Explainable AI(XAI) using LIME and Disease Detection in Mango Leaf by Transfer Learning Approach

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 11 Issue: 02 | Feb 2024 www.irjet.net p-ISSN: 2395-0072
© 2024, IRJET | Impact Factor value: 8.226 | ISO 9001:2008 Certified Journal | Page 72
Explainable AI(XAI) using LIME and Disease Detection in Mango Leaf by
Transfer Learning Approach
Vishakha Mistry
Head of Department, Department of Information Technology, 360 Research Foundation, Tumkaria, Bihar, India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - Agriculture is an essential means of earning
income for a significant percentage of the worldwide
population. As a result, the productivity of crops has become
crucial all over the world. Farmers will get more benefitsfrom
using modern digital tools for autonomous disease detection.
Because agriculture is a complex field, it is particularly
essential to improve the interpretability of agricultural ML
models. First, the article proposes to identify leaf disease in
Mango plants via pre-trained deep-learning architecture.
Secondly, for the purposeofdemonstratingthe interpretability
of my model's choice, I made use of Local InterpretableModel-
Agnostic Explanations (LIME), an explainable AI (XAI) tool.
Key Words: leaf disease detection, Transfer learning,
explainable Artificial Intelligence (XAI), Local Interpretable
Model-agnostic Explanations (LIME).
1. INTRODUCTION
Artificial intelligence and machine learning are now used in
diverse agricultural applications. The mango fruit tree is a
highly cultivated crop that is economically important across
the majority of the world. In general, it is highly prized
because of its beneficial nutrient contentandmouthwatering
taste, and it plays an important part in the lives of millions of
farmers. However, this crop is susceptible to a variety of
illnesses, which can reduce the yield and overall quality of
the plants. The quick identification of these diseases in
mango plants is needed to prevent thespreadofdisease.The
original method for identifying mango leaf disease was
visual examination by agricultural professionals. Whichhas
time commitment and knowledge-dependent limitations.
Automatic leaf disease identification is accomplishedvia the
use of machine learning and computer vision disciplines.
The goal of this research is to identify mango leaf disease
using a pre-trainedtransferlearning-basedmachinelearning
system. Model interpretability must also be researched as
the number of convolution neural networks grows and the
framework black box interpretability problem becomes
more relevant. In order to make the models more
transparent and interpretable, explanatory artificial
intelligence (XAI) reveals its importance, which is
considered to be at the highest level of explainability,
accuracy, and performance [1]. Researchers are better able
to comprehend the reasoning behind the outcomes of deep
learning frameworks when the output of the LIME model is
visually represented.
2. LITERATURE SURVEY
Several academics offered numerous machinelearning(ML)
and deep learning (DL) methodologies for detecting various
diseases in plant leaves.
Adi Dwifana Saputra , Djarot Hindarto, Handri Santoso [2]
have presented rice leaves disease classification using the
Convolutional Neural Network algorithm with DenseNet
architecture. The accuracyofDenseNet211was91.67%,that
of DenseNet169 was 90%, and that of DenseNet201 was
88.33%. The training duration of the model was 24 seconds.
Authors in [3] have used PlantVillage and PlantDoc dataset
to identify leaf disease in the corn plant. EfficientNetB0
architecture was used in this research article. The
performance of the proposed architecture is compared with
Inception V3, VGG16, Resnet50, Resnet101, and
Densenet121. The proposed approach achieved an accuracy
of 98.85% and a precision of 88% and it is more
computationally efficient.
H. Amin et al. [4] have used two pre-trained CNN
architectures namely EfficientNetB0 and DenseNet121. The
Authors have applied feature fusion techniques features
extracted from two models. The proposed model achieved
98.56% accuracy which is the highest amongst Resnet152,
InceptionV3, and Densenet121.
Authors in [5] have performed classification of leaf disease
on different fruit leaves. The average accuracy of VGG,
GoogLeNet and ResNet are compared and ResNet have
shown best accuracy amongst all. Explainability testing is
done using GradCAM, LIME, and SmoothGrad techniques on
convolution-based neural networks.
3. DATASET DESCRIPTION
My dataset was gathered from Kaggle. The dataset includes
pictures of 32 different types of Indian mango leaves. The
collection includes 768 photos of 32 Indian mango leaf
species, with 24 photos of each species taken at various
orientations and angles. The dataset's sample images are
seen in Figure 1.

Fig -1: sample dataset diseased images
With a ratio of 70:20:10, I broke down the data into train,
validation, and test sets. My model has been trained and
validated using the training and validation set. To achieve
pixel normalization, I divide each image pixel by 255. In
order to make sure that my network, once trained, sees new
variants of data at each and every epoch, I applied on-the-fly
data augmentation over the training samples.
4. PROPOSED TRANSFER LEARNING APPROACH
DENSENET169
For image detection, one of the recent developments in
neural networks is the DenseNet architecture. ResNet and
DenseNet are fairly similar, yet there are several key
distinctions. While DenseNet concatenates (.) the output of
the previous layer with the future layer, ResNet employs an
additive approach (+) to combine the previous layer
(identity) with the future layer [2]. There are various
variants of DenseNet, including DenseNet-121, DenseNet-
169, and DenseNet-201. The numberofDensenetrepresents
the neural network's layer count. The convolution layer,
pooling layer, and fully linked layer are the three layers that
make up the DenseNet-169 Architecture. By applying the
convolution layer, pooling layer, batch normalization, and
nonlinear activation layer, an output fromthepreviouslayer
serves as an input for the second layer. DenseNet (Dense
Convolutional Network) has several advantages: they
alleviate thevanishing-gradient, problem,strengthenfeature
propagation, encourage feature reuse, and substantially
reduce the number of parameters [2].

Performance analysis of the proposed architecture is
presented in Figure 2. Line plots of the accuracy of train and
test sets and loss of train and test sets show that there is no
sign of underfitting or overfitting. Researchers now choose
to assess their models in conjunction withaccuracyusingthe
F1 score, which combines precision and recall using their
harmonic mean. As a result, I also computedtheF1scoreasa
performance metric. My model has a good separability as
seen by the AUC of 0.9984 in Figure 2.
Fig -2: Graph of Accuracy, AUC, Precision, F1-score
5. EXPLAINABLE DEEP LEARNING
FRAMEWORK(LIME)
5.1 Why LIME
Shapley and gradient-based explainability are two
alternative approaches to black box explanation. These
techniques explain individual input samples by taking the
partial derivatives of a model's output with regard to its
inputs [7]. Such gradient-basedtechniquesoffera significant
amount of output features, which contributes to the high-
dimensionality issue. In addition to being computationally
expensive, Shapley values produce a large number of output
features with complex explanations [8]. LIME is more
focused on providing individualized predicted explanations
than these two approaches, and it is also easier to
understand whilerequiringlesscomputingpowerandeffort.
5.2 LIME interpretation of DenseNet169
Explainable AI is a collection of procedures and techniques
designed to make judgments made by AI and machine
learning models intelligible to humans [6]. There are
numerous methods for interpreting the machine Learning
results. Here, Local Interpretable Model-agnostic
Explanations (LIME)—a tool for interpreting machine
learning models—have been utilized. This paper examines
the classification of mango leaf diseases with further
justifications based on Local Interpretable Model-agnostic
Explanations (LIME).
LIME's main concept is to divide the input image into
superpixels, which can be considered as new examples. We
generated these superpixels by applying the Quickshift
method [9]. Perturbations are produced by randomly
sampling the superpixels and the prediction for each
perturbation is computed. To ensure explainability,
weights—a measure of the cosine distance—are calculated
between each randomly generated perturbation and the
input image that is given to the LIME. The weighted linear
regression model is fitted with these weights,perturbations,
and predictions and the model yields interpretable
coefficients. Finally, LIME output givessuperpixels thathave
a greater influence on the label prediction.
Table 1 displays the output that LIME generated to explain
the predictions made by the Machine Learning Model.
Images in the first column are the original image with the
disease in Mango leaves. The features that enable the
DenseNet model to predict diseased mango leaves are
visualized in the second column of Table 1. These features
contribute positive results to the prediction. Wecanobserve
from the prediction interpretations that the model mostly
focuses on the leaf regions that are impacted when making
predictions.

Table -1: LIME interpretation of DenseNet169 on diseased Mango leaves
Original Image LIME output

6. CONCLUSION
In this paper, I implemented a pre-trained Deep learning
model DenseNet169 for Mango leaf Diseasedetection. Ihave
evaluated the model based on Accuracy, Precision, AUC, and
F1 score. The model had the best performance, with a
97.41% accuracy rate. In the future to improve the
generalization of my model, I will try to increase my dataset
to cover a larger spectrum of plant leaf diseases.
REFERENCES
[1] Daglarli, Evren, “Explainable Artificial Intelligence(xAI)
Approaches and Deep Meta-Learning Models for Cyber-
Physical Systems”, Artificial Intelligence Paradigms for
Smart Cyber-Physical Systems, edited by Ashish Kumar
Luhach and Atilla Elçi, IGI Global, 2021, pp. 42-67.
[2] Adi Dwifana Saputra , Djarot Hindarto, Handri Santoso,
“Disease Classification on Rice Leaves using
DenseNet121, DenseNet169, DenseNet201”, Sinkron :
Jurnal dan Penelitian TeknikInformatika Volume8,Issue
1, January 2023, DOI :
https://doi.org/10.33395/sinkron.v8i1.11906
[3] Fathimathul Rajeena, Aswathy S, Mohamed A. Moustafa
and Mona A. S. Ali, “Detecting Plant Disease in Corn Leaf
Using EfficientNet Architecture—An Analytical
Approach”, Electronics 2023, https://doi.org/10.3390/
electronics12081938
[4] HASSAN AMIN, ASHRAF DARWISH (Member, IEEE),
ABOUL ELLA HASSANIEN AND MONA SOLIMAN. “End-
to-End Deep Learning Model for Corn Leaf Disease
Classification”, IEEE Access, “, Volume 10, 2022, Digital
Object Identifier 10.1109/ACCESS.2022.3159678
[5] Kaihua Wei, Bojian Chen, JingchengZhang,ShanhuiFan,
Kaihua Wu, Guangyu Liu and Dongmei Chen,
“Explainable Deep Learning Study for Leaf Disease
Classification”, Agronomy 2022, 12, 1035,
https://doi.org/ 10.3390/agronomy12051035
[6] E. Tjoa and C. Guan, “A Survey on Explainable Artificial
Intelligence (XAI): Toward Medical XAI,” IEEE Trans.
Neural Networks Learn. Syst., vol. 32, no. 11, pp. 4793–
4813, 2021
[7] D. Baehrens, T. Schroeter, S. Harmeling,M.Kawanabe, K.
Hansen, and K. R. Müller, “How to explain individual
classification decisions,” J. Mach. Learn. Res., vol. 11, pp.
1803–1831, 2010
[8] S. M. Lundberg and S. I. Lee, “A unified approach to
interpreting model predictions,” Adv. NeuralInf. Process.
Syst., vol. 2017-Decem, no. Section 2, pp. 4766–4775,
2017
[9] A. Vedaldi and S. Soatto, “Quick Shift and Kernel
Methods for Mode Seeking”, Lecture Notes in Computer
Science, pp. 705–718, 2008,
https://link.springer.com/chapter/10.1007%2F978-3-
540-88693-8_52.
BIOGRAPHY
Vishakha is working as a HOD of the
Department of IT at 360 Research
Foundation, she establishes a research
agenda in AI/ML, helps researchers inthe
field of empowerment and livelihood,and
guides research work to engineering and
computer science students. She earned
her M.Tech. from (NIT), Surat, Gujarat.
She is an experienced Assistant Professor
with a variety of engineering educational institutesincluding
NIT-Surat and R V College of Engineering, Bangalore. She is
passionate about AI/ML, speechandaudiosignal processing,
and Digital Signal Processing.

Explainable AI(XAI) using LIME and Disease Detection in Mango Leaf by Transfer Learning Approach

More Related Content

Similar to Explainable AI(XAI) using LIME and Disease Detection in Mango Leaf by Transfer Learning Approach

More from IRJET Journal

Recently uploaded

Explainable AI(XAI) using LIME and Disease Detection in Mango Leaf by Transfer Learning Approach