Using Machine Learning to Identify Crop Diseases with ResNet-18

Rihan Rahman

Skyline High School, 228

Ave SE, Sammamish, WA, U.S.A.

Keywords: ML, Machine Learning, Agriculture, Machine Learning Applications, Machine Learning in Agriculture,

ResNet-18, Deep Learning, Computer Vision.

Abstract: Plant diseases are a highly prevalent issue in agriculture, causing countless farmers annually to face career

threatening damages such as diminished profits and crop yields and environmental damages. Consequently,

it is imperative that these diseases are quickly detected and treated against. An increasingly effective solution

is to train convolutional neural networks (CNNs) using deep learning (DL). DL has several effective

applications in a variety of major fields such as healthcare and fraud detection and has a high potential to

solve issues of global significance. This research’s goal is to create a machine learning (ML) model with DL

to identify plants’ diseases using photos of infected leaves. Many farmers in rural areas struggle to treat

blights due to limited access to technology and information regarding them. Therefore, an ML model which

can automatically identify these diseases would be highly useful for these people. After sourcing a

comprehensive dataset with images of 88 types of plants and diseases, I used it to train a CNN model using

several data augmentation techniques. With the model architecture ResNet-18, while evaluating its

performance with a validation dataset, the model achieved a loss of 4.541%. This value demonstrates ResNet-

18’s applicability to the task of identifying plant diseases and illustrates the potential for classification-based

DL networks to support rural farmers and the field of agriculture. If a superior model is created to identify

blights more accurately, it should be used to help the billions of farmers who would greatly benefit from such

technology.

1 INTRODUCTION

1.1 Background and Significance

Machine Learning has a nearly infinite number of

possible applications and can have large effects on

areas such as in data mining (Maglogiannis et al.,

2007). With its Deep Learning capability, ML has the

potential to solve many complex problems of global

concern. DL is created through many hidden layers,

with iterations of transformations and abstractions

(Vargas, 2017). Consequently, DL models can adapt

to new knowledge extremely quickly, making them

the ideal solution for a variety of scenarios.

For instance, Mohsen et al. used DL to identify

brain tumours using scanned images of the brain,

demonstrating the potential to save lives by detecting

cancer in early stages with ML (Mohsen et al., 2018).

In another instance, Lee and Yoo used DL to predict

domestic market prices based on information about

foreign markets (Lee & Yoo, 2019).

1.2 Research Question/Objective

Many famers lack access to technology and

information to identify and effectively treat plant

diseases. As a result, I believe that creating a

convolutional neural network (CNN) to identify these

diseases would greatly assist these farmers. My goal

is to explore the question of if ML can be used to

identify blights from images of infected plant leaves.

When several trees in my local community

became infected, identifying and treating the diseases

became a long and tedious process. As a result, I was

inspired to use a CNN to automate this. Creating an

ML model would not only be more efficient, but also

make disease identification and treatment far more

accessible to farmers with limited experience and

resources, enabling new growers to grow a variety of

crops without fearing potential diseases.

Rahman, R.

Using Machine Learning to Identify Crop Diseases with ResNet-18.

DOI: 10.5220/0013123500003911

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

In Proceedings of the 18th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2025) - Volume 1, pages 311-315

ISBN: 978-989-758-731-3; ISSN: 2184-4305

311

2 LITERATURE REVIEW

2.1 Overview of Deep Learning

DL is a type of ML which utilizes artificial neural

networks to replicate the human brain, producing a

more capable model which can be used to solve

problems of greater complexity. DL is used to address

many modern issues in a variety of fields for this

advantage, such as medicine, economics, and agri-

culture. It functions utilizing “neurons”, which oversee

processing data and are critical to the functionality of a

DL model (Nielsen, 2015). Using ML models to

classify objects within images is called Object

Classification, an extremely popular branch of DL.

Furthermore, DL networks use several “layers” to

separate the processes required for object

classification. For instance, CNNs uses convolutional

layers to process image input and scale it down using

techniques such as Max Pooling, aiding networks to

process information faster by simplifying images.

This helps models understand what exact features to

search for in an image during classification. As a

result, object detection and classification are some of

DL’s major strengths (Zilong, 2018).

2.2 Related Studies and Projects

Deep learning for classification is a rapidly growing

field and is becoming highly prevalent in our society.

One example of image-based ML classification being

utilized is in cancer research. Lung cancer

classification research uses DL networks to analyse

lung scans and determine if they are cancerous

(Asuntha & Srinivasan, 2020).

In Asuntha and Srinivasan’s study, the researchers

used factors such as gradient, texture, and shape, to

differentiate between images of cancerous and

healthy lungs. The capability to precisely analyse data

while focusing on factors like these demonstrates the

flexibility of DL classification to be effectively used

for many different purposes.

In Zhu et al.’s research, the researchers described

many types of DL that can benefit the field of

agriculture and educated readers about its functional

benefits (Zhu et al., 2018).

3 METHODOLOGY

3.1 Dataset Description

This study uses the Kaggle dataset Plant Disease

Classification Merged Dataset by Aline Dobrovsky,

which contains pictures of whole infected and healthy

leaves (Dobrovsky, 2023). This is a collection of 14

sub datasets and has 88 classes with over 17,000

images. Each subset contains images with varying

backgrounds. None of the images were generated

using pre-processing techniques such as flipping or

zooming in (Dobrovsky, 2023). However, in this

research, a reduced version is used to train the model

due to technical limitations.

Figure 1: Sample labelled images from dataset.

(Dobrovsky, 2023).

3.2 Model Selection and Architecture

For the task of identifying diseases from images of

plant leaves, I chose a pre-trained ResNet-18 model

due to its image classification capability. An example

of this model’s usage on a similar scenario is an ML

study by Al-Falluji et al. (Al-Falluji et al., 2020). In

this study, the researchers were able to use ResNet-18

to identify if a patient had COVID-19 by feeding X-

ray images through the model. Finally, with ResNet-

18, the researchers were able to classify patient’s

diseases with an accuracy of 96.73%.

Figure 2: Diagram of ResNet-18 architecture (Ramzan et

al., 2020).

ResNet-18 uses 18 layers to process data,

supporting the complexity required for DL, with the

activation function ReLU (Rectified Linear Unit).

ReLU determines if a neuron should be activated and

is critical for CNNs. By selectively activating only the

necessary neurons, the model can run with much

higher efficiency. ResNet-18 has about 11 million

parameters which contribute to its precision and

adaptability.

BIOIMAGING 2025 - 12th International Conference on Bioimaging

312

3.3 Data Pre-Processing and

Augmentation

According to Wang and Perez. “Data augmentation

has been shown to produce promising ways to

increase the accuracy of classification tasks.” (Wang

& Perez, 2017). Therefore, I used pre-processing and

data augmentation to artificially create more data

while improving the model’s accuracy and its ability

to tolerate a greater variety of images. During pre-

processing, I first conformed all the images to a

format of (244x244) pixels, helping the network to

quickly analyse each image without having to adapt

to multiple resolutions.

The augmentation Horizontal Flip, which feeds

the model a reflection of the original image,

significantly improved the dataset’s variety, helping

to reduce the model’s angular bias and preparing it to

classify real photographs. Additionally, the Random

Rotation augmentation addressed images which are

not horizontally centered, improving the model’s

ability to analyse realistic photographs, irrespective

of the angle. Another mechanism I used was feeding

randomly cropped images to the network, which

prepared the model to analyse non-perfect images, in

which leaves are not fully captured. Finally, centering

the images’ leaves made the data more consistent and

helped the model to adapt easier.

3.4 Model Training Process

Post-processing the data, I chose the optimizer

ADAM (Adaptive Movement Estimation) to

accelerate training speed, shorten training windows,

and increase acquired accuracy. According to Liu et

al., ADAM improved tested accuracy by roughly 3

percent (Liu et al., 2021). Also, with a learning rate

value of 0.001, the model spends more time on each

image, thus improving its growth. Additionally, using

a batch size of 400 images and several epochs

expedited the training process by effectively utilizing

the machine’s processing power.

4 EXPERIMENTAL RESULTS

4.1 Experimental Setup

The CNN was developed using Google Colab with a

T4 GPU, which helped to increase the model training

speed and assisted with running the code faster.

Machine configuration as follows: Windows 11 OS,

RTX 3080 GPU, and Ryzen 9 5900x CPU. For ML, I

used the ResNet-18 model from Pytorch’s ML

libraries and Anaconda Python Environment tools.

ResNet-18 is a CNN that is 18 layers deep and

pretrained with more than 1 million images from the

ImageNet Database.

I used a large dataset from Kaggle, which

encompassed multiple sub datasets. Both RestNet and

the dataset played a key role in learning rate and

accuracy of the training. With Torchvision from

Pytorch, I was able to augment the data, expanding

the data set artificially. This helped to increase the

volume and variety of dataset, helped to train the

model with a varied set of images and increased

accuracy on the non-training dataset. For translations,

I used random horizontal flips, random rotations, and

random crops.

The CrossEntropyLoss function from Pytorch

helped to calculate the loss during training, which was

then used to update the model’s neural parameters to

decrease the loss and improve its accuracy. Adjusting

in response to the loss value calculated by the loss

function was critical to the model's growth and helped

to improve the model’s performance. The loss

function showed similar effectiveness in the

validation loop as well.

Pytorch CrossEntropyLoss Function:

(

p,q

)

=−p

(

)



logq(x)

(1)

Where x is the number of classes, p(x) is the

actual label for class x, and q(x) is the predicted

robabilit

for class x.

4.2 Performance Evaluation and

Analysis

The CrossEntropyLoss loss function from Pytorch,

also known as Softmax or Log Loss, was used as the

main evaluation metric in all accuracy calculations.

After running the validation and test loops, the loss

function returned a loss value of 4.541%. This result

matched my expectations given the limitations of my

experiments; however, it demonstrates the viability of

using ML to identify plant diseases. This loss value

indicates that the model is working well, although

there is room for improvement. According to

Muhammad et al., relevant state-of-the-art

classification models can achieve accuracies in the

range of 96-97.5% (Muhammad et al., 2023). While

my model’s accuracy was below this, its development

demonstrates and supports the spread of ML and DL

technology to be applied to agriculture. On the test

set, my model showed that its accuracy increased with

the number of epochs until the epoch count reached

Using Machine Learning to Identify Crop Diseases with ResNet-18

313

15. After reaching 15 epochs, the accuracy on the test

set started to decrease because of the model becoming

too over-specified for the training dataset. This

matches the findings of Komatsuzaki’s research,

stating that too many epochs lead to models becoming

not generalizable and losing accuracy due to being

over-specified (Komatsuzaki, 2019).

4.3 Discussion of Findings

The model’s loss value of 4.541% shows that there is

a large area for growth in terms of overall accuracy.

However, it is reasonable given I was only able to use

a fraction of the overall training data specified earlier

due to technical issues. However, the fact that the

model still scored a loss of 4.541% despite this

demonstrates that my training methodology was

correct and can be utilized to create a more accurate

model. Given better testing conditions and software,

the technical issues which prevented the use of the

entire training dataset to train the model could have

been avoided. In this case, the model would be much

more accurate and applicable to helping farmers

identify diseases. However, it is clear that a strength

of the model is its quickness to learn. Given such a

small amount of training data, it still managed to

achieve a decent accuracy, showcasing the

effectiveness of ResNet-18 to train classification

models.

When testing the data, the constant growth in

response to more epochs demonstrated the benefits of

iterations. I initially believed that the model would

become over-specified at 11 epochs. However, the

consistent growth until 15 epochs impressed me and

demonstrated ResNet-18’s strength in being able to

grow consistently without becoming over-specified. I

believe with 15 epochs of a large dataset, a model

using ResNet-18 could become very accurate and

possibly rival state-of-the-art models. Therefore, if

this experiment were to be replicated, I would

recommend 15 epochs to be used to train the ResNet-

18 model. This would lead to the greatest accuracy

with the lowest loss and produce the most effective

model.

5 CONCLUSIONS

5.1 Summary of Findings

In summary, the model I created using ResNet-18 and

Plant Disease Classification Merged Dataset

performed well and identified plant diseases

accurately. This research demonstrates the

applicability of ML in addressing global issues such

as plant diseases in agriculture and aligns with the

research of Liu et al. (Liu et al., 2021) and

Huang (Huang, 2007). There is a great opportunity

for ML to be used to support farmers worldwide, and

with improved designs, data, and computing power,

this goal is highly achievable in the near future.

5.2 Applications, Limitations, and

Future Works

With its loss value of 4.541%, this model can be

improved by using the entirety of the original dataset.

Additionally, the technical limitations can be

addressed with a better design environment and

Compute power. With similar methods on a larger

scale, the model could further be tuned for wider

range of diseases and can be integrated with a tool to

identify plant diseases. A model capable of detecting

plant diseases accurately could drastically improve

crop yield and help to feed the millions of people

lacking access to food. I suggest my research be used

and fine-tuned to aid the millions of farmers who lack

information about plant diseases worldwide by

creating a physical tool implementing ML

technology.

Such innovation could immensely improve

farmers’ crop yields, potentially providing billions of

people across the world access to a consistent food

source. Similarly, further developed classification

models could be used for a wider range of purposes

beyond the field of agriculture. Classification as a tool

can be used in numerous ways to help billions of

people worldwide. Using DL classification

effectively could be the key solution to a multitude of

significant global issues faced by billions of people

across the world.

5.3 Limitations and Future Work

Although the model functioned decently well, it can

be improved with superior methods and technology.

For instance, as previously mentioned, the technical

issue preventing the use of the entire dataset could be

mitigated with advanced software and technology.

Additionally, the capability for experimentation was

shortened due to limited compute units on Google

Colab which inhibited my ability to further refine the

model. Furthermore, with more diverse data, the

model could be improved to include a wider range of

plants rather than just 88 total varieties. Implementing

these methods could develop a model far more

practically applicable for the task of aiding rural

farmers. If a ML framework is created specifically to

BIOIMAGING 2025 - 12th International Conference on Bioimaging

314

identify plant diseases, this would also drastically

improve the accuracy of the model, as it would look

for more specific details in the leaves to determine

diseases. Or if multiple other model architectures,

such as GoogLeNet and EfficientNet, were used to

determine which model performed the best, this could

also contribute to the development of a more accurate

model.

I recommend my research be studied, utilized, and

implemented, to aid the millions of farmers affected

annually by plant diseases. In the future, with

technological advancements, this research could be a

stepping stone in the path toward the globalization of

ML technologies in agriculture.

REFERENCES

Al-Falluji, R. A., Katheeth, Z. D., & Alathari, B. (2020).

Automatic detection of COVID-19 using chest X-ray

images and modified ResNet18-based Convolution

Neural Networks. Retrieved from https://www.tech

science.com/cmc/v66n2/40667

Asuntha, A., & Srinivasan, A. (2020). Deep learning for

lung cancer detection and classification - multimedia

tools and applications. Retrieved from https://link.

springer.com/article/10.1007/s11042-019-08394-3

Dobrovsky, A. (2023). Plant Disease Classification merged

dataset. Retrieved from https://www.kaggle.com/

datasets/alinedobrovsky/plant-disease-classification-

merged-dataset

Huang, K.-Y. (2007). Application of artificial neural

network for detecting phalaenopsis seedling diseases

using color and texture features. Retrieved from

https://www.sciencedirect.com/science/article/abs/pii/

S0168169907000385

Komatsuzaki, A. (2019). One epoch is all you need.

Retrieved from https://arxiv.org/abs/1906.06669

Lee, S. I., & Yoo, S. J. (2019). Multimodal Deep Learning

for Finance: Integrating and forecasting international

stock markets - The Journal of Supercomputing.

Retrieved from https://link.springer.com/article/

10.1007/s11227-019-03101-3

Liu, Z., Shen, Z., Li, S., Helwegen, K., Huang, D., &

Cheng, K.-T. (2021). How do adam and training

strategies help BNNS optimization. Retrieved from

https://proceedings.mlr.press/v139/liu21t.html

Maglogiannis, I. G. (2007). Emerging artificial intelligence

applications in computer engineering: Real word AI

systems with applications in eHealth, HCI, information

retrieval and Pervasive Technologies. Amsterdam: IOS

Press.

Mohsen, H., El-Dahshan, E.-S. A., El-Horbaty, E.-S. M., &

Salem , A.-B. M. (2017). Classification using deep

learning neural networks for brain tumors. Retrieved

from https://www.sciencedirect.com/science/article/

pii/S2314728817300636

Nielsen, M. (2015). Retrieved from https://www.ise.ncsu.

edu/fuzzy-neural/wp-content/uploads/sites/9/2022/08/

neuralnetworksanddeeplearning.pdf

Ramzan, F., Khan, M. U. G., Rehmat, A., Iqbal, S., Saba,

T., Rehman, A., & Mehmood, Z. (2019). A deep

learning approach for automated diagnosis and Multi-

class classification of alzheimer’s disease stages using

resting-state fmri and residual neural networks - Journal

of Medical Systems. Retrieved from https://link.

springer.com/article/10.1007/s10916-019-1475-2

Shoaib, M., Shah, B., EI-Sappagh, S., Ali, A., Ullah, A.,

Alenezi, F., … Ali, F. (2023). An advanced deep

learning models-based Plant Disease Detection: A

review of recent research. Retrieved from

https://www.frontiersin.org/journals/plant-

science/articles/10.3389/fpls.2023.1158933/full

Vargas, R., Mosavi, A., & Rulz, R. (2017). (PDF) Deep

learning: A review. Retrieved from https://www.

researchgate.net/publication/328285467_Deep_Learni

ng_A_Review

Wang, J., & Perez, L. (2017). The Effectiveness of Data

Augmentation in Image Classification using Deep

Learning. Retrieved from http://vision.stanford.edu/

teaching/cs231n/reports/2017/pdfs/300.pdf

Zhu, N., Liu, X., Liu, Z., & Hu, K. (2018). (PDF) Deep

Learning for smart agriculture: Concepts, tools,

applications, and opportunities. Retrieved from

https://www.researchgate.net/publication/327016437_

Deep_learning_for_smart_agriculture_Concepts_tools

_applications_and_opportunities

Zilong, H., Jinshan, T., Ziming, W., Kai, Z., Ling, Z., &

Qingling, S. (2018). Deep learning for image-based

cancer detection and diagnosis − A survey. Retrieved

from https://www.sciencedirect.com/science/article/

abs/pii/S0031320318301845

Using Machine Learning to Identify Crop Diseases with ResNet-18

315