SugarCane Disease Classifier Using CNN

RESNET50 Architecture
Luisito Lolong Lacatan Isaac Angelo M. Dioses Danielle Bagaforo Meer
Computer Engineering Department School of Information Technology, School of Information Technology,
College of Engineering Polytechnic MAPUA University, Makati Campus MAPUA University, Makati Campus
University, Philippines Makati, Metro Manila, Philippines Makati, Metro Manila, Philippines
[email protected] [email protected] [email protected]

John Christian Diamante Jesusimo L. Dioses Jr. Shiva Cruz Jr.

School of Information Technology, College of Computing Studies, School of Information Technology,
MAPUA University, Makati Campus Information and Communication MAPUA University, Makati Campus
Makati, Metro Manila, Philippines Technology, Isabela State University, Makati, Metro Manila, Philippines
[email protected] San Fabian Echague Isabela, [email protected]
Philippines
[email protected]

Abstract— This study explores the use of deep learning disease. Typically, the diagnosis needs to be validated and
methods, more especially the RESNET50 architecture, for the confirmed by ongoing observation. Large farms, however,
image-based classification of sugarcane illnesses. Using cannot use this strategy. In underdeveloped nations such as
RESNET50, a deep neural network, photos from five different the Philippines, consulting an expert is an expensive and
classes (Healthy, Mosaic, RedRot, Rust, and Yellow) were time-consuming process [1, 2]. Machine learning is widely
analyzed. 80% of the images in the dataset were utilized for used in the field of precision agriculture. AI technology is
training, while the remaining 20% were used for testing. To used to classify, and detect plant diseases, seeds, and avian
minimize biases and guarantee a fair distribution of data,
eggs in some studies [3,4,5,6,11]. With online applications,
preprocessing was done in MATLAB. A 7x7 convolutional
this strategy necessitates more complex calculations.
filter, max-pooling, and several residual blocks were among
the optimized parameters used in the training of the
Consequently, the implementation of these techniques
RESNET50 model, which is well-known for its efficient might only yield an acceptable outcome. Unlike traditional
utilization of residual blocks to capture complicated neural network topologies, deep learning uses an artificial
characteristics. After 30 epochs of training with an RTX 2060 neural network architecture and typically has multiple
GPU and Intel Core i5 processor, 93.51% accuracy was layers to process [7]. Significant advancements in deep
attained. With accuracy rates above 90%, the results showed learning have been made in the classification of images
that RESNET50 performed remarkably well in classifying through feature extraction and effective capacity for
Yellow, Rust, RedRot, and Mosaic illnesses. But the model learning [12,13] One deep learning technique included in
had trouble correctly predicting the Healthy class, suggesting the CNN approach makes use of artificial neural networks.
that it still needed to be improved. Convolutional neural networks have been utilized in a
variety of ways to combat plant disease issues. In this study,
Keywords—ResNet50, CNN, MATLAB, Deep Learning, Researchers will apply RESIDUAL NETWORK or
Sugar Cane RESNET50 architecture in classification of Sugar Cane
diseases.
I. INTRODUCTION
Sugarcane is regarded as one of the most valuable crops II. RELATED LITERATURE
in the Philippines. Over many years, molasses and Several studies have used CNN in the classification of
sugarcane have been used as raw ingredients to produce different plant diseases. For instance, in the study conducted
food, especially drinks. The Sugar Production Bulletin for by the authors of [8]. The researchers stated that utilizing
CY 2021–2022 states that as of August 2022, there were picture samples of 649 and 550 infected and healthy soya
397,524 hectares of sugarcane plantations in the leaves, CNNs for AlexNet and GoogleNet were trained to
Philippines. About 20,841,174 metric tons of plantation detect three soya diseases. The strategy used here makes use
material were milled to produce raw and refined sugar from of the 5-fold cross-validation technique, and the models
these areas. If diseases are not identified in time, they might named AlexNet and GoogleNet have achieved 98.75% and
have a negative economic impact on sugarcane yield and 96.25% exactness, respectively. An improved deep-
production, which is a serious worry and threat to farmers. learning method for the accurate localization and
If these crops are not grown as much, it won't be good for identification of plant leaf diseases is presented by
the economy. When it comes to the management of input Alqahtani et al. [9].
resources like seed, water, soil, and fertilizers, the
sustainability of output is crucial. If these crops sustain the The study addresses a crucial issue in agriculture by
disease while growing, agricultural production and focusing on using deep learning techniques to improve the
preservation will become less flexible in the competitive precision and effectiveness of disease detection in plant
market. The identification and diagnosis of sugarcane leaves. An end-to-end deep learning model designed
illnesses have been essential in determining their avoidance. specifically for the classification of corn leaf diseases was
Experts utilize a conventional method that involves using introduced by H. A. Min et al. [10]. Their research focuses
their naked eyes to identify and detect sugarcane leaf on the accurate classification and identification of illnesses

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE

affecting maize crops with deep learning, a type of artificial III. METHODOLOGY
intelligence. Since early disease diagnosis is made possible
by this research, agricultural practices could be greatly A. Image Acquisiton
impacted. Support vector machines (SVMs) were used by For this study, the images used to train and test the deep
the authors in related work [11] to identify photos of learning models were sourced from a publicly available
eggplant leaves. Two-spotted spider infestation and dataset on Kaggle. Kaggle is a well-known platform that
Cercospora leaf spot were the two types of leaf diseases they hosts a variety of datasets contributed by the global data
were attempting to classify. The data is split into two sets science community, providing high-quality data suitable
with an 80:20 split ratio for categorization purposes. To for machine learning research. The dataset specifically
enable ten-fold cross-validation, the dataset is arbitrarily focused on sugarcane leaves, comprising images across
divided into ten subgroups. It is said that the classification five distinct classes that represent different statuses of the
success rate is 78.32%. sugarcane crop: Healthy with 474 images, Mosaic 462
Tamilvizhi et. al. presented a highly accurate model for images, RedRot with 518 images, Rust with 514 images,
the detection and classification of sugarcane leaf disease and Yellow with 505 images. After the image acquisition,
based on deep transfer learning and quantum-behaved the images were imported in MATLAB for image
particle swarm optimization [12]. Militante & Gerardo [12]. preparation. In Figure 1. Shows the framework of this
Applies and compares three deep learning models LeNet, study. The dataset was divided into two parts which are the
StridedNet, and VGGNet to detect and identify sugarcane training and testing dataset. For the training dataset 80% of
leaf diseases. Among the models, VGGNet achieved the the images were used while the testing dataset 20% of
highest accuracy in disease identification, while StridedNet images were used.
performed the lowest. For future work, the study suggests
increasing the dataset size, adjusting learning rates and
optimizers, and experimenting with new architectures to
improve model performance. The study aims to serve as a
decision-making tool to help farmers recognize and
understand sugarcane diseases. The same author utilized
deep learning to detect and classify sugarcane leaves as
either diseased or healthy. A simple convolutional neural
network (CNN) with seven different classes was used,
achieving an accuracy of 95%. The model effectively
categorized sugarcane leaves based on leaf patterns,
demonstrating the potential of computer vision and machine
learning to assist farmers in disease detection and
classification. For future research, the study suggests
exploring different models, learning rates, and optimizers to
further enhance model performance [13]. Another study
researchers, uses AMRCNN for accurately and effectively
classifying sugarcane leaf diseases in real time.
Furthermore, the model's running time was greatly
decreased by using TensorFlow Lite tools for quantization,
which qualifies it for use in mobile-based systems. Users
Figure 1. Framework of Study.
can engage with the model using the built application, which
provides a workable way to categorize diseases in sugarcane B. PreProcessing
crops with 86.53% accuracy [14]. In preprocessing, images were input in Matlab2023b. as
shown in figure 2. It displays the dataset distribution for
every class. It is demonstrated that there are no disparities
CONVOLUTIONAL NEURAL NETWORKS in the data among the classes, indicating that the data was
Deep learning models, particularly neural networks, are distributed correctly and could not have produced a biased
composed of layers of nodes (also called neurons) that are model. Also, in Figure 3. Shows the Testing dataset
interconnected. Each node processes a piece of input data distribution for every class and it shows the equal
and passes the information to the next layer, allowing the distribution.
network to learn complex patterns and representations. The
strength of deep learning lies in its ability to manage and
process vast amounts of data, making it ideal for tasks like
image and speech recognition where traditional algorithms
might struggle [15,16,17].

Figure 2. Training Dataset.

 Figure 3. Testing Dataset
Figure 5. Residual Block Diagram

Figure 6. RESNET 50 Architecture

The first stage applies a 7x7 convolution filter with 64

output channels and a stride of 2. The large filter size (7x7)
is used to capture broad spatial features from the input
image. The stride of 2 reduces the spatial dimensions of the
input, effectively downsampling the image.
Figure 4. Training Options Following the initial convolution, a max-pooling operation
In Figure 4. Shows the training options for deep learning. is performed. Max pooling with a filter size, typically 3x3,
with SGDM as solver, 0.01 Initial Learn Rate, 64 batch size further reduces the spatial dimensions of the feature maps
and 30 max epochs with validation frequency of 50. Also, while retaining the most significant features. This helps to
in this study researcher applies early stopping techniques to make the network more invariant to small translations of the
maintain the stability of training model and to avoid input image.
overfitting. The third stage represents the first residual block in the
ResNet-50 architecture. The block contains three layers
C. RESNET50 Architecture which are the 1x1 convolution, reduces the number of
This study uses RESNET 50 for SugarCane Disease channels to 64. 3x3 convolution performs the main spatial
classification. With the invention of the RESNET50 convolution. And 1x1 convolution increases the number of
architecture, He et al. [18] made it possible to train networks channels to 256. To address the vanishing gradient issue and
much more deeply than before. Training a neural network enable deeper networks, the block is repeated three times
becomes increasingly challenging the deeper it becomes. using a shortcut connection that neglects these layers and
Consequently, the residual block, or remaining learning adds the input directly to the output.
function layer, was recast by He et al. This block is the basis The fourth stage is another residual block, like the previous
of the design and is present in every layer of the CNN one but with more channels (128 and 512). This block is
ResNet architecture. In Figure 5. It shows the residual block repeated four times and increases the depth of the network,
diagram. ResNet-50 is made up of 50 layers. ResNet-50's allowing it to learn more complex features.
architecture may be broken down into multiple stages, each
of which has a downsampling operation and multiple The fifth stage is a deeper residual block with even more
convolutional layers. channels (256 and 1024), repeated six times. As the network
goes deeper, it learns more abstract features from the input
images.
The final residual block has the highest number of channels
(512 and 2048) and is repeated three times. This block
captures the most complex features and is crucial for the
final classification.
After the residual blocks, an average pooling operation is 97.4% and it is noticeable that RESNET had difficulties
applied. This global average pooling reduces each feature predicting the Healthy class
map to a single value by taking the average of all elements
in the feature map. This helps to condense the information
while reducing the risk of overfitting. Finally, the condensed
feature vector is passed through a fully connected layer and
then a softmax function, which converts the output into a
probability distribution over the five possible classes. And
the final output of the softmax layer is the predicted class,
indicating whether the sugarcane leaf is healthy or which
disease it might have.

IV. RESULTS AND DISCUSSION Figure 9. Confusion Matrix of the Model

After preprocessing and training setup. The training was
Accuracy Metrics
started. Using RTX 2060, Intel Core I5 Processor, and 24
gigabytes of ram, Deep learning training time and Precision, Recall, and F1 Score are essential metrics for
performance was enhanced. In figure 7 and 8 shows the evaluating the performance of a classification model
training progress of RESNET50 in sugarcane diseases. Precision measures the accuracy of the positive predictions
made by the model. Specifically, it is the proportion of true
positive predictions (TP) among all the instances that the
model has classified as positive. Recall assesses the model’s
ability to identify all relevant positive instances. It is the
proportion of true positive predictions among all actual
positive instances. F1 Score is the harmonic mean of
Precision and Recall. It provides a single metric that
balances both Precision and Recall, particularly useful when
you need to consider both false positives and false negatives
in the model evaluation. Their formula was stated in eq.
1,2,3 below

PRECISION = TP/TP+FP (1)

RECALL = TP/TP+FN (2)

F1 SCORE = (2) PRECISION * RECALL / PRECISION +RECALL (3)

Figure 7. Accuracy progress of RESNET50
Table 1. Accuracy Matrix and Class Wise Classification

Class Precision Recall F1 Score Class

wise %
Healthy 99% 80% 88% 79.7%
Mosaic 88% 97% 92% 97.3%
RedRot 93% 95% 94% 94.5%
Rust 94% 96% 95% 96%
Yellow 95% 99% 97% 99.2%
Figure 8. Loss Diagram of RESNET50

In the figure above shows the diagram and training

progress of RESNET50 throughout the time. It shows that For the Healthy class, the model obtained a Precision of
the RESNET50 performs well in training process and the 99%, meaning that the great majority of cases that are
validation of accuracy and loss was accurate. With iteration classified as Healthy are indeed healthy. This high Precision
of 3462 and 30 epochs, the RESNET 50 obtained 93.51% indicates how well the model works to reduce false positives
accuracy in classification of different status and diseases of in this particular category. The model is accurate at
sugarcane. predicting Healthy, however it only captures 80% of real
In Figure 9, shows the confusion matrix of the model. Healthy cases, according to the Recall for Healthy. With an
Which is the predicted outcome compared to the true F1 Score of 88%, this class exhibits good overall
outcome based on labeled data. It shows that the model performance, while there is still space for improvement in
performs well on the Yellow virus with 99.2% accuracy, terms of capturing all real Healthy cases. This is due to the
Rust with 96.1% accuracy, RedRot with 94.6, Mosaic with class's high Precision and relatively poor Recall.
With a Precision of 88%, the Mosaic class shows that 88%
of cases that are predicted to be Mosaic are really
categorized as such. This indicates that although there is
some need for improvement in terms of reducing false ACKNOWLEDGMENT
positives, the model performs a respectable job of Isaac Angelo Dioses would like to give thanks to his
recognizing Mosaic instances. With a Recall of 97%, mentors, CBI MAPUA and KADBIN. And to God
Mosaic stands out as having a good ability to recognize Almighty
almost all genuine Mosaic cases. An F1 Score of 92%,
which emphasizes a great balance between Precision and REFERENCES
Recall with a small emphasis on developing Recall,
complements this high Recall. [1] S. V. Militante, B. D. Gerardo, and R. P. Medina, “Sugarcane
Disease Recognition using Deep Learning,” 2019 IEEE Eurasia
For the RedRot class, the model achieved a Precision of Conference on IOT, Communication and Engineering (ECICE), Oct.
93% and a Recall of 95%. These metrics indicate that the 2019, doi: 10.1109/ecice47484.2019.8942690.
model is highly effective in both identifying RedRot [2] H. Park, J.-S. Eun, and S.-H. Kim, “Image-based disease diagnosing
instances and minimizing false positives. The high Recall and predicting of the crops through the deep learning mechanism,”
value reflects the model's strong performance in capturing IEEE, Oct. 2017, doi: 10.1109/ictc.2017.8190957.
nearly all instances of RedRot, while the high Precision [3] J. L. Dioses, R. P. Medina, A. C. Fajardo, A. A. Hernandez, and I.
confirms the accuracy of these predictions. The F1 Score for A. M. Dioses, “Performance of egg sexing classification models in
Philippine native duck,” IEEE, Aug. 2021, doi:
RedRot is 94%, demonstrating a robust balance between 10.1109/icsgrc53186.2021.9515279.
Precision and Recall and indicating overall excellent
[4] E. R. Arboleda, A. C. Fajardo, and R. P. Medina, “An image
performance for this class. processing technique for coffee black beans identification,” IEEE,
May 2018, doi: 10.1109/icird.2018.8376325.
The Rust class shows a Precision of 94% and a Recall of
[5] D. F. R. Anandhi and S. Sathiamoorthy, “Deep Learning based
96%, suggesting that the model is very effective in Automated Rice Plant Disease Recognition and Classification
predicting Rust instances with high accuracy and capturing Model,” IEEE, Oct. 2023, doi:
almost all actual cases. The F1 Score of 95% reflects an 10.1109/icaeeci58247.2023.10370809.
outstanding balance between Precision and Recall, [6] S. Sarkar, J. A. Ray, C. Mukherjee, S. Ghosh, Jayanthi. N, and C. L.
illustrating that the model performs exceptionally well in K. R, “Plant Leaf Disease Classification Based on SVM Based
identifying Rust while minimizing errors. Densenets,” IEEE, Nov. 2023, doi:
10.1109/icaiccit60255.2023.10465886.
And the Yellow class has a Precision of 95% and an [7] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol.
impressive Recall of 99%. This indicates that the model not 521, no. 7553, pp. 436–444, May 2015, doi: 10.1038/nature14539.
only predicts yellow instances with high accuracy but also [8] I. Y. Purbasari, B. Rahmat, and C. S. P. Pn, “Detection of Rice Plant
identifies almost all true yellow cases. The F1 Score of 97% Diseases using Convolutional Neural Network,” IOP Conference
Series Materials Science and Engineering, vol. 1125, no. 1, p.
underscores a superb balance between Precision and Recall, 012021, May 2021, doi: 10.1088/1757-899x/1125/1/012021.
demonstrating that the model excels in both correctly
[9] Y. Alqahtani, M. Nawaz, T. Nazir, A. Javed, F. Jeribi, and A. Tahir,
classifying and identifying yellow instances with minimal “An improved deep learning approach for localization and
false positives and negatives. recognition of plant leaf diseases,” Expert Systems With
Applications, vol. 230, p. 120717, Nov. 2023, doi:
The classification metrics reveal that the model performs 10.1016/j.eswa.2023.120717.
exceptionally well across most classes, with the highest F1 [10] H. Amin, A. Darwish, A. E. Hassanien, and M. Soliman, “End-to-
Score observed in the yellow class. The high Precision and End Deep learning model for corn leaf disease classification,” IEEE
Recall for Yellow suggest that it is the most effectively Access, vol. 10, pp. 31103–31115, Jan. 2022, doi:
classified class. Conversely, while the model shows strong 10.1109/access.2022.3159678.
performance in classes like Rust and RedRot, there is [11] A. S. Alon, “Machine Vision Recognition System for Iceberg
Lettuce Health Condition on Raspberry Pi 4b: A Mobile Net SSD v2
variability in Recall and Precision across different classes. Inference Approach,” International Journal of Emerging Trends in
Engineering Research, vol. 8, no. 4, pp. 1073–1078, Apr. 2020, doi:
V. CONCLUSION 10.30534/ijeter/2020/20842020.
This research demonstrates how well deep learning models [12] A. K. Rangarajan, R. Purushothaman, M. Prabhakar, and C.
more especially, RESNET50 classify diseases of sugarcane Szczepański, “Crop identification and disease classification using
traditional machine learning and deep learning approaches,” Journal
based on picture data. Deep learning has shown to be a of Engineering Research, Nov. 2021, doi: 10.36909/jer.11941.
reliable solution for challenging tasks like picture [13] S. V. Militante and B. D. Gerardo, “Detecting Sugarcane Diseases
classification because of its capacity to process and learn through Adaptive Deep Learning Models of Convolutional Neural
from enormous volumes of data. The residual blocks that Network,” 2019 IEEE 6th International Conference on Engineering
define the RESNET50 architecture made it possible to Technologies and Applied Sciences (ICETAS), Dec. 2019, doi:
10.1109/icetas48360.2019.9117332.
successfully train a deep network that can identify minute
[14] S. D. Daphal and S. M. Koli, “Enhanced deep learning technique for
details and patterns in the data. Reliable model testing and sugarcane leaf disease classification and mobile application
training were achieved by the methods used, which integration,” Heliyon, vol. 10, no. 8, p. e29438, Apr. 2024, doi:
included careful MATLAB preprocessing and the use of a 10.1016/j.heliyon.2024.e29438.
well-balanced dataset from Kaggle. The model's overall [15] A. G. Howard et al., “MobileNets: efficient convolutional neural
networks for mobile vision applications,” arXiv (Cornell
accuracy of 93.51% is impressive, indicating its ability to University), Jan. 2017, doi: 10.48550/arxiv.1704.04861.
accurately categorize sugarcane in various situations. [16] K. P. Ferentinos, “Deep learning models for plant disease detection
Remarkably, with accuracy rate of 93.51%, RESNET50 and diagnosis,” Computers and Electronics in Agriculture, vol. 145,
performed exceptionally well in identifying the Yellow pp. 311–318, Feb. 2018, doi: 10.1016/j.compag.2018.01.009.
virus, Rust, RedRot, and mosaic illnesses.
[17] A. Kamilaris and F. X. Prenafeta-Boldú, “Deep learning in
agriculture: A survey,” Computers and Electronics in Agriculture,
vol. 147, pp. 70–90, Apr. 2018, doi: 10.1016/j.compag.2018.02.016.