This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2019.2899128, IEEE
Internet of Things Journal

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) 1
<

AgriTalk: IoT for Precision Soil Farming of

Turmeric Cultivation
Wen-Liang Chen, Yi-Bing Lin, Fellow, IEEE, Yun-Wei Lin, Robert Chen, Jyun-Kai Liao, Fung-Ling
Ng, Yuan-Yao Chan, You-Cheng Liu, Chin-Cheng Wang, Cheng-Hsun Chiu, Tai-Hsiang Yen

1
crop growth, we are able to mitigate the yield risks. It is
Abstract—This paper proposes AgriTalk, an inexpensive IoT estimated that the smart agriculture market will grow 12.7%
platform for precision farming of soil cultivation. We conduct every year. However, several issues significantly affect the
experiments on turmeric cultivation, which indicates that the progress of smart agriculture industry. Many smart farming
turmeric quality is significantly enhanced through AgriTalk.
Specifically, the curcumin concentration is up to 4500-5500
studies were conducted for proof of concept or small-scale
mg/100g, which is 5 times more than existing products. We technical trials (see [1] and the references therein). In [1],
demonstrate how to intuitively configure the connections the authors developed a smart hydroponics system to
between the sensors and the actuators with the desired farming automate the growing process of the crops using Bayesian
intelligence, and to effectively maintain AgriTalk for precision Network algorithm. 60% improvement on lettuce
farming. We conduct measurement, analytic analysis, and cultivation was reported. It is not clear if small-scale
simulation experiments to investigate the IoT message delays of
experiments can be directly extended to large-scale
AgriTalk. Our study indicates that the delays for automatic
control and automatic-manual control switching with long commercial operations. Large-scale soil farming has been
distances (more than 30 Km) are very short (less than 0.2 advertised by commercial solutions [2]-[8]. Some
seconds) and AgriTalk can easily respond to quick and dynamic commercial IoT products for agriculture are too expensive.
change of the field environment conditions in soil cultivation. Furthermore, some products are difficult to install and the
Index Terms—Turmeric, Soil cultivation, Internet of maintenance costs are high. These issues prevent the
Things, Artificial Intelligence, Precision farming
farmers from accepting the smart agriculture deployment
[9].
I. INTRODUCTION
This paper proposes AgriTalk, an inexpensive IoT

T ODAY, most lands suitable for farming are already in

use. To increase the volume with the limited crop
cultivation resources, we need to improve production
platform for precision farming of soil cultivation. The name
“AgriTalk” is originated from our belief that the nature
(agriculture) always passes messages to human. However,
efficiency through precision farming. Many IoT techniques human cannot catch the meaning of the messages through
have been utilized to improve crop cultivation. Specifically, normal perception. Therefore, we develop AgriTalk that
IoT sensors collect data to monitor soil quality, weather interprets the messages of the nature environment delivered
conditions, crop growth, and so on. The IoT switch devices to the farmers.
control the agriculture actuators such as spray, drip In [10], a comprehensive overview of IoT technology use
irrigation, repellent lights and so on. Through the cases are given for precision farming, including some good
interaction between the sensors, the controllers and commercial examples for agriculture IoT devices [2]-[6].
actuators, an IoT system automates the irrigation, Most of these products can be accommodated in AgriTalk
fertilization, and pest control processes, which assists to as long as they provide appropriate application
provide better crop growth for accurate crop product programming interfaces (APIs) for porting. Although we
distribution. For example, by detecting anomalies in the are not able to access the products of all these commercial
solutions, we will briefly compare the cost of our
The corresponding author is Yi-Bing Lin. Wen-Liang Chen, Fung-Ling implementation with one of the commercial products.
Ng, Yuan-Yao Chan, You-Cheng Liu, Chin-Cheng Wang, and
Cheng-Hsun Chiu are with the College of Biological Science and Based on AgriTalk, we have established several
Technology, National Chiao Tung University, Taiwan (e-mails: IoT-based farming fields in Mountain Bao in Taiwan. This
wenurea@gmail.com, ngfungling9@gmail.com, terrychan999@gmail. paper uses turmeric cultivation as an example to show how
com, ycliu1983@nctu.edu.tw, lakewang01@gmail.com, and
chjhs920109@hotmail. com). Yi-Bing Lin, Yun-Wei Lin, Jyun-Kai Liao, AgriTalk provides precision farming. We utilize different
and Tai-Hsiang Yen are with the Department of Computer Science, farming methods in three fields to grow turmeric. Fig. 1
National Chiao Tung University, Taiwan (e-mails: liny@cs.nctu.edu.tw, shows the aerial photos of three turmeric Bao-fields of the
jyneda@gmail.com, liao85128@ gmail.com, and ksoyam95@gmail.com).
Robert Chen is the CEO of TanoSecure, Taiwan (e-mails: sizes 112.5m², 307.5m², and 600m², respectively. In the
robert@tanosecure.com). This work was supported in part by Ministry of photos, some IoT devices (weather station sensors and
Science and Technology (MOST) 106N490 and 107R491, smart irrigation system) can be seen from the air. Field
106-2221-E-009-049-MY2, 107-2221-E-009-039, and "Center for Open
Intelligent Connectivity" of National Chiao Tung University and Ministry
Bao-2 exercises bag cultivation in an open field. Bao-3 is an
of Education, Taiwan, R.O.C. open soil-based field. Bao-4 exercises bag cultivation in a

2327-4662 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2019.2899128, IEEE
Internet of Things Journal

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) 2
<

shaded field. In each field, we have fully or partially the IoT technology. Our experience indicates that it is easier
exercised precision farming to study the effects of to deploy IoT technology for hydroponic cultivation than
AgriTalk. soil cultivation. Unfortunately, rhizome plants such as
turmeric do not grow well with hydroponic cultivation. We
will discuss this issue in Section II.A. Then we describe the
factors to be monitored and controlled for turmeric
cultivation in Section II.B.
A. Hydroponic vs Soil Cultivation: from the IoT Viewpoint
This subsection compares hydroponic and soil cultivation,
and points out that deploying IoT technology for soil
cultivation is more difficult than that for hydroponic.
Hydroponic cultivates plants in mineral nutrient solutions
that are determined by the plants to be grown [11].
Typically, a hydroponic solution includes essential anions
and cations such as magnesium, sulphates, nitrate, calcium,
potassium, and so on. A hydroponic system does not require
soil, which flexibly supports cultivations in locations where
in-ground agriculture is impossible. As a closed system, the
losses of nutrient and media (water, sand, gravel, and so on)
in hydroponic cultivation are low, and therefore is more
environmental friendly with less pollution. Due to intensive
practices, the yield in hydroponic cultivation is usually
higher than normal soil cultivation. When the medium is
liquid, the materials (e.g., nutrients) are more uniformly
distributed in the medium, and can be detected by the IoT
sensors more easily than soil cultivation. A major
disadvantage of hydroponic cultivation is that its buffering
capacity is lower than the soil due to high moisture levels,
and therefore is vulnerable to pathogen attacks that result in
rapid death.
Crops grown using a hydroponics system does not have a
diverse microbiota and therefore lacks of many species of
the genus nitrobacter required to oxidize nitrite (NO2 -) to
nitrate (NO3 - ). The oxidizing process is important as
accumulation of nitrite is toxic to plants. Increasing light
intensity may help to reduce said nitrite. Additionally, the
solution of hydroponic systems only consists of 10~20
chemical components, and important metabolites normally
found in soil due to microbiota including but not limited to
Fig. 1. Magnetization The Mountain Bao fields for precision farming humic acid, amino acids, phospholipids, auxins, and
This paper investigates farming from the viewpoint of vitamins are missing, which thus impair plant health. The
IoT, and would like to encourage the farmers to accept the water-nutrient solution is also susceptible to environmental
IoT technology by reducing the costs of AgriTalk change, and requires constant attention to successfully
implementation with high reliability and availability. provide buffer from any sort of stimuli that maintain pH and
Section II describes several issues in crop cultivation and ionic concentration in this solution. Hence, it is most
then we focus on how to grow turmeric. Section III important to monitor nutrient, pH and electric conductivity
proposes AgriTalk platform for precision farming. Section (EC) values in water [10].
IV elaborates on how AgriTalk enhances turmeric In soil cultivation, the medium is the soil that has rich
cultivation. Through measurements, analytic analysis, and elements in it, such as nutrients, minerals, microorganisms,
simulation experiments, Section V investigates the time decayed stuff, water, air, gases, etc. The soil types should be
complexity of the IoT communications for farmer carefully selected to cultivate different crops, and chemical
interaction with the automatic farm control. fertilizers are typically added to the field. A great advantage
of growing outdoors is that natural soil is self-regulated, and
II. TURMERIC CULTIVATION an outdoor environment provides most of the necessary
stimuli and adjusts accordingly. The cost for soil cultivation
We use turmeric as an example for soil cultivation with is typically lower than hydroponic cultivation. The above

2327-4662 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2019.2899128, IEEE
Internet of Things Journal

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) 3
<

statement is particularly true for growing rhizome plants fN(x) = 63.2526x2+14.2131x +0.1797 (1)
[12]. The downside of natural soil is that it depends heavily
fP (x) = fK(x)= 31.6263x2+7.1065x+ 0.1798 (2)
on environmental factors and natural condition of the
Equations (1) and (2) are used in a three-layer neural
soil. Therefore, the IoT techniques for soil cultivation network model to predict the turmeric yield (rhizosphere
should respond to environmental factors much faster than weight). The input layer is the volumes (1 ≤ ≤ 3) of
that for hydroponic cultivation. Also, for soil cultivation in
N, P, and K added to the soil (unit: kg/ha). The hidden
open space, the transmission distances for IoT are typically layer contains three neurons. The output layer contains the
longer than 10 km with limited data network bandwidth and predicted productivity (unit: MT/ha). We omit the
the sensors in the soil farms are more likely to be damaged standard (and tedious) derivation, and show the final
than those in the hydroponic environment. equation iterated in the model as follows:
B. Factors for Turmeric Cultivation
Turmeric is a tropical perennial herb which grows in ( ) = linear ( )× ( )−
warm and humid climate. Its rhizomes are harvested
annually after cultivation. Some growing factors affect the where ( ) is the predicted turmeric yield obtained at the
yield and quality of turmeric production. The suitable -th iteration, ( ) is the input for the j-th neuron in the
temperature to cultivate turmeric is between 20 °C and 30 hidden layer, ( ) is the weight of the j-th neuron in the
°C with rainfall of 1500 mm or more per annum. In terms of hidden layer, and is the bias of the output layer. We
soil moisture, the appropriate range is between 60 % and iterate the above equation until converges. The
70 %. These factors directly affect the health of plants and rhizosphere weigh y is then mapped to the curcumin content
the activities of microbial population at rhizosphere. In z (in %) for Bao-3 field by the following equation [45]
particular, the soil environment must be controlled at = 0.0956 + 1.0496
appropriate pH levels and in well aeration [13]. In AgriTalk, When crops are harvested, we compare actual
the pH value in soil is a long-term monitored value that productivity with predicted productivity to calibrate
reflects the health of soil. A low pH value implies Equations (1) and (2). After a few times of calibration, the
acidification of farm field, which is probably caused by equations for the relationship between EC and productivity
abuse fertilizing. In this case, the farmer should stop will be more specific for a particular farm. In other word, y
fertilizing even if the amount of fertilizers in soil is lower (and actually, z) is used to evaluate the growth rate, and
than the threshold. According to [44], when pH is less than then based on its value (by comparing the goal, say,
5, the nitrification process cannot proceed. When pH is 6, = 4500~5500 mg/100g ), AgriTalk responds by
7.5 and 8.6, the nitrification proceeds at more rapid rates, controlling fertilizer through Equations (1) and (2). For the
and is able to release 116, 123 and 118 mg/kg of nitrate-N in Bao-3 field, the threshold value of the nitrogen fertilizer is
the soil, respectively. Following [44] and [13], we maintain about 130 g/m2 per week, and each of the phosphorus and
the soil pH value between 6.1 and 6.5 for well turmeric the potassium thresholds is about 65 g/m2 per week.
cultivation. This pH range ensures that nitrite accumulation During cultivation of turmeric, diseases that will erupt
can be minimized and will not affect the growth of turmeric. and spread such as leaf blotch caused by Taphrina
An equilibrium soil ecology plays an important role to maculans, leaf spot caused by Colletotrichum capsica, root
give high yield of turmeric products. Various rot caused by Fusarium solani, and so on [18]. Moreover,
microorganisms such as plant growth-promoting pests of turmeric including Conogethes punctiferalis (order:
rhizobacteria (PGPR) and ectomycorrhizal fungi have Lepidoptera), Aspidiella hartii (order: Hemiptera) and so on,
symbiotic relationship with plant roots and normally aid in will cause shoot borer and rhizome scale problem [19]. Pest
transferring insoluble minerals for plant absorption distribution and eruption of diseases in farm can be
[14]-[16]. Through real-time data collection of electrical predicted based on the environment factors including
conductivity (EC), pH, and soil moisture, a fertilizer temperature, humidity, wind speed, pest number, and so on.
regulatory model can be trained to fit a specific farm. This By considering these factors, we established an initial
model then provides appropriate fertilizer solution for the model to describe the relationship between insect egg
farmer to optimize the yield of turmeric throughout the hatching period, temperature t and humidity h. This model
whole process of cultivation. Relationship between was then calibrated through egg hatching experiments to
concentration of salt and EC can be described as a quadratic develop the final model specific to Spodoptera litura (order:
equation [17]. By using this concept, we first derived Lepidoptera). The equation for the final model is listed
general equations to describe the relationship between the below, which will be used in Section IV.A.
EC as the input and the Nitrogen (N), the Phosphorus (P) fE(t,h) = 0.0169Í(0.03413t2 -2.162t +
and the Potassium (K) values as the outputs. Then we 36.56)Í(-0.007421h2 + 1.123h -37.57) + 2.049 (3)
constructed experiments in Bao fields with various
combinations of N, P, K and measured the EC values to III. PRECISION FARMING BY AGRITALK
calibrate the general equations. Finally, the equations of N, This section proposes AgriTalk, an IoT platform that
P, K were generated as below. provides precision farming by allowing the farmer to

2327-4662 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2019.2899128, IEEE
Internet of Things Journal

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) 4
<

automatically and remotely manage the irrigation and the sensors as well as soil sensors (to be described) are
pest control systems. AgriTalk includes several agriculture connected to the SnsrCtl boards through wires (Fig. 3
IoT devices (Fig. 2 (1) and (2)) connected to the control (6)).
boards called “AgriCtls” (Fig. 2 (3) and (4)). An AgriCtl Soil and Insect Sensors (Fig. 2 (1)): Besides the weather
board is an Arduino microcontroller board that controls the station sensors for detecting the conditions in the air,
connected IoT devices. An AgriCtl board connected to the several soil sensors are used to grow turmeric in
sensors is called SensrCtl (sensor control board; Fig. 2 (3)). AgriTalk. These sensors are inserted in the soils of the
In the cloud, the AgriTalk server (Fig. 2 (5)) including the field to collect data specific to turmeric cultivation,
IoTtalk engine (Fig. 2 (6)) and the AgriGUI (Fig. 2 (7)) is including the 3-in-1 sensor for moisture, temperature,
deployed to configure the IoT devices with analytical and electrical conductivity (EC) (Fig. 4 (a)), and the pH
capabilities, in-built accounting/reporting features, and sensors (Fig. 4 (b)). Operational ranges for these sensors
powerful dashboards/controls that can be accessed remotely are given in Section II.B. Also, we have developed insect
through smartphones (Fig. 2 (8) and (9)). Details of IoTtalk trappers (Fig. 4 (c)) that are installed at the pole of the
and AgriTalk will be given in Section III.B weather station (Fig. 3 (7)). The soil sensors and the
insect trapper are connected to the SnsrCtl board through
wires (Fig. 3 (8)). From these sensors, we monitor the
turmeric growth and any anomalies to effectively prevent
any insects that can harm the yield. Our sensor hardware
solution is similar to commercial solutions [21][22],
which can serve as good representations of how these
sensors can be applied in real-life use cases.

Fig. 2. The AgriTalk architecture

A. IoT Devices for AgriTalk

To develop IoT-based precision farming, it is essential to
select appropriate sensors and actuators. Depending on the
crops to be grown, the types of information to be collected Fig. 3. The weather station of AgriTalk
determine the IoT devices to be deployed. The quality of the Actuators for Soil Cultivation (Fig. 2 (2)): Three major
IoT devices is critical to the accuracy of the collected data actuators are used in AgriTalk for turmeric cultivation.
and its reliability. Several IoT devices are utilized in The pest sprayers (Fig. 5 (1) ~ (3)) spray biopesticides for
AgriTalk, including the weather station and the soil sensors pest regulation. These sprayers are connected to the
for precision monitoring, and the cultivation actuators for biopesticide tank (Fig. 6 (5)). The biopesticide liquid is
irrigation, pest control and fertilization. Although we have sent to the sprayers by the pump (Fig. 6 (3)) controlled by
developed these IoT devices, AgriTalk can flexibly the relays (Fig. 6 (1)). The relays are connected to the
accommodate commercial IoT devices if they have better ActrCtl board (Fig. 2 (4)) that executes the instructions
performance and/or lower prices. This paper focuses on given by the AgriTalk sensors or the smartphones
turmeric growth, and the AgriTalk IoT devices for turmeric through the IoTtalk engine (Fig. 2 (6)). The biopesticide
cultivation are elaborated as follows. liquid is a low drug-resistant and eco-friendly pesticide
Weather Station (Fig. 2 (1)): Located across the field, the invented by our team, which is highly specific to the
sensors in the weather station collect various data from pests but non-toxic to mammals and pollinators [23][24].
the environment, which are sent to the IoTtalk engine in The repellent bulbs (Fig. 5 (4) ~ (6)) are used to keep
the cloud. The provided measurements are used to map flying insects away. Many types of insects have trouble
the climate conditions to drive the actuators for precision perceiving light with a lower color temperature, and the
farming. Fig. 3 illustrates the weather station of AgriTalk. repellent bulb shifts the color temperature to be nearly
On the top, there are CO2, temperature (-40°C~65°C), invisible to many flying insects. The repellent bulbs are
humidity (1~100% RH) and atmosphere pressure sensors connected to the relays (Fig. 6 (1)) for automatic or
(Fig. 3 (1)), the rain gauge (0~2438mm/hr; Fig. 3 (2)), the manual control. The drippers (Fig. 5 (7)) are connected to
ultraviolet sensor (0 ~ 1800 W/m2; Fig. 3 (3)), and the the irrigation tanks (Fig. 6 (4) and (6)) to form a drip
wind gauge (0 ~ 89 m/s; Fig. 3 (4)). Multiple SnsrCtl irrigation system with the control mechanism similar to
boards (Fig. 2 (3)) are installed in a waterproof box (Fig. that for pest control. Together with the water-resist bleed
3 (5)) mounted in the pole, and the weather station Tyvek clothes (Fig. 5 (8)) that preserves the water in the
soil, AgriTalk drip irrigation integrates water and

2327-4662 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2019.2899128, IEEE
Internet of Things Journal

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) 5
<

fertilizer for precise irrigation that saves more than 70% without the need of installing any mobile app in the
of water from the observation of the water consumption smartphone.
in the tanks (Fig. 6 (4)-(6)), which approaches to the AgriTalk provides a timer to periodically activate an
performance of hydroponic cultivation. actuator (e.g., a dripper). The timer supports automatic
farming, which is the first step toward smart farming. In
AgriTalk, a timer is implemented by a computer program
that is considered as a “cyber” IoT device. More details of
the timer will be elaborated in the next subsection.
We also monitor the power consumption of every device.
Fig. 9 (1) illustrates the dashboard for power consumption
of various devices. Fig. 9 (2) shows the time series of the
electric current for the repellent light and the irrigation
Fig. 4. The soil sensors; (a) a 3-in-one sensor for moisture, temperature systems at Field Bao-2, which indicates that the lights are
and EC detection; (b) a pH sensor; (c) different kinds of insect trappers turned on between 19:00 in the evening and 5:00 in the next
morning. In the time series, the sharp peaks occurring every
10 minutes are caused by pump pressurization for dripping.

Fig. 5. The repellent bulbs, sprays and drips

Fig. 7. A smartphone as the AgriTalk controller

Fig. 6. Irrigation, fertilization and pest control

Dashboard and Controller (Fig. 2 (8) and (9)): In

AgriTalk, a standard smartphone can be used to offer
remote farm monitoring capabilities without installing
any mobile application. The smartphone in Fig. 2 (8) is a
controller that provides soft switches (Fig. 7 (1)) to turn
on/off the relays in Fig. 6 (1). Optionally, a camera may
be installed in the field (e.g., mounted on the weather
station) if the cost is justified. In this case, the AgriTalk
smartphone may offer a monitor screen (Fig. 7 (2)) that
Fig. 8. A smartphone as the AgriTalk dashboard (one sample per 15
allows the farmer to remotely zoom in/out and rotate the seconds)
camera in the field. Fig. 7 (3) shows that one of the
authors turned on the repellent bulbs of a farm of B. Configuring AgriTalk Application
Mountain Bao in Taiwan from London Heathrow Airport. An IoT device management platform called IoTtalk
Similarly, a smartphone in Fig. 2 (9) can serve as the [25]-[28] is used to deploy AgriTalk. IoTtalk is centered at
AgriTalk dashboard illustrated in Fig. 8 (1). To see the concepts of device feature (DF) and device model. An
detailed sensor data, the farmer clicks the “More Info” input DF (IDF) sends the data to the IoTtalk engine and an
button (Fig. 8 (2)), and the time series chart of the sensor output DF (ODF) receives the data from the IoTtalk engine.
pops up as illustrated in Fig. 8 (3). Similar remote AgriTalk is built on top of IoTtalk, where every sensor is
solutions are represented by [7] and [8] through mobile mapped to an IDF (e.g., in Fig. 2, (1) is mapped to (11)), and
apps. Based on a technique called smartphoneTalk [25], a
every actuator is mapped to an ODF (e.g., in Fig. 2, (2) is
major advantage of our solution is that the web-based
mapped to (12)). In IoTtalk a device model is a set of DFs,
approach is implemented so that every smartphone can
access AgriTalk resources through the web browser which is partitioned into two subsets. The subsect of IDFs is

2327-4662 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2019.2899128, IEEE
Internet of Things Journal

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) 6
<

called the input device, and the subset of ODFs is called the agriculture device modules are pre-built by the AgriTalk
output device. In AgriTalk, a SnsrCtl board (Fig. 2 (3)) designer. Following the procedure described in [27], the
connected to multiple sensors is mapped to an input device device modules are created and are listed in the “Model”
model, and an ActrCtl board (Fig. 2 (4)) connected to drop-down menu list (Fig. 10 (7)). The AgriTalk engineer
multiple relays is mapped to an output device model. Every (with the assistance of the farmer) deploys precision
AgriCtl board is installed two software units. The Device farming of the fields in Mountain Bao by creating the Bao
Application (DA; see Fig. 2 (11) and (12)) communicates project (Fig. 10 (8)) with the following simple steps. When
with the IoTtalk engine by sending the IDF values or the AgriTalk engineer selects a device model (e.g.,
receiving the ODF values. The IoT Device Application WeatherSTA) form the “Model” list, the icon for that
(IDA) is a software driver for the IoT device hardware. For device is shown in the AgriGUI window, and the AgriTalk
example, the IDA for the SnsrCtl board (Fig. 2 (10)) engineer can manipulate the device following the procedure
receives the sensor signals and translates them into the described below.
meaningful values in the JSON format. On the other hand,
the IDA for the SnsrCtl board (Fig. 2 (13)) translates ODF
values to control signals that will be sent to the actuators.
The communications between the AgriCtl boards and the
IoTtalk engine is specified through the AgriTalk Graphical
User Interface called AgriGUI (Fig. 2 (7)). The GUI allows
the designer to create a project by configuring the sensors,
the actuators and their connections for precision farming.

Fig. 9. Power consumption monitoring

Fig. 10 illustrates the AgriGUI web window for an

AgriTalk project “Bao” (Fig. 10 (8)) of the Mountain Bao
fields. In this window, the IoT devices and the DFs are
represented by the icons. The input device icons are placed Fig. 10. Configuring precision farming for the fields in Mountain Bao
on the left-hand side (Fig. 10 (1)-(3)) and the output device To connect a sensor to an actuator, we only need to drag a
icons are placed on the right-hand side (Fig. 10 (4)-(6)). line between the corresponding IDF icon and the ODF icon,
Inside a device icon, there are one or more small icons that and the IoTtalk engine automatically generates a program to
represent the DFs. For example, inside WeatherSTA (the handle the interaction between them. A Join line has two
weather station icon), there are four IDF icons for rain segments. Every segment connects a DF to a circle. By
gauge, humidity, temperature and insect trapper, clicking the circle, the AgriTalk engineer can write a
respectively. For the readability purpose, every IDF name is function for the connection. For example, when the circle of
appended with “-I” and every ODF name is appended with Join 2 is clicked, a window (Fig. 11) pops up. In this
“-O”. In the Bao project, the IDF icons within the window, we write a Python function “n-demand” based on
SoilSensor icon (Fig. 10 (1)) represents the soil sensors fN(x) (i.e., Equation (1) in Section II.B). This function
connected to the SnsrCtl board. The IDF icons within the estimates the amount of nitrogen fertilizer (Nitrogen -O that
WeatherSTA icon (Fig. 10 (2)) represents the sensors receives the fN(x) values), which is controlled by the EC
connected to the SnsrCtl board of the weather station. The sensor (EC-I that produces the x values). At Line 1 in Fig.
Timer icon (Fig. 10 (3)) represents the AgriTalk timer that 11, the input argument *args of the Python function run()
can be set to turn on/off of the relays. The Irrigation icon stores the EC and the pH values of soil sent from EC-I and
(Fig. 10 (4)) includes the ODFs for the relays connected to pH-I, respectively (i.e., EC=args[0] and pH=args[1]). Line
the water and fertilizer ActrCtl board. Similarly, the 2 sets the threshold (for example, the amount of demand
PestSpray icon (Fig. 10 (5)) and the RepllentBulb icon (Fig. nitrogen fertilizer amount is about 130 g/m2 per week for
10 (6)) include the ODFs for the pesticide and the repellent the Bao-3 field, which is represented by 13 units). Line 3
bulb relays. With the help of an IoTtalk engineer, all evaluates the amount of nitrogen fertilizer in the soil by

2327-4662 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2019.2899128, IEEE
Internet of Things Journal

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) 7
<

using Equation (1) in Section II.B. When the soil EC value controlled by the humidity and the temperature through Join
continuously increases throughout cultivation period, it 7. It is also controlled by the insect trapper (the details are
means that there is excessive fertilizer in soil. When the EC given in Section IV.A). Then the amount of pesticide to be
value decreases below the threshold (the EC value x when, sprayed is computed at PestSpray (Fig. 10 (5)), which is
e.g., fN(x)=13 units in the Bao-3 field), it means that there is proportional to fE(t,h) in Equation (3) (where t is obtained
depleted fertilizers in soil. The amount of specific fertilizers from Temperature-I and h is obtained from Humidity-I) and
(Nitrogen, Phosphorus and Potassium) in soil can be the number of captured bugs obtained from InsectTrapper-I.
predicted based on fN(x), fP(x), fK(x). Therefore precise Also, UV-I (the values of ultraviolet sensor) connected to
addition of fertilizers can achieve. If the amount of nitrogen Join 7 is used to determine the bio-pesticides degradation
fertilizer is depleted (the EC value decreases) and the soil is rate. The details are given in Section IV.A.
not acidified (i.e., pH>5; see Line 5), AgriTalk makes up
the shortage quantity (Line 6). If the amount of nitrogen IV. PRECISION PERFORMANCE OF AGRITALK
fertilizer is excessive (the EC value increases above This section studies the performance of AgriTalk for
fN(x)=13 units) or the soil is about to be acidified, AgriTalk precision farming. Section IV.A shows how AgriTalk
turns off the switch by returning 0 (Line 7). significantly enhances turmeric growth. Section IV.B
Similarly, the functions fP(x) (for phosphorus) and fK(x) investigates how to detect sensor failure and conduct
(for potassium) are implemented at the circles of Joins 3 and calibration for the sensors. Section IV.C shows the effect of
4 based on Equation (2) in Section II.B. the frequency for sending the sensor data.

Fig. 12. Effect of repellent bulbs

A. Turmeric Growth Performance

For pest control, we put the sex pheromone of spodoptera
litura at the center of the insect trapper, and use the repellent
bulbs to drive the bugs away from the turmeric field to the
trapper. Fig. 12 plots the number of insects captured in the
trapper. In this figure, the red “square wave” curve shows
the on/off status of lights, where the repellent bulbs are
Fig. 11. Implementing the nitrogen control function “n-demand” in turned on in the nighttime before and at , and
AgriTalk (see the black bullets). The bulbs are turned off in the
daytime at , and . Therefore the repellent light
Drip-O in the irrigation system (Fig. 10 (4)) is driven by
system is activated before , during [ , ], [ , ], and
Joins 1, 5 and 6. When the humidity level is below a
threshold (indicated by Moisture-I and RainGauge-I), the after . The yellow curve plots the counter value of the
relay is turned on. In the Bao project, the timer (Fig. 10 (3)) insect trapper, which shows that the trapper does not catch
is used to perform automatic farming. Specifically, dripping the bugs when the bulbs are turned off. The trapper
is controlled by the timer that regularly activates the dripper continues to count the number of captured insects until no
at 17:30 and deactivates it at 18:00 every day through Join 5 more insects are captured (at points , and ). The
(for activation) and Join 6 (for deactivation). Through Join trapper waits for a while and then resets the counter to 0 (at
5, the drippers are also turned off if RainGauge-I and points , and ). According to the number of adult
Humidity-I indicate that it rains. The repellent light system insects captured by the trapper and the life cycle of bugs
(Fig. 10 (6)) is controlled by the timer through Joins 8 (for such as in lepidoptera order, the hatching period of pest
activation) and 9 (for deactivation). In this configuration eggs is about 2 to 14 days. The variation of hatching period
example, the light bulbs are turned on at 19:00 in every is affected by various combinations of temperature and
evening and are turned off at 5:00 in the morning of the next humidity factors, and the details can be found in [31]. By
day. The results are reflected in the power consumption monitoring the number of captured adult insects (in our
chart in Fig. 9 (2)). example, the threshold is 60 pests per 10 days), the system
For precision farming, the fertilizer in the liquid is alerts and suggests when to spray bio-pesticides before
controlled by Joins 2~4, where the ingredients are those larvae occur in farm through the Random Forest
dynamically adjusted by Equations (1) and (2). Following based machine learning model. The AI mechanism has been
the theory in [29][30], pastSpray-O (Fig. 10 (5)) is implemented in IoTtalk [32] to provide automatic training

2327-4662 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2019.2899128, IEEE
Internet of Things Journal

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) 8
<

of the pest control models. Through monitoring the EC value changes in real time,
The insect counter together with the humidity and the when to fertilize can be determined through Equations (1)
temperature sensors in WeatherSTA (to compute Equation and (2), and thus effectively increase the chlorophyll of
(3)) dynamically control the pest sprayers through Join 7 in turmeric plants. The chlorophyll contents in leaves were
Fig.10. In the pest control system, a model of bio-pesticides measured with Soil-Plant Analyses Development (SPAD)
degradation rate based on ultraviolet (UV) radiolytic values. Fig. 14 shows the effects of chlorophyll amount
oxidation was developed. Our study [33] proved that UV increase with dynamic fertilizing regulation through
radiolysis has the highest impact on degradation of AgriTalk (Joins 1~4 in Fig. 10). With AgriTalk, the SPAD
bio-pesticides. Therefore, by monitoring the UV intensity, is increased by 40~60%. With AgriTalk, the curcumin
AgriTalk determines the next spraying time after concentration in turmeric rhizome is up to 4500~5500
bio-pesticides are degraded below the effective level. Thus, mg/100g, which is 5 times more than existing products.
the problem of bio-pesticides over-usage can be prevented
B. Failure Detection and Calibration for the Sensors
to achieve agricultural sustainability [34]. In our model, the
concentration of bio-pesticides is derived from the Maintenance of IoT hardware is of primary importance
following equation: for precision farming. In turmeric cultivation, the sensors
d[ ] used in the field may be easily damaged. Thus, we need to
− = ensure that the hardware failure can be quickly detected and
d
where KUV is the rate constant of UV radiolytic oxidation to appropriately replaced.
proteins, which is set to 44 mol-1 s-1 at the beginning [35],
and is the total concentration of UV generated radicals,
which is derived from the following equation
d[ ]
= −
d
where is the dose rate of absorbing γ-ray (radiation
energy absorption rate per mass; for water, is 1.42 Gy/s),
is the number of radicals created per energy (for water,
is 0.045 μmol/J) [36]. In this equation, I is the intensity
of UVB from sunlight measured by the UV sensor
(UVM30A), where UVB is radiation in about 280~320 nm
in wavelength to effect on promoting protein degradation
activity. RT is the rate constant of radicals termination,
which is equal to 2.365×10-7 mol-1 s-1 [35]. When is
below the threshold and when there are bugs as indicated in
Fig. 12, the pest sprayers are turned on. Fig. 13 shows the
effects of pest regulation. Before May 15, 2018, without
using the pest control system, the plants were invaded by
pest and leaves breakage as shown in Fig. 13 (a). After
using the pest control system and applying bio-pesticides
when alerting, pest invasion became much less serious on
June 4, 2018. In Fig. 13 (b), the white circle marks a worm
that is hungry to death on leaf due to the bio-pesticide.

Fig. 14. Chlorophyll SPAD performance with and without AgriTalk

In AgriTalk, there are two simple ways to detect

abnormality of the IoT devices. First, AgriTalk
automatically compares the data of the nearby sensors of
the same type. It is clear that the measured results of these
Fig. 13. Effects of pest regulation
sensors are different. However, in the normal situation, the

2327-4662 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2019.2899128, IEEE
Internet of Things Journal

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) 9
<

variance of the differences should be stable. If the variance connect the IDFs of StdSensor through Joins 5 and 6 such
dramatically changes, the AgriTalk operator will be alerted. that the standard sensor can calibrate the moisture sensors
Fig. 8 (3) and (4) show the time series of the moisture data by moving the portable sensor next to the sensor under test.
for two sensors of the same brand located 50 meters away. The smartphone (Fig. 16 (7)) displays the results for the
The variance of the differences between these two sets of comparison of the sensor devices (Fig. 16 (8) and (9)).
time series is stable before the time point at Fig. 8 (5). The Specifically, the values of both the sensor under test and the
alert, for example, was triggered at this time point. standard sensor at the same time are paired and become an
Although the change of variance may not due to sensor entry stored in a mapping table in the Moisture-I’s IDA in
failure, such alert is very useful for quick failure detection. the SnsrCtl board for the sensor device.
Fig. 15 (a) compares the time series for two moisture
sensors of different brands (one is wired and the other is
wireless). The chart indicates that from the viewpoint of
statistics, the two sets of time series are similar. However,
they are “systematically” different due to hardware
implementation, measure location differences and
measured time delays (the delays for the data sent to the
AgriTalk server may not be the same for these two sensors).
Similarly, Fig. 15 (b) shows the time series for a wireless
and a wired temperature sensors. They are also similar.

Fig. 15. Comparing the sensors of different brands at the same location
(one sample per hour): (a) soil moisture sensors; (b) air temperature
sensors

Second, we monitor the power consumption of every

device. As illustrated in Fig. 9, when the power
consumption levels dramatically change, the alert is sent to
the AgriTalk operator for attention.
A sensor may need calibration after abnormality is
detected. If the failure is caused by the sensor material, then Fig. 16. Configuring sensor failure detection and calibration in AgriGUI
the senor hardware must be replaced. On the other hand, if
If there are enough entries in the mapping table, a
the sensor accuracy can be calibrated by the software, then
mapping curve can be generated through regression. Then
it can be done by modifying the sensor’s IDA (Fig. 2 (10))
the IDA can use this mapping table to calibrate the values
through the following procedure. In Fig. 16, the moisture
measured from the sensor in the soil. In this way, AgriTalk
sensors in SoilSensor1 (Fig. 16 (1)) and SoilSensor2 (Fig.
automatically calibrates the sensor. The beauty of our
16 (2)) are the same as that in Fig. 10 (1), which are
solution is that the firmware of the off-the-shelf sensor
connected to the Irrigation icon (Fig. 16 (3)) similar to that
needs not be modified (which cannot be done without the
in Fig. 10 (4). To test these moisture sensors, we add three
original manufacture’s special equipment). Instead, we can
devices to the Bao project. The standard moisture sensor
flexibly modify the calibration (mapping) table of the IDA
(Fig. 16 (4)) is used to compare with the sensors under test,
stored in the SensrCtl board. Not that we do not need to
which is a portable sensor device that can be plugged in the
adjust the sensor for its full operation range. We only need
smartphone (Fig. 17). This portable sensor device is called
to calibrate the sensor under the Bao field’s specific
MorSensor chip developed by our team [37], which is
environment, which is a sub-range for the sensor’s
wireless with smartphone mobility, and therefore can be put
operation.
closely to the sensor under test for comparison. The cyber
device “Test” consists of the output device (that receives C. Sampling Rate Analysis
the data from the sensors under test and the standard sensor;
The sensors report their measured data periodically. The
see Fig. 16 (5)) and the input device (that sends out the test
reporting frequency is critical to precision farming. In soil
results; see Fig. 16 (6)). To compare two nearby sensors
cultivation, the environment conditions may rapidly change,
under test, we simply connect both IDFs to the Test device
and therefore the sensor data should be reported sufficiently
through Joins 3 and 4. Then the Test device will produce
frequently. With the Ethernet communication, AgriTalk is
time series charts as those shown in Fig. 15. When some
designed for sensor reporting at the frequency up to 4 per
unusual situations are detected by the Test device, we can
minute. For a commercial solution with reasonable battery

2327-4662 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2019.2899128, IEEE
Internet of Things Journal

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) 10
<

Following the same approach in [28], the IoT message

delays are measured using the Network Time Protocol
(NTP). That is, each of the AgriCtl boards and the IoTtalk
server is installed a timer module to synchronize with the
NTP. We have obtained the histograms for through
1000 measurements for each of the ADSL and the 4G
transmission scenarios (Fig. 19). From the measured
samples, we computed the expected value E[tm] = 42.638
ms and the variance V[tm] = 0.018 E[tm]2 for ADSL
transmission, and E[tm] =85.958 ms and V[tm] = 0.057E[tm]2
for 4G transmission. It is clearly that the delays for 4G are
about twice of that for ADSL. Also, the variance for 4G is
higher than ADLS, which means that wireless
communications is not as stable as wired communications.
Fig. 17. Portable sensors used for AgriTalk sensor calibration In AgriTalk, all actuators are automatically driven by the
sensors. However, for some major events, the farmer should
budget, the wireless-based sensors connected to AgriTalk be able to step in to abort the actions activated by the
cannot perform sensor reporting with the frequency higher sensors. In AgriTalk, the farmer’s smartphone is alerted of
than one per hour. Fig. 18 illustrates two sets of the time the major events through the IoTtalk engine. In Fig. 20,
series of a moisture sensor, where the black curve is the suppose that the EC sensor reports a unreasonable low EC
time series at the one per hour frequency and the gray curve value, which is received by the IoTtalk engine at time
is the time series at the 4 per minute frequency. The figure (through path (1)->(3)->(6) in Fig. 2). The IoTtalk engine
indicates that the black curve cannot capture the dynamic executes the join function and decides to activate the
change of the measured target, and the difference can range fertilizer drippers by sending the instruction to the dripper
from 50mm to 65mm. Such inaccuracy may cause problem relay through path (6)->(4)->(2) in Fig. 2. The relay
in the irrigation system. For example, after the sensor receives the instruction at . The IoTtalk engine also sends
reports that the humidity is low, the irrigation system may an alert to the farmer through path (6)->(9) in Fig. 2. The
be triggered to turn on the drippers. But after a few minutes, farmer receives the alert at . If the farmer suspects that the
heavy rainfall may occur (which is typical in Taiwan), and EC sensor fails and the decision based on the EC value
the irrigation system will not stop the drippers until the next should not be carried out, then he/she may decide to cancel
sensor report arrives at the IoT server one hour later. The the action by sending the abort request to the IoTtalk engine
sampling frequency in AgriTalk is much higher than the through path (8)->(6) in Fig. 2 (note that in this case, the
existing smart farming solutions [1]-[8][10]. In the future, smartphones (8) and (9) are the same). The message arrives
we will use the method in [38] to investigate the at the IoTtalk engine at . The IoTtalk engine sends the
relationship between the sampling frequency and instruction to turn the dripper relay off through path
agriculture data accuracy. (6)->(4)->(2) in Fig. 2. The instruction arrives at the relay at
. Note that in Fig. 20, we do not consider the farmer’s
think time that starts at because this study aims to
investigate the AgriTalk delays.

Fig. 18. Effect of the sample frequency on sensor data reporting

V. IOT MESSAGE DELAY ANALYSIS

This section conducts the measurement experiments, and
then proposes an analytic model and the simulation
experiments to evaluate the message delays among the IoT
devices. Let be the delay of a message delivered
between an IoT device and the AgriTalk engine, that is, the
delays for paths (1)->(3)->(6) , (6)->(4)->(2) , (6)->(9) and
(8)->(6) in Fig. 2. Note that the sensors and the actuators are
Fig. 19. The histograms
connected to the AgriCtl boards through wires. We consider
both ADSL (wired) and 4G (wireless) communications for Let be a random variable with the density
connections (3)->(6), (6)->(4), (6)->(9) and (8)->(6). function ( ) . In Fig. 20, let , = − , , =

2327-4662 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2019.2899128, IEEE
Internet of Things Journal

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) 11
<

− , , = − , and , = − . It clear 2 1 (2λ)

that = +
, , , , , , and , have the same density λ 2 λ 2
function , for 1 ≤ ≤ 4 . Let = , + , + 17
, which is a random variable with the density function = (9)
, 8λ
( ). If the action needs to be cancelled, then the delay The mean value analysis indicates the impact of the mean
between when the actuator is activated and when the for . Equation (6) suggests that the probability that
actuator is deactivated is expressed as = 0 is small. Similarly, Equation (8) suggests that is
0, , > roughly twice of . To see the impact of other moments of
= (4)
− , , , < the random variable, we need to use its distribution.
Let ( )=∫ (t) t be the cumulative distribution From the histogram illustrated in Fig. 19, can be
approximated by an Erlang distribution with the density
of . Then we have
function and the cumulative function
Pr[ = 0] = ( ) , , (5) λ
( )= and
,
( − 1)!
= 1− ( ) , (6)
λ
For the mean value analysis [39], we assume that ( )=1− (10)
!
, =λ , is an Exponential density function.

Therefore, the summation of three i.i.d. exponential random where the shape parameter is = 56 and the scale
variables has an Erlang-3 density function ( ) = parameter is λ = 1.313 for ADSL and, = 17 and λ =
(λ )/2, and (5) is expressed as 0.198 for 4G. The approximation is validated by the
Kolmogorov-Smirnov test for goodness of fit. Then from
the convolution of the Erlang distribution, the density
function of are
λ
( )= (11)
(3 − 1)!
Substitute (10) into (6) to yield
λ
Fig. 20. Timing diagram for action deactivation Pr[ = 0] = ( )
!
Pr[ = 0]
λ λ
= λ , = ( ) (12)
2 , !
,
λ 1 From the frequency-domain general derivative of the
= = (7) Laplace transform, for a function f(t) we have
2 8 ∗( )
From (4), the expected value of is derived as ( )
( ) = (−1) (13)
d
E[ ] = − , ( ) , , From (12) and (13), we have
,
∗( )
(−λ) ( )
= ( ) ( ) Pr[ = 0] = (14)
! d

( ) From (11), the Laplace Transform of ( ) is

− , , , (8)
, ∗(
λ
)= ( ) = (15)
In the mean value analysis, (8) is expressed as (s + λ)
3 λ Substitute (15) into (14) to yield
E[ ] = −
λ 2 Pr[ = 0]
λ λ 3 (3 + 1) … (3 + − 1)λ
− , λ ,
, =
,
2 ! (s + λ)
3 λ
= − 3 + −1 1
λ 2 = (16)
2
λ 1
− λ ,
,
, Equation (16) indicates that for = 1, Pr[ = 0] =1/8,
2 λ ,
which is the same as (7), and lim → Pr[ = 0] = 0.

2327-4662 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2019.2899128, IEEE
Internet of Things Journal

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) 12
<

For the derivation purpose, E[ ] in (8) can be expressed as ADSL scenario, and Pr[ = 0] = 0.00001 and E[ ] =
E[ ] = − (17) 171.916 ms for the 4G scenario. The above results indicate
where that AgriTalk with ADSL has excellent message delay
performance, and has reasonably good performance with
= ( ) ( ) (18) 4G.
and We further investigate the effect of the message delay’s
variance. Fig. 21 plots the Pr[ = 0] and the E[ ] curves
= , ( ) , , (19) against variances (i.e., the shape parameter n) with fixed
, E[tm] = 42.638 ms for the ADSL scenario, and E[tm] =
From (10), (11) and (18), we have 85.958 ms for the 4G scenario. Note that for the same E[tm]
λ λ values, the variance V[tm] decreases as the shape parameter
= 1− n increases. The figure indicates that the Pr[ = 0] curves
(3 − 1)! !
are consistent with Equation (16), which approaches to 0 as
n increases. The figure also indicates that as n increases, the
3 λ variance decreases and E[ ] decreases. The E[ ] value is
= − almost fixed for n>6. Therefore, to achieve good E[ ]
λ ! (3 − 1)!
performance the communication technology should be
3 3 + −1 3 + 1 selected such that n>6. For AgriTalk, = 56 for ADSL
= − (20) and = 17 for 4G, and good performance is achieved.
λ 2λ 2
From (10), (11) and (19) we have
λ
= ,
,
(3 − 1)!
λ ,
,
× ,
( − 1)!
λ λ
= 1−
(3 − 1)! λ !

λ λ
= 1−
λ (3 − 1)! !
Fig. 21. Effect of n (V[tm]) on Pr[ = 0] and E[ ] (E[tm] = 42.638 ms
for the ADSL scenario, and E[tm] = 85.958 ms for the 4G scenario).
3 + −1 1
= 1− (21)
λ 2 VI. CONCLUSION
Substitute (20) and (21) into (17) to yield This paper developed AgriTalk, an IoT platform for
outdoor soil cultivation. We experimented precision
2 3 + −1 3 + 1 farming for turmeric cultivation to show that with
E[ ] = −
λ 2λ 2 intelligent IoT technologies, the problems of soil cultivation
can be significantly mitigated. Through AgriTalk, the
3 + −1 1
+ turmeric quality has been significantly enhanced.
λ 2 Specifically, with AgriTalk, the curcumin concentration in
2 4 −1 1 turmeric leaf after 6 months is 685.6 M as compared with
= + normal cultivation that is 72.1 M. Our turmeric products
λ λ 2
are illustrated in Fig. 22. The curcumin concentration is up
3 + −1 + 1
− (22) to4500~5500 mg/100g, which is 5 times more than existing
2λ 2 products.
For = 1, (22) is simplified as
2 3 1 1 2 1 1 17
E[ ] = + − =
λ 1 λ 2 0 2λ 2 8λ
which is the same as the result of the mean value analysis
given in (9). Equations (16) and (22) are further validated
against the simulation experiments, and the errors are
within 0.5%. The simulation model follows the same
approach used in [20][28], and the details are omitted. In Fig. 22. Turmeric products by AgriTalk
Fig. 19, Pr[ = 0] = 0 and E[ ] = 85.275 ms for the AgriTalk provides a user-friendly GUI called AgriGUI

2327-4662 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2019.2899128, IEEE
Internet of Things Journal

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) 13
<

that allows the designer to quickly and intuitively configure hatching. The regression model developed in this paper has
the connections between the sensors and the actuators with shown good results. To further predict the rate and direction
the desired intelligence. Novelty of AgriGUI is that one can of diseases eruption, wind speed and wind direction in field
immediately figures out how the sensors/controls affect the will be collected in the future. Also, other machine learning
actuators through graphical connections. Such feature has models will be built besides the technique used in Section
not been found in the existing solutions. For the farmer, IV.A. We have deployed 4 more fields with total 1.1
AgriTalk implements the web-based dashboard and the hectares, which will grow medical plants such as
controller so that the farmer can use any smartphone to Astragalus, Bletilla striata, Codonopsis pilosula, etc. These
transparently access AgriTalk resources through the web fields are located in different altitudes to verify the
browser without the need of installing any mobile app. accessibility of Agritalk. We aim to reduce the cost of
Maintenance of IoT precision farming is effective by AgriTalk as much as possible. Quanta, one of the largest IT
using the IoTtalk features. In particular, we showed how to companies in the world is manufacturing our sensor/
configure connections for quick detection of failure devices actuator design for mass production with low costs. The
and calibration. Sensor calibration for farming has not been purpose is to allow the farmers to use affordable IoT
addressed in the literature. In AgriTalk, there are two technology for smart soil farming.
simple ways to automatically detect potential abnormality
of the IoT devices. First, AgriTalk automatically compares APPENDIX A. A COST ANALYSIS
the data of the nearby sensors of the same type. Second, This appendix elaborates on the costs of precision
AgriTalk monitors the power consumption of the sensors. If farming. In traditional soil cultivation, the cost for the basic
potential problems rise, AgriTalk alerts the farmer. Then we hydroelectric installation, and the deployment of the drip
use portable standard sensors to further investigate the irrigation system, the spray system, the light system (with
problems. Through AgriGUI, the calibration process can be 72 waterproof repellent bulbs), leveling land, and planting
conveniently established. turmeric in 2000 square meter area is about 44,000 USD.
Because acidification of soil, after turmeric harvest, the The cost for AgriTalk IoT hardware and software is about
land typically lies fallow for about 2 years. AgriTalk uses 16,000 USD. For the comparison purpose, we have also
the pH and the EC sensors to prevent acidification of farm installed a commercial solution, and the price is 42,000
field, and the land can continuously grow turmeric without USD higher than AgriTalk. In particular, the cost for an
land fallowing. AgriTalk weather station (including soil temperature,
We conducted measurement, analytic analysis, and moisture, pH and CE sensors with wired connection) with
simulation experiments to investigate the IoT message the pole that can survive typhoons in Taiwan is 1,300 USD,
delays of AgriTalk for automatic-manual control switching. while the price for the commercial wireless product
This issue has not been addressed in the smart farming (including soil temperature, moisture and CE sensors with
literature. Our study indicated that the delays are very short wireless connection) is 3,200 USD.
(less than 0.2 seconds) and AgriTalk can easily respond to By exercising AgriTalk on planting turmeric in a
quick and dynamic change of the field environment 2000-square meter area, the investment for the deployment
conditions in soil cultivation. and maintenance in the first year (2018) was about 60,000
USD, and the revenue generated by selling the turmeric was
AgriTalk was the winner of three major awards in both 140,000 USD. The investment for the next year (for
2016, 2017, and 2018 iGEM held at MIT [40]-[42], and was maintenance) is estimated to be 14,000 USD with the same
the gold medal winner of 2017 Teco Green Tech [43]. An revenue as the first year (i.e., 140,000 USD). The revenue
indoor version of AgriTalk won the silver award of 2018 we generated was based on a B-to-B model. Our customers
Acer Enterprise contest. Although we found that AgriTalk resold the turmeric products with the (double) revenue of
is more cost effective than a commercial solution, we 280,000USD.
continue to look for appropriate commercial IoT devices to
be accommodated in our solution. We have conducted a
cost analysis in Appendix A. REFERENCES
As the final remark, several important AgriTalk features
[1] M. I. Alipio, A. E. M. D. Cruz, J. D. A. Doria, and R. S. Fruto, “A
are briefly described in this paper but will be addressed in smart hydroponics farming system using exact inference in
depth in a future paper. First of all, our approach provides bayesian network,” in 2017 IEEE 6th Global Conference on
toxic free agricultural products. To do so, we have Consumer Electronics, Oct 2017.
[2] allMETEO, 2018. [Online]. Available: https://www.allmeteo.com/
developed novel biopesticides that are low drug-resistant [3] Smart Elements, 2018. [Online]. Available:
and eco-friendly to the environment. These biopesticides https://smartelements.io/
are highly specific to pests and diseases but is not harmful [4] Pycno, 2018. [Online]. Available: https://www.pycno.co/
[5] Farmapp, 2018. [Online]. Available: https://farmappweb.com/
to mammals and pollinators. To apply these biopesticides [6] Growlink, 2018. [Online]. Available: https://growlink.com/
precisely before diseases occurring and pest damage, we [7] FarmLogs, 2108. [Online]. Available: https://farmlogs.com/
need to monitor the period of spore germination and egg [8] Cropio, 2108. [Online]. Available: https://about.cropio.com/#agro

2327-4662 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2019.2899128, IEEE
Internet of Things Journal

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) 14
<

[9] F. Paul, "Big trouble down on the IoT farm", Network World, 2018. Cutworm, Spodoptera Litus Fabricius", Korean Journal of Applied
[Online]. Available: Entomology, vol. 36, no. 1, pp. 48-54, 1997
https://www.networkworld.com/article/3262631/ [30] B. B. Fand, N. T. Sul, S. K. Bal, and P. S. Minhas, "Temperature
internet-of-things/big-trouble-down-on-the-iot-farm.html Impacts the Development and Survival of Common Cutworm
[10] M. Aleksandrova, "IoT in Agriculture: 5 Technology Use Cases for (Spodoptera litura): Simulation and Visualization of Potential
Smart Farming (and 4 Challenges to Consider)", Estern Peak, 7 Population Growth in India under Warmer Temperatures through
June, 2018. [Online]. Available: Life Cycle Modelling and Spatial Mapping", PLoS One, vol. 10, no.
https://easternpeak.com/blog/iot-in- 4, 2015
agriculture-5-technology-use-cases-for-smart-farming-and-4-chall [31] S. Farahani, A. A. Talebi, and Y. Fathipour, "Life Cycle And
enges-to-consider/ Fecundity of Spodoptera Exigua (Lep.: Noctuidae) on Five
[11] M. D. Sardare and S. V. Admane, "A Review on Plant without Soil Soybean Varieties", Journal of Entomological Society of Iran, vol.
– Hydroponics", International Journal of Research in Engineering 30, no. 2, pp.1-12, 2011
and Technology, ISSN: 2319-1163, vol. 2, no. 3, 2013 [32] Y.-W. Lin, Y.-B. Lin, C.-Y Liu, "AItalk: A Tutorial to Implement
[12] O. L. Oke, "Nitrite Toxicity to Plants", Nature, vol. 212, pp. 528, AI as IoT Devices", Accepted and to appear in IET Networks.
1966 [33] Modelling of bio-pesticides degradation rate, 2016. [Online].
[13] S. Temteme, E. Getachew, B. Mekonnen, and T. Shimber, Available:
"Influence of Type of Planting Material, Plant Density and Method http://2016.igem.org/Team:NCTU_Formosa/Model#title3
of Planting on the Yield and Quality of Turmeric (Curcuma [34] H. Jeyanthi and S. Kombairaj, "Pesticide Use in Vegetable Crops:
domestica L.) at Tepi Southwestern Ethiopia", Advances in Crop Frequency, Intensity and Determinant Factors", Agricultural
Science and Technology, vol. 5, no. 5, pp. 305, 2017 Economics Research Review, vol. 18. Jul. 2005
[14] G. Gupta, S. S. Parihar, N. K. Ahirwar, S. K. Snehi and V. Singh, [35] T. S. Bachari, "Theoretical Investigation on The Kinetics of Free
"Plant Growth Promoting Rhizobacteria (PGPR): Current and Radical Reactions of Styrene Emulsion Polymerization", Journal
Future Prospects for Development of Sustainable Agriculture", of Al-Qadisiyah for Pure Science, vol. 16, no. 4, pp.46-60, 2011
Microb Biochem Technol, vol. 7, pp.096-102, 2015 [36] M. E. Dzaugis, A. J. Spivack, and S. D'Hondt, "A Quantitative
[15] P. Bonfante and A. Genre, "Mechanisms Underlying Beneficial Model of Water Radiolysis and Chemical Production Rates Near
Plant – Fungus Interactions in Mycorrhizal Symbiosis", Nature Radionuclide-Containing Solids", Radiation Physics and
Communications, vol. 1, no. 48, 2010 Chemistry, vol. 115, pp. 127-134, 2015
[16] H. Kheyrodin, "Plant and Soil Relationship between Fungi", [37] Y.-B. Lin, C.-M. Huang, L.-K. Chen, G.-N. Sung and C.-C. Yang,
International Journal of Research Studies in Biosciences, vol. 2, no. "MorSocket: An Expandable IoT-based Smart Socket System",
9, pp. 42-49, Oct. 2014 IEEE Access, vol. 6, no. 1, pp. 53123-53132, Dec. 2018.
[17] T. Brinkmann, C. H Specht, and F. H Frimmel, "Non-Linear [38] C.-W. Chang, Y.-B. Lin, and J.-C. Chen, "Reporting Mechanisms
Calibration Functions in Ion Chromatography with Suppressed for Internet of Things", EAI WiCOM, Taipei, October 15, 2018 (the
Conductivity Detection Using Hydroxide Eluents", Journal of best paper award).
Chromatography A., vol. 957, no. 2, pp. 99-109, Jun. 2002 [39] E. D. Lazowska, J. Zahorjan, G. S. Graham, and K. C. Sevcik,
[18] T G Nageshwar Rao, "Diseases of Turmeric (Curcuma Longa L.) "Quantitative System Performance: Computer System Analysis
and Their Management", Journal of Spices & Aromatic Crops, vol. Using Queueing Network Models", Prentice-Hall, Inc.,
4, no. 1, pp. 49-56, 1995 Englewood Cliffs, New Jersey, 1984
[19] Y. K. Kotikal and K. A. Kulkarni, "Insect Pests Infesting Turmeric [40] 2016 major awards of iGEM, 2016. [Online]. Available:
in Northern Karnataka", Journal of Farm Sciences, vol. 13, no. 4, http://2016.igem.org/Team:NCTU_Formosa/Demonstrate
pp. 858-866, 2000 [41] 2017 major awards of iGEM, 2017. [Online]. Available:
[20] Y.-W. Lin, Y.-B. Lin, M.-T. Yang, and J.-H. Lin, "ArduTalk: An http://2017.igem.org/Team:NCTU_Formosa/Demonstrate
Arduino Network Application Development Platform based on [42] 2018 major awards of iGEM, 2018. [Online]. Available:
IoTtalk", IEEE System Journal, pp: 1-9, 28 Nov. 2017. http://2018.igem.org/Team:NCTU_Formosa/Demonstrate
[21] ARBLE, 2018. [Online]. Available: https://arable.com/ [43] Gold medal winner, Teco Green Tech, 2017. [Online]. Available:
[22] Semios, 2018. [Online]. Available: https://semios.com/ http://www.tecofound.org.tw/greentech-contest/2017/download/2
[23] M. J. Windley, V. Herzig, S. A. Dziemborowicz, M. C. Hardy, G. F. [44] A.E. Ghaly and V. V. Ramakrishnan, “Nitrification of Urea and
King, and G. M. Nicholson, "Spider-Venom Peptide as Assimilation of Nitrate in Saturated Soils Under Aerobic
Bioinsecticide", Toxins, vol. 4, no. 3, pp. 191-227, 2012 Conditions,” American Journal of Agricultural and Biological
[24] E. Fitches, M. G. Edwards, C. Mee, E. Grishin, A. M. R. Gatehouse, Sciences 8 (4): 330-342, 2013.
J. P. Edwards, J. A. Gatehouse, " Fusion Proteins Containing [45] A. Akbar, et al., "Development of Prediction Model and
Insect-Specific Toxins As Pest Control Agents: Snowdrop Lectin Experimental Validation in Predicting the Curcumin
Delivers Fused Insecticidal Spider Venom Toxin To Insect
Content of Turmeric (Curcuma longa L.)," Front Plant Sci.7:
Haemolymph Following Oral Ingestion", Journal of Insect
Physiology, vol. 50, pp. 61-71, 2000 1507, 2016.
[25] Y.-B. Lin, L.-K. Chen, M.-Z. Shieh, Y.-W. Lin, and T.-H. Yen,
"CampusTalk: IoT Devices and Their Interesting Features on
Campus Applications", IEEE Access, vol. 6, no. 1, pp.
26036-26046, Dec. 2018.
[26] Y.-B. Lin, Y.-W. Lin, C.-M. Huang, C.-Y. Chih, and P. Lin,
"IoTtalk: A Management Platform for Reconfigurable Sensor
Devices", IEEE Internet of Things Journal, vol. 4, no. 5, pp. 1152–
1562, 2017
[27] Y.-B. Lin, Y.-W. Lin, C.-Y. Chih, T.-Y. Li, C.-C. Tai, Y.-C. Wang,
F.-C. Lin, H.-C. Kuo, C.-C. Huang, and S.-C. Hsu, "EasyConnect:
A Management System for IoT Devices and Its Applications for
Interactive Design and Art", IEEE Internet of Things Journal, vol.
2, no. 6, pp. 551–561, 2015
[28] Y.-W Lin., Y.-B. Lin, C.-Y. Hsiao, and Y.-Y. Wang, "IoTtalk-RC:
Sensors as Universal Remote Control for Aftermarket Home
Appliances", IEEE Internet of Things Journal, vol. 4, no. 4, pp.
1104–1112, 2017
[29] S. D. Bae, K. B. Park, and Y. J. Oh, "Effect of Temperature and
Food Source on the Egg and Larval Development of Tobacco

2327-4662 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2019.2899128, IEEE
Internet of Things Journal

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) 15
<

Wen-Liang Chen received the Ph.D. degree in Ng Fung Ling received the B.S. degree in Biological
biotechnology at National Chiao Tung University, Science and Technology from National Chiao Tung
Hsinchu, Taiwan, in 2006. He joined National Chiao University, Taiwan, in 2017. She is pursuing the Ph.D.
degree in Biological Science and Technology since
Tung University in 2008 where he is an Associate
2018. Her current research interests are bio-pesticides
Professor in the Department of Biological Science and application and precision farming management.
Technology. His research interests include protein
chemistry, protein engineering, synthetic biology,
cancer biology, antibody engineering and development of biomaterial
based drug repositioning platforms for cancer treatment.
Yuan-Yao Chan is pursuing the B.S. degree in
Biological Science and Technology from National
Yi-Bing Lin (M’96-SM’96-F’03) received his
Chiao Tung University, Taiwan since 2017. His
Bachelor’s degree from National Cheng Kung current research interest is precision fertilization
University, Taiwan, in 1983, and his Ph.D. from application.
University of Washington, USA, in 1990. From 1990
to 1995 he was a Research Scientist with Bellcore. He
then joined National Chiao Tung University (NCTU)
in Taiwan, where he remains. In 2010, Lin became a
lifetime Chair Professor of NCTU, and in 2011, the Vice President of
NCTU. During 2014 - 2016, Lin was Deputy Minister, Ministry of Science You-Cheng Liu received the B.S., M.S. and Ph.D.
and Technology, Taiwan. Since 2016, Lin has been appointed as Vice degree in soil and environmental science from
Chancellor, University System of Taiwan (for NCTU, NTHU, NCU, and National Chung Hsing University, Taichung, Taiwan,
in June 2006, 2010 and 2017, respectively. He has
NYM). Lin is an Adjunct Research Fellow, Institute of Information
been an Assistant Research Fellow at National Chiao
Science, Academia Sinica, Research Center for Information Technology Tung University since 2018. His current research
Innovation, Academia Sinica, and a member of board of directors, interests include soil microbiome, microbial
Chunghwa Telecom. He serves on the editorial boards of IEEE Trans. on biochemistry, microbial electrochemistry, and
Vehicular Technology. He is General or Program Chair for prestigious anaerobic biology.
conferences including ACM MobiCom 2002. He is Guest Editor for
several journals including IEEE Transactions on Computers. Lin is the
co-author of the books Wireless and Mobile Network Architecture (Wiley, Chin-Cheng Wang received the B.S. and M.S. degree in
2001), Wireless and Mobile All-IP Networks (John Wiley, 2005), and Public Administration from National Chen Chi
University, Taipei, Taiwan, in 1997 and 2001,
Charging for Mobile All-IP Telecommunications (Wiley, 2008). Lin
respectively. He has worked in Institute for Information
received numerous research awards including 2005 NSC Distinguished Industry since 2002 and was a Researcher at Office of
Researcher, 2006 Academic Award of Ministry of Education, 2008 Award Food Safety, Executive Yuan, Taiwan from 2014 to 2017.
for Outstanding contributions in Science and Technology, Executive Yuen, His current research interests include agriculture project
2011 National Chair Award, and TWAS Prize in Engineering Sciences, management, information system integration and apply analysis of
2011 (the Academy of Sciences for the Developing World). He is Chair of bio-pesticides.
IEEE Taipei Section. Lin is AAAS Fellow, ACM Fellow, IEEE Fellow,
and IET Fellow.
Cheng-Hsun Chiu received the B.S. and M.S. degree
in BioBusiness Management from National Chiayi
Yun-Wei Lin received the B.S. degree in computer University, Taiwan, in 2013 and 2015. His current
and information science from Aletheia University, position is the research assistant of Department of
Taipei, Taiwan, in 2003, and the M.S. and Ph.D. Biological Science and Technology in National Chiao
degrees in computer science and information Tung University, Taiwan. His current research
engi-neering from National Chung Cheng University, interests include agriculture project management,
Chiayi, Taiwan, in 2005 and 2011, respectively. He is information system integration.
currently an Assistant Research Fellow with National
Chiao Tung University, Hsinchu, Taiwan. His current
research interests include mobile ad hoc networks, wireless sensor
Tai-Hsiang Yen received the bachelor's degree in
networks, vehicular ad hoc networks, and IoT/M2M communications.
application mathematics from National Chiao Tung
University, Taiwan, in 2011, where he is currently
Jyun-Kai LIAO received the B.S. degree in pursuing the M.S. degree with the Department of
Department of Electronics Engineering from National Computer Science. His current research interests
Chiao Tung University, Taiwan, in 2018. He is include Internet of Things, machine learning, and
pursuing the Ph.D. degree in Institute of Computer artificial intelligence.
Science and Engineering since 2018. His current
research interests include internet of things and
embedded system.

2327-4662 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.