4616 IEEE SENSORS JOURNAL, VOL. 22, NO.

5, MARCH 1, 2022

RiceBioS: Identification of Biotic Stress in Rice

Crops Using Edge-as-a-Service
Pranjal Joshi , Graduate Student Member, IEEE, Debanjan Das , Senior Member, IEEE,
Venkanna Udutalapally , Senior Member, IEEE, Monoj K. Pradhan,
and Sudip Misra , Fellow, IEEE

Abstract —The identification of biotic stress of rice crops

using handheld sensing devices is a challenge, as computa-
tionally intensive machine learning models are difficult to be
executed in these devices. This challenge is exacerbated in
farmers’ fields located in remote regions with limited internet
connectivity. Thus, an individualiased plant-specific solution
to detect biotic stress due to crop infections is required
in farms adopting digital agricultural practices. The existing
biotic stress detection solutions are deficient in their ability
to make decisions in real-time. It is required to have a system
that is capable of making decisions at the edge in handheld
devices having limited computational capability. This paper
proposes RiceBioS, an AI-based deep learning-enabled hand-
held device for identifying biotic stress in rice crops using the computational capabilities of handheld devices. RiceBioS
adopts Edge-as-a-Service (EaaS) as an approach for classifying rice crop images into two categories – healthy and
stressed. The biotic stress condition is further diagnosed into two types of infections, fungal (rice blast) and bacterial
(bacterial leaf blight of rice) by pruning the shrunk deep learning classification model and incorporating an automated RoI
detection and feature extraction workflow, which makes use of adaptive thresholding and hierarchial masking techniques
to perform dimensionality reduction. While RiceBioS demonstrates a test accuracy of 93.25%, it exhibits a negligible
tradeoff on a smartphone after deployment. This cutting edge solution helps the farmers make informed decisions based
on real-time insights provided by the user-friendly mobile application interface of RiceBioS.
Index Terms — Biotic stress identification, convolutional neural network (CNN), digital agriculture, edge-as-a-service
(EaaS), edge computing, Internet-of-Agro-Things (IoAT), machine learning.

I. I NTRODUCTION ture change, soil nutrient condition, water sufficiency, pests

and diseases. One third of the total affected crops are due
T HE factors which directly or indirectly affect the pro-
ductivity of crops are climatic conditions, tempera- to biotic stress crop infections [1]. The untimely pest attacks
and inaccurate crop insights that farmers receive lead to wrong
Manuscript received December 27, 2021; accepted January 12, agricultural inputs resulting in biotic stress in crops as depicted
2022. Date of publication January 20, 2022; date of current version in Fig. 1. This establishes the fact that when biotic stress leads
February 28, 2022. This work was supported by the Ministry of Elec- to crop infections is a problem that needs to be addressed. The
tronics and Information Technology (MeitY), Government of India, under
Grant 13(10)/2020-CC&BT. The associate editor coordinating the review stress undergone by rice crops during every calendar year not
of this article and approving it for publication was Prof. Huang Chen Lee. only hits the production globally at the macro level but majorly
(Corresponding author: Debanjan Das.) affects the life of a farmer at the micro level. The relevance
Pranjal Joshi is with IIIT Naya Raipur, Naya Raipur, Chhattisgarh
493661, India (e-mail: [email protected]). of this research can directly be correlated with the impact that
Debanjan Das is with the Department of Electronics and Communica- rice crop yield and its productivity has on human population
tion Engineering, IIIT Naya Raipur, Naya Raipur, Chhattisgarh 493661, (on the demand side) and on the farmer community (on the
India (e-mail: [email protected]).
Venkanna Udutalapally is with the Department of Computer Science supply side).
and Engineering, IIIT Naya Raipur, Naya Raipur, Chhattisgarh 493661, The infrastructure required for digital agriculture, crop
India (e-mail: [email protected]). monitoring and biotic stress management suffers because of
Monoj K. Pradhan is with the Department of Computer Science and
Engineering, Indira Gandhi Krishi Vishwavidyalaya, Raipur 492012, India uncontrollable factors such as low network bandwidth, absence
(e-mail: [email protected]). of high computing resources and lack of support for low
Sudip Misra is with the Department of Computer Science and Engi- end user devices. A user friendly, fast and accurate rice
neering, Indian Institute of Technology Kharagpur, Kharagpur 721302,
India (e-mail: [email protected]). crop biotic stress identification system which overcomes the
Digital Object Identifier 10.1109/JSEN.2022.3143950 challenges shown in Fig. 1 will help the farmer with early

1558-1748 © 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: KGiSL Institute of Technology. Downloaded on March 11,2024 at 11:43:37 UTC from IEEE Xplore. Restrictions apply.
JOSHI et al.: RiceBioS: IDENTIFICATION OF BIOTIC STRESS IN RICE CROPS USING EDGE-AS-A-SERVICE 4617

• Edge-as-a-service for real-time reporting of infection sta-

tus with type.
• An automated dimensionality reduction, feature extrac-
tion and enhancement workflow for image pre-processing
is proposed.
• The edge based DL model is trained and tested on an
indigenously built rice crop image dataset.
• A mobile application for farmers that makes fast decisions
on the edge.
• An end to end biotic stress management solution that
connects a farmer to the farm intelligently and seamlessly.

Fig. 1. Challenges in identification of biotic stress in rice crops.

II. R ELATED W ORKS
A. Related Prior Art
Traditionally, biotic stress is estimated in plants using
laboratory based methods such as studying the membrane
integrity by tracking electrolyte leakage, examining cell half
cycle by trypan blue staining and measuring the depth of
callose deposition by aniline blue staining [13]. However,
Fig. 2. Proposed RiceBioS solution through EaaS.
these methods are lab based and best suited for studying the
different causes of biotic stress but may not be affordable and
diagnosis of stress condition in crops through a smartphone. accessible to the farmer. If non laboratory based or modern
The gross productivity of any farm can be increased by regular sensing methods are to be considered, thermography is used
monitoring and prompt action on any anomalies [2]. Thus this for detection of both abiotic and biotic stresses, apart from
paper dives into the digital agriculture ecosystem to provide methods such as hyperspectral imaging, multispectral imag-
an end-to-end solution in terms of an edge based biotic stress ing, flouroscence spectrography and combination of sensors
management in rice crops. The most important categories [14]–[16]. However, lab based methods or different imaging
of rice crops and their deadly fungal and bacterial biotic mechanisms are difficult to be deployed at the edge. Therefore,
stresses (bacterial leaf blight of rice and rice blast) have been some studies to overcome the implementation challenges faced
considered for the study and experiment [3]. in networked node based monitoring of crop vegetation have
Existing technologies in the field of precision and smart also been undertaken [17].
agriculture are focused on solving problems which have In the literature, there are several research works to predict
already occurred and have affected the crop. Early diagnosis rice crop diseases, [5] and [6] based on image recognition
and on-device identification of biotic stress in crops is our framework with an accuracy of 95.74%. However, these mod-
main contribution. The prevention of infections in crops leads els depend on high computing resources and large dataset.
to reduced water consumption during irrigation which is a Similarly, Wang et al. [7] and Dwivedi et al. [18] propose a
scarce resource [4]. RiceBioS offers an end-to-end solution network for disease recognition in wheat and grape crops,
for farmers, as depicted pictorially in Fig. 2. which also relies on cloud service. Though giving high accu-
racy on the used dataset, an attempt to use edge computing
A. Proposed Solutions Through RiceBioS and run on mobile platforms is not made. Xie et al. [19] uses
digital image processing techniques and statistical machine
This paper proposes a deep learning framework RiceBioS
learning methods to predict leaf diseases in wheat crop. While
paired up with a mobile application which:
the average recognition rate of the system was 88.89%, the
• Provides an interactive user interface to the farmer to
diagnosis relies on the server for decisions and the user
capture images of rice crops through a handheld device. equiment is not designed to handle the computational load.
• Performs image preprocessing to extract features that are
On similar lines, Hasan et al. [10] proposes a fusion of
indicators of biotic stress. SVM and DCNN methods where the output of the DCNN
• Makes on-device prediction of biotic stress conditions
model in terms of extracted features is fed to the SVM
employing Edge-as-a-Service (EaaS). classifier to give a prediction among 12 classes of rice infec-
• Suggests intuitive measures to the farmers in order to
tions. However, the portability of this solution is difficult
boost crop yield or productivity. considering the size of the learning model. A microscopy
image classification method was proposed by Yang et al. [11]
B. Novel Contributions of RiceBioS where texture and shape analysis of infected rice crops is done
The major contributions of this paper are listed below: through histogram equalization technique and a microscope.
• A smartphone-based sensing system utilizing an opti- A microscope to study the rice crop at the micro level is a
mised deep learning framework RiceBioS for identifying prerequisite to this experimental setup.
biotic stress in rice crops intended to minimize the While the results in detecting the blast disease in rice crops
training and computation time, and storage requirement. are quite promising in RiceTalk [9], this paper uses sensor

Authorized licensed use limited to: KGiSL Institute of Technology. Downloaded on March 11,2024 at 11:43:37 UTC from IEEE Xplore. Restrictions apply.
4618 IEEE SENSORS JOURNAL, VOL. 22, NO. 5, MARCH 1, 2022

TABLE I
C OMPARISON OF E XISTING S OLUTIONS W ITH R ICE B IO S

based input fetching methods and leverages IoT for making

AI predictions on sensed inputs. This relies on statistical data.
Reference [12] makes use of a vital parameters on top of
disease prediction scores to predict rice blast disease specific
to considered area.

B. Research Gaps
The solutions compared above have in their own specific
way contributed in rice crop biotic stress monitoring and
disease prediction by either proposing an accurate classifica-
Fig. 3. Conceptual view of RiceBioS sensing system provisioning
tion model, image processing analysis or a fusion of sensed edge-as-a-service.
data and images to make decisions. A feature wise tabulated
description of the discussed prior art in subsection II-A is TABLE II
given in Table I. The major issues that were not found to be R ICE C ROP I MAGE D ATASET D ISTRIBUTION
diligently addressed in most of the prior art are:
• Few attempts made to perform AI-driven crops’ biotic
stress identification.
• Automated biotic stress detection and classification meth-
ods were not provided.
• Absence of edge-based solutions focusing on crop health.
• Require high computational resources to deploy classifi- fungal biotic stress conditions are two among the many stress
cation models. conditions or diseases that normally infect any rice crop.
To create the three way classification image dataset of the
III. R ICE B IO S: T HE P ROPOSED S ENSING S YSTEM above mentioned rice diseases, the whole plant and its infected
We propose a system RiceBioS which enables the rice images were taken into consideration. The entire work of
growing farmer to conveniently identify the current state of building and collection of dataset was carried out with the help
the cultivated rice crop. The algorithm segregates the state of our collaborator - Indira Gandhi Krishi Vishvavidyalaya,
of the rice crop in three different prominent categories. The Raipur. To ensure that the framework does not completely rely
crop can either be healthy or stressed. In case the status on the image capturing capabilities of the device the dataset
predicted by the algorithm is stressed, the undergone biotic has a mix of images taken from smartphone camera (Samsung
stress is broadly categorised into two types - bacterial and Galaxy M11) and professional digital camera (Nikon D5600
fungal. This classification of a crop into healthy state, biotic DSLR) with different aperture, focusing abilities and camera
stressed - bacterial or biotic stressed - fungal, helps the farmer quality. The proposed framework is versatile enough to comply
to preempt future consequences to the crop and thus take to variations in image size, capturing environment such as
suitable action. These predictions are done on the device brightness and illumination values. The image dataset was
without having to push any data to a centralized server. The validated by agriculture experts at IGKV, Raipur. A brief detail
functional representation of the process is given in Fig. 3. about the image dataset is given in the Table II.

A. Dataset for Training B. Image Preprocessing and Augmentation

The major contribution in this model has been building The proposed architecture primarily takes an image as input.
our own dataset of images of rice crop (paddy). Bacterial and Image as an input to any learning model defines the input lines

Authorized licensed use limited to: KGiSL Institute of Technology. Downloaded on March 11,2024 at 11:43:37 UTC from IEEE Xplore. Restrictions apply.
JOSHI et al.: RiceBioS: IDENTIFICATION OF BIOTIC STRESS IN RICE CROPS USING EDGE-AS-A-SERVICE 4619

Fig. 4. Sample of image preprocessing results include (a) Original Fig. 5. (a) Original image, (b) Perspective transformation, (c) Affine
image, (b) Grayscale converted image, (c) ‘a’ channel of LAB color space, transformation, and (d) Rotation.
(d) Histogram equalization output, (e) Dilated resultant and (f) 2D Otsu
adaptive threshold output.

Fig. 4(f) is a resultant of the iterative 2D Otsu adaptive thresh-

olding technique [8] applied on the dilated output obtained
as height × wi dth × channel. Here, the image dimensions
in Fig. 4(e).
are 256 × 256 × 3 (3 for RGB color space).The acquisition
Normally, whenever a deep learning model is trained with
process involves capturing the leaf from the image capturing
input having parameters on the higher side makes the dimen-
device (smartphone in this case).
sionality huge. When the decision depends on too many
The ideal condition that the algorithm would relish as an
parameters and more importantly many of them being irrel-
input is a leaf image with no or near to zero background
evant (due to background noise), the model tends to learn
noise with ideal lighting conditions in addition to the minimum
things very specific to the input provided and is often only
ensured distance at the optimum angle. While some of these
receptive to training images. This phenomenon is known as
conditions can be enforced and made as a part of the practice,
overfitting. Having created a dataset of raw images on field,
conditions such as ensuring zero background noise are quite
figuring out the best possible network for them to be trained
difficult to meet. Thus the acquisition process comprises steps
is a challenge. Therefore in this scenario, the preprocessing
such as image capturing (with above conditions), storing the
workflow ensures a targeted approach for background removal
image locally and then making it available for the application
and RoI feature extraction using hierarchial masking and
to access the same for classification purpose. Image enhance-
adaptive thresholding.
ment is done in order to eliminate as much noise possible from
As a subsequent step to this input pipeline to RiceBioS
the digital image.
model, data augmentation is done on the entire dataset of
The several iterations in the process of enhancing the leaf
rice crop images. The modus operandi of this practice is
images are shown in Fig. 4. The image shown in Fig. 4(a)
to add distortions and variations to training images which
is converted into gray scale (in Fig. 4(b)) and different color
cumulatively improves the versatility and heterogeneity of the
spaces such as RGB, HSV and CIELAB (in Fig. 4(c)) and
dataset in order to rule out overfitting instances. Few results of
several operations are done to find out the best possible
various augmentation techniques used on the rice crop image
combination in order to eliminate background noise and thus
dataset (sample image: Fig. 5) include perspective trans-
be able to give an image to the deep learning model that is
form (Fig. 5(b)), affine transform (Fig. 5(c)) and geometric
clutter free and has no unwanted objects apart from the RoI
rotation (Fig. 5(d)).
(region of interest, i.e. leaf).
An histogram equalized image is obtained using hierarchial
masking as shown in Fig. 4(d). Eq. 1 shows how the total C. Proposed RiceBioS CNN Model
variance σ 2 is calculated from the weighted class variance Artificial intelligence has narrowed down the approach of
σv2 (n), class probabilities s(n) and class means r (n), replicating a human mind and training a machine to function
σ 2 = σv2 (n) + s1 (n)[1 − s1 (n)][r1 (n) − r2 (n)]2 . (1) in that manner. Computer vision and artificial intelligence have
found their applications in solving many pressing problems.
σ 2,class variance, is used to find the outliers according to the Convolutional Neural Networks (CNN) have proven to be a
pixel intensities of each corresponding regions of the image. preferable modeling framework for most purposes. A CNN is

Authorized licensed use limited to: KGiSL Institute of Technology. Downloaded on March 11,2024 at 11:43:37 UTC from IEEE Xplore. Restrictions apply.
4620 IEEE SENSORS JOURNAL, VOL. 22, NO. 5, MARCH 1, 2022

Fig. 6. Layered architecture of proposed RiceBioS CNN Model.

an algorithm based on the deep learning techniques. It takes Algorithm 1 Rice_BioS- To Identify Biotic Stress and
image as an input and materialises on learned weights which Category of Stress
ultimately helps the algorithm in differentiating between dif- Input: Img_Leaf
ferent objects in a single frame. We have an RGB image which Inference: Healthy or Biotic Stressed - Bacterial or Fungal
has been separated by its three color planes — Red, Green,
and Blue. The proposed network architecture is graphically 1: Img_Leaf = Rice Crop Leaf Image
demonstrated in Fig. 6. 2: OUTPUT = Final prediction
The convolutional layer primarily helps in capturing the 3: TEMP = Resultant consequent image from intermediate
low-level features such as edges, color, gradient orientation, layers
etc. Here, minute details such as rice crop leaf texture, *: Computation is done at the edge (Smartphone)
stress causing pests image characteristics, texture and color **: Communication with cloud
coding of infected portions of leaf. The initial convolution 4: for each Img_Leaf do
layers extract high level features and subsequent layers extract 5: if ReadFromSmartphoneCamera(Img_Leaf) == 1 then
these low level features. The input image is made to slide 6: SendToApp(Img_Leaf)*;
over a considered kernel matrix to perform the convolution 7: end if
operation [20]. Forward propagation is the straight forward 8: Enhanced_Img_Leaf = Preprocess(Img_Leaf)*;
adopted input driven approach which is explained in Eq. 2 9: TEMP = InitializeInputFeatures(Enhanced_Img_Leaf)*;
(k)

m 
n 10: for i in layer_n do
qi j = Wrk y(i+r)( j +n) + b, (2) 11: TEMP = OUT(layer_n)*;
i=1 j =1 12: end for
where qi j constitutes the resultant of the sum of weight bias b 13: if layer_n == FC or actication then
and product of Wrk , propagation path, and y(i+r)( j +n) , the input 14: Result = LabelBinarizer(TEMP)**;
features. 15: end if
A feedback based closed loop backward propagation helps 16: if Result == 2’b00 then
in eliminating redundant irregularities in the network and make 17: OUTPUT = Healthy;
is responsive to a versatile input. Backward propagation path 18: else
and gradient is calculated in Eq. 3 and Eq. 4 19: if Result == 2’b01 then
20: OUTPUT = Biotic Stressed-Bacterial;
∂V  N−n+1
M−m+1  ∂ V ∂qi j
k
21: else
= , (3)
∂ Wr
i=1 j =1
∂qikj ∂ Vq(k) 22: if Result == 2’b11 then
23: OUTPUT = Biotic Stressed-Fungal;
∂V  N−n+1
M−m+1  ∂V 24: end if
= y(i+r)( j +n) , (4)
∂ Wr ∂qikj 25: end if
i=1 j =1
26: end if
where V is the resultant vector (which is to be given as an 27: end for
input to the pooling layer) with respect to propagation path
Wr aggregated over a range of M, N discrete time intervals
taking into consideration forward propagation vector qi j [21].
The pooling layer shown in Fig. 6 takes care of the reduc- for the model to understand the features of the captured image
tion of features that are convolved in the previous layers. such as texture, color, shape and size etc. Once the model
In order to enable edge computing, keeping less computational is trained through the entire dataset, it is able to spot the
power as a requisite, reducing the dimensionality of the variations in the above mentioned features and thus classifies
features becomes quintessential. Noise suppression is done the image into a particular kind of irregularity (biotic stress).
through max pooling. Average pooling in some cases is used The Algorithm 1 (Ri ce_Bi oS) explains in the order of
for this purpose but after evaluating the initial results, max precedence image acquisition, transfer to mobile applica-
pooling was preferred. The above process designs a blueprint tion, preprocessing (explained in Section III B), forward and

Authorized licensed use limited to: KGiSL Institute of Technology. Downloaded on March 11,2024 at 11:43:37 UTC from IEEE Xplore. Restrictions apply.
JOSHI et al.: RiceBioS: IDENTIFICATION OF BIOTIC STRESS IN RICE CROPS USING EDGE-AS-A-SERVICE 4621

Fig. 8. Test accuracies obtained at different epochs.

Fig. 7. Graph showing average stress classification accuracy of

RiceBioS across three considered states - healthy, stressed (bacterial),
and stressed (fungal).

backward propagation through layers and label binarization

(explained in Eq. 2, 3) at the end to obtain a one hot encoded
classification label - Healthy, Bacterial or Fungal stress.
When the feedback from the output stage in considered
and given back to optimise the input, backward propagation
boosts the overall iterative precision of the convolutional
neural network. The training and testing steps involved are
discussed in Section IV.

IV. E XPERIMENTAL R ESULTS AND D ISCUSSION

A. Experimental Setup
For training the model, an HP Z4 G4 CPU with Intel i7
9700, 16 GB RAM, 500 GB HDD and 2GB Nvidia GT730
VRAM was used with Ubuntu 18.04 LTS. The software Fig. 9. RiceBioS: Training and validation accuracy at (a) 100 epochs
and (b) 250 epochs, loss at (c) 100 epochs and (d) 250 epochs.
requirements include Android Studio IDE 4.3.2, Anaconda
Spyder and Jupyter interface. The frameworks used in Python
include keras, tensorflow, and tensorflow lite. The chipsets The training and validation accuracy and loss curves for
used for testing the mobile application prototype are Medi- different epochs are given in Fig. 9. Since there is not a
atek Helio G80, Qualcomm Snapdragon 450 and Qualcomm significant improvement in accuracy after 250 epochs, it has
Snapdragon 765G. been considered as the final combination. The distribution of
the dataset of images into training, validation and test sets is
B. Performance and Error Analysis of RiceBioS an important step leading to the model training. While having
The proposed deep learning framework RiceBioS was a large pool of training data always helps it is often seen
trained and then tested on a range of sample images belonging that the chances of model overfitting are high. Overfitting is
to each class. The final results post training the model is a scenario where the model is trained to be so specific to
discussed in Fig. 7. It can be seen that there is a test accuracy the training data that it is unable to produce similar results
upwards of 92% in all categories the model is tailor made for. on a real time test sample. In order to ensure that the model
Among all the rival frameworks, the RiceBioS model performs well on real time test data as well and not only on
seemed to be performing best. The closest matching model training data, the dataset is split into three different sets –
in terms of accuracy was Resnet50 architecture. Therefore, training set, validation set (to avoid overfitting) and test data
all models were considered for comparison while RiceBioS (for performance evaluation). The ratio at which dataset is
was used for final deployment. The model was trained with split and its corresponding performance in terms of accuracy
different combinations of number of epochs, learning rate is given in Table IV.
and the dataset distribution. The image dataset was trained The combination of training-validation-test split of 80:10:10
at different number of epochs to analyse the optimal epoch gives the best accuracy for the tested number of epochs. The
number. Training and validation accuracy, training and valida- thin margin of deviation in the accuracy of different dataset
tion loss are the measures considered along with test accuracy splits indicates that the model is not overfit and performs well
for accurate analysis. The comparison of test accuracies for the on test data. A detailed analysis along with model perfor-
model trained at different number of epochs is given in Fig. 8. mance of the different trained networks, AlexNet, VGG16,

Authorized licensed use limited to: KGiSL Institute of Technology. Downloaded on March 11,2024 at 11:43:37 UTC from IEEE Xplore. Restrictions apply.
4622 IEEE SENSORS JOURNAL, VOL. 22, NO. 5, MARCH 1, 2022

TABLE III
T RAINING PARAMETERS OF D IFFERENT M ODELS

TABLE IV
T EST FOR OVERFITTING OF D ATA

Fig. 10. RiceBioS mobile application real time testing results with
Inception v3, MobileNet v2 and Resnet50 are given in (a) Varying Lighting Conditions: gives higher prediction accuracy in well
Table III. The results shows that with the same dataset, the lit environment and (b) Changing Background Noise: versatile in adopting
performance of RiceBioS is best among AlexNet, VGG16, to background noise in image samples.
Inception v3, MobileNet v2 and Resnet50 [22].

C. Prototype Level Deployment of Mobile Application octa core CPU. The application was tested to examine the
computational performance of RiceBioS (Table V), while the
The model, post training was stored as a Keras saved
test accuracy of sample images is shown in Fig. 7. The
model. In order to make it deployable in a mobile application
application does not require an internet connection to compute
environment, it was first freezed and compressed in form
the results but has a requirement of approximately 160MB
of a Tensorflow Lite saved model and then exported to the
of obb files to be downloaded once installed. This helps
android application development environment [23]. The model
the android application to be packaged well and scalable
RiceBioS had to be freezed to enable the predictions to be
to more number of low and mid-range smartphones which
done at the device itself and not at the cloud server. The
are expected to be affordable by the targeted rural farm-
compressed, pruned and light weighted model is inscribed
ing audience. An extensive test of the mobile application
into the user interface of the application developed. The
illustrated in Fig. 11 was done and the results (as shown
model accompanied with the mobile application to enable
in Fig. 10) demonstrate its performance in varying image
edge computing makes it a complete product with minimum
capturing scenarios, thus giving the optimal usage conditions
compromise in performance.
for attaining best results. The application not being heavy and
Figure 11 presents working schematics of the application
requiring negligible background activity does not slow down
that depict the real time functionality of the RiceBioS frame-
the already burdened low range chipset. The performance
work in an android smartphone. Chronologically, the app test-
of the application in low end or budget smartphones shows that
ing done by farmers, launching and image capturing interfaces
the end user experience is bankable. The user interface of the
shown in Fig. 11(a), (b) and (c) are followed by real-time pre-
application has been carefully designed to cater to the needs
dictions of biotic stress conditions. While Fig. 11(d) predicts
of farmers. The graphical user interface that has less clutter
that the rice crop is healthy, different samples when tested
on the screen, more intuitive and comprehensible information
show infected or stressed status of bacterial biotic stress and
makes it a plug and play solution for any farmer irrespective
fungal biotic stress as presented in Fig. 11(e) and Fig. 11(f)
of the language or technological skill of the concerned user.
respectively.
The application occupies a ROM of 176MB and RAM
of approximately 512MB while in use. For testing purpose D. Field Testing and Deployment
the application was tested with a variety of smartphones. The experimental study and deployment of RiceBioS was
These smartphone models were Redmi 9 Prime, Samsung done in a remote village Fulkarra located in Gariyaband dis-
Galaxy M11 and Oneplus Nord. The processors on which the trict, Chhattisgarh, India. The village does not have consistent
computational requirement of this application was prototyped LTE internet connectivity and is majorly connected by EDGE
were Mediatek Helio G80, Qualcomm Snapdragon 450 and networks. 4 rice crop fields of varied maturity levels were
Qualcomm Snapdragon 765G. Both these processor chips considered for the testing of the solution. In order to get
are based out of 14nm finfet technology having a 1.8 GHz the end user (farmers) perspective, 35 farmer volunteers were

Authorized licensed use limited to: KGiSL Institute of Technology. Downloaded on March 11,2024 at 11:43:37 UTC from IEEE Xplore. Restrictions apply.
JOSHI et al.: RiceBioS: IDENTIFICATION OF BIOTIC STRESS IN RICE CROPS USING EDGE-AS-A-SERVICE 4623

Fig. 11. (a) Farmer capturing image using smartphone camera to test developed application; RiceBioS deployed in an Android application (b) app
launched from homescreen, (c) first screen or capturing screen, (d) result for sample 1, (e) result for sample 2, and (f) result for sample 3.

TABLE V
R ICE B IO S M OBILE A PPLICATION T ESTING P ERFORMANCE ON H ANDHELD D EVICES

on the performance front, it lays a foundation for developing

downsized, pruned and streamlined deep learning model for
edge based smartphone deployment specific but not limited to
rice crop biotic stress prediction.

V. C ONCLUSION AND F UTURE W ORKS

Digital agriculture is at the forefront of the farming
infrastructure development in which RiceBioS will play a
Fig. 12. Field testing: (a) raw stressed samples and (b) training and pivotal role for increasing rice crop yield. We have been able
demo of RiceBioS for farmers. to train, test and deploy our optimised, downsized and pruned
proposed framework RiceBioS which does on-device biotic
trained to download, install and make use of RiceBioS in their stress identification at a test accuracy of 93.25%. The paper
smartphones as shown in Fig. 12 (b). Fig. 12(a) shows a raw has not only proposed a deep learning framework but a real
sample of rice crop which has undergone bacterial biotic stress. time deployable android application that makes decisions with
RiceBioS performed seemingly well on raw images from the minimum cross entropy loss without requiring connectivity to
field and was able to bring an impact by early identification cloud or internet. Remote areas where internet connectivity is
of biotic stress. still a challenge in terms of affordability, coverage and band-
width, will benefit the most from this contribution. An end-to-
E. Comparison With Prior Art end smartphone based platform that enables early detection of
The proposed solution RiceBioS makes most of its novel plant stress through on-device fast computing frameworks is
framework to overcome the drawbacks of [9]–[12] which the way forward. The future work includes implementation of
are inability to make on-device decisions, internet connec- a decentralised learning mechanism for rare crops on lines of
tivity requirement and resource rich components for high federated learning. This will help in building a large dataset
performance computing. These shortcomings are addressed at the edge and only transfer the learnt weights to the central
through a solution that incorporates an end-to-end biotic server for optimised results.
stress identification workflow provisioning Edge-as-a-Service
(EaaS). The focus on end user experience by smooth deploy- R EFERENCES
ment on budget android smartphones distinguishes this paper [1] M. Mekonnen, T. Abeje, and S. Addisu, “Integrated watershed manage-
ment on soil quality, crop productivity and climate change adaptation,
from its predecessors. The same is achieved with minimum dry highland of northeast Ethiopia,” Agricult. Syst., vol. 186, Jan. 2021,
tradeoff in accuracy. While RiceBioS does not compromise Art. no. 102964.

Authorized licensed use limited to: KGiSL Institute of Technology. Downloaded on March 11,2024 at 11:43:37 UTC from IEEE Xplore. Restrictions apply.
4624 IEEE SENSORS JOURNAL, VOL. 22, NO. 5, MARCH 1, 2022

[2] V. Udutalapally, S. P. Mohanty, V. Pallagani, and V. Khandelwal, [13] M. Bach-Pages and G. M. Preston, “Methods to quantify biotic-induced
“SCrop: A novel device for sustainable automatic disease prediction, stress in plants,” in Host-Pathogen Interactions. New York, NY, USA:
crop selection, and irrigation in internet-of-agro-things for smart agri- Humana Press, 2018, pp. 241–255.
culture,” IEEE Sensors J., vol. 21, no. 16, pp. 17525–17538, Aug. 2021. [14] A. V. Zubler and J.-Y. Yoon, “Proximal methods for plant stress detection
[3] Y. Zhu et al., “Genetic diversity and disease control in rice,” Nature, using optical sensors and machine learning,” Biosensors, vol. 10, no. 12,
vol. 406, no. 6797, pp. 718–722, 2000. p. 193, Nov. 2020.
[4] S. K. Roy, S. Misra, N. S. Raghuwanshi, and S. K. Das, “AgriSens: IoT- [15] N. Mastrodimos, D. Lentzou, C. Templalexis, D. I. Tsitsigiannis,
based dynamic irrigation scheduling system for water management of and G. Xanthopoulos, “Development of thermography methodology
irrigated crops,” IEEE Internet Things J., vol. 8, no. 6, pp. 5023–5030, for early diagnosis of fungal infection in table grapes: The case
Mar. 2021. of aspergillus carbonarius,” Comput. Electron. Agricult., vol. 165,
[5] S. Kaur, S. Pandey, and S. Goel, “Semi-automatic leaf disease detection Oct. 2019, Art. no. 104972.
and classification system for soybean culture,” IET Image Process., [16] E.-C. Oerke, P. Fröhling, and U. Steiner, “Thermographic assessment
vol. 12, no. 6, pp. 1038–1048, 2018. of scab disease on apple leaves,” Precis. Agricult., vol. 12, no. 5,
[6] N. Goel, D. Jain, and A. Sinha, “Prediction model for automated leaf pp. 699–715, Oct. 2011.
disease detection & analysis,” in Proc. IEEE 8th Int. Advance Comput. [17] P. Pal, R. P. Sharma, S. Tripathi, C. Kumar, and D. Ramesh, “2.4 GHz
Conf. (IACC), Dec. 2018, pp. 360–365. RF received signal strength based node separation in WSN monitoring
[7] H. Wang, G. Li, Z. Ma, and X. Li, “Image recognition of plant diseases infrastructure for millet and Rice vegetation,” IEEE Sensors J., vol. 21,
based on backpropagation networks,” in Proc. 5th Int. Congr. Image no. 16, pp. 18298–18306, Aug. 2021.
Signal Process., Oct. 2012, pp. 894–900. [18] R. Dwivedi, S. Dey, C. Chakraborty, and S. Tiwari, “Grape disease
[8] G. Zhou, W. Zhang, A. Chen, M. He, and X. Ma, “Rapid detection detection network based on multi-task learning and attention features,”
of Rice disease based on FCM-KM and faster R-CNN fusion,” IEEE IEEE Sensors J., vol. 21, no. 16, pp. 17573–17580, Aug. 2021.
Access, vol. 7, pp. 143190–143206, 2019. [19] X. Xie, X. Zhang, B. He, D. Liang, D. Zhang, and L. Huang, “A system
[9] W.-L. Chen, Y.-B. Lin, F.-L. Ng, C.-Y. Liu, and Y.-W. Lin, “RiceTalk: for diagnosis of wheat leaf diseases based on Android smartphone,”
Rice blast detection using Internet of Things and artificial intelligence Proc. SPIE, vol. 10155, Oct. 2016, Art. no. 1015526.
technologies,” IEEE Internet Things J., vol. 7, no. 2, pp. 1001–1010, [20] S. Joshi, D. K. Verma, G. Saxena, and A. Paraye, “Issues in training a
Feb. 2020. convolutional neural network model for image classification,” in Proc.
[10] M. J. Hasan, S. Mahbub, M. S. Alom, and M. Abu Nasim, “Rice disease ICACDS. Singapore: Springer, 2019, pp. 282–293.
identification and classification by integrating support vector machine [21] N. Skatchkovsky, H. Jang, and O. Simeone, “Spiking neural networks—
with deep convolutional neural network,” in Proc. 1st Int. Conf. Adv. Part II: Detecting spatio-temporal patterns,” IEEE Commun. Lett.,
Sci., Eng. Robot. Technol. (ICASERT), May 2019, pp. 1–6. vol. 25, no. 6, pp. 1741–1745, Jun. 2021.
[11] N. Yang, Y. Qian, H. S. El-Mesery, R. Zhang, A. Wang, and J. Tang, [22] E. C. Too, L. Yujian, S. Njuki, and L. Yingchun, “A compar-
“Rapid detection of Rice disease using microscopy image identification ative study of fine-tuning deep learning models for plant dis-
based on the synergistic judgment of texture and shape features and ease identification,” Comput. Electron. Agric., vol. 161, pp. 272–279,
decision tree–confusion matrix method,” J. Sci. Food Agricult., vol. 99, Jun. 2019.
no. 14, pp. 6589–6600, 2019. [23] J. Park, Y. Kwon, Y. Park, and D. Jeon, “Microarchitecture-aware
[12] Y. Kim, J.-H. Roh, and H. Kim, “Early forecasting of Rice blast disease code generation for deep learning on single-ISA heterogeneous multi-
using long short-term memory recurrent neural networks,” Sustainability, core mobile processors,” IEEE Access, vol. 7, pp. 52371–52378,
vol. 10, no. 2, p. 34, Dec. 2017. 2019.

Authorized licensed use limited to: KGiSL Institute of Technology. Downloaded on March 11,2024 at 11:43:37 UTC from IEEE Xplore. Restrictions apply.