See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/359118440

Forest fire prediction using IoT and deep learning

Article · March 2022

CITATIONS READS
20 3,905

5 authors, including:

Sengottaiyan Nallasamy Kamal Upreti

Alliance University Christ University, Bangalore
53 PUBLICATIONS 194 CITATIONS 143 PUBLICATIONS 806 CITATIONS

SEE PROFILE SEE PROFILE

Animesh Kumar Dubey

Madhyanchal Professional University, Bhopal India
7 PUBLICATIONS 64 CITATIONS

SEE PROFILE

All content following this page was uploaded by Kamal Upreti on 09 March 2022.

The user has requested enhancement of the downloaded file.

 International Journal of Advanced Technology and Engineering Exploration, Vol 9(87)
ISSN (Print): 2394-5443 ISSN (Online): 2394-7454
Case Study
http://dx.doi.org/10.19101/IJATEE.2021.87464

Forest fire prediction using IoT and deep learning

J Ananthi1, N Sengottaiyan2, S Anbukaruppusamy3, Kamal Upreti4 and Animesh Kumar Dubey5*
Assistant Professor, Department of Computer Science and Engineering, Hindusthan College of Engineering and
Technology, Tamilnadu, India1
Professor, Department of Computer Science and Engineering, Sri Shanmugha College of Engineering and
Technology, Sankari, Salem, India2
Professor and Head, Department of Electronics and Communication Engineering, Excel Engineering College,
Komarapalayam, India3
Associate Professor, Department of Computer Science & Engineering, Dr. Akhilesh Das Gupta Institute of
Technology & Management, Delhi, India4
Assistant Professor, Department of Computer Science, Madhyanchal Professional University, Bhopal, India 5

Received: 03-October-2021; Revised: 27-January-2022; Accepted: 29-January-2022

©2022 J Ananthi et al. This is an open access article distributed under the Creative Commons Attribution (CC BY) License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Abstract
Forests are the most important part of the human life as it maintains an environmental balance to get proper rain and
sufficient resources accordingly. The major threat raising in forest areas is a fire, in which the forest fire scenario is the
most important cause to destroy many trees and animals within a few hours. The technologies such as deep learning, IoT
and smart sensors provide a lead to design a smart forest fire prediction scheme to support nature to manage the
ecosystem in the proper way. This paper is intended to design a forest fire prediction mechanism. Learning-based forest
fire prediction scheme (LBFFPS) based on deep learning has been proposed for the prediction in the timely manner. This
approach identifies the forest fire based on the sensor unit associated with the system with respect to the learning logics.
A digital camera with 1020-megapixel has adapted for the surveillance. The sensor unit consists of two different and
powerful sensors such as smoke identification sensor and the temperature and humidity monitoring sensor. Based on
these two sensors the surrounding smoke presence, temperature and the humidity level have been identified and reported
using the NodeMCU controller. In this application, internet of things (IoT) is associated, to provide a wireless
communication alert ability. It collects and maintain the information regarding the forest provided by the sensor unit to
the remote cloud server environment. The NodeMCU microcontroller has an inbuilt WiFi to acquire the internet signals
and provides a constant bridge between the sensor unit and the server end for remote data maintenance. The proposed
logic is helpful to identify the fire signals and inform the respective person to take appropriate action to prevent the forest
fire.

Keywords
Deep learning, Forest fire prediction, Internet of Things (IoT), LBFFPS.

1.Introduction The identification and surveillance of such forest fire

Forests are valuable concerns for human existence as scenarios have become a worldwide problem for
well as societal progress because they help maintain communities dedicated to forest fire scenario
the universe's entire ecosystem stability [1–3]. management. As a consequence of environmental
Unfortunately, forest fire scenarios occur regularly as temperature, the chance of igniting a fire rises
a result of certain unregulated human activity and exponentially. Forest fire scenarios are rising in
erratic environmental circumstances [4, 5]. Such fires frequency and it will continue to do so. To assist
are by far the most destructive to environmental firefighters on the battlefield, a technique for early
assets as well as the human ecology [6–9]. In this identification of forest fire scenarios is provided in
situation, forest fire scenarios have significantly this paper. This technique seems to be more exact
increased in regularity as a result of global warming, than other means of communications, including such
mortal activity and certain other things [10, 11]. observation structures as well as surveillance
systems. This paper is based on the collection of
*Author for correspondence atmospheric wireless data communication from the
246
 International Journal of Advanced Technology and Engineering Exploration, Vol 9(87)

forest area and the prediction of forest fire scenario Satellite based forest fire surveillance technologies'
occurrence using deep learning strategy with respect usable spectrum is also constrained by a number of
to artificial intelligence logic. The evaluation and characteristics, reducing their usefulness in detecting
comparison of many models produced from deep forest fires. For instance, an intelligence gathering
learning enabled us to demonstrate the viability of an system works on a continuous screening period and
independent as well as legitimate ecological its overloaded pixel dotted images have a few
surveillance infrastructure for evolving risk features. The additional issue is that cloud coverings
associated with forest fire scenarios. might obscure photographs throughout the screening
phase and quantitative characterization of forest fire
In this paper a deep learning-based forest fire parameters in a timely manner is exceedingly hard to
prediction logic have been applied using internet of perform [3, 4]. Due to these challenges, IoT enabled
things (IoT) and smart sensor unit assisted device devices with wireless communication network-based
called a smart forest monitoring kit (SFMK) for the data transmission technology as a monitoring system
forest fire prediction. In this SFMK, different smart are needed [5]. It is capable in analysing real
sensors and devices are gathered to provide a associated factors, such as temperature and relative
sufficient service to acquire the data from forest humidity and transmitting the data directly to the
environment. These details are forwarded to the remote cloud server monitoring computer. It will be
cloud server for further processing. In the server end, helpful in the organization and analysis of the
a deep learning model called learning-based forest information gathered. In comparison to conventional
fire prediction scheme analysed the incoming data to data and fundamental sustainable forest information,
check the forest fire indications. If any fire the technology is capable of making an instant
indications are identified, immediately that will be evaluation of a possible fire threat. The gathered
alerted to the respective person to take an appropriate information will subsequently be forwarded to the
action. appropriate authority as the foundation for setting
regulatory regarding firefighting or rescue operations.
The main aim of this research is to provide a
sufficient methodology to avoid forest fire accidents, The objectives of this paper are as follows:
in which it causes a huge destruction to people 1.To develop a framework based on deep learning
around the forest as well as the living beings and model for the accurate forest fire prediction
animals found in the forest region. This research is scenario and report to avoid disasters in a timely
helpful to identify the forest fires at an earlier stage manner.
and report it accordingly to the respective authority to 2.To establish a proper communication and
alert them to take appropriate action to preserve the monitoring based on IoT device.
disaster further. Mainly the intention of this research 3.To integrate the SFMK for the intellectual
is to save the living creatures and trees present in the monitoring scheme with respect to sensors and
forest area from this kind of disaster. This kind of camera associated with it.
forest fires causes a huge loss in the natural
environment and spoil the trees presented into the In this paper an efficient learning-based forest fire
forest. So that the natural environment balancing is prediction scheme (LBFFPS) have been proposed to
mismatched and even the continuous cause of such identify the forest fire. The logical evaluation of such
things leads to an environmental collapse. schemes provides a good prediction logic and
abilities to monitor the forest-oriented information
Forest fire prediction systems presently rely heavily from anywhere in the globe by using wireless
on processions, inspection of guard posts and more communication medium.
recently, surveillance systems [1, 2]. While
monitoring from guard posts is simple as well as The rest of the paper is organized as follows.
feasible, it does have a number of drawbacks. To Literature review has been discussed in section 2.
begin, this approach necessitates significant technical Section 3 contains the method discussion. Results
and logistical resources, and also a highly skilled and discussion have been explored in section 4.
labour workforce. Furthermore, numerous issues with Concluding remarks have been pointed to the final
fire prevention workforce exist, including inattention, conclusion section.
leave from duty, a shortage of actual surveillance
capability and restricted coverage area [1–8].

247
J Ananthi et al.

2.Literature review AUC, specificity, negative predictive value (NPV),

In this section latest related work has been discussed positive predictive value (PPV), and, Kappa index.
with the proper elaboration of advantages and The LEDT model offered the best results, both in
limitations. terms of training and in terms of results. The LEDT
Lin et al. [12] used the fuzzy prediction algorithm in model outperformed the others on both the training
Jiangsu Province, China, which has a dense and validation datasets, with a prediction accuracy of
population and frequent forest fires. Jiangsu Province 92%.
is located in the southern region of China. To assess
fire danger and quantify the quantitative potential fire In 2019, Zhang et al. [15] proposed that utilization of
risk, a fuzzy inference and large data analysis convolutional neural network (CNN) to develop a
approach are developed. Fuzzification and output fire geographic prediction model for forest fire
rating level are used to transform these factors to vulnerability. A geographic information system was
triangular fuzzy numbers. The rechargeable wireless used to map past forest fire locations in Yunnan
sensor network collects continuous 24-hour weather Province, China, from 2002 to 2010, as well as a set
data that can indicate the high accuracy status of the of 14 forest fire affecting elements. The suggested
forest environment. This system can detect a high model's prediction performance was evaluated using
danger of probable forest fires, indicating that there is a variety of statistical measures, including the
a need to pay greater attention to forest fire receiver operating characteristic curve (ROC), area
prevention. under curve (AUC) and Wilcoxon signed-rank test.
The results showed that the proposed CNN model
In 2019, Zhu et al. [13] described and analysed the outperformed the kernel logistic regression, SVM,
machine learning-based forest fire prediction multilayer perceptron neural network and RF as
approaches that use wireless sensor networks (WSN) benchmark classifiers in terms of accuracy of AUC
technology and perceptron algorithms to give a of 0.86.
reliable and speedy detection of probable forest fires.
The sensors were used to collect weather data, which In 2021, Mohajane et al. [16] used the historical data
is subsequently sent to a server, where a fire hazard collected from
index can be generated. From June to September in http://www.eauxetforets.gov.ma/ProtectionForet/Ince
the year 2015, they erected a collection of sensors in ndies/Pages/Incendies.aspx and used it for preparing
Nanjing to record weather parameters such as wind the forest fire inventory map. The research area is in
speed, temperature, humidity and rainfall, in order to the Tanger-T'etouan-Al Hoceima region in northern
evaluate the efficacy of the perceptron algorithm for Morocco, and it includes Fahs-Anjra province and
forest fire prediction. Forest fire prediction may be Tanger-Asilah prefecture, with a total area of 1685
implemented in a more convenient method using the km2. For mapping forest fire susceptibility in
perceptron model and Lora technology. The proposed Morocco's north, they developed five new hybrid
system still has limits due to the minimal data machine learning algorithms: frequency ratio-random
acquired, and it will need to be evaluated for a longer forest (FR-RF), frequency ratio-classification and
period of time before being deployed in reality. regression tree (FR-CART), frequency ratio-logistic
regression (FRLR), frequency ratio-multilayer
In 2019, Tehrany et al. [14] provided a comparative perceptron (FR-MLP), and frequency ratio-support
case study in the Lao Cai region of Vietnam, which vector machine (FR-SVM). To evaluate the model's
evaluates the usage of the LogitBoost ensemble- effectiveness, the area under the ROC and AUC was
based decision tree (LEDT) machine learning calculated. In the forecasting of forest fires, the
algorithm for forest fire susceptibility mapping. In findings showed that the RF-FR model had the best
the comparison, support vector machine (SVM), performance with an AUC of 0.989, followed by
random forest (RF), and kernel logistic regression AUC of 0.959 in SVM-FR, AUC of 0.858 in MLP-
(KLR) was employed as benchmarks. Based on data FR, AUC of 0.847 in CART-FR and AUC of 0.809 in
from previous forest fire occurrences, a fire inventory LR-FR. Thirteen input variables in a well-studied
database for the selected region was built, and forest-fire dataset from Portugal's Montesinho
relevant conditioning factors were developed from a Natural Park have been tested [17].
variety of sources. Following that, forest fire
probability indices were computed using each of the The 517 burn events in that dataset have a very
four modelling techniques, and results were positive skewed total burned area distribution. The
compared using the overall accuracy, sensitivity, two-stage prediction process determines the relative
248
 International Journal of Advanced Technology and Engineering Exploration, Vol 9(87)

influences of the input factors on burned-area used to assess the fire safety of Russian railways
forecasts through informative feature selection. infrastructure installations.
Optimizing with independent objective functions for
mean absolute error (MAE) and root mean square In 2021, Preeti et al. [20] obtained the data set from
error (RMSE) gives more information with which to Kaggle, which comprises of meteorological data, and
data mining of each total burned area incident was then conducted an exploratory analysis, which
possible. Beyond regression and correlation-based included pre-processing and transforming categorical
methods, the proposed method might be easily data to numerical data in order to make the dataset
modified to anticipate and data mine generic more understandable. Following the preprocessing
agricultural systems. They are dependent on complex procedure, hotspot locations are identified using
interactions between meteorological and meteorological data from the data set, and models are
environmental variables. used to forecast the likelihood of a fire occurrence
and send a notification to the nearest station. The
In 2021, Singh et al. [18] collected the data from the Montesano Natural Park in the European country
Indian Meteorological Department. This data is made provided 517 observations and 13 variables for the
up of meteorological data from India's forest regions dataset. By using the RandomizedSearchCV
for the last 15 years. The university of california, algorithm, they performed RF regression and
irvine (UCI) and Indian Meteorological datasets, hyperparameter tuning on several sub-samples of the
respectively, contain information regarding forest dataset, it fits to numerous decision trees and provide
fires and meteorological data. The UCI dataset averaging to increase the predicted accuracy and
yielded 517 instances of data, with 155 instances of control over-fitting. Forest fire incidents can be
data being used for validation. The parallel SVM represented based on the analysis of models with all
algorithm is used to construct a reliable Forest Fire of the relevant meteorological parameters. They
Prediction system. In the prediction procedure, compared various models for predicting forest fires,
SPARK and PySpark were used to do data including ANN, Decision Tree, SVM and RF
segmentation and feature selection. To successfully algorithms. The use of RandomizedSearchCV
train the meteorological data and predict the forest coefficient calculation utilising Hyperparameter
fire, a parallel SVM model is constructed. For the tuning yields the best results. It obtained the Root
forecast of a forest fire, Parallel SVM uses the mean squared error (RMSR) 0.07, Mean absolute
forecast weather index (FWI) and some weather error (MAE) 0.03, Mean squared error (MSE) 0.004.
characteristics. To determine the model's efficiency,
the parallel SVM model is tested on data from India In 2021, Vikram et al. [21] proposed the
and Portugal. In the Portugal data, the parallel SVM classification model which is based on a traditional
model has an RMSE of 63.45 while the SVM dataset taken from Montesano Natural Park in
approach has an RMSE of 63.5. Portugal during a forest fire. They proposed a
strategy for using predictive analytics to detect forest
In 2021, Baranovskiy et al. [19] used the fires early. The forest is separated into different zones
development of information and computer systems using this method. The semi-supervised classification
for anticipating the fire safety of Russian railways' technique is used to anticipate the status of a zone,
infrastructure. It facilitates an important link in the such as high active (HA), medium active (MA), and
battle against fires in WUI areas. They used the low active (LA). Static sensors, mobile sensors, and
parallel computer technology, a numerical an Initiator node are all present in each zone. Using
investigation of heat transmission mechanisms in the the Random trajectory generation (RTG) technique,
enclosing framework of a wooden building near the initiator nodes in the LA and MA zones transmit their
forest fire front. The complexity of the algorithm O mobile nodes (MN) to the closer HA zone for speedy
(2N2 + 2K) is used to implement the deterministic forest fire prediction. This strategy creates the MN
model of heat transmission in the enclosing structure. movement path by generating intermediary sites
The NumPy and Concurrent libraries are used to between the LA/MA and HA zones. The anchor
implement the programme in Python 3.x. The nodes track the movement of MN using a compressed
calculations were done at the Sirius University of sensing-based Gradient descent (GD) localization
Science and Technology on a multiprocessor cluster. algorithm. This approach lowers MN's energy usage,
The acceleration coefficient and calculation results resulting in a longer network lifetime. The accuracy
for operating modes 1, 2, 4, 8, 16, 32, 48, and 64 of MN's path detection is improved by analysing its
processes were shown. The proposed method can be localization error as it travels towards the HA zone.

249
J Ananthi et al.

In 2022, Si et al. [23] used the forest fire statistics,

vegetation type, climatic data, and digital evaluation
In 2021, Zhang et al. [22] created two new datasets model (DEM) data of the fire occurrence. Forest fire
for segmentation and classification, respectively, by data from Yunnan Province's Forest Fire Prevention
using on four well known public databases i.e., Cair, Headquarters cover the years 1997 to 2018. The
Bilkent, Corsican and Foggia. The Corsican fire vegetation type data comes from Yunnan Province's
dataset contains 1135 fire photos recorded in various National Forest Resources inventory data from 2004
situations, all of which have been manually to 2009. The China Meteorological Science Data
segmented as ground truth. Foggia is a video Sharing Service Network (http://cdc.cms.gov.cn)
collection that includes 14 fire scenes and 17 non-fire provides meteorological data. The scientific data
scenes. Normal photos and photographs with fire center of cold and arid Area provided DEM data with
from a variety of scenarios and diverse fire settings, a spatial resolution of 90m. They used the spatial
such as intensity, luminance, environment, size, and overlay analysis, Kriging interpolation, analysis, and
many more, are included in the Cair dataset. The a Logistic regression model to investigate the
Bilkent dataset is a publicly accessible collection of association between the forest fire occurrence and
40 video clips, including 13 fire films, that has been elevation, vegetation types, and meteorological
used to assess fire detection frameworks. They parameters using a fire dataset from 1997 to 2018 in
created the two datasets from above mentioned the Lijiang area. To estimate the likelihood of forest
available datasets. The first dataset is utilised for fire occurrence and fire hazard rating in the Lijiang
segmentation tests, so they chose some fire pictures area, an optimum logistic model was chosen. They
from all datasets that provide ground truth. In this found that the forest fire occurrence was substantially
dataset, 1135 fire photos from the Corsican fire linked with slope, the daily average temperature and
dataset are used, while 5000 images are chosen at relative humidity, elevation, temperature and aspect.
random from Foggia's 17 non-fire videos. In general, The Logistic model has an average prediction
the 6135 images used for segmentation are separated accuracy of 86.9%.
into 70:30 training and testing sets. The second
dataset used a total of 3690 training photos, including It is clearly indicated from the above literature and
110 fire and 110 non-fire images from the Cair analysis that the machine learning and CNN
dataset and 455 fire and 1880 non-fire images from algorithms are capable in the forest fire prediction.
the Foggia dataset. By resampling using linear But there is the need of a proper alert system in the
interpolation with a resolution of 224x224 pixels, timely manner to avoid it.
1631 input images are randomly selected from 40
Bilkent videos for the testing phase. They adopted 3.Methods
method Attention U-Net and SqueezeNet (ATT In this paper, a deep learning-based approach called
Squeeze U-Net) is a SqueezeNet-based asymmetric LBFFPS has been proposed. It provides a capability
encoder-decoder U-shape architecture that primarily to the proposed approach to monitor the forest region
works as a forest fire extractor and discriminator. with respect to a smart surveillance device called
They initially use a Channel Shuflle operation as a SFMK. The surveillance kit SFMK consists of
feature communicator in the Fire module of classical several intelligent sensors, camera section and an IoT
SqueezeNet, replacing the classical convolution layer enabled controller. Based on these specifications the
with a depth wise one. The updated SqueezeNet is proposed approach can easily identify the forest fire
then used to replace the Attention U-Net encoder, and and report it to the respective authority accordingly.
a similar DeFire module is incorporated into the The sensors associated with the SMOKE are capable
decoder as well. Finally, they used a portion of the in smoke identification, temperature and humidity
ATT Squeeze U-Net encoder to classify genuine fire. monitoring. A high definition 1020-megapixel
The improved SqueezeNet incorporated Attention U- camera is placed into the SFMK to monitor the forest
experimental Net's findings reveal a competitive region and report it accordingly to the server entity
accuracy of 0.93 and an average prediction time of without any intervention. All these specifications are
0.89. illustrated in Table 1. Table 1 classifies fire hazard as
a consequence of the forest fire scenario based on
weather index [24].

250
 International Journal of Advanced Technology and Engineering Exploration, Vol 9(87)

Table 1 Forest fire hazard classification with respect to environmental weather index
Weather Index Ratio Category Hazard
Low indexing 0 to 5 Forest fire surface that is encroaching. Perhaps the forest fire is self-extinguishing.
Moderate indexing 6 to 10 Forest region fires with a less intensity. Simple to extinguish with hand instruments.
High indexing 11 to Forest fire surface that is moderately Pumps with electric motors and pipes are essential.
20A aggressive.
Very high indexing 21 to 30 Extremely hot forest surface fire. Complex to Control.
Extreme indexing 31 and Extreme fire propagation. It is vital to take swift and decisive action.
above

Temperature and Humidity Sensor liquids such as propane, petroleum-based fluids and
In this approach, a powerful and robust temperature hydrogen.
and humidity monitoring sensor is utilized, that is Learning-based Forest fire prediction scheme
called as DHT11. This sensor accumulates the (LBFFPS)
temperature by using the standard DHT11 library and In this paper, a deep learning based LBFFPS has
converts the accumulated value from Celsius to been introduced. It is executed on the server end to
Fahrenheit by means of the standard formulation as process the data acquired from smart device SFMK.
below (Equation 1): This learning process trains the machine based on the
data acquired from the forest region and tests the
T°F  T°C × ( )+32 (1) real-time information with respect to the trained
model. The trained model is created from the real-
time data acquired from the SFMK, in which the real
Where T indicates the temperature, F indicates
data accumulating from the device are undergone into
Fahrenheit and C indicate Celsius. The relative
cross-validation with respect to the created model.
humidity is estimated based on the following
The data is considered to be problem free means; it
formulation as below (Equation 2):
will be stored into the remote server end for two
purposes such as: (i) maintenance and monitoring
rH ( ) × 100% (2) from remote end as well as (ii) the processed data is
maintained in the remote cloud server end for
Where rH indicates relative humidity, Pw indicates the validating the further testing records. The camera
density of water vapor and Ps indicates the density of module associated with the device is used to monitor
the saturation level of water vapor. the forest regions in a pictorial way to maintain the
forest region without any misuses and accidents
Gas Sensor occurring. The algorithm of LBFFPS illustrates the
This is a limited expense semi-conductor based MQ2 specification and process flow of the proposed
smoke/gas identification sensor module in associated model.
with analog as well as digital outputs that is
extremely simple to operate. Like a gaseous Algorithm: LBFFPS
substance sensing element, this component makes use Input: Forest weather data and smoke ratio, which is
of the MQ2 gas, smoke and flammable gas gathered from the SFMK
monitoring sensor. It does not involve any additional Output: Fire prediction ratio and accuracy metrics
features; simply connect the power input to voltage Step 1: Importing the required libraries to process the
common collector (VCC) and ground pins to ground learning scheme data.
(GND), in which it is good enough to go. The Step 2: Acquire the input from the smart device
threshold level for digital and analog output can be SFMK and maintain that into the array variable.
simply changed through an on-board resistor. This Step 3: Split the received values based on the comma
MQ2 module enables users to simply connect the separated value and process accordingly.
specified sensor to the respective controller such as Step 4: Define the formulation to estimate the
NodeMCU, Arduino and so on. Due to the service temperature with respect to the Equation 1.
module sensitivity to gas, it could be used to identify Step 5: Define the formulation to estimate the relative
forest fires. Additionally, the MQ2 smoke humidity with respect to the Equation 2.
identification sensor is sensitive to combustible Step 6: Load the trained model for further processing.

251
J Ananthi et al.

Step 7: Cross validate the estimated temperature and Step 15: Return the prediction accuracy and metrics
humidity values with the trained model values, in to the user end for verification.
which it is acquired from the step 6. Figure 1 illustrates the working block diagram of the
Step 8: Intimate the processed specification of the SFMK. The block diagram of the monitoring kit
respective authority by means of storing it into the shows different controller and sensors. DHT11 is a
remote server end. humidity and temperature sensor. It is capable of
Step 9: Acquire the video streams from the SFMK sensing the surrounding air. MQ2 sensor is the MQ
and accumulate that into the server end video series gas sensor. It is a highly sensitive gas sensor. It
accumulation unit. is attached with the IoT interfacing for the proper
Step 10: The video streams are collected and communication and alert services. It is connected to
maintain it into the stream ratio of 1020 pixel rates. the remote cloud server. It is enabled with digital
Embed the video stream into the server end for camera of 1020 pixels. Based on these sensors the
maintenance. surrounding smoke presence, temperature level and
Step 11: Acquire the MQ2 sensor readings from the the humidity level have been identified and reported
respective array index. using the NodeMCU controller. In this application,
Step 12: Cross validate the acquired MQ2 sensor IoT is associated, to provides a wireless
values with the trained model values to identify communication alert ability. It collects and maintain
whether it exceeds the threshold level or stay in a the information regarding the forest provided by the
safe state. sensor unit to the remote cloud server environment.
Step 13: Inform the reading ratio to the respective The NodeMCU microcontroller has an inbuilt WiFi
authority once the threshold level exceeds the normal to acquire the internet signals and provides a constant
level. bridge between the sensor unit and the server end for
Step 14: Estimate the accuracy levels of prediction remote data maintenance.
and time constraints accordingly.

Figure 1 SFMK block diagram

4.Results and discussion consumption scenario to process the data acquired

Figure 2 shows the data accumulation efficiency from the forest region by using SFMK. In which the
using LBFFPS from the smart device located in the X-axis indicates the number of data accumulated
forest region. These ratios are accumulated based on from the forest region and the Y-axis indicates the
the 10 continuous day surveillance over the real-time time required to process the data. The time
forest environment. In which the X-axis indicates the indications are mentioned in milliseconds over the y-
number of days the device is located in the forest axis. Figure 4 shows the accuracy comparison based
region and the Y-axis indicates the data accumulation on SVM and LBFFPS. Figure 5 shows the accuracy
efficiency. Figure 3 shows the LBFFPS time comparison based on RF and LBFFPS. The

252
 International Journal of Advanced Technology and Engineering Exploration, Vol 9(87)

comparison clearly indicates the effectiveness of our approach in comparison to the SVM and RF.

3500
Data accumulation efficiency

3000

2500
Data sent (bps)
2000
Data receive (bps)
1500

1000
1 2 3 4 5 6 7 8 9 10
Number of days

Figure 2 Data accumulation efficiency based on the sent and received cases

3500

3000
Time (ms)

2500

2000 Time (ms)

1500

1000
2460 2700 2772 2956 3112 3268 3300 3332 3350 3470
Data accumulated

Figure 3 Time consumption ratio based on data accumulated

110
100
90
80
70
Accuracy

60
50 SVM
40 LBFFPS
30
20
10
0
1 2 3 4 5 6 7 8 9 10
Number of days

Figure 4 Accuracy comparison based on SVM and LBFFPS

253
J Ananthi et al.

110
100
90
80
70
Accuracy

60
50 RF
40 LBFFPS
30
20
10
0
1 2 3 4 5 6 7 8 9 10
Number of days

Figure 5 Accuracy comparison based on RF and LBFFPS

Limitations using weather monitoring sensors such as DHT11

The proposed methodology is designed based on the and MQ2. The proposed LBFFPS mechanism is
deep learning strategy as well as the sensor unit capable of gathering the real-time data acquired from
consists of two sensors such as DHT11 and the MQ2 the forest region by using the number of sensors
smoke identification sensor. This DHT11 sensor associated with the smart device called SFMK.
identifies the temperature and humidity level in the Additionally, it doesn't require more computing
correct manner, but over the forest region the capabilities than a solitary wireless sensor network to
temperature and climate constraints are unpredictable form a real-time adaptive network. The concentration
as well as the conditions may change within a period on a small number of measures in conjunction with
of instance. So that a motorized robot-like moving carbon monoxide monitoring implies that the
object is required to monitor the forest region in proposed solution is less expensive, real-time
intense manner and report accordingly to the functional and effective.
authorities without any delay. These kinds of smart
enhancements can add the boosted evaluation to the Acknowledgment
methodology defined over this paper as well as the None.
proposed methodology need to be refined according
to the time constraint to prevent the forest region in Conflicts of interest
good condition. The other limitation is the real time The authors have no conflicts of interest to declare.
environment. This experimentation was performed
with the small sample. So, it should be validated on Authors contribution statement
the large-scale forest with longer time with the J Ananthi: Conceptualization, methodology, data curation,
formal analysis, investigation, validation, writing-original
performance possibilities. Powerful internet facility is draft preparation. N Sengottaiyan: Conceptualization,
also important in this research. methodology, data curation, formal analysis, investigation,
A complete list of abbreviations is shown in visualization, writing- original draft preparation, writing-
Appendix I. review and editing. S Anbukaruppusamy:
Conceptualization, methodology, data curation, formal
4.Conclusion analysis, investigation, writing- original draft preparation,
This paper establishes and develops a wireless writing- review and editing. Kamal Upreti: Validation,
supervision and project administration. Animesh Kumar
communication model using an IoT enabled network Dubey: Validation, supervision and project administration.
with intelligent sensors. It is capable of measuring
forest region data in real-time with global area
coverage options. This level of responsiveness is
achieved by attaching a simultaneous elevated data
capturing module with powerful internet connectivity
254
 International Journal of Advanced Technology and Engineering Exploration, Vol 9(87)

References [13] Zhu H, Gao D, Zhang S. A perceptron algorithm for

[1] Faroudja AB, Izeboudjen N. Decision tree based forest fire prediction based on wireless sensor
system on chip for forest fires prediction. In networks. Journal on Internet of Things. 2019;
international conference on electrical engineering 1(1):25-31.
(ICEE) 2020 (pp. 1-4). IEEE. [14] Tehrany MS, Jones S, Shabani F, Martínez-Álvarez F,
[2] Zope V, Dadlani T, Matai A, Tembhurnikar P, Kalani Tien Bui D. A novel ensemble modeling approach for
R. IoT sensor and deep neural network based wildfire the spatial prediction of tropical forest fire
prediction system. In 4th international conference on susceptibility using LogitBoost machine learning
intelligent computing and control systems (ICICCS) classifier and multi-source geospatial data. Theoretical
2020 (pp. 205-208). IEEE. and Applied Climatology. 2019; 137(1):637-53.
[3] Sinde RS, Kaijage S, Njau KN. Cluster based wireless [15] Zhang G, Wang M, Liu K. Forest fire susceptibility
sensor network for forests environmental monitoring. modeling using a convolutional neural network for
International Journal of Advanced Technology and Yunnan province of China. International Journal of
Engineering Exploration. 2020; 7(63):36-47. Disaster Risk Science. 2019; 10(3):386-403.
[4] Latifah AL, Shabrina A, Wahyuni IN, Sadikin R. [16] Mohajane M, Costache R, Karimi F, Pham QB,
Evaluation of Random Forest model for forest fire Essahlaoui A, Nguyen H, et al. Application of remote
prediction based on climatology over Borneo. In sensing and machine learning algorithms for forest fire
international conference on computer, control, mapping in a Mediterranean area. Ecological
informatics and its applications (IC3INA) 2019 (pp. 4- Indicators. 2021; 129:107869.
8). IEEE. [17] Wood DA. Prediction and data mining of burned areas
[5] Singh BK, Kumar N, Tiwari P. Extreme learning of forest fires: optimized data matching and mining
machine approach for prediction of forest fires using algorithm provides valuable insight. Artificial
topographical and metrological data of vietnam. In Intelligence in Agriculture. 2021; 5:24-42.
women institute of technology conference on electrical [18] Singh KR, Neethu KP, Madhurekaa K, Harita A,
and computer engineering (WITCON ECE) 2019 (pp. Mohan P. Parallel SVM model for forest fire
104-112). IEEE. prediction. Soft Computing Letters. 2021; 3:100014.
[6] Majhi AK, Dash S, Barik CK. Arduino based smart [19] Baranovskiy NV, Podorovskiy A, Malinin A. Parallel
home automation. ACCENTS Transactions on implementation of the algorithm to compute forest fire
Information Security. 2021; 6 (22): 7-12. impact on infrastructure facilities of JSC russian
[7] Bhadoria RS, Pandey MK, Kundu P. RVFR: Random railways. Algorithms. 2021; 14(11):333.
vector forest regression model for integrated & [20] Preeti T, Kanakaraddi S, Beelagi A, Malagi S, Sudi A.
enhanced approach in forest fires predictions. Forest fire prediction using machine learning
Ecological Informatics. 2021; 66:101471. techniques. In international conference on intelligent
[8] Yang X, Xiong S, Li H, He X, Ai H, Liu Q. Research technologies (CONIT) 2021 (pp. 1-6). IEEE.
on forest fire helicopter demand forecast based on [21] Vikram R, Sinha D, De D, Das AK. PAFF: predictive
index fuzzy segmentation and TOPSIS. In 9th analytics on forest fire using compressed sensing
international conference on fire science and fire based localized Ad Hoc wireless sensor networks.
protection engineering (ICFSFPE) 2019 (pp. 1-8). Journal of Ambient Intelligence and Humanized
IEEE. Computing. 2021; 12(2):1647-65.
[9] Labellapansa A, Syafitri N, Kadir EA, Saian R, [22] Zhang J, Zhu H, Wang P, Ling X. ATT squeeze U-
Rahman AS, Ahmad MB, et al. Prototype for early Net: a lightweight network for forest fire detection and
detection of fire hazards using fuzzy logic approach recognition. IEEE Access. 2021; 9:10858-70.
and Arduino microcontroller. International Journal of [23] Si L, Shu L, Wang M, Zhao F, Chen F, Li W, Li W.
Advanced Computer Research. 2019; 9(44):276-82. Study on forest fire danger prediction in plateau
[10] Kalantar B, Ueda N, Idrees MO, Janizadeh S, Ahmadi mountainous forest area. Natural Hazards Research.
K, Shabani F. Forest fire susceptibility prediction 2022.
based on machine learning models with resampling [24] Ely I. Droeg,“The Osborne Fire finder the basic
algorithms on remote sensing data. Remote Sensing. lookout tools: firemans guide california region”. us
2020; 12(22):3682. department of agriculture forest service, available on:
[11] Tiwari A, Shoab M, Dixit A. GIS-based forest fire www. socalfirelookouts. org/Osborne% 20Fire-
susceptibility modeling in Pauri Garhwal, India: a finderUsersGuide. Pdf.
comparative assessment of frequency ratio, analytic
hierarchy process and fuzzy modeling techniques.
Natural Hazards. 2021; 105(2):1189-230.
[12] Lin H, Liu X, Wang X, Liu Y. A fuzzy inference and
big data analysis algorithm for the prediction of forest
fire based on rechargeable wireless sensor networks.
Sustainable Computing: Informatics and Systems.
2018; 18:101-11.

255
 J Ananthi et al.

J Ananthi is working as an Assistant Animesh Kumar Dubey is working as

Professor in the Department of Assistant professor with the department
Computer Science and Engineering at of Computer Science and Engineering,
Hindusthan College of Engineering and at Madhyanchal Professional
Technology, Coimbatore. She University, Bhopal, India. He has
completed her ME CSE at the College completed his Bachelor of Engineering
of Engineering in Guindy, Chennai. (B.E.) and MTech. degree with
She has published research articles in Computer Science Engineering from
Scopus, the Web of Science, and international journals. In Rajeev Gandhi Technical University, Bhopal (M.P.). He
addition, she has published books on e-commerce, IoT, and has more than 15 publications in reputed, peer-reviewed
advanced database systems. Her field of interest is wireless national and international journals and conferences. His
networks. research areas are Data Mining, Optimization, Machine
Email: [email protected] Learning, Cloud Computing and Artificial Intelligence
Email: [email protected]
Dr N Sengottaiyan is working as a
Professor and Head of the Department Appendix 1
in CSE Sri Shanmuga College of S. No. Abbreviation Description
Engineering and Technology, 1 AUC Area Under Curve
Thiruchengode. He completed his 2 CNN Convolutional Neural Network
Ph.D. in 2011 at Anna University of 3 DEM Digital Evaluation Model
Technology and has 30 years of 4 FR-CART Frequency Ratio-Classification
experience in teaching. He has Regression Tree
published more than 20 articles in refereed journals and 5 FR-RF Frequency Ratio-Random Forest
6 FRLR Frequency Ratio-Logistic
more than 75 national and international journals. He has
Regression
also published books in Compute on computer 7 FR-MLP Frequency Ratio-Multilayer
programming and organized many national and Perceptron
international level technical events. The number of students 8 FR-SVM Frequency Ratio-Support Vector
pursuing a Ph.D. under his guidance is 12, and the number Machine
of scholars completed is 6. He is an influential reviewer for 9 FWI Forecast Weather Index
several journals. His domain of interest is wireless sensor 10 GD Gradient Descent
networks. 11 GND Ground Pins to Ground
Email: [email protected] 12 HA High Active
13 IoT Internet of Things
Dr. S Anbukaruppusamy completed 14 KLR Kernel Logistic Regression
his Ph.D. at Anna University, Chennai. 15 LA Low Active
He is working as Professor and Head of 16 LBFFPS Learning-Based Forest Fire
the Department of Electronics and Prediction Scheme
Communication Engineering at Excel 17 LEDT LogitBoost Ensemble-Based
Decision Tree
Engineering College, Namakkal. He is
18 MA Medium Active
guiding 12 scholars in doing research. 19 MAE Mean Absolute Error
He has published 43 articles in national 20 MN Mobile Nodes
and international journals and eight in refereed journals. 21 NPV Negative Predictive Value
His field of interest is wireless networks. He has done a lot 22 PPV Positive Predictive Value
of research based on real-time problems and scenarios. 23 RF random forest
Email: [email protected] 24 RMSE Root Mean Square Error
25 ROC Receiver Operating Characteristic
Dr. Kamal Upreti is currently working Curve
as an Associate Professor in the 26 RTG Random Trajectory Generation
Department of Computer Science & 27 SFMK Smart Forest Monitoring Kit
Engineering , Dr. Akhilesh Das Gupta 28 SVM Support Vector Machine
Institute of Technology & 29 UCI University of California, Irvine
Management, Delhi. He completed is 30 VCC Voltage Common Collector
B. Tech (Hons), M. Tech (Gold 31 WSN Wireless Sensor Networks
Medalist), PGDM(Executive) and PhD
in Computer Science & Engineering. He has published
many patents, books, magazine issues and research papers
in various reputed international conferences and journals.
His areas of research interest include Machine Learning,
Wireless Networking, Embedded System and Cloud
Computing.
Email: [email protected]
256

View publication stats