skip to main content
research-article
Open Access

An Effective Forest Fire Detection Framework Using Heterogeneous Wireless Multimedia Sensor Networks

Authors Info & Claims
Published:16 February 2022Publication History

Skip Abstract Section

Abstract

With improvements in the area of Internet of Things (IoT), surveillance systems have recently become more accessible. At the same time, optimizing the energy requirements of smart sensors, especially for data transmission, has always been very important and the energy efficiency of IoT systems has been the subject of numerous studies. For environmental monitoring scenarios, it is possible to extract more accurate information using smart multimedia sensors. However, multimedia data transmission is an expensive operation. In this study, a novel hierarchical approach is presented for the detection of forest fires. The proposed framework introduces a new approach in which multimedia and scalar sensors are used hierarchically to minimize the transmission of visual data. A lightweight deep learning model is also developed for devices at the edge of the network to improve detection accuracy and reduce the traffic between the edge devices and the sink. The framework is evaluated using a real testbed, network simulations, and 10-fold cross-validation in terms of energy efficiency and detection accuracy. Based on the results of our experiments, the validation accuracy of the proposed system is 98.28%, and the energy saving is 29.94%. The proposed deep learning model’s validation accuracy is very close to the accuracy of the best performing architectures when the existing studies and lightweight architectures are considered. In terms of suitability for edge computing, the proposed approach is superior to the existing ones with reduced computational requirements and model size.

Skip 1INTRODUCTION Section

1 INTRODUCTION

IoT infrastructures, smart sensors, and wireless sensor networks (WSNs) are some of the most promising technologies of our time, especially for fully autonomous systems. Traditional WSNs are generally used to transmit scalar data such as temperature, humidity, and light-related information. The size of scalar data is usually much smaller than that of multimedia data. However, the accuracy of systems based solely on scalar data may not be sufficient in some critical applications that require real-time monitoring and control. Environmental monitoring for disaster management, such as the detection of forest fire and floods, and medical applications, such as wireless patient monitoring, are two popular examples of applications in which multimedia sensors are commonly used with various scalar sensors.

Wireless Multimedia Sensor Networks (WMSNs) have access to multimedia sensors. They are used effectively with new image processing and machine learning (ML) techniques in various types of applications, including surveillance systems, traffic and transport, smart healthcare delivery, surveillance of environment and localization systems [19, 37, 45, 63]. One of the main constraints of WMSNs is energy consumption, since the transmission of multimedia data is expensive. WMSN nodes are limited not only in terms of battery but also in terms of memory and processing power. It is therefore crucial to use the available resources as efficiently as possible.

Recent studies on environmental disasters, including forest fires, show that the autonomous systems have great potential for improvement [19]. The forest fires in Athens, where at least 90 people died, 164 adults and 23 children were injured, the California wildfires in which 185,800 hectares of forest were burned in 2018 [59], and the very recent fires in Australia, where around 12 million acres were burned [50], are events that early and accurate forest fire detection systems could have a significant impact on these fires.

The main objective of this study is to present an accurate and energy-efficient framework for emergency applications combining energy-efficient scalar sensors, multimedia sensors and advanced deep learning (DL) algorithms. Since forest conditions are not suitable for continuous maintenance, self-configurable, inexpensive, accurate, easy-to-deploy, and energy-efficient systems are required.

To achieve energy efficiency, we present a hierarchical framework in this article. Two different types of sensors are considered for the hierarchy introduced:

  • Scalar sensors: These sensors have limited computing, memory, and sensing capabilities.

  • Multimedia sensors: These sensors are equipped with multimedia sensors, and they have significantly better computing and memory capacities. This is why they are also referred to as smart sensors.

Scalar sensors are used at the first level of detection. Measurements from scalar sensors are thus used to trigger multimedia sensors for more accurate detection of fires. In addition, deep learning algorithms are used in nodes with multimedia sensors (also called smart sensors). In this sense, it is possible to say that edge computing is employed; however, unlike many of the existing studies, classification is performed on end nodes instead of a dedicated sink node close to the deployment area. To be able to deploy and run the deep learning algorithms for detection at the end devices, the model size and the computational requirements of the architecture employed are reduced. Although in terms of computational requirements and model size, the proposed model is superior, the accuracy of the proposed architecture is better than most of the existing studies and quite close to the architectures with highest detection accuracy. The edge computing is employed to prevent unnecessary data transmission, which increases the efficiency of the system, because data transmission is the most energy-intensive activity. By reducing the transmission rate, the traffic load of the wireless channel is also reduced. This results in an increase in the packet reception rate, since the probability of collision decreases. For the accuracy of the system, a lightweight DL model is proposed and tested. The proposed framework is implemented using a real testbed. The results obtained from the testbed are used in conjunction with OMNET++-based [48, 62] results to verify the energy efficiency and quality of service (QoS) measures. To train and test the DL model, in particular, a Convolutional Neural Networks (CNN) model, a new image dataset1 is defined for forest fire applications. According to the results of the experiments, an accuracy of about 98% is obtained and validated to adjust the parameters of a classifier. The main contributions of this study are summarized below.

  • A novel, energy-efficient and accurate heterogeneous wireless multimedia sensor network framework is proposed for real-time surveillance applications.

  • Although there are solutions based on ML in the literature, these approaches require significant computing resources and memory space. We propose a significantly more lightweight CNN model in this study for classification.

  • To the best of our knowledge, this is the first study in which the CNN works in the smart sensors at the edge of the network for the detection of forest fires.

  • A real-time testbed implementation is presented, and simulations are used for verification and extrapolation purposes.

  • Our approach is as accurate as other competitive detection mechanisms and more efficient in terms of energy consumption, as well as computational and memory requirements.

The rest of the article is organized as follows. Existing studies are discussed in Section 2. Section 3 explains the proposed forest fire detection framework. Section 4 presents our approach. The evaluation of our proposed system is presented in Section 5. Finally, Section 6 concludes the article.

Skip 2RELATED WORK Section

2 RELATED WORK

With efficient communication infrastructures and support of network protocols at various levels, WSN and WMSN have been widely integrated with multi-modal sensors and actuators to actively detect any abnormal event in terms of safety, surveillance, and security.

2.1 Forest Fire Detection

Traditionally, forest fire detection has been conducted using human observations from watchtowers located in high-altitude areas. However, the working conditions of the surveillance staff are poor and the accuracy of this information is open to discussion [61]. Other previous implementations include video surveillance systems and use of satellites, which suffer from problems such as accuracy, delay, and energy efficiency.

Given the constraints of traditional methods, one of the most promising technologies for fire detection is wireless sensor networks with IoT-enabled applications [47]. In Reference [14], a new architecture is presented for public protection and disaster relief networks. The architecture is based on cognitive communication and uses WSNs to establish what is called as “Low-Tier Application Network.” As in our study, the simulation results are presented to demonstrate the effectiveness of the proposed architecture, considering a disaster scenario in which emergency situations such as a fire or bombing occurs in the center of an area. The results of the simulation presented are mainly related to QoS measures, such as cognitive network throughput. Another study of disaster scenarios is presented in Reference [17]. A framework is presented for an industrial field based on the industrial WSN (IWSN). Fire emergencies in general and forest fires are presented as application areas of the new framework. The importance of energy efficiency is emphasized throughout the study. The proposed framework is not evaluated in terms of energy efficiency or QoS.

In Reference [19], the fire sensing technologies and differences in the development of hardware as well as algorithms are considered critically. A modified fire sensing and control system concept is also proposed. The proposed framework makes use of an array of sensors; however, multimedia related features that could have improved the detection accuracy significantly are not considered. Furthermore, the proposed framework is at the stage of development and the quantitative results of evaluation are not presented. Similarly, in Reference [18] WSNs are employed for disaster scenarios such as severe earthquakes, collapsing buildings and fires. The message synchronization capability of WSNs is used for new opportunistic network infrastructures to achieve low delivery latency and high delivery rate.

In a recent study [30], a framework is presented for forest fire surveillance using WSNs, as IoT devices together with fog and cloud computing paradigms. Instead of multimedia data, datasets of various meteorological factors such as temperature, relative humidity, precipitation, wind speed, and spatial coordinates of a potential fire site are used together with machine learning algorithms. The collected information is used more proactively to specify the susceptibility classes of an observation area (very-low, low, moderate, high, and very-high) instead of real-time high accuracy detection through multimedia data. Artificial Neural Network (ANN), K-Nearest Neighbour (KNN), Decision Tree, and Support Vector Machine (SVM) techniques are employed for classification. ANN outperformed the other classification techniques with an accuracy of 95.45%. The lifetime of the framework is presented in terms of battery levels for an observation of up to 300–400 min.

When reviewing existing work on forest fire detection, some studies fail to address energy efficiency, since appropriate assessment methods are not employed. However, although some studies use simulation tools to show the energy efficiency of the proposed architectures, they do not consider the multimedia data with the assessment of the detection accuracy. Furthermore, systems that use machine learning or deep learning models do not perform classification at the edge of the network.

2.2 Multimedia Sensor Networks

Studies in the environmental monitoring application literature including forest fire detection, can be categorized as scalar sensor-based systems, multimedia sensor systems, and systems using both technologies as shown in Table 1.

Table 1.
StudySensors UsedEvaluation
[32]MultimediaEnergy
[68]MultimediaAccuracy
[10]MultimediaEnergy & Accuracy
[16],[35],[5],[28],[51]ScalarOnly System Design
[49],[4],[67],[41]ScalarEnergy
[42]ScalarReview
[25]ScalarSecurity
[33]Multimedia & ScalarEnergy & Accuracy

Table 1. Literature Categorization

A study of a system in which Raspberry Pi and Arduino boards are used for environmental monitoring applications is proposed in Reference [16]. Scalar sensor readings such as temperature and humidity are used to monitor the environment. A systematic review is presented for the wireless sensor network applications used to monitor coal mines in Reference [42]. WSN-based forest fire monitoring is used to test the algorithms for detecting anomalies in Reference [25]. The proposed method estimates the maximum number of malicious acts tolerated by the application. The design of the deployment of the sensors is automated using the proposed method. Similar studies are presented in Reference [51] where in-depth analysis and validation real-time scheduling is performed in resource constrained hardware platforms such as Raspberry Pi and Arduino boards, and in Reference [35], where body temperature, respiratory rate, heart rate and body movements are monitored using Raspberry Pi cards. A novel computing architecture is introduced [51], while a detailed system design is proposed in Reference [35]. The study presented in Reference [35] lacks the analysis of accuracy and energy efficiency. A traffic monitoring system is offered in Reference [32]. A Raspberry Pi card and an HD camera are used for object detection and the system is analyzed in terms of energy consumption.

Classification of moving objects, real-time detection, recognition and understanding of activity from multiple sensors are considered in studies such as References [10, 55]. While a novel fusion framework is presented in Reference [55] to combine data from multiple video sensors, a system architecture with scalar and multimedia sensors is presented in Reference [10] for WMSNs to facilitate classification of moving objects. For both studies, experimental setups are implemented, and tests are performed to show the accuracy of the frameworks. Multimedia and scalar sensors are used together in Reference [33] as well for the object detection and classification with visual and auditory data fusion. A practical implementation of a system is presented, and results of performance and energy consumption are discussed. Multimedia data, particularly in the form of surveillance video, is extensively used for event recognition. In most of the studies using surveillance video, while prioritising the accuracy of detection, the underlying frameworks become quite complicated. For example, in Reference [38], a Spatio-temporal deep residual network with hierarchical attentions is employed for video event recognition. Residual Neural Network (ResNet), Faster Region-based Convolutional Neural Networks (R-CNN), and Long Short-term Memory (LSTM) are used with a multi-layer architecture to improve the accuracy by 5.6% compared to similar existing architectures.

The literature presents numerous examples of monitoring and/or surveillance applications using multimedia sensors. Some studies focus on the accuracy of surveillance information. Some other studies consider energy efficiency alongside accuracy. However, none of these studies propose an autonomous system for forest fires. To the best of our knowledge, our study presented here is the first to attempt to combine the two approaches and evaluate their effectiveness and accuracy using simulations as well as real-world testbed for wildfire surveillance applications.

Recently, unmanned aerial vehicles (UAV) have become popular as an enabling technology of certain environment monitoring applications. In Reference [2], a UAV-based forest fire detection system is proposed. The battery life of the UAV is set as 18 min, which means that the system can work in an emergency for alert, instead of detecting the emergency. Similarly, in Reference [56], an approach based on fog computing and deep learning with drones is proposed. A deep convolutional network architecture is employed, which achieves up to 95% accuracy for fire detection. In Reference [8], a CNN-based forest fire detection system is presented where images of the forest are collected with a drone. The approach presented tests different CNN architectures, reaching up to 86% accuracy.

UAV-based environment detection systems have their advantages; however, there are many limitations to using UAVs both proactively and reactively for forest fire detection. First, due to its limited battery, UAVs can patrol the given area at specific intervals and cannot provide continuous monitoring. In that sense, if a proposed UAV-based system employs a large number of ready-to-use batteries for certain applications, then this will significantly affect the cost of the entire system. In addition, a UAV has limitations on the payload to maintain a safe and proper flight, and has less adaptability to different weather conditions. Therefore, they may be preferred for certain specific applications, but not for continuous monitoring for the early detection of possible disasters.

2.3 WMSNs and Machine Learning for (Forest) Fire Detection

Studies on forest fire detection using multimedia data typically includes studies of image processing, video processing, and, rarely, machine learning techniques. However, following their success in classification, object recognition, and detection, approaches based on machine learning (especially deep learning) have also proven successful in the recognition of events. As a result, many of the existing studies on event analysis rely heavily on deep learning architectures [1].

A wildfire detection framework based on image processing is presented in Reference [20]. Scale invariant feature transform is used for feature selection, and SVM and KNN classifiers are applied for fire classification, as fire and non-fire images. A combination of CNN and Recurrent Neural Networks (RNN) is used to create a DL model for detecting fires from video sequences in Reference [21]. On average, an accuracy of 92% is obtained. In Reference [43] a CNN-based early detection method is proposed where the processing capabilities of CCTV cameras are used. The proposed method is applicable to indoor and outdoor environments. However, the model size is quite large (238 MB) for using the model in WMSN applications. Another CNN-based fire detection system is proposed in Reference [36] where UAV images are used for training of the network. The system proposed in Reference [36] achieves an accuracy of 99% with GoogLeNet.

When the existing studies are analysed in detail, the cooperative use of multimedia and scalar sensors have been investigated and explored in various surveillance-based studies. However, there aren’t as many studies focusing on forest fires and trying to maximise the energy efficiency and similar QoS measures together with forest fire detection accuracy. Considering the WMSN applications of forest fire detection in the literature, the majority of studies rely solely on image and video processing to detect forest fires. Only a few studies, such as Reference [21], use machine learning models, including ANN, SVM, and RNN. However, the proposed machine learning models are not suitable for use in IoT edge devices, because they require relatively large memory and computing power. In this study, a lightweight CNN model is developed, which can be used on edge devices with limitations in terms of computing, memory, and power resources. The proposed lightweight CNN model and the hierarchical forest fire detection framework provide a much more efficient solution in terms of energy consumption without compromising detection accuracy. To the best of our knowledge, this is the first study in which CNN is used in the smart sensors at the edge of the network for the detection of forest fires. The proposed model is trained using a newly created forest fire dataset.

Skip 3FOREST FIRE DETECTION FRAMEWORK Section

3 FOREST FIRE DETECTION FRAMEWORK

The proposed effective forest fire detection framework uses scalar sensors for energy efficiency and multimedia sensors (for capturing snapshots) to improve the accuracy with the aid of DL and edge computing. XM1000 sensors and Raspberry Pi (RPi) 3s are used in this study as scalar and multimedia sensor nodes, respectively. An example scenario for the proposed framework is shown in Figure 1(a), and a flowchart is presented in Figure 1(b) to illustrate the flow of information. In Figure 1(b), T, H, \(Th_{Temperature}\), and \(Th_{Humidity}\) are used to denote the current temperature, the current humidity, the threshold of temperature and the threshold of humidity values, respectively.

Fig. 1.

Fig. 1. Forest fire detection system architecture and information flow chart for first and second phase of detection.

The area to be monitored is divided into regions. Each region has one multimedia sensor node with Raspbian OS and a number of scalar sensors using Contiki OS [13]. Multi-hop communication is considered between scalar and multimedia sensors, and, the built-in CC2420 transceiver is employed with ZigBee 802.15.4 protocol in the radio layer. In Radio Duty Cycling layer, ContikiMAC is used for low-power listening with energy efficiency. In the MAC and network layers, CSMA/CA and Routing Protocol for low power and lossy networks (RPL) are used, respectively, to deal with possible packet collisions.

In the first phase of the detection, scalar sensors are used to monitor the forest environment by measuring the temperature and humidity. The measured values are compared to the threshold values. The seasonal mean temperature and humidity of the deployment area (Cyprus in our case) are used in calculating the thresholds for humidity and temperature. We can denote the seasonal mean temperature and humidity of the deployment area as \(Mean_{temperature}\) and \(Mean_{humidity}\), respectively. For the temperature, the threshold is computed by considering the mean seasonal temperature with a tolerance of an additional 10% (\(Th_{Temperature} = Mean_{temperature} \times 1.1\)). Similarly, for humidity, seasonal mean humidity \(-10\%\) is considered as the threshold (\(Th_{Humidity} = Mean_{humidity} \times 0.9\)). If the measured temperature is greater than the threshold value of temperature or the measured humidity is less than the humidity threshold, then all the measured values are transmitted to the multimedia sensor node in the region via intermediate nodes. The transmitted packet includes values of temperature, humidity, light, voltage, id, number of hops, sequence number, packet size, date, and time. The nearby sensors detect the fire quicker and transmit the triggering packet to the cluster head (multimedia sensor). The id of each sensor is assigned prior to the deployment and can be used to estimate the sensor’s location and its cluster. The second phase of detection is initiated by the receiving node (the smart sensor, which is RPi in this scenario). In the second phase, RPi activates the pi camera to capture an image of the environment. The camera of the multimedia sensor is turned on even in case a single activation packet is received from the scalar sensors for introducing a robust fire detection framework. The captured image is processed by the proposed lightweight CNN model at the edge to detect forest fire. If the result indicates a forest fire, then the fire department is informed. With the proposed framework, multimedia data transmission is not necessary, since the lightweight CNN model runs on the edge devices. In addition, snapshots are captured as many times as the scalar sensor notifications. Since the snapshots are not transmitted to the sink (processed on the edge with lightweight CNN), their effect on the lifetime of the sensors is not very significant. According to our experiments, compared to the idle state, RPi consumes 0.12 W extra power while running the proposed lightweight CNN. However, compared to the idle state, the additional power consumption of the RPi increases to 1.07 W while transmitting the images. Furthermore, in case the scalar sensors generate multiple notifications, the probability of false negative is reduced, since each warning is taken into account. This, in turn, increases the reliability of the framework.

To achieve energy efficiency, three new features are being used in the proposed framework. The first feature is the hierarchy. Scalar sensor nodes are used at the first level of the hierarchy, because they consume much less power than the multimedia sensor nodes. In the implemented testbed, scalar sensors monitor the environment by taking measurements in every 10 s. Instead of sending associated information (packets) every 10 s, the measurements are examined in each individual sensor node. No transmission is carried out by scalar sensors except in the cases where measured values are not within the threshold values.

The second feature used is heterogeneity. Heterogeneous wireless networks are more appropriate solutions for surveillance and disaster management applications. In our case, the conditions of the forest environment can be difficult to deploy, manually configure, and maintain sensor nodes. In case of fire, it is also quite possible to lose a large number of sensors. In this sense, it is preferable to use scalar sensors that are inexpensive, easy-to-deploy, self-configurable, and energy-efficient. In studies such as those in References [14, 17, 25, 27, 30, 42], large numbers of scalar sensors are employed for environmental monitoring including detection of forest fires. However, the accuracy of fire detection is also crucial. By applying heterogeneity, the energy efficiency is achieved by scalar sensors and the accuracy is obtained by multimedia sensors.

The other approach used to achieve energy efficiency and accuracy includes the use of edge computing. For the second phase of detection, a lightweight CNN-based model is developed. The model developed uses the images captured by the RPi. Applying DL within the RPi reduces the number of transmissions needed from the multimedia nodes to the sink.

Skip 4FOREST FIRE DETECTION USING DEEP LEARNING Section

4 FOREST FIRE DETECTION USING DEEP LEARNING

The main objective of the proposed lightweight CNN model is to provide a robust approach for forest fire detection. As discussed in Reference [44], CNN-based image classification, especially in fire detection, provides substantial improvement in terms of detection accuracy, which in turn minimizes disastrous results of fires. However, the main drawback of CNN-based classification is high resource requirements, which can become an important problem for resource limited edge devices. To mitigate these limitations, a lightweight CNN model is proposed in this study that can be used at edge devices with considerably high detection accuracy. Well-known image classification neural networks, and the best combination of hyper-parameters are investigated to adapt the most suitable architecture for the model.

Another important task for preparation of the model is the dataset creation.

Existing literature is rather limited in terms of datasets, and the ones available are typically used to detect fires from smoke images. However, our study specifically focuses on forest fires in various forest environments while aiming for detection from still images considering both smoke and fire. For the generation of datasets, close-up videos of forest fires are examined and the necessary frames are extracted and labeled. Labeling process is verified by re-checking the labeled images manually to avoid incorrect labeling.

4.1 Dataset

Since public image datasets for forest fire detection are not available, wildfire videos are collected on YouTube and sampled according to camera movement. Manual removal is performed to avoid incorrect labeling. The dataset generated includes 1,111 fire images of 16 videos and 2,289 non-fire images from 9 videos, for a total of 3,400 images. Furthermore, different seasons and lightning conditions are taken into account. Fireless forest images are obtained for different seasons. While winter forest images are around 13% of the dataset, fall, spring and summer forest images are around 24%, 38%, and 25% of the dataset, respectively. However, forest fire images are obtained in different lighting conditions where 40% of images are at nighttime and 60% of images are at daytime. In addition, while 72% of the set used as fire images has both fire and smoke, 12% has only fire, 5% has only smoke. Various forest fires in Canada, California and Turkey are taken into account. Some examples of images in the dataset are shown in Figures 2 and 3. The division of training and testing data is performed with random sampling from each class with a specific ratio based on the validation method selected. To illustrate, in the case of an 80–20% split, for validation set, 20% of each class is randomly selected and added into the validation set, and the remaining 80% is considered as a training set.

Fig. 2.

Fig. 2. Fire Image Examples.

Fig. 3.

Fig. 3. Non-fire Image Examples.

Considering some of the publicly available datasets, such as References [31, 36, 60], or some studies that use images extracted from videos similar to our approach, such as References [20, 21, 56, 65], our proposed dataset contains more images than any existing dataset for forest fire detection. In Reference [36] aerial videos are used for image collection, since the main focus is fire detection using aerial vehicles. However, considering the system and infrastructure that we propose for forest fire detection, aerial images are not suitable to train our lightweight CNN. It is possible to further expand the dataset with additional images from different video streams. However, the size of the dataset is sufficient to represent the capability of the CNN model placed at the secondary level, since the primary focus is early detection unlike the other forest-fire-related attempts, which try to predict forest fires, the direction of the fire, or their resulting damage [11].

4.2 CNN Model

A lightweight CNN model is proposed to detect forest fires accurately. Although DL is popular in various image classification tasks, forest fire detection as a specific domain offers only limited work with DL architecture. In addition, the computation time requirements of hardware platforms that can be used for forest fire detection systems should be considered when developing the model. Therefore, this study proposes a DL model that can work on small computers such as Rasperry Pi, Orange Pi, and Hikey 970. The model is developed and tested using PyTorch and Keras libraries [9].

The proposed CNN model uses \(64\times 64\times 3\) input images. The CNN Architecture contains four convolutional layers each followed by a max pooling layer with 2 \(\times\) 2 kernels for feature extraction, and uses three fully connected layers for classification. In addition, each fully connected layer contains a dropout layer with the rate of 0.25. As shown in Figure 4, the proposed CNN model is kept as shallow as possible, because the deeper networks have higher time complexity and memory related requirements.

Fig. 4.

Fig. 4. CNN Model Architecture.

The lightweight CNN model presented here is trained using the newly created dataset on forest fires. A distribution ratio of 80–20% of the train-test is applied. Stochastic gradient descent is used as an optimizer, since it is widely used in CNN applications and studies such as References [39, 64, 66]. As an activation function, Rectified Linear Unit (RELU) function is used.

Skip 5EVALUATION OF THE PROPOSED SYSTEM Section

5 EVALUATION OF THE PROPOSED SYSTEM

The proposed framework is evaluated in terms of energy efficiency, detection accuracy and packet delivery rate using simulations and testbed implementations. OMNET++ is employed for simulation, and the newly introduced lightweight CNN is deployed in RPi for the testbed. Three different scenarios are compared. In the first scenario, only scalar sensors are used to detect a forest fire by measuring the temperature. If the measured temperature exceeds the specified threshold, then the sink node alerts the fire department. Note that the fire threshold may be specified based on geographical location and the monitoring season. The second scenario uses only multimedia sensors. The Raspberry Pi hardware and the pi camera are used as multimedia sensors. The images captured by RPis are transmitted to the sink node via intermediate multimedia sensors. Then, the fire department is alerted if a fire is detected. The last scenario is the hierarchical and heterogeneous system proposed in this study. Please note that for WSNs, the correct deployment strategy depends on the application considered, as discussed in References [3, 15]. In Reference [3], the results presented show that the uniform random deployment strategy provides better performance than the triangle grid, the hexagon grid and the tri-hexagon tiling (THT) deployment strategies in terms of coverage. Furthermore, in studies such as Reference [54], random scattering of sensor nodes is identified as one of the most common deployment strategies for forest fires, since the monitoring regions are considered in the category of large-scale open areas. As a result, for the three scenarios considered, it is assumed that the sensors are uniformly deployed to achieve maximum possible field coverage and connectivity between sensor nodes. The scalar sensors are, in turn, clustered around multimedia sensors. This assumption applies to simulation analysis as well as testbed-based experiments. Once deployed, the scalar sensors periodically (every 10 s to allow early observation) capture the data (temperature and humidity reads) to monitor the environment while multimedia sensors capture the data (image) when scalar sensors trigger them.

5.1 Simulation

The approach introduced in Reference [49] is used to calculate the energy consumption of scalar and multimedia sensors. It is possible to calculate the power consumption as (1) \[\begin{equation} P_{cons} = V I_m, \end{equation}\] where V is the supply voltage, \(I_m\) is the weighted average of current consumption, and \(P_{cons}\) is the consumed power. To calculate the weighted average of the current used, the states of the sensor nodes and the time fractions for each state are considered. A sensor node can be in one of three different states at a time. The idle state is the state where the sensor node neither receives nor transmits. In tx state, the sensor node transmits packets to other nodes in the communication range. In the rx state, the sensor nodes listen to the communication medium for incoming packets. In each state, there is a constant current consumption that is available in the data sheet of the sensor node. Then, the fractions of time for each node are calculated by experiments. The weighted average of current consumption is calculated as follows: (2) \[\begin{equation} I_m = \delta _{idle} I_{idle} + \delta _{tx} I_{tx} + \delta _{rx} I_{rx}, \end{equation}\] where \(\delta _{idle}\), \(\delta _{tx}\), and \(\delta _{rx}\) are time fractions for idle, tx, and rx states, respectively. Moreover, \(I_{idle}\), \(I_{tx}\), and \(I_{rx}\) illustrates current consumption for the idle, tx, and rx states. According to Reference [49], it is possible to assume that the sum of the time fractions is equal to one, which can be expressed as follows: (3) \[\begin{equation} \delta _{idle} + \delta _{tx} + \delta _{rx} = 1. \end{equation}\]

This energy consumption model is used in the simulations. The results of the simulations are presented comparatively with the results obtained from the testbed implementation.

The duration of the simulations is different for each of the scenarios explained (scalar only, multimedia only, our approach), since the simulation runs until the first node dies. Sensors are deployed uniformly on a 50\(\times\)50 \(\text{m}^2\) square field, and there are six nodes in total, including the sink node for each scenario. The scalar sensors sense the environment once in every 10 s to allow early observation of potential calamities. The XM1000 packet size is 200 bytes to transmit the temperature value, while the size of the packets transmitted by RPi is 2 MB, since the captured images are communicated. The initial energy of each node is 30,780 joules for XM1000 sensor node, which corresponds to the initial energy of two AA batteries, and 234,000 J for RPi, which corresponds to the initial energy of 13,000 mAh power bank. The output power tx of the nodes in each scenario is given as 0 dBm.

To calculate the power consumption of the XM1000, the radio duty cycle is calculated using the approach introduced in Reference [49]. The supply voltage and the mean current consumed by the sensor are calculated. To calculate the mean drawn current, experiments are conducted on the XM1000 mote using “powertrace” library in Contiki OS and the time fractions for each state are extracted. Please note that in scalar-only scenario, since the transmission of packets is periodical, the time fractions for idle, tx, and rx states are not fire scenario dependent. The mean drawn current and the time fractions for idle, rx and tx states are computed as 2 \(\text{mA}\), 17.4 \(\text{mA}\), 18.8 \(\text{mA}\), and 0.9619, 0.0075, 0.0306, respectively.

To compute the power consumption of the RPi, experiments are conducted using the available testbed. The results show that the RPi consumes 1.68 W power at the idle state (pi camera not activated) and 2.75 W power at tx/rx state. Additional experiments are conducted to obtain the fraction of time required and the amount of power consumed for image processing as well. The experimental results obtained show that multimedia sensors consume additional 0.12 W power while processing images with proposed lightweight CNN. Multimedia nodes process packets of around 2 MB once to decide whether there is a fire or not. The proposed lightweight CNN requires 0.3 s to process an image.

In the proposed framework, scalar sensors continue to periodically measure the environment every 10 s and they last much longer than the RPi multimedia node. RPi continuously listens the serial port to receive notifications. If it is triggered by a high temperature and low humidity reading, then the pi camera is activated to capture an image. As a case study, the data on forest fires in Cyprus are used to find the fraction of the times the pi camera is activated. The data available in Reference [12] includes a number of forest fires for the period of 2000–2017. Using the average number of forest fires, it is possible to calculate the fraction of time of the occurrence of a fire during the lifetime of RPi. The number of fire events that can occur during the lifetime of RPi is used in the configuration files of OMNET++ Castalia framework to generate fire events randomly for the specific simulation period considered. This way, since the number of events generated are taken from a real forest fire dataset, the number of times different functionalities of our framework are activated is as realistic as possible. The Castalia simulation tool [46] is used to simulate the three scenarios considered. Castalia is based on OMNET++ platform and it is generally used for the networks of low power sensor devices. It is possible to model different types of sensors in Castalia providing the corresponding configurations and protocols for each individual unit such as battery unit, communication unit, sensing unit, and processing unit. The XM1000 (scalar) and RPi (multimedia) sensors are modeled in the Castalia simulator providing appropriate configurations for each unit. Additionally, simulation results are presented along with actual benchmark results comparatively for validation.

In all simulation scenarios, the sensors are uniformly deployed in the fields considered. In both only-scalar and only-multimedia scenarios, all sensors are homogeneous, with the exception of one sensor selected to act as the sink. Packets are sent periodically in both scenarios, while scalar sensors send temperature and humidity readings and multimedia sensors send captured images. In the framework scenario proposed with the hierarchical setup, the sink node is a multimedia sensor and the source nodes are scalar sensors. The system is event driven, where scalar sensors take temperature and humidity measurements and require transmission only in case the readings are not within the pre-specified threshold values. Scalar sensor transmissions trigger the multimedia sensor to capture an image and use the proposed lightweight CNN model for forest fires detection. The packet reception rates are also presented using the OMNET++ simulation, which keeps the records of numbers of packets transmitted, received and dropped.

When the first scenario (scalar sensors only) is considered using simulation implemented with the XM1000 sensor specifications, the lifetime (the first node dies) of the network is on average 43.55 days with two AA batteries. Since only scalar data is transmitted in this scenario, the system works well in terms of energy efficiency.

The results of the simulations for the second (only multimedia sensors) and third scenarios (proposed framework) show that the lifetime of the system considered in the second scenario is approximately 11.16 h, whereas our proposed model has a lifetime of about 15.93 h with a power bank of 13,000 mAh. According to the simulation results, our proposed framework is 29.94% more efficient than the scenario in which only multimedia sensors are used. The heterogeneous nature of the proposed framework in which the scalar sensor readings are used to trigger the camera dramatically improves the lifetime of the multimedia nodes. In addition, the introduction of a lightweight CNN model that enables edge computing on RPi reduces the unnecessary transmission of video data and increases the system efficiency. Furthermore, additional simulation experiments show that using 33, 100, and 200 Ah batteries (each one of 12 V), it is possible to achieve the lifetimes of 4.70, 14.25, and 28.50 days, respectively.

Please note that the false positive information from scalars where the multimedia sensors are switched on although there is no fire, can shorten the lifetime of the system. The effect of these false alarm rates on the lifetime is studied in Figure 5. The lifetime of the proposed system is shown for systems using 200 Ah batteries with different rates of false alarms. Scenarios are considered for cases where \(0\%\), \(5\%\), \(10\%\), \(15\%\), \(20\%\), and \(50\%\) of the incoming notifications to activate the cameras are incorrect. In addition, the scenario where only multimedia sensors are employed is illustrated for a comparative analysis. The results presented in Figure 5 clearly show that the proposed system outperforms multimedia scenario even when 50% of the camera activation notifications are incorrect. This is reasonable, since the camera is always-on in the multimedia scenario. However, in the proposed system, the camera does not operate unless it is triggered by an event. Please also note that the values used for threshold can also affect the false alarm rates. Therefore, the threshold values should be decided according to the temperature and humidity profile of the environment that will be monitored. In this work, we have conducted experiments considering Cyprus as a case study, and we have used available forest fire, temperature, and humidity data [12] to specify the threshold values.

Fig. 5.

Fig. 5. Effect of False Alarm Rate on Lifetime.

Instead of sending all the messages to the base station for decision making, the proposed system uses edge computing, which can reduce the overall traffic load significantly. Therefore, QoS metrics should also be considered when evaluating the system. Instead of sending all the messages to the base station for decision making, the proposed system uses edge computing, which can reduce the overall traffic load significantly. Therefore, QoS metrics should also be considered when evaluating the system.

Since the unnecessary packet transmissions are avoided, the channel traffic decreases and the packet loss rate is reduced. Thus, by applying heterogeneity, machine learning and edge computing, the proposed system becomes more accurate, more energy-efficient and more effective in terms of QoS compared to the existing systems. The scalability of the proposed system is studied in Figure 6. The lifetime and the successful packet reception rate of the multimedia nodes are presented as functions of the number of nodes considered in each region. The scenario in which only the multimedia sensor nodes are used is considered, as well as the scenario with the proposed framework.

Fig. 6.

Fig. 6. Lifetime and packet reception rate of the proposed framework.

For forest fire detection applications, successful delivery of transmitted packets is very critical. The results clearly show that the proposed approach is suitable for emergency applications, since the reception process is more reliable and energy efficient.

5.2 Testbed Implementation

The three scenarios considered for the energy efficiency assessment are used to conduct benchmarking experiments. The main purpose of the testbed is to show the effectiveness of the proposed approach comparatively and to validate the simulation results presented for the lifetimes of the sensor nodes. The detection accuracy is considered in the section on the evaluation of the proposed deep learning model. The testbed-based evaluation approach employed is similar to the testbed based scenarios (referred to as environment simulation) presented in studies such as References [10, 16, 33, 51]. A small-scale representative is deployed for development and testing purposes. Forest fire warnings are injected to the system by considering forest fire probability similar to the approach used for the simulation to show that with close numbers of forest fire notifications, the simulation and the testbed results confirm each other in terms of energy efficiency. Figure 7 shows the scalar and multimedia sensors used in the testbed implementation. The multimedia sensor with the attached scalar sensor is the sink node as a whole in our setup. The scalar sensor attached is not used for sensing. Instead, it is used as a communication interface to collect incoming packets from the scalar sensors to the sink.

Fig. 7.

Fig. 7. Experimental setup.

To monitor the voltage levels, a library “battery-sensor” is used in Contiki OS. According to the results of the testbed, when the first scenario is considered (only scalar sensors), the XM1000 sensor nodes discharges two AA batteries in 42.50 days, which is very close to the results of our simulation computed as 43.55 days. The discrepancy is less than 3.5%. For the second scenario, (only multimedia sensors), the lifespan of the multimedia sensors is approximately 11.88 h, which is very close to the simulation results obtained as 11.16 h. The difference between the simulation results and those of the testbed is this time less than 6.1%.

The power sources used for the proposed framework are the same as those of previous experiments for scalar and multimedia sensors. Voltage levels are monitored to specify the lifetime of the nodes. The testbed validates the simulation results for the third scenario as well. While the lifetime of the proposed model is 15.93 h in the simulation, it is 15.66 for the testbed. The difference between the results is less than 1.8%. Results obtained for all three of the scenarios can be summarized as presented in Table 2.

Table 2.
ScalarMultimediaProposedFramework
Testbed42.50 days11.88 h15.66 h
Simulation43.55 days11.16 h15.93 h
Discrepancy\(\lt \!3.5\%\)\(\lt \!6.1\%\)\(\lt \!1.8\%\)

Table 2. Summary of Results

The evaluation carried out using the testbed assumes that the deployment of the sensors is performed in a way to allow the framework to function as effectively as possible. However, please note that although similar assumptions are used for testbeds in some studies, such as References [10, 16, 33, 51], the correct deployment of the scalar and multimedia sensors is very important for all the frameworks proposed for wildfire detection. The results presented comparatively for the testbed employed, and the simulation shows a good agreement with less than 3.5%, 6.1%, and 1.8% discrepancies for scenarios with scalar sensors only, multimedia sensors only, and the proposed framework, respectively. The results obtained from the testbed implementation also show that systems based on scalar sensors consume less energy than multimedia sensors. However, as stated in many studies dealing with disaster scenarios, we live in a time where multimedia data can be processed autonomously for much greater accuracy, especially when the correct classification is essential. However, since the proposed framework uses the efficiency of scalar sensors together with the accuracy of multimedia sensors, the power consumption of the system is reduced by about 29% when compared to the solutions that are solely based on multimedia sensors.

5.3 Evaluation of the Proposed Deep Learning Model

The new dataset that contains the forest images with and without fire is used to evaluate the proposed lightweight CNN model. The model proposed is created by testing different hyperparameter settings with the consideration of the model’s lightness. The best performing CNN model proposed has convolution with three filters in the first and 32 filters for the rest of the layers. To ensure that the model presented is not overfitting, 10-fold cross validation is performed. In the preliminary tests (a single pass with mixed train-test split), the model reaches to an accuracy of 99.12%. For single pass and 10-fold cross validation processes, stochastic gradient descent is used as an optimizer. RELU is used as an activation function. Each layer contains a dropout at the rate of 0.25. The model is trained for 100 epochs with a learning rate of 0.01 using cross-entropy loss, and early stopping with patience of 10 epochs. In other words, if the model no longer improves for 10 consecutive epochs, the training process is terminated and latest version of the model is saved.

The representation used for the captured images is critical for the whole system as it directly affects the complexity of the model. Complexity, in turn, specifies the computational and storage resources required in smart sensors. Although the input sizes can be reduced, the detection accuracy must be at acceptable levels. In the proposed framework, the images are resized to 64 \(\times\) 64 \(\times\) 3 to reduce the computational cost. This reduction in the input size does not significantly affect the learning process, since the proposed framework mainly focuses on early detection. It is sufficient to use the input for classification as fire or non-fire. Other features needed for localization of the fire or for the detection of other factors such as the direction of the fire, are not necessary for early detection. While the input size is reduced as explained, the model proposed in this study comes with a detection accuracy of 98.28%. There are similar studies in the literature that use well-known models, such as References [21, 36, 43] (GoogLeNet architecture) with fire detection accuracies of around 92%, 94.39%, and 99%, respectively. Compared to these approaches, our proposed model works almost as accurately as the best performing model, but requires reduced amounts resources in terms of computation and storage.

When the characteristics of the model we propose are compared with the characteristics of the models used in existing machine learning-based fire detection systems in more detail, we see that only a few of the studies such as Reference [20] focus on forest fire detection. Studies such as References [21, 36, 43] mainly consider fire detection in various environments. Similarly, while Reference [20] uses SVM and KNN as machine learning algorithms, References [21, 36, 43] use CNN-based approaches. In Reference [21], RNNs are employed in addition to the CNN-based architectures. All of these studies mainly focus on detection accuracy with well-known architectures such as GoogLeNet. They do not consider the QoS or energy efficiency related issues of the underlying systems. Instead, in our proposed system, by using a hierarchical approach, we can improve the underlying infrastructure in terms of QoS and energy efficiency. The size of the employed machine learning architecture can be quite critical to make sure that we can use the presented architecture in various types of multimedia-enabled smart devices. The model size of the deep learning architecture used is provided only in Reference [43] as 238 MB. The CNN architecture proposed in this study comes with a model size of 1.4 MB. Therefore the improvement in terms of edge computability is quite significant.

To show that the proposed model is lightweight, it is compared to well-known CNN architectures in terms of float-point operations (FLOP) [6]. While the proposed model have less than 0.55 GFLOPs, the majority of well-known models have more than 1 GFLOPs in terms of computational requirements. For example, RESNET-50 has about 4 GFLOPs and VGG-13, 16, and 19 have more than 10 GFLOPs [6]. To further emphasize the efficacy of the proposed model, we have also tested the well known light weight models such as Sufflenet [40], Squeezenet [24], MNasnet [58], Alexnet [34], Mobilenet [53], Resnet [22], and Inception [57] using our forest image dataset. Table 3 shows the number of parameters, FLOPS, model size, and accuracy for each model comparatively with the model proposed in this study.

Table 3.
ModelNameParameters(G)FLOPS(G)Accuracy (%)ModelSize(MB)
Proposed0.00400.536298.281.40
Sufflenet0.02300.451698.819.02
Squeezenet0.01201.736830.54.89
MNasnet0.00440.928464.0517.33
Alexnet0.06115.017999.00238.68
Mobilenet0.00350.899297.5813.84
Resnet180.01174.783399.0045.72
Inception0.0272183.42599.72106.35

Table 3. Comparison of the Proposed Model with other Light Weight Approaches

When the results presented in Table 3 are considered in terms of the numbers of parameters, the MNasnet, and Mobilenet models are comparable to the proposed model. However, the proposed model is superior to these models in terms of accuracy, since Mnasnet and Mobilenet have accuracies of 64.05% and 97.58%, respectively, as opposed to 98.28% accuracy of the proposed model. In terms of the FLOPS, the only model comparable to our proposed model is the Sufflenet. Although the accuracy of the Sufflenet is slightly better than our proposed model, in terms of the model size, while our proposed model is around 1.4 MB, the the Sufflenet model is around 9 MB. When the model size is considered, our proposed model is significantly more lightweight than the rest of the models. In terms of size, the closest model to the one we propose is Squeezenet. However, its size is still more than three times the size of the proposed model (4.89 MB), and its accuracy is significantly low (30.5%).

The size of the model and the required computational resources are critical for the proposed framework as these are the main factors affecting the ability to perform early detection at smart end nodes. The use of deep learning architectures with various types of boards for edge computing is also becoming very popular [7, 23, 52]. However, one of the most significant limitations in using various boards to run lightweight deep learning algorithms is the availability of resources, especially in terms of memory requirements. Therefore, these devices are referred to as “Memory-Constrained Edge Devices” [7]. For example, STM32F2 series comes with up to 120 MHz CPU speed, 1 MB of Flash, and up to 128 kB of SRAM [29], whereas Arduino MEGA 2560 supports clock speeds up to 16 MHz, with 256 KB of flash program memory and 8 KB of SRAM [26]. We, therefore, think that all of the metrics presented for the model evaluation play crucial roles. In terms of FLOPs, the proposed approach is at least 1.6 times better than all other approaches except Sufllenet, which performs similar to ours. The proposed approach is superior to all other models in terms of model size, which can be a significantly limiting factor depending on the type of end devices being used. Our proposed architecture is more than three times smaller than the next smallest architecture (Squeezenet) and more than six times smaller than Sufflenet. Finally, in terms of accuracy, the best models are Inception, Resnet18, and Alexnet with accuracies of 99.72%, 99.00%, and 99.00%, respectively, as opposed to the accuracy of the proposed model, which is 98.28%. Although the accuracy of the proposed model is quite close to the highest accuracy, in terms of the number of parameters, FLOPS and model size, the proposed approach is significantly better especially compared to Inception, Resnet18, and Alexnet approaches. Please also note that for relatively deeper models such as Inception, images of size \(300\times 300\times 3\) should be used, since \(64\times 64\times 3\) images vanish during the process of the forward pass due to the pooling and padding operations.

We end this section with a summary of the performance of the proposed model in terms of accuracy. The results of the evaluation carried out show that with an accuracy of about 98.28%, the proposed model is superior to studies such as References [21, 30, 43, 56], which report accuracies of 94.39%, 95.45%, 95%, and 86%, respectively. The study presented in Reference [36] performs slightly better than the proposed architecture with an accuracy of 99%; however, it is primarily based on GoogLeNet, which is a 22-layers convolutional neural network. The well-known architectures Mnasnet, Mobilenet, Sufflenet, Inception, Resnet, and Alexnet provide accuracies of 64.05%, 97.58%, 98.81%, 99.72%, 99.00%, and 99.00%, respectively. In terms of accuracy, the proposed approach performs close to the approaches with the highest accuracies.

Skip 6CONCLUSION Section

6 CONCLUSION

In this study, an effective and energy-efficient framework is proposed for the detection of forest fires. Unlike existing studies, multimedia sensors employing deep learning models are used alongside scalar sensors, and detection is performed using the fusion of information at different levels of the system architecture. In addition, the efficiency of communication is improved by the introduction of edge computing for decision making. The proposed framework uses the efficiency of scalar sensors and the accuracy of multimedia sensors. Through the establishment of a hierarchical infrastructure, the new framework balances energy efficiency and detection accuracy and provides a sustainable emergency surveillance system.

The results of the simulation and testbed experiments are presented and discussed in a comparative way. According to the experimental results, an improvement of 29.94% is obtained for energy efficiency (up to 28.50 days). In addition, a new lightweight CNN model is proposed to enable processing on edge devices. According to the results of the experiments, an accuracy of 98.28% is achieved. When the existing fire detection studies are taken into account, our approach is able to perform better than any other in terms of accuracy with the exception of the study included in Reference [36]. Although Reference [36] uses GoogLeNet, a 22-layer deep convolutional neural network, the performance of our approach is close to this architecture. Since the existing approaches do not take into account the limitations of edge computing, in this study, the existing lightweight architectures are also used in the experiments performed for comparison. The proposed approach performs very similar to the lightweight architectures with the best performance in terms of accuracy. However, the proposed approach’s main advantage is its suitability for edge computing on resource limited sensor nodes. To be able to use the CNN architecture in multimedia-enabled smart end devices, the model presented is set to be as light as possible in terms of the required computations and the model size. In terms of FLOPs, the proposed approach is at least 1.6 times better than all other approaches except Sufllenet, which performs similar to ours. The proposed approach is superior to all other architectures in terms of the model size, which can be a significant limiting factor depending on the type of end devices being used. The efficiency in terms of FLOPs and model size makes the proposed architecture the best candidate for smart nodes with limited capacities enabling early detection at edge nodes.

For future work, we plan to use our lightweight, high-accuracy detection mechanisms in other application areas. We believe that the lightweight architectures presented in this article can be practical and useful, especially in eHealth domain.

Footnotes

  1. 1 The new image dataset established for forest fire detection is publicly available at https://data.mendeley.com/datasets/g5nzp6j3bt/1.

    Footnote

REFERENCES

  1. [1] Ahmad Kashif and Conci Nicola. 2019. How deep features have improved event recognition in multimedia: A survey. ACM Trans. Multimedia Comput. Commun. Appl. 15, 2 (2019), 127. Google ScholarGoogle ScholarDigital LibraryDigital Library
  2. [2] Al-Kaff Abdulla, Madridano Ángel, Campos Sergio, García Fernando, Martín David, and Escalera Arturo de la. 2020. Emergency support unmanned aerial vehicle for forest fire surveillance. Electronics 9, 2 (2020), 260.Google ScholarGoogle ScholarCross RefCross Ref
  3. [3] Alablani Ibtihal and Alenazi Mohammed. 2019. Performance evaluation of sensor deployment strategies in WSNs towards IoT. In Proceedings of the IEEE/ACS 16th International Conference on Computer Systems and Applications (AICCSA’19). IEEE, 18.Google ScholarGoogle ScholarCross RefCross Ref
  4. [4] Aslan Yunus Emre, Korpeoglu Ibrahim, and Ulusoy Özgür. 2012. A framework for use of wireless sensor networks in forest fire detection and monitoring. Comput. Environ. Urban Syst. 36, 6 (2012), 614625.Google ScholarGoogle ScholarCross RefCross Ref
  5. [5] Bhosle Anand S. and Gavhane Laxmikant M.. 2016. Forest disaster management with wireless sensor network. In Proceedings of the International Conference on Electrical, Electronics, and Optimization Techniques (ICEEOT’16). IEEE, 287289.Google ScholarGoogle ScholarCross RefCross Ref
  6. [6] Bianco S., Cadene R., Celona L., and Napoletano P.. 2018. Benchmark analysis of representative deep neural network architectures. IEEE Access 6 (2018), 6427064277. https://doi.org/10.1109/ACCESS.2018.2877890.Google ScholarGoogle ScholarCross RefCross Ref
  7. [7] Capotondi Alessandro, Rusci Manuele, Fariselli Marco, and Benini Luca. 2020. CMix-NN: Mixed low-precision CNN library for memory-constrained edge devices. IEEE Trans. Circ. Syst. II: Express Briefs 67, 5 (2020), 871875.Google ScholarGoogle ScholarCross RefCross Ref
  8. [8] Chen Yanhong, Zhang Youmin, Xin Jing, Wang Guangyi, Mu Lingxia, Yi Yingmin, Liu Han, and Liu Ding. 2019. UAV image-based forest fire detection approach using convolutional neural network. In Proceedings of the 14th IEEE Conference on Industrial Electronics and Applications (ICIEA’19). IEEE, 21182123.Google ScholarGoogle ScholarCross RefCross Ref
  9. [9] François Chollet et al. 2015. Keras. Retrieved from https://fchollet.com/.Google ScholarGoogle Scholar
  10. [10] Civelek Muhsin and Yazici Adnan. 2017. Automated moving object classification in wireless multimedia sensor networks. IEEE Sensors J. 17, 4 (2017), 11161131.Google ScholarGoogle ScholarCross RefCross Ref
  11. [11] Cortez Paulo and Morais Aníbal de Jesus Raimundo. 2007. A data mining approach to predict forest fires using meteorological data. In Proceedings of the Portuguese Conference on Artificial Intelligence, J. Neves, M. F. Santos, and J. M. Machado. Springer, 512–523.Google ScholarGoogle Scholar
  12. [12] Department Forests. 2018. Forest Fire Statistics for 2000–2017, Ministry of Agriculture Rural Development and the Environment. Retrieved from http://www.moa.gov.cy.Google ScholarGoogle Scholar
  13. [13] Dunkels Adam, Schmidt Oliver, Finne Niclas, Eriksson Joakim, Österlind Fredrik, and Durvy Nicolas Tsiftesand Mathilde. 2011. The Contiki OS: The operating system for the internet of things. Retrieved from http://www.contikios.org.Google ScholarGoogle Scholar
  14. [14] Ergul Ozgur, Shah Ghalib A., Canberk Berk, and Akan Ozgur B.. 2016. Adaptive and cognitive communication architecture for next-generation PPDR systems. IEEE Commun. Mag. 54, 4 (2016), 92100. Google ScholarGoogle ScholarDigital LibraryDigital Library
  15. [15] Farsi Mohammed, Elhosseini Mostafa A., Badawy Mahmoud, Ali Hesham Arafat, and Eldin Hanaa Zain. 2019. Deployment techniques in wireless sensor networks, coverage and connectivity: A survey. IEEE Access 7 (2019), 2894028954.Google ScholarGoogle ScholarCross RefCross Ref
  16. [16] Ferdoush Sheikh and Li Xinrong. 2014. Wireless sensor network system design using raspberry pi and arduino for environmental monitoring applications. Procedia Comput. Sci. 34 (2014), 103110.Google ScholarGoogle ScholarCross RefCross Ref
  17. [17] Fu Jun-Song, Liu Yun, Chao Han-Chieh, and Zhang Zhen-Jiang. 2016. Green alarm systems driven by emergencies in industrial wireless sensor networks. IEEE Commun. Mag. 54, 10 (2016), 1621. Google ScholarGoogle ScholarDigital LibraryDigital Library
  18. [18] Fu Xiuwen, Fortino Giancarlo, Li Wenfeng, Pace Pasquale, and Yang Yongsheng. 2019. WSNs-assisted opportunistic network for low-latency message forwarding in sparse settings. Future Gen. Comput. Syst. 91 (2019), 223237.Google ScholarGoogle ScholarDigital LibraryDigital Library
  19. [19] Gaur Anshul, Singh Abhishek, Kumar Ashok, Kulkarni Kishor S., Lala Sayantani, Kapoor Kamal, Srivastava Vishal, Kumar Anuj, and Mukhopadhyay Subhas Chandra. 2019. Fire sensing technologies: A review. IEEE Sensors J. 19, 9 (2019), 31913202.Google ScholarGoogle ScholarCross RefCross Ref
  20. [20] Ghassempour Nargess, Zou Ju Jia, and He Yaping. 2018. A SIFT-based forest fire detection framework using static images. In Proceedings of the 12th International Conference on Signal Processing and Communication Systems (ICSPCS’18). IEEE, 17.Google ScholarGoogle ScholarCross RefCross Ref
  21. [21] Han JiSeong, Kim GwangSu, Lee ChanSeo, Han YeongKwang, Hwang Ung, and Kim SungHwan. 2019. Predictive models of fire via deep learning exploiting colorific variation. In Proceedings of the International Conference on Artificial Intelligence in Information and Communication (ICAIIC’19). IEEE, 579581.Google ScholarGoogle ScholarCross RefCross Ref
  22. [22] He Kaiming, Zhang Xiangyu, Ren Shaoqing, and Sun Jian. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on computer Vision and Pattern Recognition. 770778.Google ScholarGoogle ScholarCross RefCross Ref
  23. [23] Hu Pan, Im Junha, Asgar Zain, and Katti Sachin. 2020. Starfish: Resilient image compression for AIoT cameras. In Proceedings of the 18th Conference on Embedded Networked Sensor Systems. 395408. Google ScholarGoogle ScholarDigital LibraryDigital Library
  24. [24] Iandola Forrest N., Han Song, Moskewicz Matthew W., Ashraf Khalid, Dally William J., and Keutzer Kurt. 2016. SqueezeNet: AlexNet-level accuracy with \(50\times\) fewer parameters and < 0.5 MB model size. Retrieved from https://arXiv:1602.07360.Google ScholarGoogle Scholar
  25. [25] Illiano Vittorio P., Paudice Andrea, Muñoz-González Luis, and Lupu Emil C.. 2018. Determining resilience gains from anomaly detection for event integrity in wireless sensor networks. ACM Trans. Sensor Netw. 14, 1 (2018), 5. Google ScholarGoogle ScholarDigital LibraryDigital Library
  26. [26] Jackson Riley, Gresl Jonathan, and Lawrence Ramon. 2021. Efficient external sorting for memory-constrained embedded devices with flash memory. ACM Trans. Embed. Comput. Syst. 20, 4 (2021), 121. Google ScholarGoogle ScholarDigital LibraryDigital Library
  27. [27] Jan Mian Ahmad, Nanda Priyadarsi, He Xiangjian, and Liu Ren Ping. 2018. A sybil attack detection scheme for a forest wildfire monitoring application. Future Gen. Comput. Syst. 80 (2018), 613626. Google ScholarGoogle ScholarDigital LibraryDigital Library
  28. [28] Kadri Benamar, Bouyeddou Benamar, and Moussaoui Djillali. 2018. Early fire detection system using wireless sensor networks. In Proceedings of the International Conference on Applied Smart Systems (ICASS’18). IEEE, 14.Google ScholarGoogle ScholarCross RefCross Ref
  29. [29] Kajmakovic Amer, Diwold Konrad, Kajtazovic Nermin, Zupanc Robert, and Macher Georg. 2019. Flexible soft error mitigation strategy for memories in mixed-critical systems. In Proceedings of the IEEE International Symposium on Software Reliability Engineering Workshops (ISSREW’19). IEEE, 440445.Google ScholarGoogle ScholarCross RefCross Ref
  30. [30] Kaur Harkiran and Sood Sandeep K.. 2019. Fog-assisted IoT-enabled scalable network infrastructure for wildfire surveillance. J. Netw. Comput. Appl. 144 (2019), 171183.Google ScholarGoogle ScholarDigital LibraryDigital Library
  31. [31] Khan Ali. 2020. Dataset for Forest Fire Detection. https://doi.org/10.17632/GJMR63RZ2R.1.Google ScholarGoogle Scholar
  32. [32] Kochláň Michal, Hodoň Michal, Čechovič Lukáš, Kapitulík Ján, and Jurečka Matúš. 2014. WSN for traffic monitoring using raspberry pi board. In Proceedings of the Federated Conference on Computer Science and Information Systems. IEEE, 10231026.Google ScholarGoogle Scholar
  33. [33] Koyuncu Murat, Yazici Adnan, Civelek Muhsin, Cosar Ahmet, and Sert Mustafa. 2018. Visual and auditory data fusion for energy-efficient and improved object recognition in wireless multimedia sensor networks. IEEE Sensors J. 19, 5 (2018), 18391849.Google ScholarGoogle ScholarCross RefCross Ref
  34. [34] Krizhevsky Alex. 2014. One weird trick for parallelizing convolutional neural networks. Retrieved from https://arXiv:1404.5997.Google ScholarGoogle Scholar
  35. [35] Kumar R. and Rajasekaran M. Pallikonda. 2016. An IoT-based patient monitoring system using raspberry Pi. In Proceedings of the International Conference on Computing Technologies and Intelligent Data Engineering (ICCTIDE’16). IEEE, 14.Google ScholarGoogle ScholarCross RefCross Ref
  36. [36] Lee Wonjae, Kim Seonghyun, Lee Yong-Tae, Lee Hyun-Woo, and Choi Min. 2017. Deep neural networks for wild fire detection with unmanned aerial vehicle. In Proceedings of the IEEE International Conference on Consumer Electronics (ICCE’17). IEEE, 252253.Google ScholarGoogle Scholar
  37. [37] Li Jian, Yan Baoqiang, Zhang Mingjiang, Zhang Jianzhong, Jin Baoquan, Wang Yu, and Wang Dong. 2019. Long-range raman distributed fiber temperature sensor with early warning model for fire detection and prevention. IEEE Sensors J. 19, 10 (2019), 37113717.Google ScholarGoogle ScholarCross RefCross Ref
  38. [38] Li Yonggang, Liu Chunping, Ji Yi, Gong Shengrong, and Xu Haibao. 2020. Spatio-temporal deep residual network with hierarchical attentions for video event recognition. ACM Trans. Multimedia Comput. Commun. Appl. 16, 2s (2020), 121. Google ScholarGoogle ScholarDigital LibraryDigital Library
  39. [39] Liu Jiawei, Zha Zheng-Jun, Tian QI, Liu Dong, Yao Ting, Ling Qiang, and Mei Tao. 2016. Multi-scale triplet CNN for person re-identification. In Proceedings of the 24th ACM International Conference on Multimedia. ACM, 192196. Google ScholarGoogle ScholarDigital LibraryDigital Library
  40. [40] Ma Ningning, Zhang Xiangyu, Zheng Hai-Tao, and Sun Jian. 2018. Shufflenet v2: Practical guidelines for efficient CNN architecture design. In Proceedings of the European Conference on Computer Vision (ECCV’18). 116131.Google ScholarGoogle ScholarCross RefCross Ref
  41. [41] Molina-Pico Antonio, Cuesta-Frau David, Araujo Alvaro, Alejandre Javier, and Rozas Alba. 2016. Forest monitoring and wildland early fire detection by a hierarchical wireless sensor network. J. Sensors 2016 (2016).Google ScholarGoogle ScholarCross RefCross Ref
  42. [42] Muduli Lalatendu, Mishra Devi Prasad, and Jana Prasanta K.. 2018. Application of wireless sensor network for environmental monitoring in underground coal mines: A systematic review. J. Netw. Comput. Appl. 106 (2018), 4867.Google ScholarGoogle ScholarCross RefCross Ref
  43. [43] Muhammad Khan, Ahmad Jamil, and Baik Sung Wook. 2018. Early fire detection using convolutional neural networks during surveillance for effective disaster management. Neurocomputing 288 (2018), 3042. Google ScholarGoogle ScholarDigital LibraryDigital Library
  44. [44] Muhammad Khan, Ahmad Jamil, Lv Zhihan, Bellavista Paolo, Yang Po, and Baik Sung Wook. 2018. Efficient deep CNN-based fire detection and localization in video surveillance applications. IEEE Trans. Syst. Man Cybernet.: Systems 49, 7 (2018), 14191434.Google ScholarGoogle ScholarCross RefCross Ref
  45. [45] Muhammad Khan, Lloret Jaime, and Baik Sung Wook. 2019. Intelligent and energy-efficient data prioritization in green smart cities: Current challenges and future directions. IEEE Commun. Mag. 57, 2 (2019), 6065. Google ScholarGoogle ScholarDigital LibraryDigital Library
  46. [46] Ngo Khoa Anh, Huynh Trong Thua, and Huynh De Thu. 2018. Simulation wireless sensor networks in castalia. In Proceedings of the International Conference on Intelligent Information Technology. ACM, 3944. Google ScholarGoogle ScholarDigital LibraryDigital Library
  47. [47] Nguyen Linh and Nguyen Hoc T.. 2020. Mobility based network lifetime in wireless sensor networks: A review. Comput. Netw. 174 (2020), 1–24.Google ScholarGoogle ScholarCross RefCross Ref
  48. [48] OMNeT++. [n.d.]. OMNeT++, Discrete Event Simulator. Retrieved from https://omnetpp.org/.Google ScholarGoogle Scholar
  49. [49] Panousopoulou Athanasia, Azkune Mikel, and Tsakalides Panagiotis. 2016. Feature selection for performance characterization in multi-hop wireless sensor networks. Ad Hoc Netw. 49 (2016), 7089. Google ScholarGoogle ScholarDigital LibraryDigital Library
  50. [50] Parsons Andy, Goldman Russell, and Abbott Matthew. 2020. Why the Fires in Australia Are So Bad. Retrieved from https://www.nytimes.com/2020/01/01/world/australia/fires.html.Google ScholarGoogle Scholar
  51. [51] Ray Partha Pratim, Dash Dinesh, and De Debashis. 2019. Internet of things-based real-time model study on e-healthcare: Device, message service and dew computing. Comput. Netw. 149 (2019), 226239.Google ScholarGoogle ScholarCross RefCross Ref
  52. [52] Sanchez-Iborra Ramon and Skarmeta Antonio F.. 2020. TinyML-enabled frugal smart objects: Challenges and opportunities. IEEE Circ. Syst. Mag. 20, 3 (2020), 418.Google ScholarGoogle ScholarCross RefCross Ref
  53. [53] Sandler Mark, Howard Andrew, Zhu Menglong, Zhmoginov Andrey, and Chen Liang-Chieh. 2018. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 45104520.Google ScholarGoogle ScholarCross RefCross Ref
  54. [54] Sharma Vikrant, Patel R. B., Bhadauria H. S., and Prasad D.. 2016. Deployment schemes in wireless sensor network to achieve blanket coverage in large-scale open area: A review. Egypt. Inform. J. 17, 1 (2016), 4556.Google ScholarGoogle ScholarCross RefCross Ref
  55. [55] Snidaro Lauro, Visentini Ingrid, and Foresti Gian Luca. 2012. Fusing multiple video sensors for surveillance. ACM Trans. Multimedia Comput. Commun. Appl. 8, 1 (2012), 118. Google ScholarGoogle ScholarDigital LibraryDigital Library
  56. [56] Srinivas Kethavath and Dua Mohit. 2019. Fog computing and deep CNN based efficient approach to early forest fire detection with unmanned aerial vehicles. In Proceedings of the International Conference on Inventive Computation Technologies. Springer, 646652.Google ScholarGoogle Scholar
  57. [57] Szegedy Christian, Vanhoucke Vincent, Ioffe Sergey, Shlens Jon, and Wojna Zbigniew. 2016. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 28182826.Google ScholarGoogle ScholarCross RefCross Ref
  58. [58] Tan Mingxing, Chen Bo, Pang Ruoming, Vasudevan Vijay, Sandler Mark, Howard Andrew, and Le Quoc V.. 2019. Mnasnet: Platform-aware neural architecture search for mobile. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 28202828.Google ScholarGoogle ScholarCross RefCross Ref
  59. [59] Taylor-Coleman Jasmine. 2018. Wildfires Around the World: The Photos that Explain the Flames. Retrieved from https://www.bbc.com/news/world-44943143.Google ScholarGoogle Scholar
  60. [60] Toulouse Tom, Rossi Lucile, Campana Antoine, Celik Turgay, and Akhloufi Moulay A.. 2017. Computer vision for wildfire research: An evolving image dataset for processing and analysis. Fire Safety J. 92 (2017), 188194.Google ScholarGoogle ScholarCross RefCross Ref
  61. [61] Towers Fire Lookout. 2018. B.C. Fire Lookout Towers. Retrieved from http://www.firelookout.com/bc.html.Google ScholarGoogle Scholar
  62. [62] Varga András and Hornig Rudolf. 2008. An overview of the OMNeT++ simulation environment. In Proceedings of the 1st International Conference on Simulation Tools and Techniques for Communications, Networks and Systems and Workshops. ICS, 60. Google ScholarGoogle ScholarDigital LibraryDigital Library
  63. [63] Wang Denghui, Liu Jian, Yao Dezhong, and Member IEEE. 2020. An Energy-efficient distributed adaptive cooperative routing based on reinforcement learning in wireless multimedia sensor networks. Comput. Netw. 178 (2020), 107313.Google ScholarGoogle ScholarDigital LibraryDigital Library
  64. [64] Wang Linnan, Yang Yi, Min Renqiang, and Chakradhar Srimat. 2017. Accelerating deep neural network training with inconsistent stochastic gradient descent. Neural Netw. 93 (2017), 219229.Google ScholarGoogle ScholarCross RefCross Ref
  65. [65] Lee Wonjae, Kim Seonghyun, Lee Yong-Tae, Lee Hyun-Woo, and Choi Min. 2017. Deep neural networks for wild fire detection with unmanned aerial vehicle. In Proceedings of the IEEE International Conference on Consumer Electronics (ICCE’17). 252253. https://doi.org/10.1109/ICCE.2017.7889305. Google ScholarGoogle ScholarDigital LibraryDigital Library
  66. [66] Xie Lingxi, Wang Jingdong, Wei Zhen, Wang Meng, and Tian Qi. 2016. Disturblabel: Regularizing CNN on the loss layer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 47534762.Google ScholarGoogle ScholarCross RefCross Ref
  67. [67] Xu Yi-Han, Sun Qiu-Ya, and Xiao Yu-Tong. 2018. An environmentally aware scheme of wireless sensor networks for forest fire monitoring and detection. Future Internet 10, 10 (2018), 102.Google ScholarGoogle ScholarCross RefCross Ref
  68. [68] Yuan Chi, Liu Zhixiang, and Zhang Youmin. 2017. Fire detection using infrared images for UAV-based forest fire surveillance. In Proceedings of the International Conference on Unmanned Aircraft Systems (ICUAS’17). IEEE, 567572.Google ScholarGoogle ScholarCross RefCross Ref

Index Terms

  1. An Effective Forest Fire Detection Framework Using Heterogeneous Wireless Multimedia Sensor Networks

        Recommendations

        Comments

        Login options

        Check if you have access through your login credentials or your institution to get full access on this article.

        Sign in

        Full Access

        • Published in

          cover image ACM Transactions on Multimedia Computing, Communications, and Applications
          ACM Transactions on Multimedia Computing, Communications, and Applications  Volume 18, Issue 2
          May 2022
          494 pages
          ISSN:1551-6857
          EISSN:1551-6865
          DOI:10.1145/3505207
          Issue’s Table of Contents

          Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from [email protected].

          Publisher

          Association for Computing Machinery

          New York, NY, United States

          Publication History

          • Published: 16 February 2022
          • Revised: 1 June 2021
          • Accepted: 1 June 2021
          • Received: 1 November 2020
          Published in tomm Volume 18, Issue 2

          Permissions

          Request permissions about this article.

          Request Permissions

          Check for updates

          Qualifiers

          • research-article
          • Refereed

        PDF Format

        View or Download as a PDF file.

        PDF

        eReader

        View online with eReader.

        eReader

        HTML Format

        View this article in HTML Format .

        View HTML Format
        About Cookies On This Site

        We use cookies to ensure that we give you the best experience on our website.

        Learn more

        Got it!