Forest fire detection system using wireless sensor networks and machine learning
Forest fires are disasters that cause extensive damage to the entire world in economic, ecological, and environmental ways. These fires can be caused by natural reasons, such as high temperatures that can create spontaneous combustion of dry fuel such as sawdust, leaves, lightning, etc., or by human activities, such as unextinguished campfires, arson, inappropriately burned debris, etc1. According to research, 90% of the world’s forest fire incidents have occurred as a result of the abovementioned human carelessness1. The increase in carbon dioxide levels in the atmosphere due to forest fires contributes to the greenhouse effect and climate change. Additionally, ash destroys much of the nutrients in the soil and can cause erosion, which may result in floods and landslides.
At earlier times, forest fires were detected using watchtowers, which were not efficient because they were based on human observations. In recent history and even the present day, several forest fire detection methods have been implemented, such as watchtowers, satellite image processing methods, optical sensors, and digital camera-based methods2, although there are many drawbacks, such as inefficiency, power consumption, latency, accuracy and implementation costs. To address these drawbacks, a forest fire detection system using wireless sensor networks is proposed in this paper.Wireless sensor networks (WSNs) are self-configured and infrastructure-free wireless networks that help monitor physical or environmental conditions and pass these data through the network to a designated location or sink where the data can be observed and analyzed3. Efficiency and low power consumption are the major advantages of a WSN. In the proposed detection system, wireless sensor nodes are deployed according to cellular architecture to cover the entire area with sensors to monitor temperature, relative humidity, light intensity level, and carbon monoxide (CO) level using a microcontroller, transceiver module, and power components. The power supply to the sensor node is provided using batteries as the primary power supply, and solar panels are used as the secondary power supply. These sensor nodes are specially designed with a spherical shape to withstand damage caused by environmental conditions as well as animals.The sensor readings for each parameter are checked with a preset threshold ratio and a ratio that is calculated continuously in the node in real time, and only the ratios that exceed the preset ratio are sent from the sensor node to the base station for further analytical processing. The network utilized for this transmission is in the architecture of tree topology considering facts such as low power consumption, reduced latency, less complexity, etc. Cluster heads are used in this network to gather data from several sensor nodes and pass them on to the base station or the gateway node. The gateway node is an interface that connects the network with the secondary analysis process.For the analysis process, a machine learning regression model was used along with threshold ratio analysis to enhance detection accuracy. For the training and testing process of the model, data were collected during the fire and no fire situations in different areas and under different climatic conditions. During the data collection process, 7000 data samples were collected, where a data sample included temperature, relative humidity, light intensity level, and CO level at a particular time. Eighty percent of the collected data were randomly used as training data for the model, and the remaining 20% were used as test data.If the outcome of the machine learning model indicates a fire in a specific area, a text message will be sent to the mobile phone numbers of the authorized officers in responsible units. As this process is designed with a minimum delay, the fire can be detected within the initial stage, and the responsible parties can take necessary actions in a shorter period, which will minimize the damage.Related workForest fire detection has been a focus of many researchers for the last decade because of increased forest fire case reports from all over the world due to severe damage to society and the environment. Many methods have been proposed to detect forest fires, such as camera-based systems, WSN-based systems, and machine learning application-based systems, with both positive and negative aspects and performance figures of detection. Due to the higher probability of accurate and early detection due to the use of multiple sensor sources and deployment of sensor nodes in areas not visible to satellites, wireless sensor networks have a more positive outlook, and they have become the more applicable technology in many fields4.Many researchers have focused on environmental parameters, such as air temperature, relative humidity, barometric pressure, sound, light intensity, soil moisture, and wind speed and direction, along with gases, such as CO, CO2, methane, H2, and hydrocarbons apart from smoke, to detect forest fire conditions by considering the variations in these parameters during a fire5,6, and sensors have been selected according to the range, sensitivity, power consumption, and cost7,8.As supplying power to a sensor node is a challenging task in forested areas, utilizing only battery options is difficult because they do not last long, and distributing power using a wire would require a higher cost to deploy in a large forest. Therefore, many researchers have proposed solar-powered systems as secondary power sources along with rechargeable batteries as the main power source4,6, while some researchers have proposed solar batteries because they last longer9. To reduce the power consumption of sensor nodes, techniques such as keeping selected components active while others are deactivated have been proposed10,11,12.Most WSN-based detection systems are centered around a base station due to the memory and processing limitations of the nodes. Important and partially processed data are transferred to the base station through wireless media for processing and enabling relevant actions, while the base station also acts as the gateway between the sensor nodes and the system user4,9.When constructing a WSN, communicating data among the relevant entities is the main objective, and star topology and mesh topology-based networks have been proposed in many papers because of the different attributes in their systems. A mesh topology was chosen over a star topology because of its ability to self-organize, self-configure, and automatically establish among nodes in a network13. As a smaller number of nodes involved for transmission results in minimum energy consumption, concepts based on cluster heads have been used14. To minimize the loss of energy and data packages during transfer, a cluster-tree network topology structure was proposed15. Considering the sensing range of a node, fault tolerance, and energy consumption, a paper has proposed applying the on-demand k-coverage technique that provides event detection using static nodes with variable sensing ranges. This technique utilizes the maximum detection performance with the minimum power consumption for an event16. A survey on rare event detection has mentioned many event detection strategies that deliver maximized detection capability, minimized detection delay and low energy consumption, such as duty cycle, component deactivation, overpopulation/node redundancy, collaboration, and energy harvesting17.To reduce deployment cost and power consumption, a paper proposes a novel localization scheme that divides the whole forest area into different grids and allocates them to respective zones with another 8 neighboring grids. One centroid node from those grids, which is called the initiator node, predicts whether the zone is highly active (HA), medium active (MA), and low active (LA). Here, HA zones send data continuously to the base node through the interior node, MA zones send data periodically, and LA zones do not transfer data in the status that manages power consumption effectively18. Another author proposed obtaining data from the sensors every 2 min if there is the potential of a forest fire or obtaining data every 15 min otherwise to reduce the energy wastage19.To place sensor nodes in the most effective configuration to detect fire conditions, a sensor node was proposed at three different heights to perform different parameter measurements, while some authors have suggested covering sensor nodes to avoid direct sunlight exposure and minimize the false alarm rate4,10. As the network connectivity of service providers in forest areas is not robust, communication techniques that use dedicated network paths such as LoRa, ZigBee, and XBee have been used as the communication infrastructure. When considering attributes such as transmission range, high security, low power consumption (LPWAN protocol), and other relevant configurations, most papers have suggested using the LoRa module for transmission13,19.Most papers have suggested having threshold value-based fire detection on a sensor node, and if the exceeded threshold has remained the same, then a sink determines the location and will send an alarm to the fire department11. Because of the environmental parameter variations according to the place and time, threshold values are configured by the user considering geographic situations, climatic changes, seasonal changes, etc. after sensors obtain the data from the surrounding10.A fusion information process was proposed, where information from multiple sources is considered in making the final decision, which is better than using those sources individually, and two algorithms based on the threshold ratio method and Dempster-Shafer theory were used4. To enhance the detection accuracy, machine learning applications have been proposed in many papers based on different machine learning approaches, such as support vector machine (SVM) classification14 and regression techniques, such as logistic regression. However, applying machine learning techniques to fire detection systems has many limitations, such as the limited amount of energy, the energy required for data processing, the short range of communication and limited computations, the complexity of ML algorithms when executing on sensor nodes, and the difficulty of being distributed on every sensor node20,21. More
