Proceedings of Papers, 4th International Conference on Science and Technology 
Faculty of Technology, South Eastern University of Sri Lanka 
ISBN 978-955-627-028-0 117

Classification of Sri Lankan Paddy Varieties using Deep Learning 

Techniques 

M.R.M. Aththas1, M.N.F. Yusra2, M.S. Sabrina3, W.A. Sanjeewa4, M. Janotheepan5 and A.R.

Fathima Shafana6 

1,2,3,4,5Department of Computer Science and Informatics, Uva Wellassa University,        Sri Lanka
6Department of Information and Communication Technology, South Eastern University of Sri Lanka, Sri Lanka 

1cst19030@std.uwu.ac.lk, 2cst19052@std.uwu.ac.lk, 3iit19062@std.uwu.ac.lk, 4aruna.s@uwu.ac.lk, 
5janotheepan.m@uwu.ac.lk, 6arfshafana@seu.ac.lk 

Abstract 

Rice is a highly consumed staple food in Sri Lanka. 
From farming phase to distribution phase of 

paddy, classification of paddy is becoming vital as 

it provides efficiency to the planning, production, 
sales and consumption. In Sri Lanka, the 

evaluation of the classification of paddy varieties 

is typically overseen by the Rice Research and 

Development Institute (RRDI). Traditionally, 

paddy identification is done manually by human 
inspectors, ensuring some level of accuracy but 

requiring significant manpower, time, and 
subjective judgment. This research seeks to 

transform the categorization of paddy varieties in 
Sri Lanka. This paper provides an approach to 

identifying and classifying paddy variety in paddy 

sample with the help of image processing and 
CNN model. For this approach, 10 varieties of 

paddy samples were collected from Rice Research 
and Development Institute. With these samples a 

dataset of more than 10,000 images were captured 

and used in this research. Image preprocessing 
involved cropping, scaling, and noise removal to 

standardize the data. Experiments were conducted 
with nine different CNN models, iterating through 

various architectures and training parameters to 

optimize performance. The experiment was 

performed on ten rice categories to evaluate the 

suggested solution. The accuracy of classification 
is of 93.69%. 

Keywords: Convolutional Neural Network (CNN), 

Depp Learning, Paddy Classification 

I.INTRODUCTION

Rice is favorable and highly consumed food in Sri 

Lanka. Rice is one of the highly consumed and 

staple food of Sri Lanka. Around 3.1 million tons 

of rough rice (paddy) are grown every year to meet 

about 95% of the country's demand. Since more 

than 1.8 million farmers and their families depend 

on rice production, rice holds a unique importance 

compared to other agricultural products in Sri 

Lanka (Anon, n.d.).  

The accuracy of identifying paddy is one of the 

most important factors when classifying rice 

varieties. The use of paddy varieties differs 

depending on the purposes. Different varieties of 

rice are used for the production of many value-

added products, including food varieties. 

Therefore, rice variety identification is very 

important for consumers (Golpour et al., 2014). In 

addition, the price and grade of rice are decided by 

its commercial value, genetic characteristics, and 

quality factors, which depend on the type of rice 

variety.  

Currently, the classification of paddy is performed 

manually, typically through visual inspection by 

experienced and well-trained individuals. 

However, this approach has significant 

drawbacks, including time consumption and 

unreliability due to inconsistencies and the 

involvement of unskilled technicians. Moreover, 

results may vary from person to person leading to 

subjective results. Therefore, there is a pressing 

need for a more efficient and accurate method of 

paddy variety classification. The review of 

literature shows that both Machine Learning (ML) 

and Computer Vision (CV) have been extensively 

employed across various domains, offering a fast, 

accurate, nondestructive, and cost-effective 

substitute for automated paddy classification 
processes. Deep learning techniques have 

superseded statistical methods in computer vision 

due to their enhanced accuracy in tasks such as 

object identification and image recognition 

(Kiratiratanapruk et al., 2020). 

While research on the classification of paddy 

varieties is limited, the identification of rice 

varieties has been extensively studied using 

external parameters such as shape, size, color, and 

texture (Cinar, 2019). For example, Singh and 



Proceedings of Papers, 4th International Conference on Science and Technology 
Faculty of Technology, South Eastern University of Sri Lanka 
ISBN 978-955-627-028-0 118

Chaudhury (2020) classified rice grains based on 

morphology, color, texture, and wavelet features, 

using image pre-processing techniques followed 

by a cascade network classifier. Similarly, Nagoda 

and Ranathunga (2018) employed support vector 

machines (SVM) and image processing methods 

to classify rice samples based on physical 

properties like color and texture, achieving a 

segmentation accuracy of 96% and a classification 

accuracy of 88%. Cinar (2019) also identified 

seven morphological features for classifying two 

different rice species. Several machine learning 

models, including Logistic Regression (LR), 

Multilayer Perceptron (MLP), SVM, Decision 

Trees (DT), Random Forest (RF), Naïve Bayes 

(NB), and K-Nearest Neighbor (KNN), were 

tested for classification accuracy, with success 

rates ranging from 88.58% to 93.02%. 

Additionally, Chatnuntawech et al. (2018) 

proposed a deep CNN algorithm for classifying 

rice varieties, using spatial–spectral data from two 

datasets, and achieved a mean classification 

accuracy of 91.09%. Their study also employed 

hyperspectral imaging to examine rice seeds in a 

consistent orientation.  

Despite the limited research on paddy 

classification, there is a clear need to focus 

specifically on classifying Sri Lankan paddy 

varieties. The review of existing literature 

highlights the scarcity of research on Sri Lankan 

paddy and the lack of application of emerging 

deep learning methodologies. Therefore, this 

study aims to evaluate the effectiveness of deep 

learning algorithms in classifying Sri Lankan 

paddy varieties.  

II.LITERATURE REVIEW

Artificial Neural Networks (ANN) have 

significant role for rice classification. For instance, 

Pazoki et al. (2014) used ANN Multi-Layer 

perceptron (MLP) and neuro-fuzzy networks to 
classify five rice varieties in Iran along with UTA 

feature selection algorithm to fine-tune the 

classifiers. The analysis used 24 color features, 11 
morphological properties, and four shape factors 

to classify rice grains. The screening is proved to 

have a rate above 99% for both approaches. 

There are numerous ML techniques that are 

available for the classification purposes. Arora et 

al., (2020) used different image processing 

algorithms and ML algorithms for rice grain 

classification using various parameters of 

individual rice grains like major axis, minor axis, 

eccentricity, length, breadth, etc. Relevant features 

of the rice grains have been extracted using 

various image processing algorithms. The rice 

grain images have been classified using different 

machine learning algorithms, such as LR, DT, NB, 

KNN, RF and Linear Discriminant Analysis 

(LDA) classifiers. They proposed future directions 

for incorporating additional features like 

chalkiness and moisture content analysis to ensure 

good quality rice is delivered. 

While various ML algorithms have achieved 

significant classification accuracy, ensemble 

learning approach is also gaining momentum for 

classification problems. Ensemble Learning can 

achieve better performance than a single model 

alone by combining various models. It can be 

applied to various ML tasks including 

classification. Setiawan & None (2024) used 

ensemble learning methods to classify rice grains 

based on image features. The study compared 

various machine learning algorithms, ultimately 

finding that Bagging meta-estimator improved 

classification accuracy by combining predictions 

from multiple base estimators. They utilized 

Bagging meta-estimator to aggregate decisions 

from multiple base classifiers, reducing model 

variance and improving classification consistency. 

By applying this approach to various grain 

features, ensemble method achieved consistent 

classification accuracy across different paddy 

varieties. 

Most studies on image-based paddy classification 

have primarily focused on color, morphology, and 

shape features. By using near-infrared 

hyperspectral imaging technology, both spatial 

and spectral information, as well as morphological 

features, can be captured. Jin et al. (2022) 

combined near-infrared hyperspectral imaging 

with traditional machine learning methods and 

deep learning models to classify rice seed 

varieties. This non-destructive imaging technique 

captures high-resolution spectra, enabling the 
detection of even subtle differences in paddy grain 

features, which leads to accurate classification 

across various rice varieties. Among conventional 

machine learning methods, SVM performed well, 

while in deep learning, LeNet, GoogLeNet, and 

ResNet models showed effective identification. 

Deep learning methods significantly outperformed 

conventional machine learning algorithms, with 



Proceedings of Papers, 4th International Conference on Science and Technology 
Faculty of Technology, South Eastern University of Sri Lanka 
ISBN 978-955-627-028-0 119

most models achieving classification accuracies 

exceeding 95%. 

In another study, Qiu et al. (2018) employed a 

near-infrared hyperspectral imaging system with 

two different spectral ranges (380–1030 nm and 

874–1734 nm) to classify four rice seed varieties. 

The study compared the performance of various 

discriminant models, including KNN, SVM, and 

CNN. Models utilizing the spectral range of 874–

1734 nm outperformed those built with the 380–

1030 nm range, with CNN outperforming both 

KNN and SVM. 

Rajalakshmi et al. (2024) achieved 97% accuracy 

in classifying 13 southern Indian paddy varieties 

— such as Yanaikomban, Swarna Masoori, Sivapu 

Kowuni, and Mapillai Samba—using a Deep 

Neural Network (RiceSeedNet) combined with 

traditional image processing techniques. They also 

demonstrated RiceSeedNet's potential to achieve 

99% accuracy in classifying eight paddy grain 

varieties from a public dataset. The study utilized 

two datasets: one containing 13,000 images of 

southern Indian paddy varieties (1,000 images per 

variety), and another with 8,000 images from an 

open-source benchmark dataset (1,000 images per 

variety). In the research of Paddy seed variety 

classification using transfer learning based on 

deep learning, Jaithavil, D. et al. (2022) used three 

pre-trained models VGG16, InceptionV3, and 

MobileNetV2 to classify three paddy varieties. 

Compared with various other two models 

Inception-v3 showed the highest accuracy and 

least test loss with 83.33% and 28.41% 

respectively 

Few other recent studies have been successful in 

classifying paddy varieties. For instance, Ansari, 

N. et al. (2021), presented a rapid inspection

method to classify three paddy varieties using

color, texture and morphological features and k-

nearest neighbors, support vector machine, and

partial least squares-discriminant analysis (PLS-

DA) algorithm. Where the classification accuracy
using PLS-DA, SVM-C, and KNN model was

83.8%, 93.9%, and 87.2% respectively. In another

study, Uddin, M. et al. (2021) proposed a

computer vision-based system for non-destructive

paddy seed variety identification, crucial for

maintaining seed purity in agriculture and

industry. To address challenges like illumination

variations during image capture, the study

introduced a modified histogram-oriented gradient

(T20-HOG) feature. Combined with Haralick and 

traditional features, these were refined using the 

Lasso technique and used to train a feed-forward 

neural network (FNN) for accurate variety 

prediction demonstrated 99.28% accuracy in 

identifying paddy grain types. 

Anami, B.S. et al. (2020) proposed a deep 

convolutional neural network (DCNN) framework 

for automatic recognition and classification of 

various biotic and abiotic stresses in paddy crops. 

The pre-trained VGG-16 CNN model was used to 

classify stressed images during the booting growth 

stage. The trained models achieved an average 

accuracy of 92.89% on the held-out dataset, 

demonstrating the technical feasibility of using the 

deep learning approach. The proposed work finds 

applications in developing decision support 

systems and mobile applications for automating 

field crop and resource management practices. 

The approach is applicable to 11 classes of biotic 

and abiotic stresses from five different paddy crop 

varieties. 

III.METHODOLOGY

The methodology applied for this study is 

illustrated in Figure 01 below.  

Figure 01: Applied Methodology 

G. Sample Preparation

Although many different varieties of paddy are 

available today, ten common paddy grain samples 

were chosen for this study using a convenient 

sampling method. 100grams of paddy grain 

samples from eight common Sri Lankan Paddy 

varieties (At 309, At 362, At 373, Bg 300, Bg 352, 

Bg 359, Bg 374, Bw 367) and two Sri Lankan 

traditional varieties (Kahawanu, Madathawalu) 

were selected for the data set preparation. They 



Proceedings of Papers, 4th International Conference on Science and Technology 
Faculty of Technology, South Eastern University of Sri Lanka 
ISBN 978-955-627-028-0 120

were obtained from Rice Research and 

Development Institute (RRDI), Bathalagoda, 

Ibbagamuwa (Figure 02).  

Figure 07: Samples images of ten paddy varieties 

H. Image Acquisition

The paddy grains were first cleaned to eliminate 

impurities, and random samples from each variety 

were selected for image acquisition. 1000 images 

of each paddy varieties were captured in the same 

lightning condition and same fixed frame by an 

iPhone 14 pro camera. Each image of a paddy seed 

was acquired with the seed placed centrally and 

horizontally, with the seed body rotated along the 

horizontal axis.   

I. Preprocessing

First, images of paddy were cropped and scaled to 

a uniform size of 500 x 250 pixels to standardize 

all the images and noise removal was done using 

bilateral and non-local filters where, bilateral 

filtering was effective for preserving edges and 

non-local filtering was effective for various noise 

types (Figure 03).  

Figure 08: Preprocessed Images a) Raw image 

captured from camera b) Image after Cropping c) 

Image after Noise Removal 

Bilateral filtering is an advanced image processing 

technique used to smooth images while preserving 

edges, making it ideal for applications where edge 

preservation is important. Bilateral filtering 

maintains the integrity of edges while reducing 
noise and smoothing the image. 

Bilateral filtering is defined by: 

𝐵𝐹[𝐼]𝑃 =
1

𝑊𝑃
∑ 𝐺𝜎𝑠

(||𝑝 − 𝑞||) 𝐺𝜎𝑟
(𝐼𝑝 − 𝐼𝑞)

𝑞𝜖𝑆

𝐼𝑞

where: 

𝐵𝐹[𝐼]𝑃  is the filtered image at pixel p

𝐼𝑝 and 𝐼𝑞 is are the intensities at pixels p and q

respectively. 

S is the spatial domain of the image. 

𝐺𝜎𝑠
(||𝑝 − 𝑞||) is the spatial Gaussian kernel.

𝐺𝜎𝑟
(||𝐼𝑝 − 𝐼𝑞||) is the range Gaussian kernel.

𝑊𝑃 is the normalization factor:

𝑊𝑃 = ∑ 𝐺𝜎𝑠
(||𝑝 − 𝑞||) 𝐺𝜎𝑟

(𝐼𝑝 − 𝐼𝑞)

𝑞𝜖𝑆

The bilateral filter operates by combining both 

spatial and range kernels, where the spatial kernel 

depends on the Euclidean distance between pixels 

p and q, with 𝜎𝑠 controlling the spatial extent of

the filter and the range kernel depends on the 

intensity difference between pixels p and q, with 

𝜎𝑟 controlling the range of intensity values that

influence the filtering. The filtered image at a pixel 

is computed as a weighted sum of neighboring 

pixels, with the weights determined by both the 
spatial distance and intensity difference. This 

method effectively smooths regions with similar 

intensities while maintaining the sharpness of 

edges. Bilateral filtering provides better edge 

preservation than other filtering methods like 

Gaussian filtering, which often blurs edges, or 

median filtering, which can lose finer details. By 

reducing noise while retaining sharp edges, 

bilateral filtering ensures that important features 

are maintained, leading to more accurate 

classification. 

J. Model Creation and Training

Since CNN has shown proven accuracy in various 

image-based classification problems due to their   

ability to capture spatial hierarchies of features 

through convolutional layers, this study employed 

CNN for the classification of paddy (Alzubaidi et 

al., 2021). Single paddy image was used to train 

the CNN model for the paddy classification. 

Dataset was split into three sets: training, 

validation and testing for accurate evaluation. 

Several own baseline CNN Models were created 

and trained on the training set, with performance 

tracking and hyperparameter optimization guided 

by the validation set. Iterative refinement was 

done using the validation set to consistently 

improve the model's performance. 

In the development of a CNN model, the initial 

attempts utilized basic architectures with dropout 

layers to prevent overfitting, followed by the 

addition of batch normalization for improved 

training stability. The third iteration introduced 

preprocessed images with noise removal. 

Subsequent models incorporated further 

refinements, including dropout, regularization 



Proceedings of Papers, 4th International Conference on Science and Technology 
Faculty of Technology, South Eastern University of Sri Lanka 
ISBN 978-955-627-028-0 121

techniques, L2 regularization, and a learning rate 

scheduler to enhance model robustness and 

performance. In the final models, the primary 

focus was on significantly expanding the dataset 

to improve the model’s effectiveness. 

The best-performing CNN model, as determined 

by our experiments, was composed of 

convolutional layers, each followed by activation 

functions, batch normalization, and pooling 

layers. Initially the input layer of the architecture 

processed images of size 500x250x3. The first 

convolutional layer applied 32 filters of size 3x3 

with ReLU activation, followed by a 2x2 max 

pooling layer. This structure was consistently 

applied across subsequent layers, with the number 

of filters progressively increasing to 64, 128, 256, 

and finally 512, enabling the model to extract 

increasingly complex features. Batch 

normalization was used after each convolutional 

layer to enhance training stability, and dropout 

layers were incorporated to mitigate overfitting. 

The model concluded with a fully connected layer 

comprising 512 neurons, regularized with L2, and 

a dropout rate of 0.5. The final softmax output 

layer classified the input into one of 10 categories. 

K. Evaluation

In the development of a CNN models, each model 

designed with different architectural complexities. 

The comparison of these models was conducted to 

identify the most effective paddy identification 

task. Performance evaluation was carried out 

using performance metrics, including accuracy, 

precision, F1 score, recall and AUC.  

Accuracy: The percentage of correctly classified 

instances in the dataset is measured generally as 

accuracy. The number of true positives (TP), true 

negatives (TN), false positives (FP), and false 

negatives (FN) is used to calculate it. 

The accuracy is defined by: 

Accuracy =  
TP + TN

TP + TN + FP + FN

Precision: Defined as the percentage of correctly 

identified positive instances among all predicted 

positive instances. It is particularly important 

when the cost of false positives is high, as it 

indicates the reliability of the positive predictions. 

The precision is defined by: 

Precision =  
TP

TP + FP

Recall: quantifies the percentage of real positives 

that the model accurately detected. When ignoring 

positive cases (false negatives) is more crucial 

than mislabeling negatives as positives, it is 

extremely significant.  

The recall is defined by: 

Precision =  
TP

TP + FN

F1 score: The average mean of recall and 

precision combined. 

The F1 score is defined by: 

F1 Score =  
2 × (Precision × Recall)

Precision + Recall

AUC: Area Under the Curve evaluates the model's 

ability to distinguish between positive and 

negative classes. It reflects the probability that a 

randomly chosen positive instance is ranked 

higher than a negative one. Higher AUC values 

indicate better performance, with 1.0 being perfect 

and 0.5 indicating no discrimination. 

IV.RESULTS & DISCUSSION

A. Model 1

The global minimum of the loss function was 

reached at the 30th epoch. Despite this, the results 

were suboptimal; the validation accuracy was 

74.2%. A significant oversight in this training 

phase was the use of raw data, rather than 

preprocessed data, which likely impacted the 

model's performance negatively.  This experience 

underscores the importance of data preparation in 

building effective machine learning models, as 

preprocessing can significantly influence the 

accuracy and efficiency of the training process. 

B. Model 2

The global minimum of the loss function was 

again reached at the 30th epoch. However, despite 

these adjustments, the results did not meet 

expectations. The validation accuracy was 

recorded at 62.5%. Similar to our initial attempt, 

the model was trained using raw data instead of 

preprocessed data, which adversely affected its 

performance. This experience further highlighted 

the critical importance of data preprocessing in 

training CNN models, as the lack of it can lead to 

significant discrepancies in performance metrics. 



Proceedings of Papers, 4th International Conference on Science and Technology 
Faculty of Technology, South Eastern University of Sri Lanka 
ISBN 978-955-627-028-0 122

C. Model 3

The global minimum of the loss function was 

achieved at the 80th epoch. Despite these 

improvements in data preprocessing, the 

performance metrics indicated that the training 

was not optimal. The validation accuracy was 

relatively low at 47.5%. This suggested issues in 

the model architecture or parameter settings that 

were not addressed merely by preprocessing the 

data. The significant loss indicates that further 

model evaluation and adjustments are necessary to 

improve its effectiveness. 

D. Model 4

with the global minimum of the loss function being 

reached remarkably early at the 10th epoch. 

Despite these extensive changes, the training 

outcomes were highly unsatisfactory. The model 

achieved a validation accuracy of only 17%. This 

performance indicates a significant misalignment 

in the model’s training process or architectural 

setup. These results underscore the need for a  

thorough review and recalibration of the model’s 

configuration and training strategy. 

E. Model 5

The model achieved a validation accuracy of 83%, 

indicating a substantial enhancement in its ability 

to generalize from the training data to unseen 

validation data. The improvements in image 

quality, along with careful preprocessing and 

effective model architecture, contributed to the 

much-improved performance metrics. This 

iteration demonstrates the critical importance of 

high-quality data and appropriate model tuning in 

developing effective deep learning systems. 

F. Model 6

The results led to a notable improvement in the 

model's performance, achieving a validation 
accuracy of 76%. This represents a significant 

enhancement, confirming the effectiveness of the 

learning rate scheduler in optimizing the training 

process and the L2 regularization in improving the 

model’s performance. This iteration underscores 

the utility of adaptive learning rate mechanisms 

and regularization techniques in boosting the 

accuracy and efficiency of machine learning 

models, especially in scenarios involving complex 

datasets and model architectures. 

G. Model 7

The increase in dataset size proved to be highly 

beneficial, as reflected by a validation accuracy of 

75.42%, the highest achieved across all above 

iterations.  

H. Model 8

Finaly the same model 7 was trained using our 

whole dataset and the model achieved a validation 

accuracy of 93.69%. The substantial improvement 

in performance with the expanded dataset 

highlights the critical role of data volume in 

training machine learning models. A larger dataset 

provides a more comprehensive representation of 

the variability and complexity inherent in real-

world data, thereby enhancing the model's ability 

to learn and generalize effectively. This milestone 

underscores the importance of both quality and 

quantity in dataset composition when aiming to 

improve model accuracy and robustness. 

The performance of our models was affected by 

the quality of the dataset, image conditions, and 

the architectural choices made during model 

development. Models 1 and 2, which used raw, 

unprocessed images captured under varying 

lighting conditions and angles, struggled with 

noise and irrelevant features due to wide 

backgrounds and inconsistent image conditions, 

leading to poor identification and suboptimal 

results. In Model 3, preprocessing steps such as 

resizing and cropping were introduced, but the 

model still underperformed, indicating that the 

presence of wide backgrounds continued to 

overshadow the seeds. To address these issues, 

Models 5 through 8 utilized a standardized image 

capture process, where all images were centrally 

aligned, uniformly cropped, and preprocessed 

using bilateral filtering for edge detection and non-

local means filtering for noise reduction. The 

architectural improvements in later models, 

including the addition of dropout, L2 

regularization, and learning rate scheduling in 

Models 5 and 6, further helped to prevent 

overfitting and enhance generalization.  

To ensure the trained model is correctly identifies 

the paddy seed, 30 paddy images were used for the 

prediction which are not used for train, test, or 

validate the model. The best performed model 

identified all images and other models identified 

few.  



Proceedings of Papers, 4th International Conference on Science and Technology 
Faculty of Technology, South Eastern University of Sri Lanka 
ISBN 978-955-627-028-0 123

The result is summarized and presented below in 

Table 01 and Table 02. 

Table 01: Summary of different CNN models which are 

trained using proper dataset 
M

o
d

e
l 

E
p

o
c
h

 

L
e
a
r
n

in
g
 r

a
te

 

V
a
li

d
a
ti

o
n

 

A
c
c
u

r
a
c
y

 

T
r
a
in

 

A
c
c
u

r
a
c
y

 

T
e
st

 A
c
c
u

r
a
c
y

 

F
1
 S

c
o
r
e
 

P
r
e
c
is

io
n

 

r
e
c
a
ll

 

A
U

C
 

5 6

0 

0.001 83% 99.

5 

78.

25 

0.1

1 

0.1

4 

0.1

1 

0.

51 

6 6

0 

Dyna

mical

ly 

chan

ged 

76% 10

0% 

55

% 

0.1 0.1

8 

0.1

1 

0.

47 

7 6

0 

Dyna

mical

ly 

chan

ged 

75.4

2% 

10

0% 

82.

25

% 

0.1

1 

0.1

1 

0.1

2 

0.

51 

8 6

0 

Dyna

mical

ly 

chan

ged 

93.6

9% 

10

0% 

89.

75

% 

0.9

4 

0.9

5 

0.9

4 

0.

99 

Table 02: Summary of different CNN models which are 

trained using improperly captured images 

M
o
d

el
 

E
p

o
ch

 

L
ea

rn
in

g
 

ra
te

 

V
a
li

d
a
ti

o
n

 

A
cc

u
ra

cy
 

1 100 0.0001 74.2% 

2 40 0.001 62.5% 

3 100 0.0001 47.% 

4 60 0.001 17% 
Figure 03 illustrates the prediction of paddy seeds 

by Model 8. 

Figure 03: Prediction of paddy seeds by model 8 

X. CONCLUSION

Deep learning technologies are now commonly 

used in various sectors of agricultural production 

and industrial food production. In this paper, we 

aim to develop CNN models to classify10 paddy 

varieties from a dataset of nearly ten thousand 

images of paddy seeds. We investigated nearly 

1000 data samples in each paddy variety for 

training and testing models. Several CNN models 

were evaluated and compared in order to obtain a 

model that had the best performance. the highest 

classification accuracy obtained was 93.69%. The 

preliminary work presented in this paper could be 

further enhanced by focusing on clustering to 

identify and classify different paddy varieties in a 

single image. 

REFERENCES 

Alzubaidi, L. et al. (2021b) ‘Review of Deep Learning: 

Concepts, CNN Architectures, challenges, 

applications, Future Directions’, Journal of Big Data, 

8(1). dio:10.1186/s40537-21-00444-8. 

Anami, B.S., Malvade, N.N. and Palaiah, S. (2020) 

‘Deep Learning Approach for recognition and 

classification of yield affecting paddy crop stresses 

using field images’, Artificial Intelligence in 

Agriculture, 4, pp. 12–20. 

doi:10.1016/j.aiia.2020.03.001.  

Ansari, N. et al. (2021) ‘Inspection of paddy seed 

varietal purity using machine vision and multivariate 

analysis’, Journal of Agriculture and Food Research, 3, 

p. 100109. doi:10.1016/j.jafr.2021.100109.

Arora, B. et al. (2020) ‘Rice grain classification using 

Image Processing & Machine Learning Techniques’, 

2020 International Conference on Inventive 

Computation Technologies (ICICT) [Preprint]. 

doi:10.1109/icict48043.2020.9112418. 

Cinar, I. and Koklu, M. (2019) ‘Classification of rice 

varieties using artificial intelligence methods’, 

International Journal of Intelligent Systems and 

Applications in Engineering, 7(3), pp. 188–194. 

doi:10.18201/ijisae.2019355381. 

Common rice diseases (no date) Department of 

Agriculture Sri lanka. Available at: 

https://doa.gov.lk/rrdi/index.php (Accessed: 6 Dec. 

2023).  

Golpour, I., Parian, J.A. and Chayjan, R.A. (2014) 

‘Identification and classification of bulk paddy, Brown, 

and white rice cultivars with colour features extraction 

using image analysis and neural network’, Czech 

Journal of Food Sciences, 32(3), pp. 280–287. 

doi:10.17221/238/2013-cjfs.  



Proceedings of Papers, 4th International Conference on Science and Technology 
Faculty of Technology, South Eastern University of Sri Lanka 
ISBN 978-955-627-028-0 124

Jaithavil, D., Triamlumlerd, S. and Pracha, M. (2022) 

‘Paddy seed variety classification using transfer 

learning based on Deep Learning’, 2022 International 

Electrical Engineering Congress (iEECON) [Preprint]. 

doi:10.1109/ieecon53204.2022.9741677.  

Jin, B. et al. (2022) ‘Identification of rice seed varieties 

based on near-infrared hyperspectral imaging 

technology combined with Deep Learning’, ACS 

Omega, 7(6), pp. 4735–4749. 

doi:10.1021/acsomega.1c04102.  

Kiratiratanapruk, K. et al. (2020) ‘Development of 

paddy rice seed classification process using machine 

learning techniques for automatic grading machine’, 

Journal of Sensors, 2020, pp. 1–14. 

doi:10.1155/2020/7041310.  

Nagoda, N. and Ranathunga, L. (2018) ‘Rice sample 

segmentation and classification using image processing 

and support vector machine’, 2018 IEEE 13th 

International Conference on Industrial and Information 

Systems (ICIIS) [Preprint]. 

doi:10.1109/iciinfs.2018.8721312.  

Qiu, Z. et al. (2018) ‘Variety identification of single rice 

seed using hyperspectral imaging combined with 

Convolutional Neural Network’, Applied Sciences, 

8(2), p. 212. doi:10.3390/app8020212.  

Rajalakshmi, R. et al. (2024) ‘RiceSeedNet: Rice Seed 

Variety Identification Using Deep Neural Network’, 

Journal of Agriculture and Food Research, 16, p. 

101062. doi:10.1016/j.jafr.2024.101062. 

Robert Singh, K. and Chaudhury, S. (2020) ‘A Cascade 

Network for the classification of rice grain based on 

single Rice Kernel’, Complex & Intelligent 

Systems, 6(2), pp. 321–334. doi:10.1007/s40747-020-

00132-9.  

Setiawan, R. and Hayatou Oumarou (2024) 

‘Classification of rice grain varieties using ensemble 

learning and Image Analysis Techniques’, Indonesian 

Journal of Data and Science, 5(1), pp. 54–63. 

doi:10.56705/ijodas.v5i1.129.  

Uddin, M. et al. (2021) ‘Paddy seed variety 

identification using T20-hog and Haralick textural 

features’, Complex & Intelligent Systems, 8(1), 

pp. 657–671. doi:10.1007/s40747-021-00545-0. 


	Sivasankari, S., V. Venkatesalu, M. Anantharaj and M. Chandrasekaran ( 2006). Effects of seaweeds extracts on the growth and biochemical constituents of Vigna sinensis, Bioresource Technology, 97: 1745-1751
	Classification of Sri Lankan Paddy Varieties using Deep Learning Techniques
	M.R.M. Aththas1, M.N.F. Yusra2, M.S. Sabrina3, W.A. Sanjeewa4, M. Janotheepan5 and A.R. Fathima Shafana6

	G. Sample Preparation
	H. Image Acquisition
	I. Preprocessing
	J. Model Creation and Training
	K. Evaluation
	A. Model 1
	B. Model 2
	C. Model 3
	D. Model 4
	E. Model 5
	F. Model 6
	G. Model 7
	H. Model 8