University of Gondar

College of Informatics
Department of Information Science
Data Science and analytics Post Graduate
Program
Computer Vision

Potato Leaf Diseases Classification Using Transfer learning

Prepared by: - Eyerus Wendmu

- Eyerus Fentahun

Submittedto Million.M (Ph.D.)

April,28, 2023
Gondar, Ethiopia
Table of Contents
Abstract ......................................................................................................................................................... 2
1. Introduction ........................................................................................................................................... 3
2. Objective ............................................................................................................................................... 4
3. Related work ......................................................................................................................................... 4
5. Methodology ......................................................................................................................................... 8
Image Preprocessing................................................................................................................................. 8
Image augmentation ................................................................................................................................. 8
6. CNN Model........................................................................................................................................... 8
7. Training and Testing split ..................................................................................................................... 8
8. Architecture........................................................................................................................................... 9
Inception Module ..................................................................................................................................... 9
Convolution......................................................................................................................... 10
Max pool layer .................................................................................................................... 10
Concat ................................................................................................................................. 10
Pooling layers ...................................................................................................................... 10
Fully connected layer .......................................................................................................... 10
Softmax ............................................................................................................................... 10
9. Experimental result ............................................................................................................................. 11
Strength ....................................................................................................................................................... 12
Conclusion .................................................................................................................................................. 13
Bibliography ............................................................................................................................................... 14
Contribution ................................................................................................................................................ 15

1
Abstract

Potato leaf blight Alternaria, also known as early blight, is a soil-borne fungal pathogen that
primarily attacks potato crops. Diagnosis can be problematic as symptoms can be mistaken for
nutrient deficiencies or Verticillium wilt. Deep Learning is the newest and current trend in the
machine learning field that paid a lot of researchers' attention in the recent few years. Deep
learning has been widely used in many applications in the field of agriculture to solve complex
problems that require extremely high accuracy and sensitivity. The dataset consists of 180 image
classes; in this study, only three classes used for classification. The three leaf image classes used
for disease classification system are infected by Late Blight, infected by Early Blight, and
healthy potato leaves. The objectives of this project is to use images of potato leaf diseases that
are less damaged and not infected, to classify the images using the healthy images and to develop
the best one in the deep learning architecture. The main aim of the proposed system is to detect
early the plant diseases to decrease the plant’s production losses due to plan diseases like early
and late blight of the potato crops. The main approach of our proposed system is to detect early
the plant diseases to decrease the plant’s production losses by using a diagnosis and detection
system based on the Convolution Neural Network (CNN). The images dataset split to 80% for
training, 10% for testing and 10% for validation. In this paper, a model is proposed that uses pre-
trained models such as InceptionV3 and EfficientNetB1 for fine-tuning (transfer learning) to
extract relevant features from a dataset. To train the database, we used Convolutional-Neural-
Network (CNN) algorithm of EfficientNetB1 88% and the accuracy is 94% than InceptionV3,
better than EfficientNetB1, InceptionV3 showed good accuracy.

2
 1. Introduction
Potato is one of the most important export products of the country and a lot of income is
generated annually through the export of this product to the target countries. Alternaria causes
lesions on the leaves which often have a target spot appearance of concentric rings. Given the
importance of exporting potatoes for the country, paying attention to the quality classification of
potatoes seems to be very important and necessary. [1] In fact, through the quality classification
of potato products, more target countries can be attracted by providing a quality product, which
will result in more revenue for the country. In general, potato crop classification methods can be
divided into two categories: hardware and software. In this research, a software method is
proposed to classify the potato crop. In the proposed method, using image processing techniques
and machine vision, the desired feature is extracted from different samples of potatoes, and based
on the values obtained for the desired feature; the classification operation is performed by CNN
and deep learning. [2] The ultimate goal of the proposed method is to classify potato disease into
several categories. A plant disease is the change that affects or modifies its vital functions. It is
mainly caused by bacteria, fungi, microscopic animals or viruses, and has a strong impact on
agricultural yields.

In this study, have emerged as one of the most promising approaches given their ability to learn
reliable visual characteristics and features from the image itself, the thing that help in detecting
specific symptoms in the plan disease like those appear in the potato Healthy, Early Blight and
Late Blight. [3] The problem of food shortage has grown in recent years in developing countries.
In many of these countries, potato is one of the major eaten throughout the year. Recently the
production of potatoes has been falling short due to various diseases like Healthy, Early Blight
and Late Blight which cause a huge loss of cropped plants. This also leads to a major loss in the
national economy as well. Potato diseases are one of the most devastating global plant diseases
because it affects the productivity and quality of potato crops and adversely affects both
individual farmers and the agricultural industry. Advances in the early classification and
detection of crop blight using deep learning technologies have increased the opportunity to
enhance and expand plant protection. [4]

3
 2. Objective
The overall objective of this study is to accurately identify and develop a classification of potato
leaf spot disease using deep learning algorithms.

3. Related work
There are many methods to detect the plant diseases and various researchers had suggested
various techniques for detecting potato leaf diseases in their work. In this section, a summary of
those approaches is presented.

4
Citation Uses of Dataset Uses of classes Accuracy Strength Gap
Techniques
[1] Convolution The dataset 5 classes (Gray Mold, They get 100% They use 5 class and Multi-class leaf disease
neural network taken Skin Spot, Common and 99%. They achieve high classification may struggle
(CNN), from databases Scab, accuracy. to correctly identify images
Gangrene and Leak)
Neural network CFIA, USDA that contain anomalies or
and Support and potatoes unique features that were
Vector Machine farms in not found in the original
(SVM). Ardebil, data set on which the model
the city of Iran. was trained.
[2] Plant Village 3 classes (Healthy, They archive They used CNN useful While the paper states that
dataset from Early Blight and Late 98.2%. for image recognition the proposed vision
Blight.)
Egypt and image classification, transformer-based classifier
as well as other computer achieves higher accuracy
CNN vision tasks. than other state-of-the-art
methods, it is unclear
whether the performance
improvement is of
agricultural value.
[3] Taken from 5 classes (Healthy, They get It showed great accuracy
convolution Databases CFIA, Black Scurf, 100% and 99% by using a good format
USDA, and
neural network from potatoes
Common Scab, accuracy in and model
(CNN) farms in Ardebil Black some of the
city of Iran. Leg, Pink Rot. classes.
[4] From Kaggle 3 classes (Early 97.8% over Multiple classifiers Not allowing us to use
Plant Village blight, Late blight the test dataset Precision, Recall, F1- phone and other simple
pre-trained
Dataset
VGG19, and Healthy) Score These models ways to find out the disease
inceptionV3,fine- allow the farmers to more easily than it is now
tuning(transfer distinguish the diseased
learning)
from the diseased and
prevent the disease from

5
 spreading to healthy
leaves.
[5] Private dataset 5class(Alternaria They get 91% The study uses advanced Better accuracy could be
Solani,Healthy, achieve. deep learning techniques achieved by adding larger
Phytophthora to provide insights into data sets and models.
VGG16 and Infestant,Virus,Insect) the diagnosis and
VGG19 identification of health
conditions such as fruit
and vegetable leaf spot
disease.

6
 4. Dataset
The dataset used in this study consist of potato leaf images from Plant Village dataset. The
dataset consists of 180 image classes; in this study, only three classes used for classification. The
three leaf image classes used for disease classification system are infected by Late Blight,
infected by Early Blight, and healthy potato leaves. The used dataset for the proposed system
consists of 2150 colored or RGB leaf images with resolution 299×299 pixels.

There are 60 leaf images with Late blight, 60 leaf images with Early blight, and 60 images of an
uninfected or healthy leaf.

The dataset needed preprocessing and preparing before the training and testing processes start.
First, reading the images from the dataset was needed, so the images with the same class were
gathered in a single file named by the class name. Second, the images resized to 299 by 299
pixels using the OpenCV library. Third, the read images were converted to arrays because the
input layer of the model entered the model as an array. Fourth, the labeling process was done
using the Skikit-learn machine learning library, by creating the label of each class based on the
file name, which represents the class name of each image. Finally, the dataset split for training
and testing. The data collected from of Kaggle

Samples of potato images dataset;

Potato Late_blight Potato healthy Potato Early_blight

7
 5. Methodology
Image Preprocessing
Image preprocessing is the steps taken to format images before they are used by model training
and inference. This includes, but is not limited to, resizing, orienting, and color corrections.
Image preprocessing may also decrease model training time and increase model inference speed.

Image augmentation
Data augmentation is a set of techniques to artificially increase the amount of data by generating
new data points from existing data. This includes making small changes to data or using deep
learning models to generate new data points. Image augmentations are manipulations applied to
images to create different versions of similar content in order to expose the model to a wider
array of training examples. Image augmentation manipulations are forms of image
preprocessing, but there is a critical difference: while image preprocessing steps are applied to
training and test sets, image augmentation is only applied to the training data. so, There are four
techniques that we used for data augmentation; Rescale,shear_range,zoom_range,horizontal_flip.

6. CNN Model
The CNN model is used for disease detection and classification, the model includes an Input
layer, Convolutional layers, Pooling layers, fully connected layers, and output layer. The model
is implemented using python programming language and Tensorflow library, which “is an
interface for expressing machine learning algorithms, and an implementation for executing such
algorithms

7. Training and Testing split

The images dataset split to 80% for training, 10% for testing and 10% for validation. The model
trained using 80% of the dataset, which is around 144 images, through 20 epochs. One Epoch
means one forward pass of all the training dataset and one reverse pass. After each epoch the
testing data enter the model, then validation accuracy of the epoch calculated. Finally, after
finishing all epochs the total training accuracy and validation accuracy calculated.To implement
the model there are different libraries used like: “Numpy” to deal with arrays.

8
 8. Architecture

Inception Module
Inception Module is an image model block that aims to approximate an optimal local sparse
structure in a CNN. Put simply, it allows for us to use multiple types of filter size, instead of
being restricted to a single filter size, in a single image block, which we then concatenate and
pass onto the next layer.

The Inception V3 is a deep learning model based on Convolutional Neural Networks, which is
used for image classification. Inception-v3 is a convolutional neural network architecture from
the Inception family that makes several improvements including using Label Smoothing,
Factorized 7 x 7 convolutions, and the use of an auxiliary classifier to propagate label
information lower down the network (along with the use of batch normalization for layers in the
sidehead).

9
 Convolution

Is an orderly procedure where two sources of information are intertwined; it’s an operation that
changes a function into something else. Convolutions have been used for a long time typically in
image processing to blur and sharpen images, but also to perform other operations.

Max pool layer

Is similar to convolution layer, but instead of doing convolution operation, we are selecting the
max values in the receptive fields of the input, saving the indices and then producing a
summarized output volume.

Concat

Creates a single vector-valued column from multiple columns. It can be performed on data
before training a model. The concatenation can significantly speed up the processing of data
when the number of columns is as large as hundreds to thousands.

Pooling layers

Pool layer is to reduce the dimensions of the hidden layer by combining the outputs of neuron
clusters at the previous layer into a single neuron in the next layer.

Fully connected layer

Multiplies the input by a weight matrix and then adds a bias vector. The convolutional (and
down-sampling) layers are followed by one or more fully connected layers. As the name
suggests, all neurons in a fully connected layer connect to all the neurons in the previous layer.

Softmax

That Extends this idea into a multi-class world. That is, Softmax assigns decimal probabilities to
each class in a multi-class problem. Those decimal probabilities must add up to 1.0. This
additional constraint helps training converge more quickly than it otherwise would.

Efficient Net Model

EfficientNetB1 uses a similar architecture to other popular convolutional neural networks, such as ResNet and
Inception, but with a few key differences. One of the most notable differences is the use of "swish" activation
functions, which have been shown to improve performance compared to traditional activation functions like
10
ReLU. EfficientNetB1 is often used as a pre-trained model for transfer learning in computer vision tasks, as it
has achieved state-of-the-art performance on a wide range of image classification benchmarks.
EfficientNetB1 is one of the versions of the EfficientNet model which has 7M parameters. It is designed based
on a compound scaling approach that uniformly scales the input image resolution, depth/width, and resolution
of all convolutional layers in an optimal manner. This allows EfficientNet to achieve better accuracy at a much
faster run time than other popular CNN architectures.

11
 9. Experimental result
The epoch specified in this study was 20 epochs with32 batch size, and a learning rate is 0.01 to
improve the performance of the model. The learning process means that the algorithm with the
proposed method seeks to come across values in the image dataset to be able to recognize new
images. Each step of the epoch must be the same as the value of the image being trained. The
results of the epochs will be recorded to determine the value of loss and accuracy. Loss is an
indication of a bad value on the model, the value of the loss obtained must be close to zero or
equal to zero and accuracy value is a parameter of the success level of the system in classifying
objects. The result model1 (inceptionv3) of training and validation accuracy are presented
bellow graphically

The result model1 (inceptionv3) of our loss function to the training and validation are give bellow

12
The result model2 (EfficientNetB1) of training and validation accuracy are presented bellow
graphically

The result model1 (EfficientNetB1) of our loss function to the training and validation are give
bellow

13
 Testing

After we build the model we test the model by giving new image to it and it classify to the
highest probability score for the class labels,

Strength
The strength of our study is that the use of pre-trained models requires less training time and data, and
the inclusion of more; Our results demonstrate the applicability and efficiency of a previously trained
CNN model in potato leaf imaging and identification. Using two different models, we have produced a
clear and cool result of strictness for the society.
14
Conclusion
Currently, it is very important to identify disease at the plant waste level to increase productivity and
product quality. generally one of the most critical problems that threaten the worlds agricultural, causing
large In this paper, potato leaf diseases diagnosis and detection system are proposed based on the
Convolutional Neural Network technology with a customized model architecture. The main aim of the
proposed system is to detect early the plant diseases to decrease the plant’s production losses due to plan
diseases like early and late blight of the potato crops. Since diagnosis of disease requires a lot of
knowledge, if we can apply this method to it, it will be very useful. But what I consider a weakness is that
farmers or anyone with a smart phone cannot take a photo of the leaf and send it to a server. The server
can immediately identify the type of disease and send the results to the Smartphone along with the
prescribed medicines and the medicine can be used immediately. So large dataset Classification accuracy
needs to be increased.

15
Bibliography
[1] Divyansh Tiwari, Mritunjay Ashish, Nitish Gangwar,Abhishek Sharma, Suhanshu Patel, Dr. Suyash
Bhardwaj, "Potato Leaf Diseases Detection Using Deep learning," International Conference on
Intelligent Computing and Control Systems (ICICCS), no. IEEE, june 2020.

[2] Samer I. Mohamed, "Potato Leaf Disease Diagnosis and Detection System Based on Convolution
Neural Network ," International Journal of Recent Technology and Engineering (IJRTE) , vol. 9, no.
4, November 2020.

[3] Rizqi Amaliatus Sholihati1,Indra Adji Sulistijono2,Anhar Risnumawan3,Eny Kusumawati4, "Potato

Leaf Disease Classification Using Deep Learning Approach ," International Electronics Symposium
(IES), no. IEEE, November 2020.

[4] Ali Arshaghi1, Mohsen Ashourin2*, Leila Ghabeli1, "Detection and Classification of Potato Diseases
Potato Using a New Convolution Neural Network Architecture," internatinal conference, December
2021.

[5] Ali Arshaghi 1 & Mohsen Ashourian2 & Leila Ghabeli1, "Potato diseases detection and classification
using deep learning methods," july 2022.

16
Contribution
Name of members Task
1. Eyerus Fentahun Abstract, Problem identification,
Literature review, Architecture,
conclusion and References

2. Eyerus Wendemu Introduction, Methodology, pre

trained Model building coding
(implementation) and result
discussion, future work