification of rice leaf disease. We use
three pre-trained models (DenseNet121, InceptionV3, and ResNet50V2)
as base models and a majority voting mechanism to combine the
predictions from the different models. The system achieves an im-
Figure 5: ResNet50V2 with CNN
pressive accuracy of 96on a dataset of 4932 images of 6 rice leaf
diseases. In the discussion of future work, we will create our dataset
by capturing photos, also we will try to cover as many classes as
possible.
4
REFERENCES Electronics Engineers in Israel.
[1] https://www.kaggle.com/datasets/dedeikhsandwisaputra/rice-leafs-disease- [11] J. T. TOWNSEND "Theoretical analysis of an alphabetic confusion matrix" Per-
dataset ception & Psychophysics, vol. 9, pp. 40–50, 1971.
[2] K. Ahmed, T. Rahman, M. Irfanul and S. Momen "Rice Leaf Disease Detection Us- [12] M. S. Kabir, M. U. Salam, A. Chowdhury, N. M. F. Rahman, K. M. Iftekharuddaula,
ing Machine Learning Techniques" 2019 International Conference on Sustainable M. S. Rahman, M. H. Rashid, S. S. Dipti, A. Islam, M. A. Latif, A. K. M. S. Islam, M.
Technologies for Industry 4.0 (STI), 24-25 December, Dhaka. M. Hossain, B. Nessa, T. H. Ansari, M. A. Ali and J. K. Biswas "Rice Vision for
[3] S. Ghosal, K. Sarkar "Rice Leaf Diseases Classification Using CNN With Transfer Bangladesh: 2050 and Beyond" angladesh Rice J. 19 (2): 1-18, 2015.
Learning" Proceedings of 2020 IEEE Calcutta Conference (CALCON). [13] R. Chauhan, K. K. Ghanshala, and R.C. Joshi "Convolutional Neural Network
[4] R. A. DL. Pugoy, V. Y. Marianoa "Automated rice leaf disease detec- (CNN) for Image Detection and Recognition" 2018 First International Conference
tion using color image analysis" Proceedings Volume 8009, Third Interna- on Secure Cyber Computing and Communication (ICSCCC).
tional Conference on Digital Image Processing (ICDIP 2011); 80090F (2011) [14] M. A. Azim, M. K. Islam, Md. M. Rahman, F. Jahan "An effective feature extraction
https://doi.org/10.1117/12.896494 Event: 3rd International Conference on Digital method for rice leaf disease classification" TELKOMNIKA Telecommunication,
Image Processing, 2011, Chengdu, China. Computing, Electronics and Control Vol. 19, No. 2, April 2021, pp. 463 470. ISSN:
[5] P. K. Sethy, N. K. Barpanda, A. K. Rath, S. K. Behera "Deep feature based rice leaf 1693-6930, accredited First Grade by Kemenristekdikti, Decree No: 21/E/KPT/2018
disease identification using support vector machine" Computers and Electronics DOI: 10.12928/TELKOMNIKA.v19i2.16488.
in Agriculture Volume 175, August 2020, 105527. [15] K.S. Archana1, A. Sahayadhas "Automatic Rice Leaf Disease Segmentation Using
[6] S. Phadikar, J. Sil, and A. K. Das "Classification of Rice Leaf Diseases Based on Image Processing Techniques" International Journal of Engineering and Technol-
Morphological Changes" International Journal of Information and Electronics ogy, 7 (3.27) (2018) 182-185. International Journal of Engineering and Technology.
Engineering, Vol. 2, No. 3, May 2012. [16] S. Phadikar, J. Sil, and A. K. Das "Classification of Rice Leaf Diseases Based on
[7] T. Akter, M. T. Parvin, F. A. Mila, and A. Nahar "Factors determining the prof- Morphological Changes" International Journal of Information and Electronics
itability of rice farming in Bangladesh" Article in Journal of the Bangladesh Engineering, Vol. 2, No. 3, May 2012.
Agricultural University · March 2019 DOI: 10.3329/jbau.v17i1.40668 [17] M. E. Pothen, and D.r.Maya L Pai "Detection of Rice Leaf Diseases Usin Im-
[8] S. Rallapalli, and M. A. S. Durai "A contemporary approach for disease identifica- age Processing" Proceedings of the Fourth International Conference on Com-
tion in rice leaf" Received: 26 February 2021 / Revised: 14 May 2021 / Accepted: puting Methodologies and Communication (ICCMC 2020). IEEE Xplore Part
27 May 2021. The Society for Reliability Engineering, Quality and Operations Number:CFP20K25-ART; ISBN:978-1-7281-4889-2.
Management (SREQOM), India, and The Division of Operation and Maintenance, [18] F. Jiang, Y. Lu, Y. Chen, D. Cai, G. Li "Image recognition of four rice leaf diseases
Lulea University of Technology, Sweden 2021. based on deep learning and support vector machine" Received 5 April 2020,
[9] R. Chauhan, K. K. Ghanshala, and R.C. Joshi " Convolutional Neural Network Revised 28 August 2020, Accepted 3 October 2020, Available online 20 October
(CNN) for Image Detection and Recognition" 2018 First International Conference 2020, Version of Record 20 October 2020.
on Secure Cyber Computing and Communication (ICSCCC).
[10] N. D. Marom, K. K. Ghanshala, and R.C. Joshi " Using the Confusion Matrix for
Improving Ensemble Classifiers"2010 IEEE 26th Convention of Electrical and