(IJACSA) International Journal of Advanced Computer Science and Applications,

Vol. 12, No. 10, 2021

Using Transfer Learning for Nutrient Deficiency

Prediction and Classification in Tomato Plant
Vrunda Kusanur1
Dr. Veena S Chakravarthi2
Research Scholar, Visvesvaraya Technological University
Professor, Department of Electronics and Communication
Bengaluru, Karnataka, India
Engineering, BNMIT, Bengaluru, Karnataka, India

Abstract—Plants need nutrients to develop normally. The the key objective of this work is to automate the identification
essential nutrients like carbon, oxygen, and hydrogen are of nutrient deficiencies in plants using Convolutional Neural
obtained from sunlight, air, and water to prepare food and plant Networks (CNN).
growth. For healthy growth, plants also need macronutrients
such as Potassium, Calcium, Nitrogen, Sulphur, Magnesium, and Artificial Intelligence has numerous applications in
Phosphorus in relatively great quantities. When a plant doesn’t multiple industries, healthcare, environment, finance,
find necessary nutrients for its growth inadequate amount, education, agriculture, etc. to solve complex problems and
deficiency of plant nutrients occur. Plants exhibit various make our daily life more secure and fast.
symptoms to indicate the deficiency. Automatic identification and
differentiation of these deficiencies are very important in the G. Madhulatha et al. [2] proposed an automatic plant
greenhouse environment. Deep Neural Networks are extremely disease detection on the plant leaves to decrease crop loss and
efficient in image categorization problems. In this work, we used increase productivity. Plant diseases are predicted and
the part of the pre-trained deep learning model i.e. Transfer classified with 96.50% accuracy based on visual symptoms
Learning model to detect the nutrient stress in the plant. We using deep CNN. The authors used a dataset from the “Plant-
compared three different architectures including Inception-V3, Village” dataset for plant leaf diseases. The model was pre-
ResNet50, and VGG16 with two classifiers: RF and SVM to trained using AlexNet. Muhammad Hammad Saleem et al. [3]
improve, classification accuracy. A total of 880 images of developed three Deep Learning meta-architectures namely;
Calcium and Magnesium deficiencies in the Tomato plant from Faster Region-based Convolutional Neural Network (RCNN),
the greenhouse were collected to form a dataset. For training, Single Shot MultiBoX Detector (SSD), and Region-based
704(80%) images are used and for testing, 176(20%) images are Fully Convolutional Networks (RFCN) to recognize plant
used to examine the model performance. Experimental results disease and healthy leaves. All three models include a feature
demonstrated that the largest accuracy of 99.14% has resulted extractor and a base network. This research used Gradient
for the VGG16 model with SVM classifier and 98.71% for Descent with its Momentum version, Adaptive Moment
Inception-V3 with Random Forest Classifier. For a batch size of Estimation (Adam), and Root Mean Square Propagation
8 and epochs equal to 10, the Inception -V3 architecture attained (RMSProp) optimization algorithms to increase the
the highest validation accuracy of 99.99% and the least
performance of the Deep Learning meta-architectures. The
validation loss of 0.0000384 on an average.
authors examined that all the Deep Learning meta-architectures
Keywords—Nutrient deficiency; plant nutrients; deep neural needed 126 epochs (200,000 iterations) for training
networks; transfer learning; random forest (RF); support vector convergence. When the SSD model was trained using Adam
machine (SVM) optimizer, the maximum means Average Precision (mAP) of
73.07% was obtained. Guan Wang et al. [4] suggested a deep
I. INTRODUCTION learning model for control plant disease application. The
A proper combination of nutrients is required for plants to authors used the apple leaf black rot images produced by the
live, develop and reproduce. So, plant analysis is a necessary fungus Botryospaeria obtuse from the PlantVillage dataset for
tool that assists farmers by providing significant information disease severity classification. The highest overall accuracy of
about the nutritional description of the growing plant to obtain 90.4% was obtained for the VGG16 model. Sharada P.
a better yield. Generally, plant analysis indicates the analysis of Mohanty et al. [5] established a smartphone-assisted
magnesium (Mg), sulphur (S), phosphorous (P), calcium (Ca), application to detect the disease using a deep convolutional
nitrogen (N), potassium (K), etc. Normally, plants show signs neural network. In this research, GoogLeNet architecture
of being unhealthy when they suffer from undernutrition. For performs better and provides 99.35% accuracy as compared to
example, yellow around the edges of its leaves are a sign of AlexNet architecture. The presently available deep learning
magnesium deficiency. Yellow spots between the leaf veins methods to identify the plant disease were reviewed by M.
and Blossom End Root denote the absence of calcium. Brown Nagaraju and Priyanka Chawla [6].
edges along the plant leaves indicate a deficiency of potassium. Many previous works have considered Image Recognition
Yellow or pale green leaves imply the need for nitrogen [1]. and Machine Learning models to classify the images into
These nutrient deficiency symptoms will help growers to healthy and unhealthy images. However, most of these
identify the nutrient status of plants for a better crop yield. algorithms require image segmentation and feature extraction.
Manually diagnosing these deficiencies is a difficult task. So, But, from the many extracted features, it is difficult to judge

784 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 12, No. 10, 2021

the important and dominant features for plant disease detection. pixels is the size of all the resized images. These sample
Moreover, under difficult background circumstances, many images are input to the convolutional neural network for
techniques fail to successfully segment the leaf and will lead to training the model. The trained model is applied for the class
unreliable deficiency recognition. So, image segmentation and prediction of unseen images. These phases are explained in
feature extraction are still challenging tasks. Therefore, detail in the following sections. Machine learning algorithms
automatic plant disease detection and nutrient deficiency including SVM, Decision-Tree and, RF are excellent in
recognition are still challenging tasks. Recently, Convolutional resolving classification problems [10]. However, they go
Neural Network (CNN) is becoming the preferred scheme to wrong in extracting the proper features from the image.
overcome few challenges. Alternatively, Convolutional Neural Networks receives the raw
pixel of the images directly as inputs instead of extracting
The main objective of this research is to diagnose nutrient certain features manually [12-14]. CNN learns how to take out
deficiency in plants and take several measures like adjusting these features from the actual image.
the pH value of water to achieve a quality yield, providing the
right amount of fertilizer, etc. using deep learning models. For B. Convolutional Neural Networks
nutrient deficiency classification, we employed the Transfer CNN's are a class of Deep Neural Networks that can
Learning method, where pre-trained models are used as the identify and categorize specific features in images and are
entry point to develop the neural network models. In this generally used for examining visual images. Significantly,
research, we have used these models to predict Calcium (Ca) CNN can yield good results than the traditional feature
and Magnesium (Mg) deficiency in tomato crops grown under extraction algorithms in plant disease diagnosis [15-18]. In
a greenhouse environment. CNN, the filters are learnable. A classic CNN consists of two
The key advantage of transfer learning is that instead of components: The Convolution Block and the Fully Connected
beginning the learning process from the scratch, the model block, which are detailed as follows.
commences from the characteristics that have been educated
when resolving other problems which are analogous to the one TABLE I. CALCIUM AND MAGNESIUM NUTRIENT DEFICIENCY
SYMPTOMS IN TOMATO
being resolved. We have used three pre-trained models-
InceptionV3, VGG16, and ResNet50 as a base model and Deficiency Symptoms
Nutrients Description
SVM or Random Forest classifier on top of it to attain better on Leaf/ Fruit
results.
Young leaves curl inwards and
The rest of this paper is structured in the following fashion. cause dry decaying areas at
Calcium (Ca)
Section II introduces the images collected to form the dataset Blossom End of the fruit (BER)
of Ca and Mg deficiencies followed by related concepts. This in Tomatoes.
section also presents Inception V3, ResNet50, and VGG16
architectures, and the proposed model to identify and classify Interveinal chlorosis (Veins of
the deficiencies. Section III dedicated to the evaluation, and the Leaf remains green whereas the
comparative analysis of results obtained in this experiment. In Magnesium areas between the leaves go
Section IV, the paper is summarized and future work is (Mg) yellow). Elder leaves drip their
mentioned. color excluding in the veins. It
does not affect the fruit.
II. MATERIALS AND METHODS
C. Convolutional Neural Networks
A. Data Acquisition
CNN's are a class of Deep Neural Networks that can
Tomato plants were grown in a greenhouse of a size 10x4 identify and categorize specific features in images and are
sq.ft. to study and gather the dataset for lack of nutrients in generally used for examining visual images. Significantly,
tomato leaves and fruits. The calcium and magnesium CNN can yield good results than the traditional feature
deficiencies were induced for the plants in different stages and extraction algorithms in plant disease diagnosis [9, 11]. In
their images were captured from the camera for training and CNN, the filters are learnable. A classic CNN consists of two
testing the performance of the model. The dataset was components: The Convolution Block and the Fully Connected
developed with two classes for classification and prediction: block, which are detailed as follows.
Calcium and Magnesium. Altogether, there are 880 images in
the dataset. Out of 880 images, 704 (80%) images are for 1) Convolution block: The convolution block contains the
training the model, and 176 (20%) images are for testing the Convolution Layer and the Pooling Layer. In this block, the
model. There are 374 calcium deficiency and 330 magnesium task of feature extraction is accomplished. The convolutional
deficiency images in the training dataset. Further, out of 176 layer produces the feature maps or activation maps by
testing images, 94 images are of calcium, and the remaining 82
applying filters to input images using the ReLU activation
images are of magnesium deficiency images. To enhance the
dataset, the data augmentation methods including image function. The ReLU function returns x for all the values of x >
resizing, flipping, random rotation, shearing, etc., are applied. 0, and returns 0 for all values of x ≤ 0 and is given in
The details of calcium and magnesium nutrient deficiency equation 1.
symptoms in tomatoes are presented in Table I. 256 x 256 F(x) = max(0,x) (1)

785 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 12, No. 10, 2021

Fig. 1. VGG16 Architecture.

The convolutional layer uses filters kernels to recognize

various features like edges, horizontal lines, vertical lines, etc., [ ] (2)
in an image. To extract more composite and thoughtful
features, the same size convolution kernel is used again and Where, y0^ represents the probability with which class 0
again multiple times. The pooling layer is enforced next to a (Ca) is predicted and y1^ represents the probability with which
convolution layer in which a down sampling operation is class 1(Mg) is predicted. The RGB image of constant size
performed on a convolved feature to scale down the number of 224x224 is the input to the conv1 layer. The image is moved
dimensions of the feature map. Commonly, the average and through several convolutional layers. Each layer uses a small
maximum values are selected by the pooling layer for this task. 3x3 or 1x1 filter. Five max-pooling layers perform spatial
pooling. A 2x2 pixel window with a stride of 2 is used to
2) Fully connected block: The Fully Connected block implement max-pooling. There are three Fully-Connected (FC)
comprises of fully connected simple neural network design layers where there are 4096 channels in each of the first two
which does classification depending on convolutional block layers and the third layer comprises 1000 channels. The
inputs. Convolutional Neural Network has one or more fully softmax layer is the terminating layer. All networks have a
connected layers at the end of it. At the end of the fully similar configuration of the fully connected layers. A non-
linear ReLu activation function is used by all hidden layers.
connected layer, there is a softmax activation function whose
output is a probability (from 0 to 1) for every classification E. Inception-V3 Model Architecture
label. Inception-V3 is the most generally used CNN architecture
D. VGG16 Model Architecture and achieved more than 78.1% accuracy for image prediction
on the ImageNet dataset. The model comprises Convolution
VGG is a pre-trained model and has 138 Million Layers, Max pooling Layers, Average pooling Layers, Concate
parameters. VGG is trained over 14 million images belonging Layers, Dropout Layers, and Fully Connected Layers. In
to 1000 classes and learned to detect generic features from Inception V3, the resized images are of size 299x299x3 pixels.
images. There are 16 and 19 weight layers in the network for The structure of Inception-V3 is analogous to Inception-V2
VGG-16 and VGG-19 respectively. with few modifications including Label Smoothing
This research work uses VGG-16 as the base model and Regularization, Batch normalization, Auxiliary Classifier. Use
altered it to create a different network. As VGG16 attains of Factorized 7x7 convolutions. Inception-V3 is a CNN with
92.7% test accuracy in ImageNet, and because of its high 48 layers in depth. The inception model is a concatenation of
performance, the pre-trained weights are retained and only the parallel convolution layers with 1x1, 3x3, 5x5, etc. sized filters
top three Fully Connected Layers or Dense Layers are and a max pooling layers of 3x3 matrix. The error rate
modified to fine-tune the neural network. In this work, the improved to 0.2 % by adding label smoothing in Inception-
features extracted from VGG16 are given as input to RF or V3 architecture. Fig. 2 describes the Inception-V3 scheme.
SVM Classifiers to reduce the training time and increase the Inception-V3 model used RMSProp optimizer that offered
classification accuracy. Fig. 1 describes the VGG16 scheme. significant results in connection with accuracy and time to
All the resized images in this model are of fixed size 244x244. achieve it. RMSProp is fixed as a default optimizer. The update
The VGG16 model used NVIDIA Titan Black GPUs and dynamics in the Inception-V3 model are given by equation 2
was trained for weeks. The VGG16 model can categorise the and equation 3.
images into 1000 classes. The VGG model handles the input ( ) (3)
image and yields the vector of 2 values. y^ denotes the
probability of classification for the corresponding class and is ( ) (4)
given by equation 2. √

786 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 12, No. 10, 2021

Fig. 2. Inception V3 Architecture.

Where, gk is gradient descent at time k, gk+1 is gradient

descent at time k+1, wk is the weight at time k, wk+1 is the
updated weight at the time k+1, „α‟ is the step size, „β‟ is
known as momentum and „ϵ‟ is small positive constant to avoid
division by zero in implementation, „ is the gradient, which
is taken of f, ‘η’ is learning rate.
Fig. 4. Proposed Model.
In RMSProp, the parameters „α‟, „β‟, and „ϵ‟ are set as
decay α = 0.9, momentum β = 0.9, and ϵ = 1.0
III. EXPERIMENTAL RESULTS AND ANALYSIS
F. ResNet50 Model Architecture In this work, image pre-processing techniques, data
ResNet (Residual Network) is presented by Kaiming He, augmentation, and implementation of Convolutional Neural
Xiangyu Zhang, Shaoqing Ren, and Jian Sum in 2015 in their Network algorithms were conducted using Jupyter
paper “Deep Residual Learning for Image Recognition”. The notebook(Python 3.9), Keras API, OpenCV library, Matplotlib
development of ResNet improved the problem of training deep visualization library, OS module, glob module, and so on. The
neural networks. The simple element in ResNet is as depicted hardware specifications in this experiment to train and test our
in Fig. 3. In the Residual network, there is a straight model includes Intel(R) Core (TM) i7-4210U CPU, 4.00 GB
connection called „skip connection‟ which skips some in RAM. In this experiment, the CNN is developed using
between layers. InceptionV3, ResNet50, and VGG16 Transfer Learning
Models.
The „skip connection‟ is used to resolve the vanishing
gradient problem and to learn the identity functions. The All three models used pre-trained weights from the
output H(X) with the introduction 4of skip connection is given ImageNet dataset by eliminating the upper layer and redefining
by the equation H(X) = F(X) + X. Table III shows the elements a fresh fully connected Softmax layer with 2 classes for
of the ResNet50 model. The ResNet model was tested on the classification [7, 8]. In this experiment, the batch size was
ImageNet set and attained a 20.47% top-1 error rate also 5.25% fixed to 8 and the number of epochs was set to 10 with Adam
top-5 error rate. optimizer. The features extracted from the Transfer Learning
technique were used by SVM and Random-Forest classifiers.
The proposed model used these transfer learning techniques 80% of the total images were used to form a training dataset to
for feature extraction and altered their basic structures by train the model and 20% were used to form a testing dataset.
adding Random Forest or SVM classifiers to improve the For the Inception-V3 model, all the images were resized to
classification ability of the models as illustrated in Fig. 4. 299x299x3, the input image size for ResNet-50 and VGG16
was 224x224. Inception V3 attained the validation accuracy of
99.99 % and the validation loss of 0.0000384 as depicted in
Table II out of the three models.
The accuracy and loss obtained from three different
Transfer Learning models are presented in Fig. 5 to 7.

Fig. 3. A Basic Residual Network.

TABLE II. ACCURACY AND LOSS OF DIFFERENT TRANSFER LEARNING MODELS AFTER 5 AND 10 EPOCHS
Transfer Learning
Training Accuracy (%) Validation Accuracy (%) Training Loss Validation Loss
Model
5 Epochs 10 Epochs 5 Epochs 10 Epochs 5 Epochs 10 Epochs 5 Epochs 10 Epochs
Inception V3 99.23 100 98 99.99 0.1364 0.000264 0.3685 0.0000384
ResNet-50 89.40 82.28 87.00 82.10 0.2632 0.5541 0.2568 0.4783
VGG16 100 100 98.86 98.86 0.000001 0 0.3572 0.2220

787 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 12, No. 10, 2021

Fig. 5. Inception V3: Accuracy and Loss Model.

Fig. 6. ResNet50: Accuracy and Loss Model.

Fig. 9. Confusion Matrix using ResNet50.

Fig. 7. VGG16: Accuracy and Loss Model.

Fig. 8 to 10 represents the confusion matrix of plant

nutrient deficiencies identification using three Transfer
Learning models with RF and SVM classifiers. The
classification accuracy chart of plant nutrient deficiency
identification in tomato plants using Transfer Learning models
with Random Forest and SVM classifiers is presented in
Fig. 11. It is noticed from the chart that the largest accuracy of
99.14% has resulted using the VGG16 model with SVM
classifier and 98.71% for Inception-V3 with Random Forest
Classifier.

Fig. 10. Confusion Matrix using VGG16.

The results of lack of nutrients predicted from three

different models with RF and SVM classifiers on few samples
are displayed in Fig. 12. From Table II, it can be observed that
almost calcium and magnesium deficiencies were detected
properly by all three Transfer Learning models with RF and
SVM classifiers. The average classification accuracy is high
for InceptionV3 and VGG16 models in various experiments.
These models could be extended for the identification of other
nutrient deficiencies.
Fig. 8. Confusion Matrix using Inception V3.

788 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 12, No. 10, 2021

SVM classifier and 98.71% for InceptionV3 with Random

Accuracy Forest Classifier.
105.00% To control these plant nutrient deficiencies, the tomato
97.85% 98.71% 99.14%
100.00% 95.71% greenhouse environmental factors such as humidity,
95.00% temperature, pH, and soil moisture need to be monitored to
88.84% find out the right quantity of fertilizer to be applied. Hence, in
90.00% the future, this work can be improved by monitoring the
84.12%
85.00% greenhouse parameters by Wireless Sensor Network (WSN) to
80.00% apply fertilizer precisely in a greenhouse.
75.00% ACKNOWLEDGMENT
We extend our heartfelt gratitude to the management of the
BNM Institute of Technology, Bengaluru for providing us with
all the necessary sources to accomplish this work and all the
support to do the subsequent publications. We are also grateful
to Visvesvaraya Technological University for giving us an
appropriate platform to complete this research.
InceptionV3 with RF InceptionV3 with SVM
REFERENCES
ResNet-50 with RF ResNet-50 with SVM
[1] Gaganjot Kaur,”Automated Nutrient Deficiency Detection In Plants: A
VGG16 with RF VGG16 with SVM Review”, Palarch‟s Journal of Archaeology Of Egypt/Egyptology, vol.
17, no. 6, pp. 5894-5901.
Fig. 11. Comparison Efficiency for Transfer Learning Models with RF and [2] G. Madhulatha, O. Ramadevi, “Recognition of Plant Diseases using
SVM Classifiers. Convolutional Neural Network”, International Conference on IoT in
Social, Mobile, Analytics and Cloud (I-SMAC), 2020, ISBN: 978-1-
7281-5464-0, DOI: 10.1109/I-SMAC49090.2020.9243422.
[3] Muhammad Hammad Saleem, Sapna Khanchi, Johan Potgieter and
Khalid Mahmood Arif, “Image-Based Plant Disease Identification by
Deep Learning Meta-Architectures, Plants 2020,9,1451, MDPI
Publication, DOI: 10.3390/plants9111451.
[4] Guan Wang, Yu Sun, and Jianxin Wang, “ Automatic Image-Based
Plant Disease Severity Estimation Using Deep Learning”,
Computational Intelligence and Neuroscience, Volume 2017, Article ID
2917536, https://doi.org/10.1155/2017/2917536.
[5] Sharada P. Mohanty, David P. Hughes, and Marcel Salathe, “Using
Deep Learning for Image Based Plant Disease Detection”, Frontiers in
Plant Science, 7, 1419, 2016.
[6] M, Nagaraju, Priyanka Chawla, “ Systematic review of deep learning
techniques in plant disease detection”, Int J Syst Assur Engg Manag.
11(3), 547-560, 2020.
[7] Nafees Akhter Farooqui and Ritika, “An Identification and Detection
Process for Leaves Disease of Wheat Using Advance Machine Learning
Fig. 12. The Predicted Samples of Plant Nutrient Deficiency Images. Techniques”, Bioscience Biotech Research Communication, vol. 12,
no.4, pp. 1081-1091, 2019, DOI: 10.21786/bbrc/12.4/31.
IV. CONCLUSION [8] Zhe Xu, Xi Guo, Anfan Zhu, Xiaolin He, Xiaomin Zhao, Yi Han, and
Roshan Subedi, “Using Deep Convolutional Neural Networks for
Quick identification of plant nutrient deficiency is Image- Based Diagnosis of Nutrient Deficiencies in Rice,
necessary for a greenhouse environment. Manual inspection of Computational Intelligence and Neuroscience, Hindawi Publications,
these deficiency symptoms in a large greenhouse requires more vol. 2020, 12 pages, DOI: 10.1155/2020/7307252.
effort. Consequently, automated plant nutrient deficiency [9] Shima Ramesh, Mr. Ramachandra Hebbar, Niveditha M, Pooja R,
diagnosis is required in greenhouse technology. With Prasad N, Shashank N, Mr. P V Vinod, “Plant Disease Detection Using
technology growth, a CNN using Transfer Learning models Machine Learning”, International Conference on Design Innnovations
for 3Cs Compute, 2018, ISBN: 978-1-5386-7523-6, DOI:
such as Inception V3, ResNet50, and VGG16 were proposed 10.1109/ICD13C 2018.00017.
along with Random Forest (RF) and SVM classifiers to [10] Jayme Garcia Anal Barbedo, “Detection of nutrition deficiencies in
improve the efficiency. These models are pre-trained on plants using proximal images and machine learning: A review”,
ImageNet dataset and are modified for our tomato dataset with Computers and Electronics in Agriculture, Elsevier Publications, vol.
images of calcium and magnesium deficiencies in this research. 162, pp. 482-492,2019, DOI:
On average, out of all the three Transfer Learning techniques, https://doi.org/10.1016/j.compag.2019.04.035.
Inception V3 attained the highest validation accuracy of 99.99 [11] Aravind Krishnaswamy Rangarajan, Raja Purushothaman, Anirudh
Ramesh, “Tomato crop disease classification using pre-trained deep
% and the least validation loss of 0.0000384 for 10 epochs. learning algorithms”, International Conference on Robotics and Smart
Further, when the experiment was conducted for Random Manufacturing (RoSMa2018),2018, pp. 1040-1047, DOI:
Forest (RF) and SVM classifiers, results show that the largest 10.1016/j.procs.2018.07.070.
accuracy of 99.14% has resulted using the VGG16 model with [12] Yan Guo, Jin Zhang, Chengxin Yin, Xiaonan Hu, Yo Zou, Zhipeng Xue,
Wei Wang, “Plant Disease Identification Based on Deep Learning

789 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 12, No. 10, 2021

Algorithm in Smart Farming”, Discrete Dynamics in Nature and [16] G.L. Grinblat, L.C. Uzal, M.G. Larese and P. M. Granitto “Deep
Society, vol.2020, no.7, pp.1-11, 2020. DOI: 10.1155/2020/2479172. learning for plant identification using vein morphological patterns”,
[13] Lili Ayu Wulandhari, Alexander Agung Santoso Gunawan, Aie Qurania, Computers and Electronics in Agriculture, vol.127, pp. 418-424, 2016.
Prihastuti Harsani, Triastinurmiatingsih, Ferdy Tarawan, and Riska [17] K. P. Ferentinos, “Deep learning models for plant disease detection and
Fauzia Hermawan, “Plant Nutrient Deficiency Detection Using Deep diagnosis”, Computers and Electronics in Agriculture, vol. 145, pp.311-
Convolutional Neural Network”, ICIC International Conference, ICIC 318, 2018.
Express Letters, pp. 971-977, vol. 13, no. 10, 2019. [18] D. Story, M. Kacira, C. Kubota, A. Akoglu and L. An, Lettuce calcium
[14] G.Chu, “How to use transfer learning and fine-tuning in keras and deficiency detection with machine vision computed plant features in
tensorflow to build an image recognition system and classify (almost) controlled environments, Computers and Electronics in Agriculture, vol.
any object”, Deep Learning Sandbox, 2017. 74, no.2, pp. 238-243, 2010.
[15] A. M. G. J. Hanson, A. Joy, and J. Francis, “Plant leaf disease detection
using deep learning and convolutional neural network”, International
Journal of Engineering Science, vol. 7, no.3, pp.5324-5328, 2017.

790 | P a g e
www.ijacsa.thesai.org