Design and Implementation of a Wireless

Sensor Network to Detect Forest Fires

Jorge Granda Cantuña1 , Dennis Bastidas1 , Santiago Solórzano1 Jean-Michel Clairand1,2
Facultad de Ingenierı́a y Ciencias Agropecuarias
1
Institute for Energy Engineering
2

Universidad de las Américas Universitat Politècnica de València

170122, Quito, Ecuador 46022, Valencia, Spain
E-mail: [email protected], [email protected], E-mail: [email protected]
[email protected],

Abstract—In this paper, an innovative design and development direction and wind speed. August, 2015 represents a critical
process is presented for a Wireless Sensor Network (WSN) case with 1361 forest fires, and September 2015 represents
prototype aimed at detecting and monitoring forest fires at the the worst case in terms of the extension of acres of forest
Guanguiltagua Park located in the Metropolitan District of Quito
(DMQ), capital of Ecuador. Every year the city of Quito is destroyed by fire [3].
affected by recurrent forest fires, especially in summer time Wireless Sensor Network (WSN) is a state of the art
where high temperatures and soil dryness contribute to the technology that is currently being applied in a wide variety of
propagation of fire in areas with vast vegetation. As a result to this fields including, but not limited to environmental [4], natural
problem, a WSN technological solution was created to address disasters [5], health [6], [7], public safety [8], e-government [9]
this problem which can detect forest fires in real-time sending
alerts immediately to the end user. This prompt response helps and national defense [10]. Environmental applications include
to attenuate the impact of forest fires such as loss of human lives, fire detection [11], flood detection [12], agriculture [13], and
loss of large areas of vegetation, economic loss and environmental many others. A sensor network has a specific set of sensor
contamination. The WSN prototype described in this document nodes and a gateway, which allows all internal sensor data
uses hardware, software, communication protocols, topology and and GPS information to reach an external communications
functionality focused on optimizing its end results. The system is
based on measuring continuously three types of gases, which are network such as a cellular communications network or public
present in a combustion process such as CO2, CO and CH4. In safety infrastructure if required.
addition to these sensors, environment temperature and humidity, Some technical solutions based on WSN have been pro-
as well as GPS location data is provided; this data is transmitted posed to prevent wild fires. In [14], a WSN was deployed
via wireless communications toward a graphical interface where for Rural and Forest Fire Detection and Verification, based
the sensors data will be interpreted and statistical information
can be generated. Finally, through a WSN prototype validation on a Linksys WRT54GL router that was able to sense fire
process, it is demonstrated that such prototype provides an by infrared radiation and smoke. In [15], authors presented a
efficient and reliable method to detect and monitor forest fires framework to detect a fire threat as soon as possible with a
in a short amount of time. simulator implementation and with low energy consumption.
Keywords—Wireless Sensor Network, Forest Fire Detection, In [16], the methodology used is to sense gas temperature and
Real-Time, Communication Protocol, Database, eGovernment.
air moisture content, to detect fire and to track the fire spread
I. I NTRODUCTION during its spatial and temporal evolution. In [17], the solution
was proposed with a BTBee module and with artificial neural
Every year wild fires across the world destroy millions networks approach for data classification and collection.
of acres of land, provoke loss of human lives, environment eGovernment services represent a critical driving force in
contamination and have a significant economic impact on the successful development and wellbeing of society; this
government’s budget [1]. can be accomplished by integrating advanced information and
Only in the year 2015 exactly on 16 September, the DMQ communications technologies (ICT) with the services provided
was declared in a state of emergency since it dealt with a by government agencies [18].
wave of fires that resulted in the death of three firefighters, 22 The WSN design and implementation results presented in
injured people and 960 acres of forest in ashes [2]. this document integrates software and hardware components
Statistical data on forest fire frequency and acres burnt used for forest fire detection and monitoring activities in
were obtained from the Ecuadorian National Meteorology the Guanguiltagua Metropolitan Park, located in the DMQ,
and Hydrology Institute for 2014, 2015, and 2016. The data Ecuador. Wildfires can propagate at a speed up to 24 kilome-
demonstrates that during summer time the percentage of forest ters per hour as specified in [19].The WSN design consists
fires increases, but it also depends on a combination of differ- of an advanced network of Wireless sensors using advanced
ent parameters such as: temperature, humidity precipitation, algorithms to detect and generate an alert when a forest fire is
978-1-5090-4830-4/17/$31.00 c 2017 IEEE started with its precise location. With this innovative design, a

15
 Fig. 1. Hadware Network Diagram

Fig. 2. Principal Node Diagram

forest fire can be detected in its early stages by wireless sensors
which generate an alert, and transmit real time data to an
external database, where it can be accessed by first responders, should measure critical data such as air hygrometry as
as well as local and government authorities. The WSN system proposed in [16].
presented in this paper contributes to the expansion of eGov- ‚ CO2 Sensor(carbon dioxide) MG-811
ernment services by collecting data from different sources such The sensor is composed by a MG-811 module that is
as temperature, humidity, CO2 (carbon dioxyde), CO (carbon very sensitive to CO2 gas. This gas is present in the air;
monoxide), and CH4 (method) sensor nodes, as well as storing its typical value is between 350 to 450 ppm (parts per
this data in databases provided by cloud computing services. million). It is also present when there is combustion, for
This WSN solution has a positive economic, technological and example when there is a fire.
social impact, due to its scalable infrastructure and adaptability ‚ CO Sensor (carbon monoxide) MQ7
to different application environments. WSN is one efficient and This sensor measures the carbon monoxide quantity
more common technique used for acquiring data [20]. present in the air. It’s able to measure from 20 to 200 ppm
The paper is organized as follows: in section 2, the design of of CO. It is also very sensitive and has a fast response.
the hardware network is presented. In section 3, user interface ‚ CH4 Sensor (methane) MQ4
is showed. Section 4 is devoted to the implementation. In This sensor measure methane (CH4) gas concentration.
section 5 results are presented and discussed. Finally, in Its detection rank is from 300 to 10000 ppm. It is also
section 6 the work is concluded with some remarks. very sensible and has a fast response.
‚ GPS NEO-6M
II. D ESIGN OF WSN H ARDWARE This GPS is ready-to-use, its size is reduced, for this
reason it is easy to use in wireless projects. It comes
The WSN Prototype described in this paper is designed
with an antenna and it does not need more accessories
with three sensor nodes and a gateway as shown in Fig
to work. It provides the information such as: latitude,
1. A sensor node is comprised by the following hardware
longitude, date and time.
components: Sensing unit, Processing Unit, Communication
‚ Solar Panel
Unit, and Power unit [17].
It generates the electricity required for WSN sensor node
A. Principal Node operation where there is no electricity available. During
the day, it will deliver energy to the prototype and it will
The Principal Sensor Node consists of the following ele- also charge the battery.
ments, as shown in Fig.2: ‚ Lipo Battery
‚ Temperature and humidity Sensor DHTT22 This battery was selected because of its lightness, small
The DHT22 sensor is able to measure ambient tem- size, capacity and durability features. The battery will
perature through a thermistor, and humidity through a be charged by the solar panel during the day. The energy
capacitive humidity sensor [21]. It has also an analog- needed for the prototype to operate is 807.5 mAh. In order
digital converter to provide the output signal in digital tu support 12 hours of the evening, when the photovoltaic
form. In the WSN prototype, this sensor obtains ambient panel doesn’t deliver power, a battery pack with 10 000
temperature and humidity in real time. A thermal sensor mAh capacity was selected.

16
 Fig. 3. Secondary Node Diagram

‚ ATmega328P Arduino
This microcontroller is based on the ATmega28, which
has 14 digital pins, 6 analog pins and a restart button.
Its reduced size makes it suitable for this project. This
microcontroller was programmed in C# with the Arduino
software.
‚ Xbee S2B Pro
This module establishes wireless communication between
devices. Its features make it easy to use and can create
point to point networks. The model S2B Pro is well
known for its low power consumption and a line of sight
of approximately 1.5 kilometers. It uses a ZigBee mesh
communication protocol and operated in the 2.4 GHz
frequency range.

B. Secondary Node
For the secondary node, it consists of the following ele- Fig. 4. Algorithm Diagram
ments, as shown in Fig.3:
‚ Temperature and humidity Sensor DHTT11
‚ Nano Atmega328P Arduino adapted and applied to this case in order to define techni-
‚ Xbee S2B Pro cal operational requirements for the WSN prototype, design
‚ Lipo Battery alternatives were also considered, design review meetings
were continuously coordinated to analyze the conceptual WSN
C. Methodology design and development, validation test planning, test data
analysis procedures, end user feedback and finally, a report
The objective is to detect fires through measurements of
summarizing results, recommendations and conclusions. Fig.
humidity, temperature and smoke levels which are continually
4 represents the algorithm of the fire detection. This algorithm
detected by the sensor node in order to monitor the forest,
was programmed in Arduino.
reducing detection time and determining the precise location
information of a forest fire scenario using WSN technology; all
D. WSN Network Configuration
this data is transmitted in real-time, applying a network star
topology. Some other real-time communication applications The WSN Network was configured with XCTU software,
such as SPEED, MM-SPEED, and RPAR have been developed which is an open source application and is compatible with all
[22], but are not suitable for routing information in critical operating systems including Microsoft Windows and Linux. To
situations such as forest fires. WSN scalability projection is configure the sensor nodes in XCTU, the firmware is updated
also considered in the design, since the number of sensor nodes from the beginning by selecting XBP24BZ7 as the product
can increase to a large scale [23] due to the flexibility of the family and firmware version 2OAO. Then some specific fields
hardware components, topology, communication protocols and that are configured are ID PAN which can have values between
mathematical algorithms to process the sensor data. Systems 0 and FFFE, SC Scan Channels which can have values
engineering design and implementation methodologies were between 0 and FFFF and is set to D in this case, Ni Node

17
 Fig. 5. Principal Window (Without University logo)

Identifier where each sensor node ID is set (Coordinator), and

finally Baud rate which is set to 9600.
To configure the gateway node, the product family to XB24-
ZB and firmware version 28A0. The ID PAN ID is set to 1723;
SC Channels to D, DH Destination Address High XBee serial
number is set to 13A200, DL Destination Address Low MAC
address as 409E6EC2, Ni node identifier Nodo3, and baud rate Fig. 6. System Management Window
9600.
III. U SER I NTERFACE
The user interface was designed in C# in Visual Studio
2012, with a database in SQL. The principal window has three
options: System Management, Database, Transparent mode.
A. System Management (Gestion del Sistema)
In this tab, as shown in Fig. 6, all the sensor nodes
information, sensors and critic levels data is populated.
In this window it will be possible to access the three
following options, as shown in Fig. 5: Fig. 7. Database Window
‚ Insert Node (Insertar Nodo): enter the name, Mac ID,
Latitude, length of the node. It is possible to enter n
nodes. For this case study, only three nodes were used
and the information is presented in the right side of the
DaraGridView.
‚ Insert Sensor (Insertar Sensor): the name and kind of
sensor is entered, and an option to assign a node to each
sensor. The information will be presented in the right side
on the DataGridView.
‚ Limits definition (Definir Lı́mites): selects the kind of
sensor and enters a value limit in order to activate an
alarm if it goes over the ceiling value. Information is
presented in the right side in a DataGridView.
Fig. 8. Transparent Mode Window
B. Database (Base de datos)
It will show, as in Fig.7, stored node information that is
received wirelessly from the sensors nodes to the gateway and When gas level goes over established limits, the fire alarm
to a PC terminal. It is also possible to filter by kind of sensor will activate beeping, and a led indicator will turn from green
and by critical levels, when measurements go over established (normal level) to red (critic status). Note that in order to
values. stop the alarm a Stop Alarm (Detener Alarma) Button was
designed.
C. Transparent mode (Modo transparente) There is also a map that can detect the location of the
This window, as shown in Fig. 8, is the most important different sensor nodes. In case of an alarm case, it is possible
because it is possible to view real time information and to to detect immediately the location and avoid fire spread.
monitor fire detection system continually.

18
 IV. I MPLEMENTATION
A. Sensor calibration
The sensors used have one analog and one digital output.
Digital output sends a 0 if there isn’t gas present and 1 if there
is. It is easy to work with this type of outputs, some sensors
require calibration, especially if data for CO gas quantity levels
are required in particles per million (ppm).
Analog output sends values between 0 and 1023, depending
on the concentration of existing gas; each MQ sensor has its
own rank of detection and sensibility.
In order to explain sensor calibration, the example of the
MQ7 sensor, which measure carbon monoxide, is presented.
To calculate voltage output, the next equation is used, where
RS and RL form a voltage divider.
RL Fig. 9. Application demonstration
V 0 “ VC . (1)
RS ` R L
where: RS is the sensor surface resistance on various V. R ESULTS
concentrations of gases A. Data Validation
RL is the load resistance
In order to validate prototype data, it was compared to
V0 is the output voltage
IBRID Mx6 data. This device is a high technology gas
VC is the input voltage in 5V
detection that uses a pump to analyze the air quality. For this
Replacing, the next equation is used.
first experimental test, sheets of paper were burned and the
pVC ´ V0 q.RL sensor nodes were exposed to gases derived from the smoke.
RS “ (2) Results are presented in Table 1.
V0
By other side, we have: Gas MQ Sensors IBRID Mx6
CO2 350 ppm 350 ppm
RS CO 200 ppm 180 ppm
R0 “ (3) CH4 10 ppm Don’t have
RCAF TABLE I
C OMPARISON BETWEEN MQ S ENSORS AND IBRID M X 6
Where:
R0 : sensor resistance at 100ppm CO in clean air
RCAF : resistance for clean air factor The MX6’s advantage is that it saves a little proportion of
Finally we have: air sample and examines it, by other side MQ sensors, and
measure the quantity that arrives in real time. Nevertheless,
pVC ´ V0 q.RL
R0 “ (4) results show that the measurements are quite close, so that the
V0 .RCAF
MQ sensors are suitable for this application.
Note that the sensor datasheet explain that if the concentra-
tion is below 100 ppm, the relation R B. Operation test
R0 =1 in 200 C and 65%
S

relative humidity. If the relation is slow, the values correspond The WSN prototype has been tested and demonstrated its
of bigger concentrations of CO in ppm. functionality in a real scenario, meaning that the equipment
With the calibration of R0 , it is possible to measure the was installed and tested in the Guanguiltagua Metropolitan
air values, calculating RS value in relation with the output Park located in Quito, Ecuador to monitor and detect a forest
voltage and comparing the relation R R0 from the datasheet in
S fire scenario. In order to validate the operation, a test was
order to obtain the CO concentration in ppm. planned and executed; the device was operating and a fire
From the datasheet sensor, the next equation is obtained: was simulated burning Eucalyptus leaves, which represents the
typical vegetation of Quito. The Fig. 9 shows the green light
RS turn to red indicating that limits went over the recommended
“ 22, 07.rCOs´ 3 q
2
(5)
R0 values. The alarm was also activated. The GPS indicates
In order to obtain CO concentration, we have: quickly the location of the fire in order to intervene as soon
as possible.
RS In order to determine the wireless communication propaga-
q´ 2
3
rCOs “ p (6)
22, 07.R0 tion loss it is recommended to use the appropriate propagation
model based on the forest features including tree density,
terrain details, topography maps, and existing obstacles which

19
can cause radio wave diffraction. Some propagation models R EFERENCES
that can be considered are Weissberger Model , ITU (ITU-
R) Model, and Cost 235(ITU-R) Model. In this case 80 [1] Y. Jin, M. L. Goulden, N. Faivre, S. Veraverbeke, F. Sun,
A. Hall, M. S. Hand, S. Hook, and J. T. Randerson, “Identification
per cent of the Guanguiltagua Park is made of eucalyptus of two distinct fire regimes in Southern California: implications
trees, with an extension of approximately 560 acres; based on for economic impact and future change,” Environmental Research
this information and type of vegetation, the most appropriate Letters, vol. 10, no. 9, p. 094005, 2015. [Online]. Available:
http://iopscience.iop.org/article/10.1088/1748-9326/10/9/094005
propagation model to calculate losses in the Guanguiltagua [2] A. AFP. (2015) Declaran emergencia en Quito por ola de incendios
Metropolitan Park is the Weissberger Model as follows: forestales. [Online]. Available: http://www.elcomercio.com/actualidad/
declaran-emergencia-quito-incendios-forestales.html
[3] C. de Bomberos. Informe de Gestion 2015 dry, cold, and
windy. [Online]. Available: https://www.bomberosquito.gob.ec/images/
LW pdBq “ 1, 33.f 0,284 .d0,588 if 14m † d § 400m (7) pdf/cuentas/fase1rrc15.pdf
[4] P. Corke, T. Wark, R. Jurdak, W. Hu, P. Valencia, and D. Moore,
“Environmental wireless sensor networks,” Proceedings of the IEEE,
vol. 98, no. 11, pp. 1903–1917, 2010.
LW pdBq “ 0, 45.f 0,284 .d0,588 if 0m † d § 14m (8) [5] M. Qiu, Z. Ming, J. Wang, L. T. Yang, and Y. Xiang, “Enabling Cloud
Computing in Emergency Management Systems,” Cloud Computing,
IEEE, vol. 1, no. 4, pp. 60–67, 2014.
Where: [6] A. Pantelopoulos and N. G. Bourbakis, “A survey on wearable sensor-
LW : Defines the power loss in decibels due to vegetation. based systems for health monitoring and prognosis,” IEEE Transactions
on Systems, Man and Cybernetics Part C: Applications and Reviews,
f : Frequency in GHz. vol. 40, no. 1, pp. 1–12, 2010.
d : Distance from transmitter to receiver. [7] H. Alemdar and C. Ersoy, “Wireless sensor networks for healthcare: A
Tree leaf density in the Guanguiltagua Metropolitan Park survey,” Computer Networks, vol. 54, no. 15, pp. 2688–2710, 2010.
[Online]. Available: http://dx.doi.org/10.1016/j.comnet.2010.05.003
also has an impact in the propagation loss calculations with [8] T. D. Räty, “Survey on contemporary remote surveillance systems for
approximately 3 to 5 dB loss. Therefore in order to im- public safety,” IEEE Transactions on Systems, Man and Cybernetics
plement a reliable WSN solution to detect forest fires, it is Part C: Applications and Reviews, vol. 40, no. 5, pp. 493–515, 2010.
recommended to calibrate the sensors properly based on the [9] V. H. González-Jaramillo, “Tutorial: Internet of Things and the upcom-
ing wireless sensor networks related with the use of big data in mapping
manufacturer datasheets to obtain precise ppm values [24]. services; issues of smart cities,” 2016 Third International Conference on
If the WSN network requires more nodes, a mesh topology eDemocracy & eGovernment (ICEDEG), pp. 5–6, 2016.
is recommended with the XBEE devices configured in API [10] I. Bayir, “A glimpse to future commercial spy satellite systems,” RAST
2009 - Proceedings of 4th International Conference on Recent Advances
mode which guarantees a better coverage area. Space Technologies, pp. 370–375, 2009.
[11] P. Ji, C. Wu, Y. Zhang, and H. Chu, “A Wireless Sensor Network Routing
VI. C ONCLUSION Protocol for Fire Estimation in Building,” Energy, 2009.
[12] W. Y. Zhuang, M. Costa, P. Cheong, and K. W. Tam, “Flood monitoring
At the present moment, Quito DMQ doesn’t have an in- of distribution substation in low-lying areas using wireless sensor
network,” Proceedings 2011 International Conference on System Science
telligent and automated system that can detect and monitor and Engineering, ICSSE 2011, no. June, pp. 601–604, 2011.
forest fires in real-time. In this paper, a Wireless Sensor Net- [13] Aqeel-Ur-Rehman, A. Z. Abbasi, N. Islam, and Z. A. Shaikh, “A
work prototype for forest fire detection in the Guanguiltagua review of wireless sensors and networks’ applications in agriculture,”
Computer Standards and Interfaces, vol. 36, no. 2, pp. 263–270, 2014.
Metropolitan Park in Quito was designed and implemented. [Online]. Available: http://dx.doi.org/10.1016/j.csi.2011.03.004
The results demonstrate that the prototype sensors detect CO2 , [14] J. Lloret, M. Garcia, D. Bri, and S. Sendra, “A wireless sensor network
Co and CH4 gases, when they are present in the environment. deployment for rural and forest fire detection and verification,” Sensors,
vol. 9, no. 11, pp. 8722–8747, 2009.
A forest fire has an exponential growth and should be detected
[15] Y. E. Aslan, I. Korpeoglu, and ö. Ulusoy, “A framework for use
and controlled in about 6 minutes to prevent it from spreading. of wireless sensor networks in forest fire detection and monitoring,”
The user graphic interface is able to present data in real time Computers, Environment and Urban Systems, vol. 36, no. 6,
pp. 614–625, 2012. [Online]. Available: http://dx.doi.org/10.1016/j.
and save this in a database. Its design makes it easy to control compenvurbsys.2012.03.002
by the end user and requires proper training to understand [16] T. Antoine-Santoni, J. F. Santucci, E. de Gentili, X. Silvani, and
all of its functionality. This technology is efficient since WSN F. Morandini, “Performance of a protected wireless sensor network in
components are expected to become even smaller with the use a fire. Analysis of fire spread and data transmission,” Sensors, vol. 9,
no. 8, pp. 5878–5893, 2009.
of Micro Electro Mechanical Systems (MEMS) technology. [17] A. Chauhan, S. Semwal, and R. Chawhan, “Artificial
In the last years Quito has suffered severe damages includ- neural network-based forest fire detection system using
ing deaths of civilians and firefighters, thousands of acres wireless sensor network,” in 2013 Annual IEEE India
Conference (INDICON), 2013, pp. 1–6. [Online]. Available:
of forest burnt, environment pollution and economic loss. http://ieeexplore.ieee.org/ielx7/6712175/6725842/06725913.pdf?tp=
This WSN solution helps authorities to use their human and tz&uarnumber=6725913tz&uisnumber=6725842$zbackslash$nhttp:
economic resources more efficiently when facing emergency //ieeexplore.ieee.org/xpls/abstz uall.jsp?arnumber=6725913
[18] L. Zheng, H. Zhang, W. Han, and X. Zhou, “Technologies,
wildfire scenarios and can reduce economic impact, loss of applications, and governance in the Internet of Things,” In
human lives, and environment pollution caused by forest fires. proceedings of the Internet of things-Global technological and societal
It allows local authorities to make critical decisions in a short trends, pp. 141–175, 2011. [Online]. Available: http://books.google.
com/books?hl=entz&ulr=tz&uid=Eug-RvslW30Ctz&uoi=fndtz&upg=
amount of time, as wells as optimizing resources to detect and PA143tz&udq=Technologies,+Applications,+and+Governance+in+the+
keep forest fires under control. Internet+of+Thingstz&uots=3Sy6ECmAEttz&usig=NsG7DZSc2Uqtz

20
 uSmQhSXB10-89kQU$zbackslash$nhttp://books.google.com/books?
hl=entz&ulr=tz&uid=Eug-RvslW30Ctz&uoi=fndtz&u
[19] N. Geographic. Wildfires dry, cold, and windy.
[Online]. Available: http://environment.nationalgeographic.com/
environment/natural-disasters/wildfires/
[20] K. Trivedi and A. Kumar Srivastava, “An energy efficient framework for
detection and monitoring of forest fire using mobile agent in wireless
sensor networks,” in 2014 IEEE International Conference on Computa-
tional Intelligence and Computing Research, IEEE ICCIC 2014, 2014.
[21] D. R. Bastidas Cevallos, “Diseño e Implementación de un Prototipo
de Red de Sensores Inalámbricos para Detectar Incendios en el Parque
Metropolitano Guangüiltagua de Quito,” Universidad de las Américas,
2016.
[22] Y. Zeng, C. J. Sreenan, and L. Sitanayah, “A real-time and robust routing
protocol for building fire emergency applications using wireless sensor
networks,” 2010 8th IEEE International Conference on Pervasive Com-
puting and Communications Workshops, PERCOM Workshops 2010, pp.
358–363, 2010.
[23] F. Ye, A. Chen, S. Lu, and L. Zhang, “A scalable solution to minimum
cost forwarding in large sensor networks,” Proceedings - International
Conference on Computer Communications and Networks, ICCCN, vol.
2001-Janua, no. 3, pp. 304–309, 2001.
[24] A. Arellano Aucancela, D. Avila-Pesántez, E. P. Jennyfer, and
C. Hervas Parra, “Evaluación de una red de Sensores Inalámbrica
para detección de Incendios Forestales,” Revista Tecnológica ESPOL,
vol. 28, no. Noviembre, pp. 115–130, 2015. [Online]. Available:
https://www.researchgate.net/publication/297208802

21