See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/357447801

Recognition of Image-Based Plant Leaf Diseases Using Deep Learning

Classification Models

Article  in  Nature Environment and Pollution Technology · December 2021

DOI: 10.46488/NEPT.2021.v20i05.031

CITATIONS READS

12 495

7 authors, including:

Sakshi Takkar Veerpal Kaur

Lovely Professional University Lovely Professional University
11 PUBLICATIONS   48 CITATIONS    15 PUBLICATIONS   52 CITATIONS   

SEE PROFILE SEE PROFILE

Manik Rakhra Manish Sharma

Lovely Professional University Himgiri Zee University
86 PUBLICATIONS   736 CITATIONS    37 PUBLICATIONS   974 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Medicaap: An online cancer detection and Analysis View project

Thesis View project

All content following this page was uploaded by Manish Sharma on 31 December 2021.

The user has requested enhancement of the downloaded file.

 p-ISSN: 0972-6268 No. 5
Nature Environment and Pollution Technology (Print copies up to 2016)
2021
An International Quarterly Scientific Journal Vol. 20 (Suppl) pp. 2137-2147
e-ISSN: 2395-3454

Original Research Paper https://doi.org/10.46488/NEPT.2021.v20i05.031

Original Research Paper Open Access Journal

Recognition of Image-Based Plant Leaf Diseases Using Deep Learning

Classification Models
Sakshi Takkar*, Anuj Kakran**, Veerpal Kaur*, Manik Rakhra*†, Manish Sharma***, Pargin Bangotra***
and Neha Verma*****
*School of Computer Science and Engineering, Lovely Professional University, Phagwara, Punjab, India
** Quantum University, Roorkee, Uttrakhand, India
***Department of Physics, School of Basic Sciences, Bahra University, Shimla Hills, Shimla, Himachal Pradesh, India
**** Department of Physics, SBS&R, Sharda University, Greater Noida, India
*****KRM DAV Nakodar, Punjab, India
†Corresponding author: Manik Rakhra; [email protected]

ABSTRACT
Nat. Env. & Poll. Tech.
Website: www.neptjournal.com
Plant diseases are spread by a variety of pests, weeds, and pathogens and may have a devastating
Received: 03-06-2021 effect on agriculture, if not handled in a timely manner. Farmers face umpteen challenges from a
Revised: 03-07-2021 proper water supply, untimely rain, storage facilities, and several plant diseases. Crops disease is the
Accepted: 13-07-2021 primary threat and it causes enormous loss to farmers in terms of production and finance. Identifying
the disease from several hectares of agricultural land is a very difficult practice even with the presence
Key Words: of modern technology. Accurate and rapid illness prediction for early illness treatment to crops
Plant diseases minimizes economical loss to the individual and further proves to be productive for healthy crops. Many
Convolutional neural network studies use modern deep learning approaches to improve the accuracy and performance of object
Support vector machine detection and identification systems. The suggested method notifies farmers of different agricultural
Artificial neural network illnesses, prompting them to take further essential precautions before the disease spreads to the whole
agricultural field. The primary objective of this study is to detect the illnesses as soon as they begin
to spread on the leaves of the plants. Super-Resolution Convolutional Neural Network (SRCNN) and
Bicubic models are employed in the system to identify healthy and diseased leaves with an accuracy
of 99.175 % and 99.156 % respectively.

INTRODUCTION productive results. Technical and proper research-based

methods are required to monitor the crops regularly. Crop
Agriculture is one of the most important economic activities disease is one of the reasons for decreasing the quantity and
of the Indian subcontinent and two-third population is direct- quality of crops in the fields. The use of toxic pathogens,
ly involved in farming and related occupations. Agriculture extreme change in climate conditions, and poor disease
has long been considered India’s backbone, dating back control are one of the factors of poor food production. Large
to the Indus Valley civilization. To earn income, mankind numbers of pesticides are available to control the diseases
established their residence land according to agricultural of crops in agriculture and to increase the production of
facilities and favorable conditions. Agriculture is important crops in the fields. Identifying the current disease of the
in most developing countries because it provides jobs and crops and finding the appropriate pesticide to control the
contributes a significant portion to GDP (Pradhan 2007). diseases is a task that requires the advice of experts in the
Bacterial growth and diseases are a primary threat for the field of agriculture that makes this task very consuming and
crops and affect the agricultural cycles and patterns too. To expensive. Accurate and timely identification of diseases of
overcome this problem, a variety of pesticides, fertilizers, crops is one of the reasons for successful agriculture. It is also
and research-based remedies are used nowadays. Agriculture very important to spend less time and money to identify the
is getting attention in every five-year plan and preference is hazardous diseases on crops. These diseases on plants and
given for the development of agriculture in India. crops can be pre-identified with some kind of initial tracks
The agricultural field needs more up-gradation due to and spots developed on leaves and fruits. Many farmers use
changes in weather and economic conditions in the country. their knowledge or seek assistance from other professionals
The crops in the fields should be healthy to get more to spot crop diseases with their naked eyes. Due to the
2138 Sakshi Takkar et al.

similarities in symptoms of crops, this method raised the these spot results in unmarketable fruits and vegetables.
possibility of human error and faulty illness diagnosis. This 4. Black Measles: Black measles occur in grapes and
type of disease diagnosis mistake leads to the overuse of is also known as grapevine measles, esca, or Spanish
pesticides and fertilizers that contains heavy metals, which measles. The term measles refers to artificial spots that
reduces crop yield and even pollutes the environment through appear on the grapes. The symptom on the leaves appear
deposition in various areas, resulting in radiological and as a ‘Tiger Stripe’ pattern and it becomes more serious
chemical exposures to humans, flora, and fauna (Bangotra from year to year. During the season, the spot merges
et al. 2019, 2021, Pandit et al. 2020, Mehra et al. 2015). over the surface of the grapes and makes the grapes
Nowadays, server-based and mobile-based technologies black. Spots on the berries can appear any time after
are used to identify the diseases of crops accurately (Sladoje- the fruit set and before some days of the harvesting.
vic et al. 2016, Huang et al. 2014). Modern approaches like 5. Cedar Apple Rust: Cedar apple rust is a fungal disease
Machine Learning and Deep Neural networks are used to that occurs in apples. The infected leaves show yellow
increase the accuracy of results in finding diseases of crops. to orange round spots on the upper surface. As soon as
There is a need for a Machine Learning Vision system to the infection grows, the spots also appear on the lower
identify the disease from the image of crop and to suggest surface of the leaves. This disease can affect the stems
the pesticide as a solution to control the disease. Various and fruits. When the disease of the fruits grows, then
researchers have conducted studies in this field by using the lesions on the fruits may crack and appear brown
traditional machine learning algorithms like Support Vector in color.
Machine (SVM), Artificial Neural Network (ANN), Random
Forest algorithm, K-Means method, and Convolutional 6. Citrus Greening: Citrus greening is one of the danger-
Neural Network. ous plant diseases in the world. Once a plant is infected
due to this disease then there is no cure. This disease
PLANT DISEASES is caused by one disease-infected insect- Asian Citrus
Psyllid. This disease is mostly found in oranges. This
1. Apple Scab: Apple scab is a serious disease of apple disease shows the symptoms like yellowing of leaves,
that attacks both fruits and leaves. Olive green spots dieback of twigs, and decline in vigor which leads to
or pale yellow spots appear on the upper surface of death of the entire plant.
the leaves. Dark and velvet spots appear on the lower Common Rust: Common rust is one of the serious fungal
surface of the leaves. Apple with this scab disease is diseases which attack the roses, corns, and tomatoes,
not fit for eating. This disease reduces the quality and etc. This disease occurs mostly in mild and moist con-
size of fruits. Apple scab can cause total failure of crops ditions. Rust is actually spread by spores from infected
without control measures. to healthy plants. The spores are generally transferred by
2. Black Rot: Black rot is a disease that is caused by bac- wind or water. This is the reason rust appears often after
teria that can infect crops. It is very difficult to control watering. Yellow or white spots appear as symptoms on
this disease by the growers. Generally, the loss of crops the upper part of leaves. This results in leaf distortion
happens in hot and humid weather conditions. These and deformation.
diseases are generally found in apples and Grapes. 7. Early Bight: Early blight disease is very common in
Disease symptoms appear as yellow and dead tissues potatoes and tomatoes which are caused by the fungus
at the edge of the leaves in older plants. The spot on name Alternaria Solani. Firstly, its symptoms appeared
leaves get larger and infect other plants and fruit bunch on old leaves as small brown spots with a pattern of
very rapidly. Bull’s eye. When it spreads than its color changes to
yellow. After the stem, fruit, and upper portion of the
3. Bacterial Spot: Bacterial spot is a dangerous disease
plant get infected and crops can be devastated. Early
of plants found in warm and humid weather conditions.
blight disease develops at moderate to warm tempera-
Bacterial spot occurs on pepper and tomatoes. Symp-
tures.
toms on the leaves appear as small yellow-green lesions
which get deformed and twisted and change into the 8. Gray Leaf Spot: Gray leaf spot is a fungal disease that
dark, water-soaked, and greasy lesions. This disease attacks the corn plants which is also known as maize.
is due to bacteria that attack the vegetation, stems, and First, symptoms of this disease are noticed in the lower
fruits of tomatoes and pepper. Once this is present on leaves. The region on the leaf begins as a small yellow
the plants it is very difficult to control the disease and dot spot. As time passes this yellow spot changes to

Vol. 20, No. 5 (Suppl), 2021 • Nature Environment and Pollution Technology
 DEEP LEARNING CLASSIFICATION MODELS FOR IMAGE-BASED PLANT LEAF DISEASES 2139

brown color and then to a gray rectangular shape. This pear, cherry, and potatoes. This may reduce the fruit’s
region appears as the shape of a matchstick which slowly number and size. This can create yellowish rings on the
results in the killing of leaves. The grayish color on the leaves if leaves ripen in warm weather.
leaves appears due to the presence of fungal spores. 15. Northern Leaf Blight: Northern leaf blight is a disease
9. Late Blight: Late Blight is a disease that attacks pota- that affects the corn leaves. This disease is caused by
toes and tomatoes. This disease is caused by the water the fungus Setosphaeriaturcica. Symptoms generally
mold of Phytophthora infestans. This disease mostly appear on lower leaves with gray-green color and then
occurs in humid regions where the temperature is rang- turn to pale or tan color. Dark gray spores are produced
ing between 4-29°C. The infected tomatoes and potatoes under moist conditions and it gives regions a dirty gray
may get rot within two weeks. This disease spreads very appearance. Spores are generally transferred by wind
quickly in fields and results in total crop destruction if or by the splashing of water.
they are not controlled. 16. Powdery Mildew: Powdery Mildew is a fungal disease
10. Leaf Blight: Leaf Blight is a fungal disease that at- that affects a variety of plants and it reduces the quality
tacks grapes. This disease is caused by a fungus named and quantity of fruits and flowers. When the fungus
Helminthosporium turcicum Pass. This disease occurs takes over on one of your plants then the spores make a
in humid conditions and it shows symptoms with red- layer of mildew on the top of leaves. The spores are then
dish-purple or tan spots and it gets bigger on the leaves. transferred to other plants by the wind. This disease can
The symptoms on the leaves first appear on older or slow down the growth of plants and reduces the quality
lower leave but after then it spreads on the younger or of crops.
upper leaves. This drastic disease gives a burnt appear- 17. Spider Mites: Spider mite is a disease that eats plants
ance to the leaves. and they look like tiny spiders. Most of the spider mites
11. Leaf Mold: Leaf mold is a disease that is found in to- get active in dry and hot conditions. Because of the
matoes. This disease causes loss in tomatoes which are feeding of spider mites, white to yellow spots appear on
found in high tunnels or greenhouses due to humidity in the upper surfaces of leaves. The eggs also stick on the
those environments. This disease is caused by a fungus leaves’ undersides. As the disease infiltration, the color
named Passalora fulva. The oldest leaves are infected of leaves appears as bronze and then becomes stiff.
first due to this disease. 18. Target Spot: Target spot is a disease that attacks the
12. Leaf Scorch: Leaf Scorch is a disease that attacks tomato leaves. Initially small, dark-brown spots appear
strawberries. It is a serious disease that is caused by on the upper parts of older leaves, and then eventually its
the bacterium Xylella fastidiosa. The first symptom size increases and makes concentric rings. This disease
which can be noticed is the browning of leaves in the is spread by air-borne spores. This is a fungal disease
mid-summer. The symptoms get worse throughout the and affects many other crops like pepper, papaya or
late summer and after then gets fall. As the disease cucumber, etc.
progresses over the years, branches and trees decline 19. Yellow Leaf Curl Virus: Tomato yellow leaf curl dis-
slowly. The symptoms first appear on the lower branches ease is caused by the yellow leaf curl virus. The leaves
and then on the upper leaves. which are infected are curled inward or upwards. The
13. Leaf Spot: Leaf spot is a serious disease that is found infected plants reduced the flowers and fruits in large
in tomatoes. This disease is a fungal disease that is numbers. This disease is not seed-borne but is spread
caused by bacteria. Leaf spots show symptoms with by whiteflies. This disease is generally found in tropical
brown color but spots can be tan or black depending and sub-tropical regions which cause economic loss.
on the type of fungus. Some Concentric rings or dark
margins are also found around the dark spots. Leaf spot BACKGROUND AND RELATED WORK
diseases actually weaken the shrub or trees by blocking
the photosynthesis process. Badage (2018) elaborated that disease in plants is caused by
insects and various pests. Plant diseases decrease the produc-
14. Mosaic Virus: Mosaic Virus is a disease that attacks tivity of crops. Farmers face a lot of problems and losses due
tomatoes. This disease affects the leaves which show to these various crop diseases. The system is proposed by the
symptoms with spots of yellow, white, light, and dark author who tells about crop diseases and actions to control
green color. After then leaves may be curled, malformed, them. This proposed system is divided into two phases: the
and reduced in size. This virus can also infect pepper, first phase includes training of the datasets of crop diseases

Nature Environment and Pollution Technology • Vol. 20, No. 5 (Suppl), 2021
2140 Sakshi Takkar et al.

and the second phase includes the identification of crop dis- Qin et al. (2003) analyzed the stresses of rising diseases
eases by using Canny’s edge detection algorithm (Badage for pest management in fields. The research was carried out
2018). Maniyath et al. (2018) proposed some techniques on a rice field, and correlations between ground data and
on the leaf-based image classification to find out the results image data were made. The experiment results show that
and plant diseases. Random Forest algorithm was used to remote sensing imagery has a very important application
identify healthy and diseased leaves from the leaf-based and ground data shows an average accuracy of more than
image dataset. Various steps have been implemented like 70% for classification (Qin et al. 2003)we first examine
the collection of the dataset, feature extraction method, and the applicability of broadband high-spatial-resolution
training of dataset and classification approach. The machine ADAR (airborne data acquisition and registration. Rothe &
learning approach gives a clear picture of training the dataset Kshirsagar (2015) proposed a pattern recognition system to
and classification of images. identify the different cotton plant diseases. This work was
Sladojevic et al. (2016) argued that Convolution Neural done on the images of cotton leaves taken from the fields.
Network achieved more accurate results in the leaf image The contour model was used for the segmentation of images
classification to identify plant or crop diseases. This new and training of the adaptive neuro-fuzzy inference method.
approach of training the dataset is a quick and easy method The accuracy for the classification is approximately 85%.
of implementation. This proposed model could find out The diseases of cotton leaves were identified by using a back
thirteen different types of plant diseases by identifying the propagation neural network.
surroundings or edges of leaves. This proposed method Gulhane & Kolekar (2015) used Principal Component
showed the experimental results with an average precision of Analysis (PCA) and K-Nearest Neighbor (KNN) method to
96.3% (Sladojevic et al. 2016). Saleem et al. (2019) analyzed diagnose the diseases of cotton leaves in fields. In a num-
that early identification of plants diseases is very promi- ber of cases, human assistance in identifying diseases may
nent for healthy crops and plants. Many machine learning be incorrect. Machine learning models have been created
algorithms were used for the detection and identification of to determine the accuracy of disease detection in cotton
plant diseases but the subset of machine learning i.e. deep leaves plants. Implementation of PCA/KNN equipped with
learning techniques showed more accurate results as com- multi-variate techniques was used to analyze the statistical
pared to other machine learning algorithms. Various deep data. The PCA/KNN bases classifier showed a classification
learning techniques were combined with other visualization accuracy of 95% (Gulhane & Kolekar 2015). Revathi et al.
techniques to identify the symptoms and diseases of plants. (2011) proposed a computing technology to help farmers all
Performance metrics were used to evaluate the deep learning over the world to take care of the crops from various diseases.
techniques (Saleem et al. 2019). The author had proposed a method to diagnose the diseases of
Sarangdhar & Pawar (2017) analyzed the particular attack cotton plants by capturing the image using a mobile camera
of diseases that decrease the production of cotton crops. In and then categorized the diseases using a neural network. The
this study, a vector machine algorithm was used to identify work is based on the image segmentation technique where
five different types of cotton leaf diseases. An android app RGB color feature image segmentation is used to identify the
will be used by the farmers where diseases after identification disease spots on cotton plants (Revathi et al. 2011).
will be informed with their remedies. This android app also Blessy & Joy Winnie Wise (2018) selected Convolutional
identifies the soil type with its moisture and humidity. This Neural Network (CNN) technique to identify the disease
system has been made with sensors and raspberry pi that spots on the plants. The image of the sample leaf was used
makes the system more effective. The accuracy achieved as an input where green pixels from the image were marked
with this proposed system is 83.26 %. in green color that represents the healthy part of the image.
Huang et al. (2014) proposed new spectral indices that Further, this green area from the image was removed to cal-
are used to identify the different diseases on wheat crops. culate the rest of the infected area from the image. Following
Optimized spectral indices were obtained by the combination the extraction of characteristics from the affected areas, the
of a single band and the difference of wavelength between CNN model is used to classify the diseases. After diagnosing
two different bands. RELIEF-F algorithm has been used by the disease, detailed information about diseases was sent to
an author to identify the wavelengths from the leaf spectral the mobile of farmers with the solutions through GSM device.
data. This algorithm is more effective as it can deal with Rastogi et al. (2015) studied the automatic detection of
multiclass classification problems. This study indicates crop diseases based on the two phases used in the proposed
new spectral indices can easily detect diseases by using system. In the first phase, pre-processing of leaf images,
hyperspectral data. feature extraction, and classification using Artificial Neu-

Vol. 20, No. 5 (Suppl), 2021 • Nature Environment and Pollution Technology
 DEEP LEARNING CLASSIFICATION MODELS FOR IMAGE-BASED PLANT LEAF DISEASES 2141

ral Network (ANN) was done to recognize the leaves. In out the accuracy and other performance parameters for crop
the second phase, K means-based segmentation was done disease prediction (Revathi et al. 2011).
to identify the defected areas, then feature extraction was Sethy et al. (2018) developed a prototype to identify the
done to find the defected portion and classification of the diseases of rice crops using computational intelligence and
disease identified by using ANN. After the identification of machine learning techniques. Numerous diseases on the rice
diseases, grading was done based on the amount of disease crops appear as a spot on the leaves and if diagnosis of the
present in the leaf. diseases has not been done on time, then it can cause great
Owomugisha et al. (2014) proposed machine learning harm to rice crops. The pesticide-based treatment of crops
techniques to identify bacterial diseases in banana plants. can cause severe environmental pollution. In this proposed
The computer vision technique was investigated to make methodology, Fuzzy logic was introduced with a K-means
an algorithm that is further divided into four phases. In the algorithm to identify the degree of the sternness of disease on
first phase, images of banana leaves were captured using a rice crops. This proposed methodology showed an accuracy
digital camera. In phase two, feature extraction techniques of 86.35%.
were used to send the data in phase three for classification. Shruthi et al. (2019) presented a comparative analysis
In phase three, different classifiers were used to identify of various machine learning techniques to identify the best
the diseases. In the last phase, the performance of all the technique for crop disease detection. It was observed that
classifiers was compared based on Area under Curve (AUC) Convolution Neural Network provides more accuracy to
parameter for evaluation. Tian et al. (2012) presented an identify the diseases from crops (Shruthi et al. 2019). Reza
SVM- based Multiple Classifier System to recognize the et al. (2016) framed a research methodology to detect jute
patterns of wheat leaf diseases. Encouraging results were plant diseases using image processing and machine learning
obtained by using this proposed methodology to identify the techniques. The proposed methodology involves an android
diseases from wheat diseases. Three different types of SVM application that helps to capture the pictures of jute plant
classifiers were used as color features, shape features, and diseases and to send the pictures to the server to identify
texture features for training sets. the jute plant diseases. The features were extracted from the
Shi et al. (2015) proposed an automatic crop disease captured image and extracted values were compared with
recognition method that takes the statistical features from the values stored in the database that helped to identify the
leaf images and meteorological data and combines them. leaf diseases. The classification of the diseases was done
The Probabilistic Neural Network (PNN) method was used by using a Multi-SVM classifier. After that, the final result
to identify the classification accuracy. Infected crop leaves was sent to the farmer within a fraction of seconds with the
images were captured by using a digital camera to extract the necessary solutions or control measures through the android
statistical features like color, shape, and texture by using the application (Reza et al. 2017).
image processing method. PNN classifier accuracy rate was
90% which is more than the accuracy rate of other classifiers. MATERIALS AND METHODS
Ashqar & Abu-Naser (2018) presented a deep learn- In this proposed system, a dataset of 54,343 images of differ-
ing technique to identify tomato leaves diseases using ent plant species was taken that involves diseased plants or
image-based recognition. Around 9000 images dataset of healthy plants images of various fruits and vegetable crops.
healthy and diseased tomato leaves were collected under The dataset was split into three sets: training set, validation
controlled conditions. Training of deep convolutional set, and testing set. Training is done by using the pre-trained
neural network was done to identify the five diseases. The model Inception V3 by fine-tuning the last layers of the net-
proposed method showed an accuracy of 99.84% that made work. Four custom convolutional and max-pooling, layers
this approach more feasible to diagnose the tomato leaves have been added on the top of transfer learning architectures.
diseases in agriculture (Ashqar & Abu-Naser 2018). Revathi At the last, two dense layers have been used with 64 neutrons
& Hemalatha (2012) analyzed a comparative study of crop and 2 neurons respectively. The last layer is used for classifi-
diseases using machine learning techniques in the field of cation with softmax as an activation function. Training of the
agriculture. The algorithms like SVM machine learning algo- model has been done by 20 epochs or iterations by changing
rithm, decision tree algorithm, and artificial neural network the various parameters like batch size, optimizer, pre-trained
were used to represent the work. Data mining technique is weights, and learning rate. To reduce the overfitting between
one of the innovative techniques and used to predict various the batch normalization and different layers, 30% dropouts
crop diseases. This study was based on the applicability of were used to reduce the internal covariate shift and this
data mining techniques with a comparative analysis to find actually helps the model to avoid getting stuck in the local

Nature Environment and Pollution Technology • Vol. 20, No. 5 (Suppl), 2021
 2142 Sakshi Takkar et al.

optimum. The evaluation metric named Multi-Class Log present study. At first, the data was labeled and cleaned for
Loss was also used in the model with some data generators prior image processing. For image processing, images with
for training data and testing data. These generators help to good resolution and angles were selected to ensure a high
load the required amount of data directly from the source accuracy algorithm detection system. After the selection of
folders with the batch sizes as per need of the detection and all the images from the database, in-depth knowledge was
help to convert the batches into training data and training gained for different plants and their diseases. Different types
models. To do a fair comparison between the results of all of plant diseases with their symptoms were studied in this
experiments, an attempt has been made to normalize hy- processing system. After the deep and detailed study, image
per-parameters across all the experimentations and research. segregation was done to label the images and the following
Pre-processing and Training the Model: Pre-processing steps were performed:
Inis this proposed
the foremost system,
step a dataset on
that is performed of the
54,343
images.images
In thisof different plant species was taken that involves diseased
1. The input test image was acquired and pre-processing
plants or healthy plants images of various
step, the database is pre-processed such as reshaping of an fruits and vegetable crops.
was done The dataset
accordingly. was split intothe
After pre-processing, three sets:
image
training set, validation
image, resizing of an image,set,
andand testingofset.
conversion Training
an image into is donewas byconverted
using theinto pre-trained
an array formmodel Inception V3 by
for evaluation.
an array form (Fig. 1). Pre-processing is also done
fine-tuning the last layers of the network. Four custom convolutional on the and max-pooling,
2. Then the dataset was segregatedlayers have been added
and pre-processing has
test images in the same way. A database consists of 54,343
on the top
images of transfer
of different plantlearning
species and architectures. At theany
out of all the images last, two dense layers have been used with 64 neutrons and
been done.
2image
neurons can respectively.
be used as a test The last The
image. layerdatabase
is usedisfor classification
used to 3. After with softmax as an
pre-processing, activation
training function.
of dataset has beenTraining
done
oftrain
thethemodel
datasethas been
using done by V3
the Inception 20model
epochs andor by using Inception V3 model and further
iterations by changing the various parameters like batch size,
it further classification
helps to identify the test image and the disease of the test has been completed.
optimizer, pre-trained weights, and learning rate. To reduce the overfitting between the batch normalization
image. After the training of the model, the software can find 4. In the next step, a comparison of the test and trained
and differentin layers,
the diseases the plants30%
thatdropouts
were already were used
stored in to
thereduce
da- the model
internal covariate shift and this actually helps the
was done and the final results have been obtained
model
tabase. to avoid
After getting stuck
the completion in theand
of training local optimum. The evaluation
pre-processing, metric
as per system named Multi-Class Log Loss was
inputs.
the evaluation is done between the test and trained
also used in the model with some data generators for training model to data
predict the disease. 5. In the lastand
step,testing data.displayed,
the software These generators
whether the help
plant to
load the required amount of data directly from the source folders with the batch sizes as per need of the
is diseased or healthy.
Collection of Database: Acquiring valid database col-
detection
lection is and help step
the initial to convert the batches
of any image into trainingClassification
processing-based data and training
Models:models. To datasets,
To classify do a fairtwocomparison
models
between the results
project. There are manyof all
waysexperiments,
to generate thean attempt
database has were used: SRCNN (Super-Resolution
as been made to normalize hyper-parameters across Convolutional Neural
all the
Network) Model and Bicubic Model.
experimentations
either to store all imagesand research.
or to collect the images from
different sources and make own database for processing. ∑ SRCNN (Super-Resolution Convolutional Neural Net-
The Kaggle Plant Village Dataset database was used in the work) Model: In deep learning, generally Convolutional

Input Test
Image Pre-processing Conversion
Diseased
Plants

SRCNN/Bicubic
Based
Classification

Healthy
Plants
Segregated Training Using
Pre-processing Inception V3 model
Data Set

Fig. 1: Flowchart for plant disease detection.

Fig. 1: Flowchart for Plant disease detection.

Vol. 20, No. 5 (Suppl), 2021 • Nature Environment and Pollution Technology
Pre-processing and Training the Model: Pre-processing is the foremost step that is performed on the images.
 1. The input test image was acquired and pre-processing was done accordingly. After pre-processing, the
imageDEEPwas converted
LEARNINGinto an array formMODELS
CLASSIFICATION for evaluation.
FOR IMAGE-BASED PLANT LEAF DISEASES 2143
2. Then the dataset was segregated and pre-processing has been done.
Neural Network (CNN) is used for image classification. computational cost increases with each pre-up sampling
3. After pre-processing, training of dataset has been done
The objective of Super-Resolution (SR) is to recuperate
by using Inception V3 model and further
step. This algorithm is mainly used in computer editing
classification
high-resolutionhas beenfrom
images completed.
low-resolution images. software or by editors for reconstructing and resampling
4. In the next step, a comparison of the test and trained model
The SRCNN network involves mainly four operations: wasDuring
the images. doneinterpolation
and the finalof anresults have
image, the been
pixels
pre-processing, feature extraction, non-linear mapping, get distorted from one grid to another grid. This is a
obtained as per system inputs.
and reconstruction. very slow algorithm as it takes time to process during
5. In the last step, the software displayed, whether
1. Pre-processing: This step means upscaling of
the plant is diseased
the resampling orimage.
of an healthy.
There are two interpolation
low-resolution to high-resolution images. algorithms: Adaptive Interpolation and Non-Adaptive
Interpolation.
2. Feature Extraction: This step extracts the set of
feature maps from the upscaled low-resolution a) Adaptive Interpolation: Adaptive interpolation
Classificationimage.
Models: To classify datasets, two models algorithm were used:dependsSRCNN
on what the(Super-Resolution
image is intro-
Convolutional

Neural Network) Model and Bicubic
3. Non-Linear Mapping: Mapping of feature maps
Model. ducing. Adaptive algorithms are used in exclusive
techniques which are used in various latest photo
that represents low-resolution to high-resolution editingModel:
software In
like deep
Adobe learning,
Photoshop and Pho-
 SRCNN patches.
(Super-Resolution Convolutional Neural Network) generally
tozoom Pro.
Convolutional
Neural Network
5. Reconstruction: (CNN)
Produces or is used
reconstructs for image classification. The objective of Super-Resolution
the high- b) Non-Adaptive Interpolation: Non-adaptive in-
(SR) is toresolution
recuperate high-resolution
image images
from high-resolution from low-resolution
patches. images.
terpolation The treats
method SRCNN network
all the involves
pixels equally.
mainly four isoperations:
SRCNN pre-processing,
a deep convolutional feature
neural network thatextraction,
is non-linear mapping,
Non-adaptive and reconstruction.
algorithms involve various other
used for end-to-end mapping of low-resolution to high-res- algorithms like the k-nearest neighbor, spline,
a.olution
Pre-processing: This step means upscaling of low-resolution to high-resolution images.
images (Fig. 2). This model is used to improve the Bicubic, and bilinear.
b.quality
Feature Extraction:
of low-resolution ThisWith
images. stepthis
extracts
approach the
of set of feature maps from the upscaled low-resolution image.
Super
c. Non-Linear Mapping: Mapping of feature mapsRESULTS
Resolution, a better quality of images can be obtained from that represents low-resolution to high-resolution
AND DISCUSSION
even a small size of input images. The performance of this
patches.
network can be evaluated by using different parameters such Dataset Description
d.as Mean Squared ErrorProduces
Reconstruction: (MSE), Peakor Signal
reconstructs
to Noisethe high-resolution image from high-resolution patches.
Ratio In this study, a plant disease village dataset of 54,343 images
(PSNR), and Structural Similarity (SSIM) Index. have been taken which are fine-tuned using a well-known
SRCNN is aInterpolation
∑ Bicubic deep convolutional
Algorithm:neural
Bicubic network that is
Interpolation usedInception
model for end-to-end mapping of and
V3 after pre-processing, low-resolution
the dataset has to
high-resolution
Algorithm isimages (Fig. 2). This
a two-dimensional systemmodel is aused
that uses pol- to been
improve
taken the
fromquality
Kaggleof low-resolution
Repository. images.
This dataset With
involves
this approach of Super Resolution, a better quality of images can be obtained from even a small size of
ynomial technique for enlarging or sharpening digital the study of different plants like Apple, Blueberry, Cherry,
images (Fig. 3). This algorithm (samples dimension: Corn, Grape, Orange, Peach, Pepper bell, Potato, Raspberry,
input 4x4
images.
and 16The performance
samples of thislow-resolution
at a time) upscale network can be evaluated
Soybean, Squash, byStrawberry,
using different parameters
Tomato, etc. such
The different sets as
Meanimages
Squared Error
before (MSE),
going to thePeak Signal
network. to Noise
However, theRatio (PSNR),
of healthy and and Structural
diseased Similarity
plant images are used.(SSIM) Index.
The research

Fig. 2: SRCNN based classification model.

Fig. 2: SRCNN Based Classification Model.

Nature Environment and Pollution Technology • Vol. 20, No. 5 (Suppl), 2021
2144 Sakshi Takkar et al.

Table 1: Dataset for image classification of leaf disease.

Sr. No. Plant Category Disease/ Healthy Number of Original Images

1 Apple Apple Scab 631
2 Apple Black Rot 622
3 Apple Cedar Apple Rust 276
4 Apple Healthy 1646
5 Blueberry Healthy 1503
6 Cherry Healthy 855
7 Cherry Powdery Mildew 1053
8 Corn Common Rust 1193
9 Corn Gray Leaf Spot 514
10 Corn Healthy 1163
11 Corn Northern Leaf Blight 986
12 Grape Black Measles 1384
13 Grape Black Rot 1181
14 Grape Healthy 424
15 Grape Leaf Blight 1077
16 Orange Citrus Greening 5508
17 Peach Bacterial Spot 2298
18 Peach Healthy 361
19 Pepper Bell Bacterial Spot 998
20 Pepper bell Healthy 1479
21 Potato Early Blight 1001
22 Potato Healthy 153
23 Potato Late Blight 1001
24 Raspberry Healthy 372
25 Soybean Healthy 5091
26 Squash Powdery Mildew 1836
27 Strawberry Healthy 457
28 Strawberry Leaf Scorch 1110
29 Tomato Bacterial Spot 2128
30 Tomato Early Blight 1001
31 Tomato Healthy 1592
32 Tomato Late Blight 1910
33 Tomato Leaf Mold 953
34 Tomato Leaf Spot 1772
35 Tomato Mosaic Virus 374
36 Tomato Spider Mites 1677
37 Tomato Target Spot 1405
38 Tomato Yellow Leaf Curl Virus 5358
53343

Vol. 20, No. 5 (Suppl), 2021 • Nature Environment and Pollution Technology
8 0.944938 0.166638 0.001 0.921461 0.267373 8 0.943688 0.173809 0.001 0.909642 0.28659
bilinear.
9 0.954747 0.139192 0.001 0.946598 0.169225 9 0.951368 0.151127 0.001 0.943534 0.188381
10 0.95025
DEEP0.148785 0.001
LEARNING 0.92096
CLASSIFICATION0.29281
MODELS 10 IMAGE-BASED
FOR 0.94875 0.15574 0.001DISEASES
PLANT LEAF 0.906891 0.306541
2145
11 0.951493 0.149138 0.001 0.931528 0.23313 11 0.950491 0.154059 0.001 0.922149 0.271082
12 0.958125 0.130086 0.001 0.957604 0.129822 12 0.958 0.129596 0.001 0.96467 0.110303
13 0.956125 0.131764 0.001 0.966483 0.120217 13 0.95675 0.131862 0.001 0.90514 0.346343
14 0.963573 0.11314 0.001 0.922961 0.291764 14 0.960381 0.120436 0.001 0.939032 0.210869
15 0.966125 0.108081 0.001 0.975175 0.074392 15 0.960438 0.126139 0.001 0.926963 0.260596
16 0.963009 0.112118 0.001 0.955103 0.145883 16 0.966201 0.103237 0.001 0.959042 0.139745
17 0.96575 0.10554 0.001 0.943034 0.19108 17 0.963875 0.110578 0.001 0.922836 0.268867
18 0.965137 0.105917 0.001 0.928589 0.264863 18 0.96545 0.108457 0.001 0.964107 0.109852
19 0.964938 0.110921 0.001 0.953789 0.142576 19 0.983813 0.046763 0.0002 0.993997 0.018768
20 0.97175 0.089099 0.001 0.96986 0.093184 20 0.988375 0.036806 0.0002 0.993684 0.017171
21 0.971209 0.087371 0.001 Fig. 3: Bicubic
0.96048 Interpolation based
0.147692 21 classification
0.988859 model. 0.034741 0.0002 0.993934 0.018891
22
has 0.985438
been done on 0.04471
different0.0002 0.99556
diseases like 0.016243
apple scab, 22
bacte- and 0.989688
learning rate by 0.031454
epochs using0.0002 0.994122
two different 0.019938
classifiers
23 0.990236 0.030135 0.0002 0.995373 0.013251 23 0.990299 0.029428 0.0002 0.995998
rial spot, black measles, black rot, cedar apple rust, citrus SRCNN and Bicubic algorithm as shown in Table 2. Out 0.012489
of
Fig. 3: Bicubic Interpolation based Classification model.
24 greening,
0.991563common rust, early
0.024861 0.0002blight, spot, late these
gray leaf 0.013701
0.996248 24 classifiers,
0.99175 the SRCNN
0.025347classifier
0.0002shows better accuracy
0.993184 0.02261
blight, leaf blight, leaf mold, leaf scorch, leaf spot, mosaic as compare to Bicubic (Table 3). The graphs of training and
virus, northern leaf blight, powdery mildew, spider mites, validation accuracy by epochs for the Bicubic and SRCNN
target spot, and yellow leaf curl virus. Table 1 describes the model have been depicted in Fig. 4.
RESULTS AND
plant category, DISCUSSION
healthy Table
or diseased 3: Shows
image, and thethe comparison report with other models.
number
of images in each category. CONCLUSION
Dataset Description Model Proposed
Experimental Results Protection of crops Classification
in an agricultureAccuracy
field is a very tedious
Bicubic 99.156
task and still, there is a need for a qualitative study to know
In this study,
In each model, a data
plant disease
SRCNN
is divided village
into dataset
two sets- ofset54,343
training images
and about thehave
cropsbeen takenlikely
and their which
99.175 are fine-tuned
weeds, pathogens, andusing a
well-known model Inception V3 after pre-processing, and
testing set. Training has been done by using the Inception V3 theThe
pests. dataset
presenthas been taken
methodology from Kaggle
identifies Repository.
diseases in plants
model and the dataset has been split into 70-30, 50-50, and to increase the productivity of crops in fields. The system
This dataset involves the study of different plants like Apple, Blueberry, Cherry, Corn, Grape, Orange, Peach,
30-70. In the training phase, we train the classifiers and in the is developed for the benefit of farmers and the agricultural
Pepper
testingbell, Potato,
phase, testingRaspberry, Soybean,
is done to analyze Squash, of
the performance Strawberry,
sector. InTomato,
this system,etc. The
deep different
learning modelssetswere
of healthy
used for and
diseased plant images are used. The research has beenthe
the classifier. Results are demonstrated using different param- done on different
detection diseases
of plant diseases like
using appleleaf
different scab, bacterial
images to
eters like accuracy, loss, validation accuracy, validation loss,
spot, black measles, black rot, cedar apple rust, citrus greening, common rust, early blight, gray leaf spot, late
identify whether the leaf is healthy or diseased. The outcome
blight, leaf blight, leaf mold, leaf scorch, leaf spot, mosaic virus, northern leaf blight, powdery mildew, spider
mites, target spot, and yellow leaf curl virus. Table 1 describes the plant category, healthy or diseased image,
and the number of images in each category.

Table 1: Dataset for image classification of leaf disease.

Fig. 4: Graphs representing training and validation accuracy for models.

Nature Environment and Pollution Technology • Vol. 20, No. 5 (Suppl), 2021
2146 Sakshi Takkar et al.

Table 2: Comparison of bicubicvs SRCNN classifier.

Bicubic Result by Epoch SRCNN Result by Epoch

Epoch Acc loss lr val_acc val_loss Epoch acc loss lr val_acc val_loss
0 0.68675 1.078107 0.001 0.359188 3.793237 0 0.685 1.076052 0.001 0.597313 1.825342
1 0.853313 0.452098 0.001 0.807466 0.675127 1 0.853563 0.456332 0.001 0.608867 1.734919
2 0.899355 0.314722 0.001 0.883504 0.393525 2 0.892846 0.326332 0.001 0.762569 0.936293
3 0.911813 0.27287 0.001 0.868497 0.433363 3 0.914438 0.266576 0.001 0.889007 0.34634
4 0.927396 0.225642 0.001 0.904577 0.324232 4 0.924454 0.234888 0.001 0.918772 0.274175
5 0.933563 0.206778 0.001 0.771073 0.849628 5 0.932875 0.209391 0.001 0.929777 0.229069
6 0.937688 0.195976 0.001 0.831978 0.864499 6 0.937688 0.196275 0.001 0.925775 0.234507
7 0.93885 0.188665 0.001 0.93972 0.185791 7 0.938537 0.189037 0.001 0.925775 0.282873
8 0.944938 0.166638 0.001 0.921461 0.267373 8 0.943688 0.173809 0.001 0.909642 0.28659
9 0.954747 0.139192 0.001 0.946598 0.169225 9 0.951368 0.151127 0.001 0.943534 0.188381
10 0.95025 0.148785 0.001 0.92096 0.29281 10 0.94875 0.15574 0.001 0.906891 0.306541
11 0.951493 0.149138 0.001 0.931528 0.23313 11 0.950491 0.154059 0.001 0.922149 0.271082
12 0.958125 0.130086 0.001 0.957604 0.129822 12 0.958 0.129596 0.001 0.96467 0.110303
13 0.956125 0.131764 0.001 0.966483 0.120217 13 0.95675 0.131862 0.001 0.90514 0.346343
14 0.963573 0.11314 0.001 0.922961 0.291764 14 0.960381 0.120436 0.001 0.939032 0.210869
15 0.966125 0.108081 0.001 0.975175 0.074392 15 0.960438 0.126139 0.001 0.926963 0.260596
16 0.963009 0.112118 0.001 0.955103 0.145883 16 0.966201 0.103237 0.001 0.959042 0.139745
17 0.96575 0.10554 0.001 0.943034 0.19108 17 0.963875 0.110578 0.001 0.922836 0.268867
18 0.965137 0.105917 0.001 0.928589 0.264863 18 0.96545 0.108457 0.001 0.964107 0.109852
19 0.964938 0.110921 0.001 0.953789 0.142576 19 0.983813 0.046763 0.0002 0.993997 0.018768
20 0.97175 0.089099 0.001 0.96986 0.093184 20 0.988375 0.036806 0.0002 0.993684 0.017171
21 0.971209 0.087371 0.001 0.96048 0.147692 21 0.988859 0.034741 0.0002 0.993934 0.018891
22 0.985438 0.04471 0.0002 0.99556 0.016243 22 0.989688 0.031454 0.0002 0.994122 0.019938
23 0.990236 0.030135 0.0002 0.995373 0.013251 23 0.990299 0.029428 0.0002 0.995998 0.012489
24 0.991563 0.024861 0.0002 0.996248 0.013701 24 0.99175 0.025347 0.0002 0.993184 0.02261

Table 3: Shows the comparison report with other models. proposed system and it can be widely used in the agricultural
sector for the help of farmers.
Model Proposed Classification Accuracy
Bicubic 99.156 REFERENCES
SRCNN 99.175
Ashqar, B.A.M. and Abu-Naser. S.S. 2018. Image-based tomato leaves
diseases detection using deep learning. Int. J. Academic Eng. Res.,
of experimental results and comparison between two models 2(12): 10-16.
Badage, A. 2018. Crop disease detection using machine learning: Indian
SRCNN and Bicubic demonstrates the accuracy to recog- agriculture. Int. Res. J. Eng. Technol., 5(9): 866-869.
nize the correct disease in plants. Out of these two models, Bangotra, P., Sharma, M., Mehra, R., Jakhu, R., Singh, A., Gautam, A.S. and
SRCNN gives an accuracy rate of 99.175 % in recognizing Gautam, S. 2021. A systematic study of uranium retention in human
plant diseases. Diseases are not the specific problem in the organs and quantification of radiological and chemical doses from
uranium ingestion. Environ. Technol. Innov., 21: 101360.
agricultural sector but crops growing in good soil and getting Bangotra, P., Mehra, R., Jakhu, R, Pandit, P. and Prasad, M. 2019. Quantifi-
nutritious food protect the plant from various pest attacks. cation of an alpha flux-based radiological dose from seasonal exposure
The experimental result represents the effectiveness of our to 222Rn, 220Rn, and their different EEC species. Sci. Rep., 9: 2515.

Vol. 20, No. 5 (Suppl), 2021 • Nature Environment and Pollution Technology
 DEEP LEARNING CLASSIFICATION MODELS FOR IMAGE-BASED PLANT LEAF DISEASES 2147

Blessy, A. and Joy Winnie Wise, D.C. 2018. Detection of affected parts of 2012 3rd International Conference on Computing, Communication and
plant leaves and classification of diseases using CNN technique. Int. Networking Technologies, ICCCNT 2012, (July).
J. Eng. Tech., 4(2): 823-829. Revathi, P., Revathi, R. and Hemalatha, M. 2011. Comparative study of
Gulhane, V.A. and Kolekar, M.H. 2015. Diagnosis of diseases on cotton knowledge in crop diseases using machine learning techniques. Int. J.
leaves using principal component analysis classifier. 11th IEEE India Comput. Sci. Inf. Technol., 2(5): 2180-2182.
Conference: Emerging Trends and Innovation in Technology, INDI- Reza, Z.N., Nuzhat, F., Mahsa, N.A. and Ali, M.H. 2017. Detecting jute
CON, 11-13 December 2014, Pune India. IEEE, India, pp.1-5. plant disease using image processing and machine learning. 2016 3rd
Huang, W., Guan, Q., Luo, J., Zhang, J., Zhao, J., Liang, D. and Zhang, D. International Conference on Electrical Engineering and Information
2014. New optimized spectral indices for identifying and monitoring and Communication Technology, (ICEEiCT), 22-24 September,
winter wheat diseases. IEEE J. Select. Topics Appl. Earth Observ. Dhaka, Bangladesh. IEEE, India, pp. 11-26.
Remote Sens., 7(6): 2516-2524. Rothe, P.R. and Kshirsagar, R.V. 2015. Cotton leaf disease identification
Maniyath, S.R., Vinod, P.V., Niveditha, M., Pooja, R., Prasad Bhat, N., using pattern recognition techniques. 2015 International Conference
Shashank, N. and Hebbar, R. 2018. Plant disease detection using ma- on Pervasive Computing: Advance Communication Technology and
chine learning. Proceedings - 2018. International Conference on Design Application for Society, ICPC, 8-10 Jan, 2015, Pune, India. IEEE,
Innovations for 3Cs Compute Communicate Control, ICDI3C, 25-28 India, pp. 1-6.
April, 2018, Bangalore, India. IEEE, India, pp. 41-45. Saleem, M.H., Potgieter, J. and Arif, K.M. 2019. Plant disease detection
Mehra, R., Bangotra, P. and Kaur, K. 2015.  222Rn and 220Rn levels of and classification by deep learning. Plants, 8(11): 32–34.
Mansa and Muktsar district of Punjab, India. Front. Environ. Sci., 3:37. Sarangdhar, A.A. and Pawar, V. 2017. Machine learning regression technique
Pradhan, R.P. 2007. Indian agriculture in the globalization era: The perfor- for cotton leaf disease detection and controlling using IoT. Proceedings
mance and determinants. J. Global Econ., 3(1): 3-12. of the International Conference on Electronics, Communication and
Owomugisha, G., Quinn, J.A., Mwebaze, E. and Lwasa, J. 2014. Automat- Aerospace Technology, 20-22 April 2017, Coimbatore, India. Institute
ed vision-based diagnosis of banana bacterial wilt disease and black of Electrical and Electronics Engineers (IEEE), India, pp. 449-454.
Sigatoka disease. International Conference on the Use of Mobile ICT Sethy, P.K., Negi, B., Barpanda, N.K., Behera, S.K. and Rath, A.K. 2018.
in Africa, 9-10th December 2014, Stellenbosch, South Africa. Stellen- Measurement of disease severity of rice crop using machine learning
bosch University, Department of Electrical & Electronic Engineering, and computational intelligence. Proc. Comp Sci., 167(1): 516-530.
South Africa, pp. 1-5. Shi, Y., Wang, X.F., Zhang, S.W. and Zhang, C.L. 2015. PNN based crop
Qin, Z., Zhang, M., Christensen, T., Li, W. and Tang, H. 2003. Remote disease recognition with leaf image features and meteorological data.
sensing analysis of rice disease stresses for farm pest management Int. J. Agric. Biol. Eng., 8(4): 60-68.
using wide-band airborne data. Int. Geosci. Remote Sens. Symp. Shruthi, U., Nagaveni, V. and Raghavendra, B. K. 2019. A review on
4(C): 2215-2217. machine learning classification techniques for plant disease detec-
Pandit, P., Mangala, P., Saini, A, Bangotra, P., Kumar, V., Mehra, R. and tion. 2019 5th International Conference on Advanced Computing
Ghosh, D. 2020. Radiological and pollution risk assessments of ter- and Communication Systems, ICACCS 2019,15-16 March 2019,
restrial radionuclides and heavy metals in a mineralized zone of the Coimbatore, India, Institute of Electrical and Electronics Engineers,
Siwalik region (India), Chemosphere, 254: 126857. India, pp. 281–284.
Rastogi, A., Arora, R. and Sharma, S. 2015. Leaf disease detection and grad- Sladojevic, S., Arsenovic, M., Anderla, A., Culibrk, D. and Stefanovic,
ing using computer vision technology & fuzzy logic. 2nd International D. 2016. Deep neural networks-based recognition of plant diseases
Conference on Signal Processing and Integrated Networks, 19-20 Feb, by leaf image classification. Comput. Intell. Neurosci., 2016: 54-79.
2016, Noida, India. IEEE, India, pp. 500-505. Tian, Y., Zhao, C., Lu, S. and Guo, X. 2012. SVM-based multiple classifier
Revathi, P. and Hemalatha, M. 2012. Advance computing enrichment eval- systems for recognition of wheat leaf diseases. World Autom. Congr.
uation of cotton leaf spot disease detection using Image Edge detection. Proc., 5(1): 2-6.

Nature Environment and Pollution Technology • Vol. 20, No. 5 (Suppl), 2021

View publication stats