AMBO UNIVERSITY

HACHALU HUNDESSA CAMPUS

DEPARTMENT OF COMPUTER SCIENCE

A Deep Learning Approach for Teff Leaf Disease Classification

Name ID.no
Dereje Tolesa……………………………………………………….Ugr/50319/13
Chala Girma ……………………………………………………...Ugr/36322/12
Chala Tesfaye ……………………………………………………...Ugr/51625/ 13
Abraham Desalegn ………………………………………………..Ugr/51629/13
Daniel Abera ………………………………………………………Ugr/51464/13
Tasloach Gattot…………………………………………………….Ugr/51391/13

Submission date:19/01/2024
Submitted to:Mrs.Lelise
 A Deep Learning Approach for Teff Leaf Disease
Classification
Abstract
Tef is often considered as a relatively healthy crop since it suffers very little from
epidemics of pests including diseases as compared to most other cereal crops grown in
the country (Ketema, 1993, Assefa et al., 2011, Assefa and Chanyalew, 2018; Chanyalew
et al., 2019). This has, amongst others, been due to the fact that the crop species being an
indigenous ancient crop originated and domesticated in Ethiopia before the Semetic
invasion of 4000-1000 BC (Ponti, 1978), it has coevolved with the pathogens and other
pests over the millennia of its cultivation (Badebo, 2013).
A comprehensive review of the pathological research on tef was made first by Amogne et
al. (2001) and later on by Badebo (2013). The main objective of this topic is to give a
summary of the body of knowledge, information, and innovation on teff diseases and
their management in Ethiopia.
This study gives the body of knowledge and information available to date on the biology,
distribution, significance, and management of the major diseases of tef in Ethiopia. To
this effect, the fisrt section presents an inventory of the diseases so far known and
recorded in teff in Ethiopia. The subsequent sections focus on description, symptoms,
spore characters, yield losses and management options of five important teff diseases
including teff rust (Uromyces eragrostidis), head smudge (Helminthosporium miyakei),
smut sp., zonate eye spot (Helminthosporium giganteum), and damping off. Accounts are
also made of recently emerging diseases such as blast and smut. Finally, it is concluded
by giving summarized highlights of the diseases and recommendations on the future
directions for the management of the diseases.

Keywords: Teff leaf Disease, Convolutional Neural Networks, Crop Viral Strains,
Artificial Intelligence
1. Introduction
Teff is a regular meal in the tropics that enriches the body with carbohydrates, or energy,
for over 200 million people. Teff produces more carbs per hectare than the major cereal
plants that can be cultivated at a cheaper cost. Teff also thrives under less-than-ideal
conditions: It may be stored underground for several months after maturation and is
tolerant of soil infertility and drought stress (Taiwo, 2006). Teff is planted in almost
every African country between 30° north and 30° south of the equator; however, it is
mostly grown in regions of humid tropics, although they can grow with minimal
agronomic inputs on low-nutrient soils. It can withstand drought and adapts well to
traditional mixed-cropping farming methods. It's simple to cultivate and propagate,
produces a lot of leaves per acre, and grows quickly (Ufuan Achidi et al., 2005). Protein,
vitamins, and minerals are abundant in Teff leaves. Teff leaves are consumed to varying
degrees in Africa's Teff-growing regions, and in some countries, they represent a key part
of the diet. Teff is grown in a variety of environmental and climatic circumstances on a
small, medium, and large scale across Africa, adding to food security and industrial
crops.
However, the main struggle known to Teff is grown almost in all regions of the country
as sole crop, mixed crop, relay crop, or in rotation with several types of crops in both
Belg (short rainy season February to May) and meher (long rainy season June to
September) seasons. It is also produced as irrigated crop in some areas. In these
production systems, teff is infected by about 25 fungal diseases and three pathogenic
nematodes (Bekele (1986), Amogne et al. (2001), (Table 1)). The composition of the
fungal species suggests that tef is attacked by different species at the seed, seedling,
vegetative, and reproductive stages as well as stored grains. However, only few of these
diseases occasionally cause economic yield losses.
Among the fungal diseases, zonate eye spot (Helminthosporium giganteum), smut, and
leaf blast are recent records(Ashenafi et. al, 2018). In the past the fungus Darluca filum
(B.V. ex Fr) Cast. (Eudarluca caricis (Biv). O. E Eriskss (teleomorph) was reported as
causative agent of disease in teff; however, this fungus is a mycoparasite of cereal rust-
fungi. Bacterial and viral diseases are not known in teff.
For tef rust (Uromyces eragrostidis), head smudge (Helminthosporium miyakei), smut
sp., Damping off, and zonate eye spot (Helminthosporium giganteum) brief description
symptoms and spore character and management options are given.

2. Objective
2.1. General Objective
This study aims to develop a CNN technique for Teff leaf disease diagnosis.
2.2. Specific Object
The aim of this study was achieved with the following objectives:
i. To obtain the Teff leaf dataset and preprocess the dataset.
ii. To design a CNN framework for the identification and classification of Teff leaf
disease.
iii. To develop a CNN model for the identification and classification of Teff leaf
disease.
3. Related Works
S/N Author Dataset Name Application Accuracy Training
% Models

1 ((Francis & Apple and tomato leaf Disease 88.7 CNN

Deisy, 2019) image dataset classification
(Author’s Generated)

2 (Ayu et al., Images of teff leaves Detection of teff 65.6 MobileNet v2

2021) ((Lab, 2021) leaf disease

3 (Mohanty et Plant Village Disease detection 99.35 AlexNet

al., 2016) Dataset(spMohanty,
 2016)

4 (Trivedi et al., Tomato leaf disease Tomato leaf 98.49 CNN

2021) dataset (Author’s disease detection
generated)

5 (Chebet et al., Plant leaf Images Plant disease 99.75 VGG 16,
2018) ((Hughes & Salathe, identification Inception V4,
2015) ResNet and
DenseNets

6 (Fuentes et Images of diseases and Detection of 83 R-CNN VGG-

al., 2017) pest in tomatoes plants tomato plant 16
(Author’s generated) diseases and pests’
recognition

Table 1 summarizes the existing deep-learning approaches utilized by other researchers to

detect plant disease.

4. Methodology
This research work involves using deep learning techniques for the diagnosis of teff leaf
disease. The proposed conceptual framework was implemented using Deep
Convolutional Neural Networks, and feature extractions were executed to minimize the
initial raw set of data to a manageable group for processing because features that are
improperly selected will hurt our model. The proof of concept was evaluated using
standard machine learning evaluation techniques such as accuracy, precision, recall and
F1Score.

4.1 Proposed Deep Convolutional Neural Network Model

An efficient solution for the detection of Teff leaf disease from the Teff leaf disease
image dataset is proposed in this study. The argumentation of data was used to address
the problem of the limited dataset, and then, Deep Convolutional Neural Networks were
fine-tuned for the classification of Teff leaf disease. The conceptual framework of the
proposed model is shown in Figure 2.

Figure 2: Proposed Model

4.2 Dataset Description

Teff leaf disease is a novel dataset aiming to provide the different classes of the aliments
that affect the production of Teff. This data set is designed for image classification and
image detection. When a disease affects a crop's leaves, it is thought to be the first
warning that the sickness will spread to the remainder of the crop. The majority of the
shots were taken by farmers and interpreted by professionals at the National Crops
Resource Research Institute (NaCRRI) in Kampala, in partnership with the AI lab at
Makerere University using a smartphone. The dataset, which includes 22,031 images
with an image size of 800 X 600 pixels, showing the different classes of the Teff leaf
disease which appeared to be already (Mwebaze et al., 2019), is being used to assess
and categorize the following diseases to build and optimize our network. The different
categories in the Teff leaf disease dataset are shown in Table 2 and Figure 3.
Table 2: Class Labelling of Dataset
Class Numeric label

Healthy 0

(CBB) Teff Bacterial Blight 1

(CGM) Teff Green Mottle 2

(CMD) Teff Mosaic Disease 3

(CBSD) Teff Brown Streak Disease 4

Figure Error! No text of specified style in document.: Sample Image of the dataset
containing the

5 classes of the Teff leaf diseases

4.3 Data preprocessing

This data was acquired in Uganda by National Crops Resources Research Institute
(NaCRRI). This data was collected over a period of one year by over 200 farmers with
the help of their smartphones. With the assistance of NaCRRIs Teff diagnostics domain
experts, each image taken was annotated and verified, to be sure of removing duplicate
images and images with inconsistent labels. The images dataset has a total of 9,436
labelled images and 12,595 unlabeled images. The annotations were divided into
training and testing sets and were distributed across 5 classes as shown in Table 3.
Table Error! No text of specified style in document.: Total Number of labelled Teff
Leaves
Teff leaf disease classes Train Test
Healthy 316 211
(CBB) Teff Bacterial Blight 466 311
(CGM) Teff Green Mottle 773 516
(CMD) Teff Mosaic Disease 2658 1773
(CBSD) Teff Brown Streak Disease 1443 963

4.4 Data Augmentation

The domain-agnostic data augmentation technique was utilized to implement the data
augmentation as expressed in Equation 1.1 and Equation 1.2.
𝑥̃ = 𝜆𝑥𝑖 + (1 − 𝜆) 1.1

𝑦̃ = 𝜆𝑦𝑖 + (1 − 𝜆) 1.2

where 𝑥𝑖, 𝑥𝑗 are the raw input data and 𝑦𝑖, 𝑦𝑗 are the one-hot label encodings. The [0, 1]
range of values that are sampled from the lambda distribution are the lambda values.
In detecting Teff leaf disease, the proposed model is interested in finding a function 𝑓 ∈
𝐹 that describes the relationship between a feature in the dataset 𝑋 and the target class 𝑌,
which follows the joint distribution (𝑋, 𝑌). The loss function 𝑙𝑓 was defined, which was
used to penalise the differences between predictions (𝑥) and the actual target 𝑦, for
instance, (𝑥, 𝑦)~𝑃. Then, the model minimizes the average of the loss function 𝑙𝑓 over the
data distribution 𝑃, also known as the expected risk:

(𝑓) = ∫ 𝑙(𝑓(𝑥), 𝑦)𝑑𝑃(𝑥, 𝑦) 1.3

The distribution 𝑃 of the training dataset , where (𝑥𝑖, 𝑦𝑖)~𝑃 for all 𝑖 = 1,
… , 𝑛 Using the training data 𝑇𝑑, the distribution 𝑃 may be approximated by the
empirical distribution in
Equation 1.4

where 𝛿(𝑥 = 𝑥𝑖, 𝑦 = 𝑦𝑖) is a mass centred at (𝑥𝑖, 𝑦𝑖), which can be used to approximate the
expected risk by the empirical risk using the empirical distribution 𝑃𝛿.

Figure 4: Augmented Images

4.5 ImageNet CNN Model

Pre-trained networks or models are previously trained models that include weight and
bias values that represent the dataset's characteristics. It has already learned the features
and parameters from a specific dataset or domain, allowing applied knowledge to be
successfully transferred to various types of data. Learned properties such as borders,
vertical lines, and edges can be transferred from one domain to another using the
network. Reduced computing time, faster learning, and convergence are some of the
benefits of pre-trained models. Pretrained Models taken from the ImageNet challenge
include (Alom et al., 2019; Celik et al., 2020; Khan et al., 2019; Ragab et al., 2021). The
detailed design is shown in Figure 3.2 and fine-tuning of the developed model is
discussed below:

i. A Cross-Entropy Loss with default parameters was utilized. ii. A learning rate of 1e-
4 with “ReduceLROnPlateau” scheduler based on average validation loss (mode=’min’,
factor=0.2, patience=5, eps=1e-6) was also adopted.
iii. Train augmentations from the Albumentations Python library were utilized. iv. A
RandomResizedCrop, Transpose, HorizontalFlip. VerticalFlip, ShiftScaleRotate, and
Normalize functions were also employed in reshaping the raw image data.
v. The MixUp hyperparameter technique was utilized to return a random index from
another image in the batch. This ensures that images from the most frequent class
(i.e., class 3 in the dataset) will always be mixed with an image from another class,
hence, ensuring that the other classes are more present in the training dataset.
Table 4 and Table 5 illustrate the raw image size, reshaped image size, the custom-
top-liner layer as well as the learning rate of the developed model.

Table 1: Reshaping and Resizing of the Raw Images

Original Image Size Reshaped Image Size Custom-top-Liner Layer Learning Rate
800*600 512*512 384*384 1e-4 / 7

Table 5: Batch Size and N_Folds and Epochs of the Proposed Model

Batch Size Validation No. Labels N_Folds Batch Size Base Epochs
139*8 5 5 16 30

The developed ImageNet mean and standard deviation include:

IMAGENET_MEAN = tf.constant([0.485, 0.456, 0.406], dtype=tf.float32)
IMAGENET_STD = tf.constant([0.229, 0.224, 0.225], dtype=tf.float32)
Table 6: Break down of the ImageNet architecture showing details of the layer as
well as the output shape and no of parameters
Layer (type) Output Shape Param #
efficientnet-b4 (Functional) (None, 16, 16, 1792) 17673816
dropout (Dropout) (None, 16, 16, 1792) 0
global_average_pooling2d (Gl (None, 1792) 0
dropout_1 (Dropout) (None, 1792) 0
dense (Dense) (None, 5) 8965

5. Results and Discussion

This section focuses on evaluating the findings of the proposed model. This section also
discussed the evaluation technique utilized in this study, and the results obtained from
the implemented model. Finally, the implemented model was benchmarked with
cutting-edge deep learning methodology in literature.

5.1 Evaluation Technique

In the study, standard machine learning evaluation indicators such as Recall, F1 score,
Precision, and mAP were used to evaluate the performance of the trained Teff leaf
disease diagnosis model. The calculation equations are as follows:
Table 7: Evaluation Metric for Machine learning

Evaluation Metrics Description Equation

Recall This is the proportion of
positive cases in the overall number of positive
instances.

Precision or This is the expected number

Positive Predictive of positive cases divided by
Value the total number of positive
cases.
F1 score The harmonic mean of

accuracy and recall is this.

This considers both, thus the

greater the F1 score, the more accurate the
model.
mAP The mAP compares the
detected box to the groundtruth bounding box to
determine a score. The accuracy of the model's
detection increases with score.

Table 8: Primary Components of Machine Learning Evaluation Technique

S/N Components Description
1 True Positive (TP) The total number of indicators in the positive class
that were successfully predicted.
2 True Negative (TN) The total number of indicators in the negative class
that were successfully predicted.
3 False Positive (FP) The total number of indicators that should have been
in the positive class but were instead placed in the
negative class.
4 False Negative (FN) The total number of indicators that should have been
in the negative category but were instead placed in
the positive category.

Results of the CNN Model

The Teff leaf disease detection model's precision is illustrated in this session. The results
demonstrate that the CNN model recorded a prediction precision of 0.9728. The result in
Figure 5 depicts that the CNN model achieved a validation loss below 0.0272. This
result shows correctly predicted samples as well as the misclassification rate, and the
result indicate that the model performed well in detecting Teff disease and it is in
synchronization with the prediction accuracy as well as the classification report.

Table 9: Classification Report

Class labels Precision Recall F1-Score Support
Healthy – 0 0.77 0.70 0.73 260

(CBB) Teff Bacterial Blight – 1 0.87 0.87 0.87 613

(CGM) Teff Green Mottle – 2 0.83 0.83 0.83 516

(CMD) Teff Mosaic Disease – 0.95 0.97 0.96 2593

3
(CBSD) Teff Brown Streak 0.77 0.74 0.76 466
Disease – 4
Accuracy - - 0.90 4448

Macro Avg 0.84 0.82 0.83 4448

Weighted Avg 0.90 0.90 0.90 4448

Figure 5: Validation Loss

5.2 Confusion Matrix

Figure 6 shows the experimental result for the CNN model in the form of a confusion
matrix. The classification model's (or classifier model's) performance on a set of test
data for which the true values are known is frequently described by a table called the
confusion matrix. A sample confusion matrix table is shown in Figure 6.

Figure 6: Confusion Matrix

5.3 Benchmarking the CNN Model with the Existing models in Literature
The CNN model approach was benchmarked with state-of-the-art approaches and
models. The CNN model efficiency was investigated using performance accuracy. Table
10 and Figure 8 illustrate the performance accuracies obtained from state-of-the-art
techniques and the CNN model. Even though (Sangbamrung et al., 2020) Study
outperforms other studies in terms of accuracy, the CNN model outperforms this work
and gets a competitive accuracy result. Based on the outcomes of analyzing the proof of
concept, it can be concluded that the method used in this study is effective for
classifying and detecting Teff leaf disease picture collection. Table 10: Comparison
with state-of-the-art techniques
S/N Authors Model Model Performance (%)
1 (METLEK, 2021) ResNet50 85.4
2 (Methil et al., 2021) Efficient Net 85.64
3 (Zhuang, 2021) Vision transformer 90
4 (Ramcharan et al., 2017) CNN with Transfer 93
Learning
5 (Sambasivam & Opiyo, 2021) CNN 93
6 (Sangbamrung et al., 2020) R-CNN 96
7 Developed CNN Model CNN 97.28

Benchmarking the Proposed Model with Existing Model in Literature

Referenced Models
(METLEK,2021)
(Methil et al.,2021)
(Zhuang,2021)
(Ramcharan et al.,
2017)
(Sambasivam & Opiyo, 2021)
(Sangbamrung et al.,
2020)
Proposed Model
75 80 85 90 95 100
Accuracy

Figure 7: Benchmarking of the Proposed Model with Existing Model in Literature

6. Conclusion
Plant infections are a type of natural disaster that inhibits plants from growing and, in
the majority of cases, causes them to die. The risk of incorrectly identifying and
incorrectly classifying plant disease with a high error rate, as well as the labour-
intensive nature of identifying and classifying diseases using artificial eyes, led to the
proposed method of employing deep learning-based methods to identify plant illness.
This study implemented a fast Teff disease detection model. The main objective is the
development of a pre-trained deep learning model to detect Teff disease and reduced the
false positive rate. A secondary data source containing a total of 21,367 data images.
The data set was trained using the pre-trained network, and its performance was
analyzed using standard evaluation techniques. The result shows that the proposed
model achieved a prediction accuracy of 97.28 which outperformed existing models in
literature. The results indicate that the proposed model performed well in detecting Teff
disease and it is capable of diagnosing Teff leaf diseases. The implemented model will
be trained on the Nigeria Teff leaf dataset. The idea is to aid farmers in Nigeria improve
detect early plant diseases and improving their harvest rate.
References
Agarwal, M., Singh, A., Arjaria, S., Sinha, A., & Gupta, S. (2020). ScienceDirect
ToLeD : ToLeD : Tomato Tomato Leaf Leaf Disease Disease Detection Detection
using using Convolution
Convolution Neural Neural Network Network. Procedia Computer Science,
167(2019), 293– 301. https://doi.org/10.1016/j.procs.2020.03.225
Alom, M. Z., Taha, T. M., Yakopcic, C., Westberg, S., Sidike, P., Nasrin, M. S., Hasan,
M., Van Essen, B. C., Awwal, A. A. S., & Asari, V. K. (2019). A state-of-the-art
survey on deep learning theory and architectures. Electronics (Switzerland), 8(3).
https://doi.org/10.3390/electronics8030292