Artoaj MS Id 556316

Research Article Agri Res & Tech: Open Access J
Volume 25 Issue 5 - April 2021

Copyright © All rights are reserved by Damian Mingle
DOI: 10.19080/ARTOAJ.2021.25.556316
Classify Rice Disease Using Self-Optimizing Models and

Edge Computing with Agricultural Implications
Damian Mingle1,2,3* and Amit Kumar1
1
LogicPlum, Inc., Franklin, Tennessee, 37067, USA
2
SwitchPoint Ventures, Nashville, Tennessee, 37215, USA
3
Lipscomb University, Adjunct of College of Computing & Technology, Nashville, Tennessee, 37204, USA
Submission: March 24, 2020; Published: April 05, 2021
Corresponding author: Damian Mingle, West McEwen Drive, Suite 300, Franklin, TN, 37067, USA
*
Abstract
Rice continues to be a primary food for the world’s population. Over its complex history, dating as far back as 8,000 B.C., there have been
agricultural challenges, such as a variety of diseases. A consequence of disease in rice plants may lead to no harvest of grain; therefore, detecting
disease early and providing expert remedies in a low-cost solution is highly desirable. In this article, we study a pragmatic approach for rice
growers to leverage artificial intelligence solutions that reduce cost, increase speed, improve ease of use, and increase model performance over
other solutions, thereby directly impacting field operations. Our method significantly improves upon prior methods by combining automated
feature extraction for image data, exploring thousands of traditional machine learning configurations, defining a search space for hyper-
parameters, deploying a model using edge computing field usability, and suggesting remedies for rice growers. These results prove the validity of
the proposed approach for rice disease detection and treatments.
Keywords: Agriculture Technology; Machine Learning Applications; Rice Production, Edge Computing; Precision Farming; Agriculture Education;
Pre-Trained Models For Image Classification; Deep Learning Applications; Farming Knowledge; Rice Disease Management
Introduction
Figure 1: Rice Plant Disease Detection Algorithm provided by Prajapati et al., 2017.
Agri Res & Tech: Open Access J 25(5): ARTOAJ.MS.ID.556316 (2021) 001
Agricultural Research & Technology: Open Access Journal
Rice supports more than half the world’s population as of neurons per layer, (d) quantity of hidden layers, and (e) dropout
a primary food source [1]. The quality and quantity of rice regularization remains unrealistic. Much of the research to date
production are significantly affected by rice disease. In general, has been concerned with many pre-processing steps and augmen-
identification of rice disease is made by visual observation of tation techniques for images to maximize model performance: (a)
experienced producers in the field. This method requires constant resize images, (b) denoise, (c) segmentation, and (d) morphology.
surveillance from manual labor, which could be prohibitively In almost all the research, model performance has suffered from
expensive for large farms. However, with the advances in image over-fitting, evidenced by high accuracy scores for training sets
processing and pattern recognition, a cost-effective method for but significantly lower accuracy for validation sets.
disease identification is demonstrated. Advances in research
Given that growers value more what is likely to happen in
continue on image processing and pattern recognition as a
day-to-day utilization, the emphasis on a practical solution sug-
result of innovations with digital cameras and the increase in
gests validation scores matter more than training scores. It will
computational capacity. These tools have been effectively applied
measure how well a solution performs. Lastly, there is little to no
in many areas [2-5]. Prajapati et al., [6] developed a rice plant
connection between identifying plant disease and what action rice
disease classification system after detailed experimental analysis
farms should do next to experience the benefit of an algorithm
of various techniques. Four techniques of background removal and
detecting a plant disease early. In this work, we studied the ben-
three techniques of segmentation were empirically evaluated. It
efits of crafting an end-to-end solution for rice farmers using an
was proposed for accurate feature extraction, a centroid feeding-
automated machine learning platform with the aim of building a
based K-means clustering for segmentation of disease from a
production-grade solution for agriculture that provides real-time
leaf image was necessary. The output from K-means clustering
decision support for rice farms. We combine several methods,
was enhanced by removing green pixels in the disease portion.
namely, employing automated feature extraction for image data,
Additional feature extraction was done under three categories:
exploring thousands of possible traditional machine learning con-
(1) color, (2) shape, and (3) texture. Ultimately, Support Vector
figurations, defining a search space for hyper-parameters, deploy-
Machines was chosen to perform a multiclass classifier (Figure 1).
ing a model built for edge computing for field usability, and sug-
Generally, rice growers identify plant disease through leaves as gesting remedies for rice growers. This journal article comprises
the first source. This can be detected automatically using comput- the following sections: methods and materials, results, discussion,
er vision techniques. Until now, there have been several research- and conclusion.
ers who have conducted experiments with very little utility for
Methods and Materials
rice farms. Considerations for farmers are cost, speed, ease of use,
model performance, and direct impact on the field. There has been Data Acquisition
little attention to structuring a useful machine learning approach
The dataset contains 120 jpeg images of disease-infected rice
that is end-to-end in agriculture. Previous investigations have suc-
leaves. There are 3 classes of images based on the type of disease,
cessfully demonstrated the potential of deep learning algorithms
each containing 40 images, and captured with a NIKON D90 dig-
in plant disease detection; yet, the cost associated with such archi-
ital SLR camera with 12.3 megapixels. This dataset was curated
tecture makes this unattainable for many rice growers. The length
by the research team at the Department of Information Technol-
of training time required for such deep learning models has histor-
ogy, Dharmsinh Desai University, and is made publicly available.
ically been lengthy, and specialty hardware is needed. Additional-
The authors gathered these leaves from a rice field in a village
ly, the expertise necessary to maintain and optimize deep learning
called Shertha in Gujarat, India, and in consultation with farmers,
network hyper-parameters, such as (a) a comparison of activation
grouped the leaves into the aforementioned-diseases categories
functions like ReLU, Sigm, and Tanh, (b) learning rate, (c) quantity
(Figure 2).
Figure 2: Three different rice disease classes contained in an image database.
How to cite this article: Damian Mingle, Amit Kumar. Classify Rice Disease Using Self-Optimizing Models and Edge Computing with Agricultural
002
Implications. Agri Res & Tech: Open Access J. 2021; 25 (5): 556316. DOI: 10.19080/ARTOAJ.2021.25.556316
LogicPlum Deep Learning for Computer Vision

As part of our research and analysis, we opted to use an A.I. in- Within research, education, and industry applications, the
novation platform named LogicPlum. LogicPlum includes a library most essential step in a computer vision process is to extract fea-
of proprietary and open-source model types, including linear, tures from the images in a dataset. In this context, a feature is a
non-linear, and deep learning approaches [7]. While manual in- tangible piece of information about a given image, such as color,
terventions are possible during model development, in this study, line, edges, etc. A model needs to observe in order to learn the
the autonomous model builder was used; specifically, the platform characteristics of a given image and thereby classify it correctly.
was provided only with the original data images, and it decided Traditional machine learning approaches allow for several dif-
appropriate configurations for machine learning models automat- ferent feature extraction methods, which require manual feature
ically. Additionally, we chose two different autonomous run types selection and engineering. This process relies heavily on domain
within the LogicPlum platform. The first was Rapid mode, which knowledge, both in computer vision and rice plant disease, to cre-
is designed for model development under 5 minutes. We also ate model inputs that make machine learning algorithms work
used Intensive mode, which is intended for model development better. To increase speed to market for the solution and eliminate
that allows for an undefined time but stops after several rounds the need for expertise in machine learning and plant pathology,
of non-improvement with a given model evaluation metric. The we explored automatically extracting features using deep learn-
software considers several families of algorithms and ranks them ing. The network automatically extracts features and learns their
according to model performance based on the machine learning importance based on the output by applying weights to its connec-
task. Lastly, a combination of base models is automatically evalu- tions. In practice, an individual feeds the raw image to the network
ated, and a subsequent composite model is tested for increased lift and, as it passes through the network layers, the network identi-
before a final solution. fies patterns within the image to create features.
Figure 3: Microarcitecural view: Organization of convolution filters in the fire module. We illustrate the convolution filters only.
Figure 4: Image featurizer becomes model inputs for ExtraTrees Classifier.
We use the SqueezeNet network to extract features from the layer comprised of a 1×1 convolution that reduces the amount of
images. SqueezeNet is a lightweight architecture that is extreme- channels, for example, from 64 to 16 in each image. The squeeze
ly useful in low bandwidth scenarios like mobile platforms and layer aims to compress the data so that the 3×3 convolution does
has ImageNet accuracy similar to AlexNet, the convolution neural not need to learn so many parameters. This is followed by an ex-
network that began the deep learning revolution in 2012 (Figure pand block with two parallel convolution layers: one with a 1×1
3). The first layer demonstrates that the first layer is a squeeze kernel, the other with a 3×3 kernel. These convolution layers also
003
increase the quantity of channels again, from 16 back to 64. Their There are zero fully-connected layers. At the end of the process is
outputs are joined together so the output of this fire module has a convolution layer that performs the classification, followed by
128 channels overall. SqueezeNet has 8 of these Fire modules in the global average [8] (Figure 4).
succession, sometimes with max-pooling layers between them.
Determined Architecture with Rapid Mode

Modeling with ExtraTrees
Figure 5: Stochastic Gradient Descent rice plant disease predictions become model inputs for Gaussian Naive Bayes Classifier.
ExtraTrees Classifier was selected as the top performer. This er [9]. LogicPlum extends the standard stacking algorithm using
classifier fits many randomized decision trees on numerous cross-validation to arrange the input data for the level-2 classifier.
sub-samples of the dataset and uses averaging to enhance the pre-
In the usual stacking procedure, the first-level classifiers fit
dictive accuracy and control over-fitting. ExtraTrees is considered
the same training set used to arrange the inputs for the level-2
a perturb-and-combine technique specifically designed for trees.
classifier, which may lead to overfitting. However, the LogicPlum
Effectively this means that a diverse set of classifiers is designed
approach uses the concept of cross-validation: the dataset is split
by introducing randomness in the classifier construction. The pre-
into k-folds, and in k successive sequences, k-1 folds are used to fit
diction of the collection of weak learners is given as the averaged
the first level classifier. The first-level classifiers are then utilized
prediction of the individual classifiers (Figure 5).
on the remaining 1 subset that was not used for model fitting in
Determined Architecture with Autonomous Mode each iteration in each round. The resulting predictions are then
stacked and provided – as input data – to the second-level clas-
For a composite model to outperform base models, some sam- sifier. After the training of the StackedCVClassifier, the first-level
ples must be better predicted by one model, and other samples classifiers are fit to the entire dataset, as illustrated in the figure
by another model. Stacking is an ensemble learning technique to below. More formally, the Stacking Cross-Validation algorithm can
bring together multiple classification models via a meta-classifi- be summarized as follows: (Table 1)
Table 1: Algorithmic steps for a Cross-Validated Stacked classifier inspired by Wolpert, 1992.
Algorithm Stacking with K-fold Cross Validation
Input : m
=
Training data D { xi , yi }i =1 ( xi ∈ R ^ nyi ∈ Y
Output : An ensemble classifier H
1 : Adopt cross-validation approach in preparing a training set for second-level classifier
2 : Randomly split D into equal-size subsets: D = {D1 , D2 , ...., Dk }

3 : for k ← 1 to K do
4 : Step 1.1: Learn first-level classifiers
5 : for t ← 1 to T do
6 : Learn a classifier hkt form D \ D

k
7 : end for
8 : Step 1.2 Construct a training set for second-level classifier
9 : for xi ∈ Dk do
10 : ' '
Get a record { xi , yi } where xi = {hk1 ( xi ), hk 2 ( xi ), ..., hkt ( xi )}
11 : end for
12 : end for
13 : Step 2: Learn a second-level classifier
004
14 : '
Learn a new classifier from the collection of { xi , yi }
15 : Step 3: Re-learn first level classifiers
16 : for t ← 1 to T do
17 :
Learn a classifier ht based on D
18 : end for
19 : '
return H ( x ) = h ( h1 ( x ), h2 ( x ), ..., hT ( x )
Modeling with Stochastic Gradient Descent
Figure 6: Flow diagram of Stacked Cross-Validation Algorithm.
This estimator applies regularized linear models with sto- Gradient Descent weights across all updates and therefore did not
chastic gradient descent learning; the loss’s gradient is estimated store the results as coefficients. We did not set class weights, and
each sample at a time. The model is revised along the way with a thus, all classes are assigned to a weight of one. Early stopping was
decreasing learning rate [10]. This implementation makes use of not engaged, forcing us to not terminate training when validation
data represented as dense or sparse arrays of floating-point val- scores did not improve. The initial learning rate set was 1.0. We
ues for the features (Figure 6). The model it fits can be monitored did not assume the data was already centered, and chose to es-
with the loss parameter; by default, it fits a linear support vector timate the intercept. We used a squared hinge loss function that
machine. The regularizer is a consequence added to the loss func- is equivalent to Support Vector Classification, but is quadratically
tion that shrinks model parameters in the direction of the zero penalized. For the exponent for inverse scaling learning rate, we
vector using the squared Euclidean norm (L2), the absolute norm used a power_t =0.1. We set the maximum number of passes over
(L1), or a mixture of both (Elastic Net). Many hyperparameters the training data to be 1,000. The L1 ratio is defined with a range
were considered in optimizing for the Stochastic Gradient Descent of 0 to 1, and we set it to 1.0. We used Elastic Net as the penal-
Classifier. The constant that multiplies the regularization term, ty term, which brought sparsity to the model. The learning rate
alpha, is set to 0.0001. In general, the higher the value, the stron- schedule used was inverse scaling,
ger the regularization. We did not compute the average Stochastic
005
eta0
(t ) Connected to farmers for usability
η
= power _t
t Results
Where eta0 and t power _ t are hyperparameters chosen by
LogicPlum. Evaluation Metrics
Modeling with Gaussian Naïve Bayes We use a ground-truth-based approach to compare the results
of various machine learning models. Ground truth is a term used
We implemented the Gaussian Naïve Bayes algorithm for clas- in multiple fields to refer to the information provided by direct
sification. The likelihood of the features is believed to be Gaussian: observation instead of the information provided by inference. We
2
1 ( xi − u y ) understood the machine learning task to be a multiclass classi-
p ( xi | y )
= exp( − )
2 2 fication problem that could be realized in a binary classification
2πσ y 2σ y
model framework.
Where the parameters σ y and µy are estimated using max- The model results were compared concerning the ground
imum likelihood. truth as follows:
The classes’ prior probabilities were not specified as part of Given the definitions of terms within table 2, we can generate
our experiment and therefore were not adjusted according to the standard evaluation metrics for machine learning classification
data. It was determined that variable smoothing should be 1e-9, models:
which was the most considerable variance of all features added to
variances for calculation stability.
Table 2: Definition of terms.
Term Definition
True Positive (TP) Items where the true label is positive and whose class is correctly predicted to be positive.
False Positive (FP) Items where the true label is negative and whose class is incorrectly predicted to be positive.
False Negative (FN) Items where the true label is positive and whose class is incorrectly predicted to be negative.
True Negative (TN) Items where the true label is negative and whose class is correctly predicted to be negative.
Accuracy is defined as the number of items correctly identified much significance is given to recall as to precision. Mathematical-
as either true positive or true negative out of the total number of ly, described as,
items. Mathematically described as, (2 * precision )
(TP + TN ) ( precision + Recall )
(TP + TN + FP + FN )
Macro Average computes the metric independently for each
Recall is defined as the number of items correctly identified as class then averages the results. Mathematically, described as,
positive out of the total actual positives. Mathematically described Pr1 + Pr2 + ... + Prk
as, Prmacro =
k
(TP )
(TP + FN ) Weighted average weights are calculated by the frequency of a
class. Mathematically, described as,
Precision is defined as the number of items correctly identi-
fied as positive out of the total items identified as positive. Mathe- #0bs1 #0bs2 #0bsk
Prweighted −macro
= + + ... + Prk
matically described as, N N N
(TP ) LogicPlum randomly selected 30 images from each class and

(TP + FP ) formed a training dataset of 90 images. The remaining 30 images
were portioned as a test set. The data in both train and test con-
F1 score is defined as the harmonic average of precision and sisted of 513 features (Table 3).
recall, measures the effectiveness of identification when just as
006
Table 3: Rapid results model (2 minutes).
Disease Type Precision Recall F1 Score Support
Bacteria Leaf Blight 0.875 1 0.933 14
Brown Spot 0.882 1 0.938 15
Leaf Smut 1 0.636 0.778 11
Accuracy 0.9 40
Macro Avg 0.919 0.879 0.883 40
Weighted Avg 0.912 0.9 0.892 40
Model Performance for model performance. This measure aims to answer the ques-
tion, “What proportion of actual positives was identified correct-
Statistics
ly?” In the context of the Intensive model, which was completed in
Our primary evaluation metric for model performance was ac- 60 minutes, we observed accuracy of 92.5% across all rice disease
curacy. We observed accuracy of 0.90 across all rice plant disease classes. However, the lowest secondary measure is recall as it re-
classifications on the Rapid model’s validation dataset. Recall, as lates to Leaf Smut (Table 4).
it relates to Leaf Smut, is the lowest secondary evaluation metric
Table 4: Intensive results model (60 minutes).
Disease Type Precision Recall F1 Score Support

Bacteria Leaf Blight 0.824 1 0.903 14
Brown Spot 1 0.933 0.966 15
Leaf Smut 1 0.818 0.9 11
Accuracy 0.925 40
Macro Avg 0.941 0.917 0.923 40
Weighted Avg 0.938 0.925 0.926 40
To completely evaluate the effectiveness of a model, we must ing the F1 score from both the Rapid and Intensive mode, we can
examine both precision and recall. Precision and recall are often in observe that the Intensive mode does significantly better at classi-
tension. That is, improving precision typically reduces recall and fying Leaf Smut than the Rapid mode, with a 15.68% increase. It is
vice versa. Thus, many machine learning practitioners rely on the worth noting that while the Intensive mode is superior in almost
F1 score, which combines the effects of both precision and recall. every respect, it does show a percentage decrease of 3.21% when
An F1 score is considered perfect if it reaches 1.0. When compar- considering the F1 score for Bacterial Leaf Blight.
Confusion Matrix
Below are the confusion matrices for both models.
Table 5: Confusion matrix for LogicPlum Rapid mode.
007
Table 5 illustrates where the Rapid mode made incorrect pre- misclassification that occurs in two classes, Brown Spot and Leaf
dictions: Leaf Smut for the True Class should be 11, and instead Smut. However, the total misclassification rate for Intensive was
we have 7. We incorrectly classified 4 Leaf Smut cases as Bacteri- lower by 25% over Rapid mode. Additionally, Bacterial Leaf Blight
al Leaf Blight in two instances and Brown Spot in the remaining offered new improvement, and Brown Spot created some minor
instances (Table 6). In the case of the Intensive mode, there was confusion for the Intensive mode.
Table 6: Confusion matrix of LogicPlum Intensive mode.
Comparative Accuracy of Models Relation to other results

Our experiment was conducted in LogicPlum cloud and only We achieved 90.0% on the test dataset with Rapid mode,
leveraged a CPU configuration. As seen in Table 4, we achieve test which builds the A.I. solution from data upload to prediction with-
accuracy of 90% with the Rapid results model, whereas with the in 2 minutes. Additionally, we achieved 92.5% accuracy on the test
Intensive results model, accuracy goes up to 92.5%. Barring one dataset, which has a training time that completes within 60 min-
image, all the test images belonging to Bacteria Leaf Light and utes. Both approaches increase detection accuracy for rice plant
Brown Spot are correctly classified. disease over the prior research, which achieved 73.33% accuracy
on the dataset [6]. As it relates to model performance, the Rapid
Discussion mode exhibits a 22.73% increase over the prior research, while
Summary of conclusions: the Intensive mode demonstrates a 26.14% percent increase.
Furthermore, we reduced the number of technical steps taken by
This paper proposed two new approaches for detecting dis-
practitioners in the preceding study, from 11 steps to 5 steps in the
ease in rice plants, Rapid mode and Intensive mode, using a mea-
case of Rapid mode, and 6 steps in the Intensive mode—a 54.54%
ger number of images for training a classifier. We provide an in-
and 45.45% decrease, respectively, over the prior research (Figure
depth analysis of our methods, which outperform the original
7).
paper results on the same dataset with significantly fewer ma-
chine learning techniques. Future work involves exploring the
edge computing capabilities of these methods.
Figure 7: Model performance by approach to identify bacterial leaf blight, brown spot, and leaf smut within rice plants.
008
Prior methods and privacy. Edge computing is at the center of several IoT agricul-
tural applications, such as pest identification, safety traceability
This paper evaluated four techniques of background removal
of farm products, unmanned agrarian machinery, agrarian tech-
by applying masks generated based on the following: (1) original
nology promotion, and in this case, classifying diseases from the
image, (2) hue component values of the image in HSV color space,
images of rice leaves purely because of its speed and efficiency
(3) value components of the image in HSV color space, and finally
compared to the cloud infrastructure. It offers a potentially tracta-
(4) saturation component values of the image in HSV color space.
ble model for mainstreaming smart agriculture [11]. Agriculture
Three techniques of segmentation were utilized: (1) LAB color
IoT systems can make informed decisions in the field when using
space based K-means clustering, (2) Otsu’s segmentation tech-
edge computing [12].
nique, and (3) HSV color space based K-means clustering. Using
various features under three categories: color, texture, and shape, We propose an approach that allows for access to our A.I.
the authors extracted 88 features from the disease portion of a solution without an internet connection in the field. Figure 8 (A)
leaf image. Finally, the paper used Support Vector Machines with illustrates the process of a farmer in a field who needs access to
Gaussian kernel for multiclass classification of the leaf images. rice plant disease classification via her smartphone and does not
have access to a network connection. The farmer can make use of
Implications the classification algorithm as it is embedded on the phone. (B)
Edge computing for smartphone users demonstrates that the trained model is converted to a LogicPlum
Lite file type, which is how the model becomes executable on a
Edge computing has the capability to address the concerns of
mobile phone device. Figure 8.C exemplifies the concept of return-
bringing machine learning approaches to the farming fields. Spe-
ing to a location that supplies network connection, and a transfer
cifically, edge computing deals with response time requirements,
occurs. If an update exists, then an update is made available.
battery life consumption, bandwidth cost savings, and data safety
Figure 8: Process makes rice plant disease detection available at the edge of the network and allows for intelligent updates when available.
Borrowing knowledge from plant experts plant disease, shifts farming further into the 21st century. Many
education areas have evolved into a self-paced process of finding
Making expert plant knowledge readily available to farmers in
information or learning skills exactly when and where they are
the field promises a meaningful impact. Edge computing allows
needed, and farming is no different. Our approach offers flexible
farmers with a mobile app to capture the image of infected rice leaf
delivery of learning to farmers with an “anytime, anyplace” frame-
and classify the disease, thereby greatly reducing the need for con-
work. This approach allows farmers to independently access in-
sultation with plant pathologists, which can be a time-consuming
formation from plant pathology databases in the context of what
process. Furthermore, once a condition is detected, appropriate
they observe within their environment. This approach is linked
expert measures can be applied with various management strate-
to the idea of lifelong learning, learner-driven learning, and proj-
gies, namely preventive methods, cultural methods, and chemical
ect-based learning. We have organized expert remedies for each of
methods (Figure 9). The Next Action Model is built on a concept of
the three rice disease classes we analyzed: Rice Leaf Blight, Brow
just-in-time learning, which meets farmers where they are instead
Spot, and Leaf Smut. According to Tamil Naud Agricultural Univer-
of requiring structured education to form concept knowledge. The
sity, each of these rice diseases has three management strategies
advent of our machine learning approach, coupled with edge com-
categorized as preventive, cultural, and chemicall (Tables 8-14).
puting and remedies for specific management strategies of rice
009
Figure 9: Rice farmers can capture images of suspected rice disease and have an expert system classify the disease. Additionally, once the
classification is made available, remedies for preventive, cultural, and chemical methods are presented in real-time.
Table 7: The contents of preventive methods for rice leaf blight were created by Tamil Nadu Agricultural University.
Methods Category Actions

Seed treatment with bleaching powder (100g/l) and zinc sulfate (2%) reduces bacterial blight.
Seed treatment - seed soaking for 8 hours in Agrimycin (0.025%) and wettable ceresan (0.05%) followed by hot
water treatment for 30 min at 52-54oC
Preventive – Rice Leaf Blight Seed soaking for 8 hours in ceresan (0.1%) and treat with Streptocyclin (3g in 1 liter)
Spray neem oil 3% or NSKE 5%
Spray fresh cow dung extract for the control of bacterial blight. Dissolve 20 g cow dung in one liter of water; allow
to settle and sieve. Use supernatant liquid. (starting from the initial appearance of the disease and another at
fortnightly interval)
Table 8: The contents of cultural methods for rice leaf blight were created by Tamil Nadu Agricultural University.
Methods Category Action
Grow Tolerant varieties (I.R. 20 I.R. 72,PONMANI , TKM 6).

Secure disease-free seed.
Grow nurseries preferably in isolated upland conditions.
Avoid clipping of seedlings during transplanting.
Cultural – Rice Leaf Blight
Balanced fertilization, avoid excess N – application.
Skip N - application at booting (if the disease is moderate).
Drain the field (except at the flowering stage of the crop).
Destruction of weeds and collateral hosts.
Avoid flow of water from affected fields.
Maintain proper plant spacing.
Table 9: The contents of chemical methods for rice leaf blight were created by Tamil Nadu Agricultural University.
Methods Category Action
Reduce bacterial leaf blight by seed treatment with bleaching powder (100g/l) and zinc sulfate (2%).
Seed treatment - seed soaking for 8 hours in Agrimycin (0.025%) and wettable ceresan (0.05%) followed by hot
water treatment for 30 min at 52-54oC.
Chemical – Rice Leaf Blight seed soaking for 8 hours in ceresan (0.1%) and treat with Streptocyclin (3g in 1 liter).
Spray Streptomycin sulphate + Tetracycline combination 300 g + Copper oxychloride 1.25kg/ha. If necessary
repeat 15 days later.
Application of bleaching powder @ 5 kg/ha in the irrigation water is recommended in the kresek stage.
Foliar spray with copper fungicides alternatively with Strepto-cyclin (250 ppm) to check secondary spread.
0010
Table 10: The contents of preventive methods for rice brown smut were created by Tamil Nadu Agricultural University.
Utilize disease-free seeds that are selected from a healthy crop.
Seed treatment with carbendazim 2.0g/kg of seeds.

Preventive – Rice Smut
Control insect pests.
Split application of nitrogen is recommended.
Removal and proper disposal of infected plant debris.
Table 11: The contents of cultural methods for rice smut spot were created by Tamil Nadu Agricultural University.
Among the cultural control, the destruction of straw and stubble from infected plants is recommended to reduce
the disease.
Use varieties that are found to be resistant or tolerant against the disease in India.
Avoid field activities when the plants are wet.
Early planted crop has less smut balls than the late crop.
Cultural Methods – Rice Smut When harvesting, diseased plants should be removed and destroyed so that sclerotia does not fall in the field.
This will reduce the primary inoculum for the next crop.
Field bunds and irrigation channels should be kept clean to eliminate alternate hosts.
Excess application of nitrogenous fertilizer should be avoided.
Regular monitoring of disease incidence during rabi season is very essential.
Proper destruction of straw and stubble.
Table 12: The contents of chemical methods for rice smut spot were created by Tamil Nadu Agricultural University.
Spraying either copper oxychloride at 2.5 g/liter or Propiconazole at 1.0 ml/liter at boot leaf and milky stages will
help prevent the fungal infection.
Seed treatment with carbendazim 2.0g/kg of seeds.
Chemical Methods – Rice Smut At tillering and pre-flowering stages, spray Hexaconazole @ 1ml/lit or Chlorothalonil 2g/lit.
In areas where the disease may cause yield loss, applying captan, captafol, fentin hydroxide, and mancozeb can be
inhibited by conidial germination.
At tillering and pre-flowering stages, spraying of carbendazim fungicide and copper base fungicide can effectively
control the disease.
Table 13: The contents of preventive methods for rice brown spot were created by Tamil Nadu Agricultural University.
Seed treatment with Pseudomonas fluorescens @ 10g/kg of seed followed by seedling dip @ of 2.5 kg or products/
ha dissolved in 100 liters and dipping for 30 minutes.
seed soak/seed treatment with Captan or Thiram at 2.0g /kg of seed

Preventive – Rice Brown Spot
Seed treatment with Agrosan or Ceresan 2.5 g/kg seed to ward off seedling blight stage
Since the fungus is seed transmitted, a hot water seed treatment (53-54°C) for 10-12 minutes may be effective be-
fore sowing. This treatment controls primary infection at the seedling stage. Presoaking the seed in cold water for 8
hours increases effectivity of the treatment.
0011
Table 14: The contents of cultural methods for rice brown spot were created by Tamil Nadu Agricultural University.
As the disease is seed-borne, use disease-free seeds.
Removal of alternate & collateral hosts.

The use of resistant varieties is the most economical means of control.
Growing Resistant varieties like ADT 44,PY 4,CORH 1,CO 44,CAUVERY,BHAVANI,TPS 4 and Dhanu.
Cultural – Rice Brown Spot
Providing proper nutrition for optimum plant growth and prevention of water stress seem to be the most important
factors in controlling brown spot.
The disease is rarely observed in normally fertile rice soils.
Soils known to be low in plant-available silicon should be amended with calcium silicate slag before planting, and the
field should be well irrigated to avoid water stress.
Data science knowledge edge commuting one of the emerging technologies staged to
transform the agricultural industry. With sensors, actuators, and
Our approach leverages an automated machine learning pro-
real-time data-driven models, digitization can help us overcome
cess that allows for rapid experimentation on real-world prob-
some of the biggest challenges of our time [14]. Autonomous trac-
lems. This approach covers the entire process from beginning to
tors and robotic machinery, often known as Agribots, can run on
end, more specifically, from uploading the data to the deployment
autopilot, communicating with nearby sensors to acquire the nec-
of a machine learning classifier, with little to no human interac-
essary data about the surrounding environment. The introduction
tion. This approach has data science expertise built into the pro-
of drones has shown great promise with agricultural implications.
cess, offering guardrails for lay users of machine learning. In this
These unmanned aerial vehicles can help in various ways, such as
approach, the emphasis is placed on the creative use of the tech-
monitoring livestock and crop growth, and increasing output with
nology rather than the details of a given algorithm.
real-time insights. Additionally, the introduction of the 5G mobile
Effects of age, educational level, and adoption of farm- network, which is designed to connect virtually everyone and ev-
ing practices erything together, including machines, objects, and devices, will
further drive the adoption of digital farming techniques.
Children who were raised on family farms are familiar with
the farming practices that have proven successful for their par- Precision farming
ents. So, even when younger family members don’t make identical
Technology has become an imperative consideration for ev-
decisions to those of their parents, their decisions will continue
ery stakeholder involved in agriculture, starting from farmer to
to be informed by years spent under their parents’ guidance [13].
agronomist. Precision farming makes farming more accurate and
This is known as multi-generational farming, which often doesn’t
controlled when it comes to growing crops and raising livestock.
involve technology in agriculture.
It can decide on and carry out the best technical intervention in
According to Moore’s law, computer processing speed doubles the right place at the best possible moment. It makes it simpler
every 18 or so months, and a generation is generally understood to plan ahead of time and to act precisely in terms of space. A vi-
to be between 20 and 30 years. This means that processing speeds tal component of the precision farming management approach
may double 20 times during a given farming generation, allowing is the use of technology with its wide array of instruments, such
for more insight and actionable machine learning models (Kole- as robotics, drones, variable rate technology, sensors, GPS-based
va,2021). Although former generations may not have been raised soil sampling, telematics, and software. A balance must be found
with digital technology, such significant enhancements in machine between precision farming, capable of determining the correct,
learning model performance, along with edge computing, should limited scale of mediation at the right time and in the right place,
encourage adoption within agriculture, requiring new behaviors and a preventive, systemic approach empowering a cultivated
and ways of thinking. We believe that just like rakes, hoes, and ecosystem to produce without the need for curative treatments.
shovels are essential for today’s farmers, machine learning will be Digital technology will make it possible for targeted interventions,
added to the basic set of farming tools in the 21st century. through data processing, forecasting and anticipating, simulating,
and safeguarding [15].
Digital farming techniques
Our approach is additive in the context of modern agricultural
Conclusion
methods. Successfully delivering productive and sustainable ag- The best prediction statistics were achieved with a Gauss-
ricultural systems worldwide will help form the foundations for ian Naïve Bayes stacked classifier that used Stochastic Gradient
overcoming food insecurity and hunger. Economic viability makes Descent Classifiers predictions as model inputs. The automated
0012
model construction approach resulted in a validation set of 92.5% 5. Santra B, Mukherjee DP (2019) A comprehensive survey on computer
accuracy. Therefore, it can be recommended for use, with little to vision based approaches for automatic identification of products in
retail store. Image and Vision Computing 86: 45-63.
no involvement from a machine learning expert or trained plant
pathologist. Our approach ranged from as much as 60 minutes in 6. Prajapati HB, Shah JP, Dabhi VK (2018) Detection and classification of
rice plant diseases. Intelligent Decision Technologies, 1-18.
total time to 2 minutes. Since our method was automated com-
pared to a manually crafted process, it is faster loading the data, 7. Logic Plum, Inc. (n.d.). A.I. Innovation Platform. LogicPlum.
model construction, optimization, and deployment. This method 8. Bhole V, Kumar A (2020) Analysis of Convolutional Neural Network
is inexpensive compared to other methods, not only in time but in Using Pre-Trained Squeezenet Model for Classification of Thermal Fruit
Images. ICT for Competitive Strategies, pp: 759-768.
economic terms, as our method only uses CPU rather than GPU ar-
chitecture. Our approach cut the number of steps in half compared 9. Džeroski S, Ženko B (2004) Is Combining Classifiers with Stacking
Better than Selecting the Best One? Machine Learning 54(3): 255-273.
to prior methods and is also self-optimizing, permitting users of
this approach to be hands-free. Additionally, our process does 10. Gardner WA (1984) Learning characteristics of stochastic-gradient-
descent algorithms: A general study, analysis, and critique. Signal
not end with the identification of rice plant disease. Instead, we Processing 6(2): 113-133.
combined management strategies for specific rice diseases from
11. Pérez-Pons ME, Plaza-Hernández M, Alonso RS, Parra-Domínguez J,
known plant experts using edge computing. This was chosen to in- Prieto J (2020) Increasing Profitability and Monitoring Environmental
crease accessibility to the machine learning approach, and allows Performance: A Case Study in the Agri-Food Industry through an Edge-
for our system to meet more farmers where they are and when IoT Platform. Sustainability 13(1), 283.
they need it [16-20]. 12. Tang Y, Dananjayan S, Hou C, Guo Q, Luo S, et al. (2021) A survey
on the 5G network and its impact on agriculture: Challenges and
Acknowledgments opportunities. Computers and Electronics in Agriculture 180: 105895.
We would like to thank the University of California, Irvine’s 13. Wójcik M, Jeziorska-Biel P, Czapiewski K (2019) Between words: A
generational discussion about farming knowledge sources. Journal of
Center for Machine Learning and Intelligent Systems, for making
Rural Studies 67: 130-141.
the images of rice plant disease available, and Logic Plum, Inc. for
14. Carmela Annosi M, Brunetta F, Capo F, Heideveld L (2020) Digitalization
supporting our study.
in the agri-food industry: the relationship between technology and
sustainable development. Management Decision 58(8): 1737-1757.
Conflict of Interest
15. Jacob PM (2018) A Comparative Analysis on Smart Farming Techniques
This research is sponsored by LogicPlum and may lead to using Internet of Things (IoT). HELIX 8(2): 3294-3302.
the development of products that may be licensed. Amit Kumar
16. Fan FAN, Roy T, Roy K (2020) Classification and Detection Rice leaf
receives a salary from LogicPlum. Damian Mingle is the founder of Diseases Using Information and Communication Technology (ICT)
LogicPlum and receives other financial benefits.. Tools. International Journal of Advanced Engineering Research and
Science 7(6): 460-470.
References 17. Koleva MK (2021) New Approach to Deep Miniaturization: A Way to
1. International Rice Research Institute (2021) (n.da.). Rice to zero Overcoming Moore’s Law. Journal of Modern Physics 12(03): 167-178.
hunger. IRRI.Org.
18. Plastiras G, Terzi M, Kyrkou C, Theocharidcs T (2018) Edge intelligence:
2. Sinha P, Balas B, Ostrovsky Y, Russell R (2006) Face Recognition by Challenges and opportunities of near-sensor machine learning
Humans: Nineteen Results All Computer Vision Researchers Should applications. 2018 IEEE 29th International Conference on Application-
Know About. Proceedings of the IEEE 94(11): 1948-1962. specific Systems, Architectures and Processors (ASAP).
3. Noble JA (1995) From inspection to process understanding and 19. Tamil Nadu Agricultural University (n.d.). Bacterial Leaf Blight. TNAU.
monitoring: a view on computer vision in manufacturing. Image and Ac.In.
Vision Computing 13(3): 197-214.
20. Wolpert DH (1992) Stacked generalization. Neural Networks 5(2):
4. Sahni S, Mittal A, Kidwai F, Tiwari A, Khandelwal K (2020) Insurance 241-259.
Fraud Identification using Computer Vision and IoT: A Study of Field
Fires. Procedia Computer Science 173: 56-63.
0013
This work is licensed under Creative

Commons Attribution 4.0 License
Your next submission with Juniper Publishers
DOI:10.19080/ARTOAJ.2021.25.556316 will reach you the below assets
• Quality Editorial service
• Swift Peer Review
• Reprints availability
• E-prints Service
• Manuscript Podcast for convenient understanding
• Global attainment for your research
• Manuscript accessibility in different formats
( Pdf, E-pub, Full Text, Audio)
• Unceasing customer service
Track the below URL for one-step submission

https://juniperpublishers.com/online-submission.php
0014

Artoaj MS Id 556316

Uploaded by

Copyright:

Available Formats

Artoaj MS Id 556316

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Artoaj MS Id 556316

Uploaded by

Copyright:

Available Formats

Research Article Agri Res & Tech: Open Access J

Volume 25 Issue 5 - April 2021

Classify Rice Disease Using Self-Optimizing Models and

Figure 2: Three different rice disease classes contained in an image database.

LogicPlum Deep Learning for Computer Vision

Figure 4: Image featurizer becomes model inputs for ExtraTrees Classifier.

Determined Architecture with Rapid Mode

Algorithm Stacking with K-fold Cross Validation

2 : Randomly split D into equal-size subsets: D = {D1 , D2 , ...., Dk }

6 : Learn a classifier hkt form D \ D

Modeling with Stochastic Gradient Descent

Figure 6: Flow diagram of Stacked Cross-Validation Algorithm.

(TP ) LogicPlum randomly selected 30 images from each class and

Table 3: Rapid results model (2 minutes).

Disease Type Precision Recall F1 Score Support

Bacteria Leaf Blight 0.875 1 0.933 14

Brown Spot 0.882 1 0.938 15

Leaf Smut 1 0.636 0.778 11

Macro Avg 0.919 0.879 0.883 40

Weighted Avg 0.912 0.9 0.892 40

Table 4: Intensive results model (60 minutes).

Disease Type Precision Recall F1 Score Support

Comparative Accuracy of Models Relation to other results

Methods Category Actions

Methods Category Action

Grow Tolerant varieties (I.R. 20 I.R. 72,PONMANI , TKM 6).

Methods Category Action

Methods Category Actions

Utilize disease-free seeds that are selected from a healthy crop.

Seed treatment with carbendazim 2.0g/kg of seeds.

Methods Category Actions

Methods Category Actions

Seed treatment with carbendazim 2.0g/kg of seeds.

Methods Category Actions

seed soak/seed treatment with Captan or Thiram at 2.0g /kg of seed

Methods Category Actions

As the disease is seed-borne, use disease-free seeds.

Removal of alternate & collateral hosts.

This work is licensed under Creative

Track the below URL for one-step submission

You might also like

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.