Project Report

Plant Disease Detection

Abstract and Random Forests. More recently, deep learn-

ing models, particularly Convolutional Neural Net-
Plant disease detection plays a crucial role in works (CNNs), have demonstrated superior perfor-
ensuring agricultural productivity and food se- mance in image-based disease detection tasks.
curity. In recent years, machine learning tech-
niques have shown promising results in au-
tomating the detection process. This paper
3 Methodology
provides a comprehensive review of various
In this section, we describe the methodology used
machine learning approaches used for plant
disease detection, highlighting their strengths, for plant disease detection using TensorFlow, a pop-
weaknesses, and future directions. ular open-source machine learning framework de-
veloped by Google. TensorFlow provides a flexible
1 Introduction ecosystem for building and training deep learning
models, making it well-suited for image classifica-
Plant disease detection plays a crucial role in safe- tion tasks such as plant disease detection.
guarding agricultural productivity and ensuring
global food security. As plants face constant threats
from various pathogens, timely and accurate identi-
fication of diseases is essential for effective disease
management. Recent advancements in technol-
ogy, such as the integration of artificial intelligence
and image processing, have revolutionized plant
disease detection methods. By leveraging these
technologies, researchers and farmers can now ef-
ficiently and accurately diagnose plant diseases,
We employed a deep learning approach based
enabling them to implement targeted and timely
on Convolutional Neural Networks (CNNs), which
interventions. This not only minimizes crop losses
have demonstrated outstanding performance in im-
but also promotes sustainable agriculture practices
age recognition tasks. The architecture of our CNN
by reducing the reliance on chemical treatments.
model consists of multiple convolutional layers fol-
As we continue to explore innovative solutions,
lowed by max-pooling layers for feature extraction.
plant disease detection stands at the forefront of
We utilized TensorFlow’s high-level API, Keras, to
efforts to enhance crop health, optimize yield, and
define and train our CNN model.
contribute to the overall resilience of our agricul-
tural systems. For data preprocessing, we collected a large
dataset of plant images containing healthy and dis-
2 Related Works eased samples across multiple plant species. We
preprocessed the images by resizing them to a stan-
Previous research has explored various ML ap- dard size, normalizing pixel values, and augment-
proaches for plant disease detection, including su- ing the data using techniques such as rotation, flip-
pervised learning, unsupervised learning, and deep ping, and zooming to improve model generaliza-
learning methods. Early studies focused on hand- tion.
crafted feature extraction combined with traditional To train the CNN model, we split the dataset into
classifiers such as Support Vector Machines (SVM) training, validation, and test sets. We used Tensor-
Flow’s built-in data augmentation and image pre-
processing utilities to efficiently handle large-scale
datasets and enhance model robustness. We em-
ployed transfer learning by fine-tuning pre-trained
CNN architectures such as ResNet, Inception, or
MobileNet to accelerate training and improve con-
vergence.
During model training, we employed various
optimization techniques such as stochastic gradi-
ent descent (SGD), Adam optimizer, and learning
rate scheduling to optimize model parameters and
minimize training loss. We monitored the model’s
performance on the validation set and employed Data splitting and model training:
early stopping to prevent overfitting. S Creating
a plant disease detection model using TensorFlow
and Keras involves several steps. • Split the dataset into training, validation, and
test sets.
Dataset Collection: New Plant Diseases
Dataset The "new-plant-diseases-dataset" sourced
from Kaggle is a comprehensive collection of im- • Train the model on the training set using the
ages capturing various plant diseases across dif- compiled model, specifying the number of
ferent plant species. This dataset was specifically epochs and batch size.
curated to facilitate research and development in
the field of plant disease detection using machine
• Monitor the validation performance to detect
learning techniques. The dataset encompasses high-
overfitting.
resolution images showcasing both healthy plants
and plants affected by a diverse range of diseases.
These classes cover a wide range of crops, includ- • Use Adam optimizer to optimize the final
ing tomatoes, potatoes, grapes, apples, corn, pep- model which we train and then change the
pers, strawberries, peaches, and more. The di- whole trained tensorflow model into tensor-
versity in diseases and healthy states makes this flow lite model.
dataset suitable for training and evaluating machine
learning models for plant disease detection across
• Use Adam optimizer to optimize the final
multiple plant species. Each class represents a dis-
model which we train and then change the
tinct challenge in identifying and managing plant
whole trained tensorflow model into tensor-
health, making it a valuable resource for research
flow lite model.
and application in agriculture Data Preprocessing:

• Resize images to a consistent size.

• Normalize pixel values to a range between 0

and 1.

Model Selection: We choose a pre-trained con-

volutional neural network (CNN) as a base model Deploying on Edge Impulse: We obtained the
MobileNet trained on Imagenet dataset. Transfer TensorFlow Lite (TFLite) file from the trained
Learning: We load the pre-trained model and freeze model and proceeded to upload it onto the Edge
its layers to retain learned features. Add custom Impulse platform. After uploading, we deployed
layers on top of the pre-trained model to adapt it the model and initiated its execution to perform
for the plant disease detection task. real-time inference tasks.
 practical applications in agriculture and crop man-
agement, contributing to improved crop yield and
food security.

5 Result

The project successfully developed a Convolutional

Neural Network (CNN) model using TensorFlow
and Keras for plant disease detection, achieving ap-
proximately 90accuracy. After training, the model
was converted into TensorFlow Lite (TFLite) for-
mat for seamless deployment on the Edge Impulse
4 Experimental Results platform.
We present the experimental results of our During testing on Edge Impulse, the deployed
TensorFlow-based plant disease detection system. model demonstrated robust performance, accu-
We evaluated the performance of the trained CNN rately identifying various plant diseases in real-
model on a held-out test set containing unseen plant time scenarios. With its high precision, recall, and
images. We report the following performance met- overall classification accuracy, the model proved
rics: effective in distinguishing between healthy and dis-
eased plants.
• Accuracy: The proportion of correctly classi- This deployment highlights the model’s adapt-
fied samples out of the total number of sam- ability to edge computing environments, showcas-
ples. ing its potential for practical applications in pre-
cision agriculture and crop management systems.
• Precision: The ratio of true positive predic-
Leveraging its efficiency and real-time capabilities,
tions to the total number of positive predic-
the model holds promise for optimizing yield and
tions.
minimizing crop losses in agricultural settings.
• Recall: The ratio of true positive predictions Moving forward, ongoing optimization efforts
to the total number of actual positive samples. can further enhance the model’s scalability and per-
formance across diverse agricultural contexts. Ad-
• F1-score: The harmonic mean of precision
ditionally, continual updates to the Edge Impulse
and recall, providing a balance between the
platform can streamline the deployment process,
two metrics.
making AI-driven solutions for plant disease de-
Our TensorFlow-based model achieved a clas- tection more accessible to farmers and agricultural
sification accuracy of X% on the test set, demon- stakeholders.
strating its effectiveness in distinguishing between
healthy and diseased plants. The precision, recall, 6 Conclusion
and F1-score of the model were Y%, Z%, and
W%, respectively, indicating its ability to mini- The CNN model’s robust performance, achieving
mize false positives and false negatives. approximately 90disease detection, highlights its
We conducted additional experiments to analyze efficacy in agricultural applications. Its success-
the impact of hyperparameters, network architec- ful deployment on Edge Impulse underscores its
ture, and training strategies on model performance. potential for real-time disease monitoring. This
Through extensive experimentation and hyperpa- advancement has promising implications for opti-
rameter tuning, we identified the optimal configu- mizing crop yield and minimizing losses in preci-
ration for our TensorFlow-based plant disease de- sion agriculture. Continued optimization efforts
tection system. can further enhance its scalability and accessibility,
Overall, our results demonstrate the efficacy of fostering sustainable practices and food security. In
TensorFlow in developing accurate and efficient summary, the CNN model represents a significant
deep learning models for plant disease detection. stride towards leveraging AI for transformative im-
We believe that our approach holds promise for pact in agriculture.
7 References
1. Camargo A. and J. S. Smith. 2008.
An image-processing based algorithm to auto-
matically identify plant disease Visual symp-
toms.
Bio.Systematic. Engineering., 102: 9 – 21

2. Camargo, A. and J. S. Smith. 2009.

Image processing for pattern classification for
the identification of dis-ease causing agents
in plants.
Com. Elect. Agr.66:121 –125.

3. Guru, D. S., P. B. Mallikarjuna and S .Manju-

nath. 2011.
Segmentation and Classification of Tobacco
Seedling Diseases.
Proceedings of the Fourth Annual ACM Ban-
galore Conference.

4. Zhao, Y. X., K. R. Wang, Z. Y. Bai, S. K. Li,

R. Z. Xieand S. J.Gao. 2009.
Research of Maize Leaf DiseaseIdentifying
Models Based Image Recognition. Crop Mod-
eling and Decision
Support.Tsinghuauni.press .Beiging. pp. 317-
324.