An Application with Jetson Nano for Plant Stress Detection and

On-ﬁeld Spray Decision

Marcos A. de Oliveira Jr.

1,2

, Gregory Sedrez

Guilherme de Souza

and Gerson Geraldo H. Cavalheiro

Programa de P

os-Graduac¸

ao em Computac¸

ao, Universidade Federal de Pelotas, Pelotas/RS, Brazil

Instituto Federal de Educac¸

ao, Ci

encia e Tecnologia Farroupilha, Santa Maria/RS, Brazil

Keywords:

Decision Support System, Agro-sensor, Time Series, Jetson Nano, Smart Farm.

Abstract:

Increasing ﬁeld productivity is not just a ﬁnancial need, but also a social issue. Several technologies converge

to promote food production and, in this context, the fog computing paradigm can support the development of

solutions for precision agriculture. This paper proposes an application of the Jetson Nano device, embedded

in an agricultural spraying implement. This device supports the decision on irrigation activity, based on data

collected by sensors distributed in the ﬁeld. The sensors read information about the plant’s stress level from

electrical signals and the Jetson Nano enables real-time analysis, through machine learning algorithms, to

manage the product spray rate, according to the condition of the crop. Initial studies validated the proposed

solution on an experimental basis, showing that the device can be an alternative for this purpose, since it can be

used efﬁciently in machine learning tasks from data collected by the sensors. The experiment also highlighted

some limitations of the proposed solution, such as the importance of observing the conditions of the system as

a whole, its context and environment, in order to improve performance in spraying process.

1 INTRODUCTION

Agriculture is the fundamental basis for human sur-

vival, however, in recent years, some factors arising

from the evolution and globalization of society are

impacting agricultural production, restricting the de-

velopment of the sector and worrying producers. As

shown in (Zhou et al., 2011), the three main reasons

for these concerns are: (i) the aging of the agricul-

tural population, migration from rural to urban areas

and, consequently, a reduction in rural labor; (ii) large

constructions invading rural spaces and, therefore, re-

ducing the production area; and (iii) increased climate

change, such as temperature, rainfall and soil mois-

ture, which affect crop growing conditions in unpre-

dictable ways.

Internet of Things (IoT) and Artiﬁcial Intelligence

(AI), among other innovative technologies, have been

promoting major changes in the agricultural scenario

(Misra et al., 2020). These technologies not only

solve the problems related to the increased demand

for food and environmental pollution caused by the

use of pesticides without proper control, but also pro-

mote intensely the continuous development of agri-

culture. The increased use of IoT devices in agricul-

ture has led to a huge growth in the use of sensors and

the growing number of data to be processed. Allied to

these technologies, the concepts of Edge Computing

(Shi et al., 2016) and Fog Computing (Bonomi et al.,

2012) emerge. These paradigms intend to meet appli-

cation requirements, such as the need for greater re-

liability, hardware weight (size, location and number

of machines), and network and power consumption

(Byers, 2017). This model allows the investigation of

data processing performance close to the place where

they are produced and/or needed (Baller et al., 2021).

Since there is less processing power in edge and

fog devices, applications in these contexts are instan-

tiated in a specialized manner, dealing with a sin-

gle problem. In recent years, one of the contexts in

which research in edge and fog computing has in-

creased was precision agriculture, due to the great

impact caused by technology in the agricultural sec-

tor. The Smart Farms concept has been leveraged in

recent years by IoT devices. Its use allows farmers

to receive near real-time data about farm conditions.

With the smart farm concept under construction, Arti-

ﬁcial Intelligence (AI) strategies (Li et al., 2018) have

been incorporated into the solutions. From the po-

tential to reduce the need for human interference in

ﬁeld management processes, these strategies in gen-

eral need some level of edge computing, so the devel-

Oliveira Jr., M., Sedrez, G., de Souza, G. and Cavalheiro, G.

An Application with Jetson Nano for Plant Stress Detection and On-ﬁeld Spray Decision.

DOI: 10.5220/0010983900003118

In Proceedings of the 11th International Conference on Sensor Networks (SENSORNETS 2022), pages 215-222

ISBN: 978-989-758-551-7; ISSN: 2184-4380

215

opment of solutions goes through the search for de-

vices capable of performing these processes. Thus,

the key is to identify edge devices that have hard-

ware resources capable of processing AI techniques.

Among the edge devices available on the market, we

have the Jetson Nano (Nvidia, 2021) as a possible so-

lution. This device is a embedded computing board

from NVidia, which contains a low power system and

hardware resources like GPU, designed for accelerat-

ing machine learning applications.

Therefore, this work investigates the feasibility of

using the edge device Jetson Nano for the implemen-

tation of AI techniques, focusing on the identiﬁcation

of the stress level in bean plants. This paper extends

previous results of (de Toledo, 2019) in which plant

stress state classiﬁcation techniques were successfully

performed using neural networks and machine learn-

ing. The main contribution intended is to validate the

use of the mentioned device for an application that

measures the amount of agricultural inputs, water, fer-

tilizers, pesticides, etc., in an agricultural implement.

This is possible through the analysis of data collected

by sensors spread across different regions of the crop,

which serve as input to classiﬁcation algorithms that

indicate the stress level of plants in that region. So,

based on that stress level, each region receives a spe-

ciﬁc amount of spray. From the experiments carried

out, it was possible to validate an application model

with the Jetson Nano device for the detection of stress

in bean plants, which enables the use of this informa-

tion for real-time decision-making in the spraying of

agricultural products in the ﬁeld.

The paper is organized as follows: Section 2

presents works related in precision agriculture; Sec-

tion 3 presents the proposed application design for

practical use and Section 4 the case study developed

and initial experiments, as well as a discussion of the

results. The conclusions and possibilities for future

work are presented in Section 5.

2 RELATED WORK

The use of IoT devices has allowed a lot of achieve-

ments, since the intelligent objects connected to sen-

sors allow their interaction with the physical and log-

ical world without the need for human intervention.

However, even if the Internet itself is not needed,

some form of communication support for the edge/-

fog devices must be available.

In (Saraf and Gawali, 2017), an intelligent agri-

cultural irrigation system, monitored via an Android

smartphone, is proposed. Field communication sup-

port is provided by the Zigbee protocol, which en-

sures communication between sensors and a base sta-

tion. The system performs real-time readings that are

presented in a web-based interface to the user. So,

the user interaction can be through their smartphone,

controlling the water distribution. In (Kamilaris et al.,

2016), a framework capable of connecting and mon-

itoring numerous sensors is presented. The collected

data, as well as the decisions made by the sensors, are

collected by a central database and then made avail-

able in a cloud. AgriSys (Abdullah et al., 2016), using

Fuzzy logic, explores the cloud in the same way, but

deals with the desert’s challenge: dust, infertile sandy

soil, constant wind, very low humidity and extreme

variations in daytime and seasonal temperatures.

In the works cited, for its full functionality, the use

of network resources is necessary, either for the col-

lection of monitored information or for sending com-

mands to perform actions on the ﬁeld. However, pro-

viding access to network services in the agricultural

context is expensive, mainly due to the use of en-

ergy for communication on farms with a large area

to be covered. Considering these aspects, solutions

that minimize network consumption have emerged in

the context of precision agriculture.

As an alternative, WSN (Wireless Sensor Net-

work) type network protocols and their low-power

and/or long-range hardware resources were devel-

oped, such as the LoRA protocol. This protocol has

been used in places where other networking tech-

nologies are not supported. However, the low band-

width offered is a characteristic of the protocol, which

must be considered when designing solutions. In (Gia

et al., 2019) a case study is presented using LoRA to

establish an edge layer connecting the sensors to each

other and to a gateway. The gateway then regroups

the information collected in its coverage region and

forwards it, in a single message, to a concentrator.

In order to reduce the need for the use of net-

work resources, recently, edge computing concepts

have been introduced as an alternative to cloud com-

puting and aim to maximize processing close to the

data generation point. Thus, a new generation of ap-

plications for precision agriculture has been exploring

the use of machine learning techniques embedded in

edge devices, to distribute data processing across ap-

plication layers and reduce the transmission of infor-

mation between them.

Also, popular ML techniques (Machine learning)

have been employed in systems. Such techniques in-

clude K-NN (K-Nearest Neighbors, SVM (Support

Vector Machines), ANN (Artiﬁcial Neural Networks)

among others, which have been shown to be of great

value. Its effects and applicability are diverse, in

(Kamilaris and Prenafeta-Bold

u, 2018), they demon-

SENSORNETS 2022 - 11th International Conference on Sensor Networks

216

strate the different algorithms that have been used in

search of the best result in the agricultural environ-

ment. Applications are focused on tasks such as an-

alyzing plant images for anomalies, fruit counting,

weed detection and even ﬁre outbreaks. Moreover,

inadequate spraying tends to mistreat plants, pollute

water near the harvest and below ground. In (Liakos

et al., 2018), he demonstrates that it is possible to have

even greater control over the amount of spraying re-

quired and the size of the sprayed area.

Few proposals provide more information about

hardware or even propose one that can serve from

small to large farms. In (Imran et al., 2020), the au-

thors present a review of the main embedded devices

found on the market with processors capable of run-

ning AI algorithms at the edge. In (Proietti et al.,

2021), the authors present a study that aims to develop

a system based on Deep Learning that will be applied

in greenhouses in order to detect anomalies in plant

growth. To develop the solution, embedded devices

NVIDIA Jetson Nano and Raspberry Pi 4 Model B

were used, capable of using a complete deep learning

framework such as TensorFlow (Abadi et al., 2015).

In (Kawai and Mineno, 2020), the validation of an au-

tomatic irrigation system in tomato plantations using

edge AI and the embedded devices NVIDIA Jetson

Nano and Raspberry Pi 3 Model B is performed.

All these possibilities for the application of new

technologies in agriculture are having promising re-

sults and enabling the growing use of IoT and AI edge

technologies. Thus, our motivation in this work is

to validate a solution for detecting the stress level in

common bean, using edge devices for the execution

of AI algorithms in model training and data classiﬁ-

cation tasks.

3 PROPOSED SOLUTION

ARCHITECTURE

The proposed solution is based on the previous work

documented in (de Toledo, 2019), which uses the

black bean BRS-Expedito species for a laboratory ex-

periment. Like any plant, beans are subject to multi-

ple environmental variations, some of which generate

stress, such as lack of water resources and soil salin-

ity. These stressful situations lead to ﬁnancial losses.

However, the manifestation of stress can be observed

by reactions involving changes in electrical activity in

the plant (Maffei and Bossi, 2006).

In (de Toledo, 2019) experiments were carried out

to monitor seedlings submitted to induction stress,

with application of solutions that generate situations

of lack of water and salinity. To induce stress in the

plants, three solutions were used: NaCl, Polyethylene

glycol and NaCl+Polyethylene glycol. For each type

of stress, 30 experiments were performed. Each ex-

periment monitored the electrical activity of 3 individ-

uals for 4 consecutive hours, two hours before appli-

cation of the solution and two hours after. The capture

of electromas from the plants was carried out with

electrodes inserted in the plant, with a reading rate

of 62.5 Hz. The time series produced, for each two-

hour range, has, therefore, 450,000 readings (elec-

tromas). These data were input to classiﬁcation al-

gorithms, using traditional computational resources.

In the methodology used, each sequence of 30,000

points, equivalent to 8 minutes of monitoring, was

represented by three values: average, maximum and

minimum of the values captured in the period. The

interval of 30 thousand points was adopted, accord-

ing to the author, as it presents a better result in terms

of accuracy. According to the discussion presented,

the results showed a high success rate in recognizing

the plant’s stress situation.

Extending this idea to the ﬁeld scenario, a solu-

tion is proposed for monitoring a crop, as well as col-

lecting and processing data through the Jetson Nano

device embedded in an agricultural implement. The

base model of the designed application is shown in

Figure 1, where, in a crop, a set of plants is selected

for monitoring, in a distributed way, allowing each

monitored point to represent the situation of the sur-

rounding plants. In detail, the monitored plant and

the monitoring device are presented. This device is

capable of storing the time series that represents the

evolution of the plant’s electromagnetic signal. This

device is also capable of responding, via a wiﬁ net-

work, to the request to send the collected data.

Eventually, this agricultural implement circulates

over this plantation, spraying agricultural inputs such

as fertilizer or pesticides, or even water. The amount

of product to be sprayed in a given area can be iden-

tiﬁed in real time by rating the stress level of the in-

dividual monitored in that region. A device on this

implement requests, via a wiﬁ network, information

from the plant being monitored in the different re-

gions to calculate the amount of product needed in

each region. Also, if available and depending on the

application, meteorological data can be used in the

decision-making process to complement the informa-

tion available for analysis.

Among the small board options to meet the objec-

tives of this work, the option chosen was to use the

NVIDIA Jetson Nano coupled to the agricultural im-

plement. The choice was based on the wide use of this

device embedded in edge solutions for agriculture, ac-

cording to related works, which motivated the acqui-

An Application with Jetson Nano for Plant Stress Detection and On-ﬁeld Spray Decision

217

Figure 1: Proposed application model.

sition of this hardware in this research project. Basic

technical speciﬁcations of the device can be found on

the supplier’s website (Nvidia, 2021). This device is

focused on meeting the recent demands of handling

data at the edges. With a GPU for processing, this

device is capable of supporting the execution of pro-

grams with great processing needs, such as image pro-

cessing, computer vision and deep learning, as is the

case in this work, with low energy consumption (be-

tween 5 to 10 watts).

Furthermore, this type of device can be easily in-

tegrated into an agricultural implement since its re-

duced dimensions do not imply adjustments to the

machinery structure. In fact, this equipment does not

necessarily need to be permanently attached to an im-

plement, which can enable the use of the same device

in different implements. This can be speciﬁcally ad-

justed according to different aspects such as machine

model and number of devices available.

Given its characteristics, the Jetson Nano meets

the four major needs, mentioned above, related to the

concept of edge computing, such as reliability, hard-

ware weight and low energy consumption. The other

items that make up the infrastructure shown in Fig-

ure 1 are not detailed in this article due to space lim-

itations. This paper documents the evaluation of the

Jetson Nano to determine if this device makes the pro-

posed solution feasible within this scenario.

4 METHODOLOGY AND

TECHNICAL FEASIBILITY

In general, the main objective of the work is to es-

tablish a proof of concept for the use of the Jetson

Nano device in agricultural applications to processing

data sensed by IoT devices. The data were stored in

a time series format and used for decision-making in

spraying tasks, through machine learning algorithms.

Thus, it is intended to identify the feasibility of using

an equipment embedded in an agricultural implement

to, during its operation, in real time, identify the stress

level of bean plants and make a decision regarding the

spraying of inputs on the ﬁeld.

To validate the proposed solution, a bench situa-

tion was simulated, using real data from a previous

agricultural experiment, seeking to observe the per-

formance of the Jetson Nano device for the avail-

able data set, as well as designing a real application

through use that device. Next, the observed perfor-

mance of the Jetson Nano as an option for on-board

equipment is presented, as well as a discussion of the

impacts of this solution. All data considered in the

present study are those collected in the experiment

documented in (de Toledo, 2019).

4.1 Implementation of Classiﬁcation

Algorithms

The time series that represent the data collected from

the plants are stored in a text ﬁle, with one sample

per line, and preprocessed, transformed into triples,

as mentioned above, to reduce the data volume. The

result obtained is a lower processing demand for the

classiﬁcation algorithms, with no loss in result qual-

ity. These values are stored in a spreadsheet in .csv

format, where each line represents a measurement

range, with the ﬁrst three columns corresponding to

the minimum, average and maximum values of the

range. In model training, the fourth column informs

if the plant has any stress (1 means stress and 0 no

stress).

The collected data were used in the training of

classiﬁer models to identify the substances. The algo-

SENSORNETS 2022 - 11th International Conference on Sensor Networks

218

rithms used were K-NN, SVM and ANN. The choice

was made based on the investigation of solutions typ-

ically used for these scenarios, and, for the purpose of

this article, for performance analysis, all algorithms

were used based on default parameters. The imple-

mentations were in Python, using two machine learn-

ing libraries: Scikit-learn (Pedregosa et al., 2011) in

the ﬁrst two and TensorFlow (Abadi et al., 2015) for

the other.

To measure the classiﬁcation accuracy, the input

data set (around 15,000 entries) was divided: 70% of

the data were used in the model training and 30% for

the accuracy test. With regard to classiﬁcation accu-

racy, the results obtained with K-NN and SVM showed

the best results, reaching 70% and 60%, respectively,

of accuracy, while the ANN had 54%. In this step, all

the algorithms were used with the standard parame-

ters of the mentioned libraries, that is, in this ﬁrst mo-

ment, the algorithms were not tuned, since the initial

results were considered interesting. The focus here is

to analyze the performance of the algorithms in terms

of execution time for training models and for classi-

fying a new input.

After measuring the performance in terms of ac-

curacy, we focused on identifying the feasibility of

the solution in relation to the time needed to clas-

sify a new sample. In this implementation, using the

Time library, two timestamps were set, marking the

start and the end of the execution. The time differ-

ence between the two points then corresponds to the

time needed for classiﬁcation. To manipulate the data

sets, Pandas data analysis and manipulation tool (Pan-

das Development Team, 2020), one of the most popu-

lar open source library in this domain for Python, was

used. Algorithm 1 presents the code used to measure

the time spent in classifying a new entry by a model

trained with the K-NN algorithm. The text1.csv ﬁle

(line 6) contains the collected data.

Within the measured time (between lines 5 and

14), the following tasks are considered: data import

(line 6), model training (lines 10 and 11) and classi-

ﬁcation of a new point (line 12). Now, 100% of the

available inputs are used for training, and only one

point is predicted (line 15). The implementation with

the SVM algorithm follows the same logic and code

structure. For the implementation of the ANN was

used also the same code structure, but adding the li-

brary Keras (Chollet et al., 2015) to the aforemen-

tioned libraries. In short, the ANN was composed of

three layers, with 3, 5 and 5 neurons each, relu ac-

tivation method, Stochastic Gradient Descent (SGD)

optimizer and trained in 100 epochs. Due to space

limitations, the complete codes are not presented in

this paper, but can be made available by contacting

Algorithm 1: K-NN measurement.

1 i mpor t p a n d a s a s pd

2 from s k l e a r n . n e i g h b o r s i mp or t

K N e i g h b o r s C l a s s i f i e r

3 i mp or t t i me

5 s t a r t = t i me . t ime ( )

6 d a t a s e t = pd . r e a d c s v ( ’ t e s t 1 . c s v ’ )

7 d a t a = d a t a s e t . i l o c [ : , 0 : 3 ] . v a l u e s

8 c l a s = d a t a s e t . i l o c [ : , 3 ]

10 knn = K N e i g h b o r s C l a s s i f i e r (

11 n n e i g h b o r s =5 ,

12 m e t r i c = ’ m in ko wski ’ ,

13 n j o b s = 2 )

14 knn . f i t ( d a t a , c l a s )

15 knn . p r e d i c t

( [ [ − 3 63 4 6 4 . 0 , − 23 7 18 . 2 , 1 2 6 3 4 3 . 0 ] ] )

17 end = t i me . t i me ( )

19 p r i n t ( ” t i me : ” )

20 p r i n t ( end − s t a r t )

the research group.

Table 1 presents the times taken by the classiﬁca-

tion algorithms, both on a desktop computer and on

the Jetson Nano. The times correspond to an average

of 30 runs, where the distribution of samples corre-

sponds to normal. After the executions, the system

was always restarted, in order to guarantee a standard

execution situation, avoiding that the consequences

of a execution, such as the temperature of the com-

ponents, could inﬂuence the other executions and,

consequently, cause differences in performance. The

desktop processing was done entirely in CPU, while

the Jetson Nano used the GPU. The desktop times (In-

tel i5, 2/4 colors/threads, 8 GB RAM) are presented

illustratively, aiming to identify the distance of the

processing times obtained with the Jetson Nano to a

standard desktop computer conﬁguration.

4.2 Performance Analysis and

Discussion

Regarding classiﬁcation accuracy, K-NN and SVM

presented better classiﬁcation of the stimuli to which

the bean plants were submitted in the bench experi-

ments. ANN, on the other hand, had lower accuracy

indices. A more in-depth study of the parameters ap-

plied to classiﬁcation algorithms is considered neces-

sary in order, in a tunning process, to identify the best

conﬁguration for the application in question. How-

ever, considering neural networks, it is a fact that a

greater amount of data in the training stage may re-

An Application with Jetson Nano for Plant Stress Detection and On-ﬁeld Spray Decision

219

Table 1: Average time of algorithm executions (in seconds).

Device Desktop Jetson Nano

Algorithms ANN K-NN SVM ANN K-NN SVM

NaCl 0.14 0.11 0.05 2.90 0.15 0.20

Polietilenoglicol 0.13 0.11 0.06 3.00 0.14 0.19

NaCl+Polietilenoglicol 0.14 0.12 0.20 3.00 0.16 0.70

ﬂect better classiﬁcation success rates. This aspect

is relevant, because the K-NN and SVM algorithms

need the training data to be available, and to be ma-

nipulated, at each new classiﬁcation. Thus, can be

considered valid to perform the prediction time anal-

ysis in the ANN algorithm, since for this network it is

not necessary to access the training base to perform a

new prediction. The consequence is a smaller amount

of memory needed on a device in the fog, a desirable

requirement in this type of application.

Regarding the practical application of the pro-

posed solution, it is necessary to analyze the impact of

classiﬁcation execution times in the real world, since

the average speed of an agricultural implement in the

ﬁeld is less than 15 km/h, or 4 m/s. It is observed that

the average execution times of the algorithms on the

Jetson Nano are, in most cases, less than 0.2 seconds,

which is equivalent to a displacement of 0.8 m. Thus,

these results validate the use of the device for the pur-

pose, as it is capable of dealing speciﬁcally with the

classiﬁcation task in almost real time.

However, a signiﬁcant problem was perceived in

the task of collecting and preprocessing the data,

which took a considerable time slice of 54 seconds,

considering a data entry of 450,000 samples, at inter-

vals of 30,000 samples, as detailed above. In terms of

displacement, the agricultural implement would have

traveled more than 200 m, moving away from the

region where the measurement was taken. In addi-

tion, the communication times between the sensor and

the device embedded in the implement must also be

added to the times presented.

Regarding communication, let’s consider a LoRa

(Long Range) (Alliance, 2015) network, quite pop-

ular as network infrastructure in smart farm applica-

tions, with an average range of 15 km in open areas

and transmission rate between 0.3 to 50 Kbps. There-

fore, from the technical speciﬁcations of this network

model, it is possible to observe that, even with the

high volume of transmitted data, the proposed solu-

tion is valid. If we consider a median value for the

transmission rate of 25 Kbps, the time to send data

(6 MB) is around 240 seconds, which represents a

displacement of less than 1 km, that is, the imple-

ment will still be within the network’s coverage area.

In this way, the machinery could start data collection

and classiﬁcation when it is at least that distance from

the spray application point.

As the proposed scenario has an experimental

character, it is possible that we have variations in the

ﬁeld area that each sensor will cover. However, given

the long-range coverage of the Lora network, there is

good scope for a project to deploy sensors within a

signiﬁcant coverage radius.

Furthermore, seeking to overcome these observed

limitations, improvements can be investigated in the

sense of: reducing the speed of the implement; de-

crease the volume of data transmitted (preprocessing);

or still, use an extra device, in the ﬁeld, as a gateway,

for data preprocessing, before the implement action.

The ﬁrst and possibly the simplest alternative to

ensure that the proposed solution works is a decel-

eration of the agricultural implement, to increase the

time available for data reception, training and classi-

ﬁcation. For this, however, it is necessary to analyze,

together with experts in the agricultural area, what is

the impact of this lower speed on the cost of the activ-

ity, in terms of fuel consumption and autonomy, given

a consequent increase in time to cover large areas of

farming.

Regarding the reduction of data volume, the fol-

lowing possibilities can be investigated: data com-

pression, through algorithms that are supported in

edge devices; or even the decrease in the frequency of

data collection. For the ﬁrst, the main challenge is to

ﬁnd information compression techniques that do not

affect the efﬁciency of classiﬁcation models, trained

from compressed data. Previous studies (de Toledo,

2019) and (Pereira et al., 2018) indicate good classiﬁ-

cation indices, by neural networks, using compressed

representations of the time series. In this case study,

where the database provided corresponds to 450,000

samples (6 MB), the compact representation has only

15 triples of numerical data, totaling 180 bytes. The

indication, therefore, is that the device coupled to the

sensor in the bean plant has enough intelligence to

compose this compact representation of the time se-

ries. As a positive aspect of this solution, the need

for storage on this device is reduced. For the sec-

ond, there is the same concern regarding the impact

of training models with a smaller data set. As the fre-

quency is reduced, less information will be available,

so it must be observed whether the classiﬁcation will

not be harmed by this situation.

SENSORNETS 2022 - 11th International Conference on Sensor Networks

220

Within the scope of our group’s research, ad-

vances are also being sought in terms of data vol-

ume reduction. Recently, the work published in

(Oliveira Jr et al., 2021) presented contributions in

the search for time series approximation techniques.

This work carried out in parallel contributed not only

with applicable approximation techniques, but also

with the investigation of which approximation tech-

niques can be used in combination with classiﬁcation

algorithms, without loss of data quality due to volume

reduction.

Another possible solution is to use a central (gate-

way), installed in the ﬁeld as well. This central would

be in charge of processing the data and passing on to

the agricultural machinery only the amount of input

to be applied and in which area it should be applied.

However, this may imply the use of a system with

GPS to control the areas, which has become quite

common in newer equipment. As the data packets to

be sent to this gateway, and from the gateway to the

implement, are small, it would not have much impact

on transmission time. A single gateway can receive

data from thousands of devices and forward it to the

network server. Depending on the topology condi-

tions of the deployed network, a single gateway can

cover a radius of kilometers away.

Finally, it is important to remember that inside the

laboratory, where the experiments were carried out, it

is a controlled environment where the plants are in a

favorable environment for their well-being and devel-

opment. Stresses are applied at controlled intervals

so that we can know exactly where the stresses have

occurred. In the ﬁeld, the plant undergoes constant

stresses, such as climate, soil, pests, which can lead to

the need for other related information, such as the du-

ration interval of each stress and weather conditions.

Improvements in this aspect can also contribute to in-

creased plant stress detection efﬁciency.

The work described in this paper was carried out

in a bench test. A second experiment, which data is

currently being compiled, was carried out in the ﬁeld,

and, in fact, was subject to several practical problems,

which will be discussed in an upcoming paper. In-

frastructure implementation aspects (battery, energy

consumption, system autonomy, sensors, communi-

cation, robustness) are also addressed in another work

front of the group.

5 CONCLUSION AND FUTURE

WORK

Computing in Fog, thanks to the technological evo-

lution of recent years and the consequent variety of

devices, has the potential to promote the development

of applications in the most different sectors of soci-

ety. New uses of computing have thus emerged and,

particularly in the context of this work, applying tech-

niques that bring intelligence to processing.

In this work, the technical feasibility of using the

NVIDIA Jetson Nano edge device in an agricultural

spray application was studied. In the device, time

series classiﬁcation algorithms were executed in the

ﬁeld, and the initial results, in an experimental na-

ture, showed that this type of device can be success-

fully used for the speciﬁc purpose. Furthermore, the

experiment allowed the identiﬁcation of limitations

of the proposed solution, highlighting aspects to be

observed and necessary improvements for the imple-

mentation of the solution in a real scenario. The issue

of data volume represented by a time series presents a

set of difﬁculties that must be addressed to enable the

successful implementation of the proposed scenario.

Given the nature of time series, the number of

samples can grow signiﬁcantly due to the time period

and frequency with which data are collected, which

will imply time series of large volumes. This charac-

teristic must be considered in the analysis of applica-

tion costs, both with regard to storage and processing

of this information. Previous studies indicated that

there is no loss in classiﬁcation quality in the studied

algorithms with reduced representations, however, the

processing time to obtain these triples for representing

periods of the time series is quite expressive. Also

considering the volume of data, whose communica-

tion times are quite high, there is a clear indication

that an alternative should be sought for this task.

Future works aim to enable the proposed preci-

sion agriculture scenario. Objectively, focusing on

the problem of data volume, alternative techniques are

being studied for the compression of time series al-

ready in the device coupled to the sensor, minimizing

the storage, communication and processing require-

ments in the device responsible for the classiﬁcation.

This activity also involves determining which alterna-

tive time series compression techniques provide suit-

able results for classiﬁcation. Furthermore, an inves-

tigation about hyper parameter tuning in classiﬁcation

algorithms is also desired, in order to improve the per-

formance for this task. Regarding communication, is-

sues related to range and security (prevention of mali-

cious access to the network) are being considered, as

well as considering data collection by sensors of other

natures, such as unmanned aerial vehicles.

An Application with Jetson Nano for Plant Stress Detection and On-ﬁeld Spray Decision

221

ACKNOWLEDGEMENTS

This study was ﬁnanced in part by the Coordenac¸

de Aperfeic¸oamento de Pessoal de N

ıvel Superior -

Brasil (CAPES) - Finance Code 001.

REFERENCES

Abadi, M. et al. (2015). TensorFlow: Large-scale machine

learning on heterogeneous systems. Software avail-

able from tensorﬂow.org.

Abdullah, A., Al Enazi, S., and Damaj, I. (2016). Agrisys:

A smart and ubiquitous controlled-environment agri-

culture system. In 2016 3rd MEC International Con-

ference on Big Data and Smart City (ICBDSC), pages

1–6. IEEE.

Alliance, L. (2015). What is LoRaWAN - A

technical overview of Lora and LoRaWAN.

https://lora-alliance.org/resource\textunderscorehub/

what-is-lorawan/. Acessed in 25/07/2021.

Baller, S. P., Jindal, A., Chadha, M., and Gerndt, M.

(2021). DeepEdgeBench: Benchmarking deep neural

networks on edge devices. CoRR, abs/2108.09457.

Bonomi, F., Milito, R., Zhu, J., and Addepalli, S. (2012).

Fog computing and its role in the internet of things. In

Proceedings of the First Edition of the MCC Workshop

on Mobile Cloud Computing, MCC ’12, pages 13–16,

New York, NY, USA. ACM.

Byers, C. C. (2017). Architectural imperatives for fog

computing: Use cases, requirements, and architectural

techniques for fog-enabled iot networks. IEEE Com-

munications Magazine, 55(8):14–20.

Chollet, F. et al. (2015). Keras. https://github.com/fchollet/

keras.

de Toledo, G. R. A. (2019). Electrophysiological charac-

terization of beans (Phaseolus vulgaris L.) cv. BRS-

Expedit under different water availabilities (In Por-

tuguese). PhD thesis, UFPel, Pelotas.

Gia, T. N., Qingqing, L., Queralta, J. P., Zou, Z., Tenhunen,

H., and Westerlund, T. (2019). Edge AI in smart farm-

ing IoT: CNNs at the edge and fog computing with

LoRa. IEEE AFRICON-2019.

Imran, H. A., Mujahid, U., Wazir, S., Latif, U., and

Mehmood, K. (2020). Embedded development boards

for Edge-AI: A comprehensive report.

Kamilaris, A., Gao, F., Prenafeta-Boldu, F. X., and Ali,

M. I. (2016). Agri-IoT: A semantic framework for

internet of things-enabled smart farming applications.

In 2016 IEEE 3rd World Forum on Internet of Things

(WF-IoT), pages 442–447. IEEE.

Kamilaris, A. and Prenafeta-Bold

u, F. X. (2018). Deep

learning in agriculture: A survey. Computers and elec-

tronics in agriculture, 147:70–90.

Kawai, T. and Mineno, H. (2020). Evaluation environment

using edge computing for artiﬁcial intelligence-based

irrigation system. In 2020 16th International Con-

ference on Mobility, Sensing and Networking (MSN),

pages 214–219.

Li, H., Ota, K., and Dong, M. (2018). Learning iot in

edge: Deep learning for the internet of things with

edge computing. IEEE Network, 32(1):96–101.

Liakos, K. G., Busato, P., Moshou, D., Pearson, S., and

Bochtis, D. (2018). Machine learning in agriculture:

A review. Sensors, 18(8):2674.

Maffei, M. and Bossi, S. (2006). Electrophysiology and

plant responses to biotic stress. In Plant Electrophys-

iology, pages 461–481. Springer.

Misra, N. N., Dixit, Y., Al-Mallahi, A., Bhullar, M. S.,

Upadhyay, R., and Martynenko, A. (2020). IoT, big

data and artiﬁcial intelligence in agriculture and food

industry. IEEE Internet of Things Journal, pages 1–1.

Nvidia, C. (2021). Jetson Nano. https://www.nvidia.com/

jetson-nano. Acessed in 16/07/2021.

Oliveira Jr, M. A. d., Sedrez, G., Monteiro, A., Puntel, F. E.,

and Cavalheiro, G. G. H. (2021). Effects of agro-

sensor time series approximation on plant stress de-

tection: An experimental study. In 2021 XIII Brazilian

Congress on Agroinformatics (SBIAGRO).

Pandas Development Team, T. (2020). pandas-dev/pandas:

Pandas.

Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V.,

Thirion, B., Grisel, O., Blondel, M., Prettenhofer,

P., Weiss, R., Dubourg, V., Vanderplas, J., Passos,

A., Cournapeau, D., Brucher, M., Perrot, M., and

Duchesnay, E. (2011). Scikit-learn: Machine learning

in Python. Journal of Machine Learning Research,

12:2825–2830.

Pereira, D. R., Papa, J. P., Saraiva, G. F. R., and Souza,

G. M. (2018). Automatic classiﬁcation of plant elec-

trophysiological responses to environmental stimuli

using machine learning and interval arithmetic. Com-

puters and Electronics in Agriculture, 145:35–42.

Proietti, M., Bianchi, F., Marini, A., Menculini, L., Termite,

L., Garinei, A., Biondi, L., and Marconi, M. (2021).

Edge intelligence with deep learning in greenhouse

management. In Proceedings of the 10th International

Conference on Smart Cities and Green ICT Systems

- Volume 1: SMARTGREENS,, pages 180–187. IN-

STICC, SciTePress.

Saraf, S. B. and Gawali, D. H. (2017). IoT based smart ir-

rigation monitoring and controlling system. In 2017

2nd IEEE International Conference on Recent Trends

in Electronics, Information & Communication Tech-

nology (RTEICT), pages 815–819. IEEE.

Shi, W., Cao, J., Zhang, Q., Li, Y., and Xu, L. (2016). Edge

computing: Vision and challenges. IEEE Internet of

Things Journal, 3(5):637–646.

Zhou, G.-X., Li, P.-q., and Zhou, X.-d. (2011). Global cli-

mate warming on climate and agriculture in western

liaoning impact. In 2011 International Conference on

Remote Sensing, Environment and Transportation En-

gineering, pages 2200–2203.

SENSORNETS 2022 - 11th International Conference on Sensor Networks

222