A Review of Technologies for the Early Detection of Wildfires

According to a 2021 report by the World Meteorological Organization [1], natural hazards of all kinds increased fivefold between 1970 and 2020. On average, 115 people are killed per day, and $202 million in losses are incurred globally each day due to natural hazards.

Wildfires are among the most destructive natural hazards and have increasingly become a global problem. Their frequency and intensity are steadily growing worldwide (Fig. 1). The six worst wildfires in U.S. history happened in the past 7 years [5]. According to California Department of Forestry and Fire Protection, nearly 10,000 wildfires burned 4.2 million acres in 2020 [6], making it the worst wildfire year in California. However, 2020 wasn't an anomaly. In 2018, 8500 wildfires burned in California, resulting in $102.6 billion in damage, according to The Ecologist [7]. A 2024 CoreLogic report estimates that over 2.6 million homes across fifteen western and southern U.S. states are in areas of moderate to severe wildfire risk, representing a potential reconstruction value of $1.2 trillion [8]. According to data from the Global Fire Emissions Database and University of California Los Angeles (UCLA) researchers in 2020, CO₂ emissions from California wildfires were 25% higher than the state's annual fossil fuel emissions, wiping out over 15 years of efforts to reduce emissions [9]. The Journal of Environmental Research & Public Health reports that between 2008 and 2012, 1500–2500 premature deaths were caused by short-term exposure to PM2.5 particles released by wildfires, and long-term exposure resulted in 8700–32,000 deaths [10].

Globally, both Canada and Greece experienced their worst wildfire seasons on record in 2023. According to a March 2024 report by Inger Anderson, Executive Director of the UN Environment Programme, “The Mediterranean basin is rapidly turning into a tinderbox, and firefighting alone will not be enough to protect it in the long run” [11]. A July 2021 report by the Washington Post stated, “Australian wildfires had a bigger impact on global climate in 2020 than COVID lockdowns” [4]. During the 2019–2020 Australian wildfire season, 42 million acres burned, and 3.5 billion animals were killed or harmed. Smoke from the fires entered the stratosphere, circled the globe, and reflected sunlight into space, causing a cooling effect, particularly over the Southern Hemisphere. The Amazon, which contains 40% of the world's remaining tropical forests, plays a critical role in the atmospheric carbon sequestration and regulating the earth's climate. At the current rate of deforestation, Amazon forest loss is predicted to reach 21%–40% by 2050, primarily driven by wildfires [2,12]. In addition, wildfire-related air pollution causes 1.5 million deaths annually, with over 90% occurring in developing countries [13].

A fire progresses through four stages [2]. The first stage is the incipient stage, where the ignition has occurred, but the fire has not yet spread. Figure 2 shows a typical heat curve throughout the life of a fire. This phase may begin as a smoldering fire, with burning fuels that typically have a high surface-to-volume ratio. In the second stage, known as the growth stage, the fire becomes self-sustaining, exhibiting rapid expansion, and beginning to consume fuels with larger volumes. At this stage, the fire's growth rate can become exponential, depending on meteorological conditions and fuel availability. The heat generated evaporates any remaining moisture in the surrounding fuel, further accelerating combustion and producing even more heat in a positive feedback loop. The third stage is the fully developed stage of fire, generating the maximum heat and burning all available fuel. The final stage is where the fire diminishes as it begins to run out of fuel.

As a wildfire progresses through various stages, appropriate suppression strategies must adapt, as they depend on the wildfire rank and stage (Fig. 3). The most cost-effective and least risky suppression method is to contain the fire during the incipient stage (either Rank 1 or Rank 2 in Fig. 3 diagram). The incipient stage begins when heat, oxygen, and a fuel source combine to initiate ignition. This initial phase is often characterized by smoldering or a small flame that may ignite nearby combustible materials. Incipient fire can be controlled or extinguished using portable firefighting equipment. It is characterized by small flames usually less than four feet in height, its temperature is lower than in later stages, and visibility remains relatively unaffected by smoke. Professional firefighters can suppress incipient fires using handheld equipment. Additionally, constructed control lines and areas cleared of combustible material will be effective in limiting the fire's growth and spread. Once a wildfire reaches Rank 3, and especially Rank 4, ground crews conducting direct attacks will require air support from fixed-wing air tankers, skimmers, or helicopters conducting bucketing or tanking operations. At these stages, the ability to control wildfire damage caused depends on various environmental factors such as wind speed and direction, availability of fuel and vegetation, humidity, and terrain. Hence, the containment is not guaranteed before significant damage occurs. Table 1 [15] demonstrates the relationship between flame length (which increases with the fire's Rank) and possible suppression strategies. Depending on the environmental factors outlined in Table 2 [15], a wildfire can escalate in rank within minutes, exhibiting exponential growth. As with many other hazardous physical phenomena that display rapid, non-linear growth patterns, early detection remains the safest and most reliable approach to containment.

As illustrated in Fig. 4 [16], the physics of wildfire growth are exponential rather than linear. Any reduction in response time during initial wildfire suppression can yield similarly exponential benefits. Under hot, dry, and/or windy conditions, every second is critical for successful wildfire mitigation. As Fig. 4 shows, even during a typical day in Southern California (this is not an extreme scenario), a wildfire with 10-foot flames can spread to 3 acres within 5 min and 10 acres within 10 min. This underscores the importance of detecting a wildfire during its incipient stage which is extremely helpful and essential.

The most significant weakness and vulnerability in the fire response chain lies in the early detection, notification, and reaction phase. According to an August 2023 report from the Gordon and Betty Moore Foundation, “An investment that helps California fire officials achieve a 15-minute reduction in response time could be expected to generate $3.5—$8.2 billion in economic benefits, and $150—$350 million in fiscal benefits annually for California” [17]. In November 2018, the Camp Fire in California traveled 7.8 miles in the first 45 min after ignition, engulfing the town of Paradise, killing 86 people, and causing $19 billion in damage [18]. The majority of the damage and fatalities occurred within the first 4 h of ignition.

Over the past 30 years, the frequency and severity of wildfire events have steadily increased, driving a corresponding rise in related research, publications, and proposed solutions both in academia and in industry [16,19,20]. This trend became particularly noticeable following the devastating global wildfires of 2018–2020, which triggered a significant surge in wildfire-related publications. In general, wildfire research publications span the following areas [20]:

• Wildfire Prevention & Preparedness;

• Wildfire Detection & Notification;

• Wildfire Growth Prediction & Monitoring;

• Wildfire Suppression; and

• Wildfire Restoration & Adaptation.

As discussed, the motivation and focus of this work center on detection, in particular, the analysis of various methodologies, technologies, and published research that enable early detection of wildfires during the initial stage of growth. In the following sections, we provide a review of the various architectures and technologies currently used for wildfire detection and compare and contrast their strengths and weaknesses. In the conclusion, we summarize our conceptions and provide recommendations based on our analysis.

The focus of this paper was the evaluation of various technologies for early detection of wildfires. Broadly, detection methodologies fall into five categories (see Table 3).

Satellite-based wildfire detection primarily relies on specialized sensors on satellites to identify heat anomalies, typically in the infrared spectrum, indicating the presence of fire This enables early detection of wildfires even in remote areas or when obscured by smoke. Two main types of satellites are used for wildfire detection:

1. Polar Orbit Satellites such as the NASA-Terra [21], NASA-Aqua [22], NASA-Landsat 8 [23], and NOAA Joint Polar Satellite System [24]. Polar satellites scan the entire Earth several times a day; allowing for fire monitoring (Fig. 6). However, each consecutive scan over the same location occurs several hours apart, resulting in a low temporal sampling rate for these satellites.

2. Geostationary Satellites like the NOAA Geostationary Operational Environmental Satellite (GOES-R series) [25]. Geostationary satellites provide much higher temporal data for a specific area but are limited in their ability to monitor other global environmental developments.

The instruments onboard these satellites collect various types of data, some of which are actively analyzed to detect and observe wildfires, track the global transport of pollutants, and asses the long-term climate impacts of fires. The imaging instruments on satellites used for wildfire detection include the following:

• Moderate Resolution Imaging Spectroradiometer (MODIS) [26] records 36 spectral bands data between 0.4 µm – and 14.4 µm and monitors fires and hot spots through visible and IR imaging (Fig. 5). It is carried by the NASA Terra and Aqua satellites.

• The Advanced Baseline Imager (ABI) [28] is the primary instrument on the GOES-R Series for imaging Earth's weather, oceans, and environment. ABI observes the Earth with 16 different spectral bands, including two visible channels, four near-infrared (NIR) channels, and ten infrared channels. Models and tools use these other channels (wavelengths) to identify various elements on the Earth's surface or in the atmosphere, such as trees, water, clouds, moisture, or smoke.

• Visible Infrared Imaging Radiometer Suite (VIIRS) [29] spans visible and infrared wavelengths, covering 22 channels between 0.41 and 12.01 µm. Similar to MODIS, VIIRS gathers data at a higher resolution and measures aerosol, among other parameters. It is carried by the Suomi National Polar-Orbiting Partnership (NPP) satellite.

• The Landsat 8 satellite payload consists of two science instruments [23]—the Operational Land Imager (OLI) and the Thermal Infrared Sensor (TIRS). These two sensors provide seasonal coverage of the global landmass at a spatial resolution of 30 meters (visible, Near infrared, and Short wave infrared), 100 meters (thermal), and 15 meters (panchromatic).

All of the instruments mentioned above utilize specific infrared bands to maximize the likelihood of fire detection. More mature high-temperature flames have a radiation footprint mainly in the mid-wave infrared (MWIR), while early-stage fires produce a substantial footprint in the long-wave infrared (LWIR) band. Additional infrared channels enhance detection by reducing false alarms caused by clouds, bright surfaces, or other environmental factors [30]. Studies [30,31] have applied various approaches to data from multiple satellite instruments mentioned above, demonstrating increased sensitivity to accurate detection of smaller and lower-temperature fires. In recent years deep learning techniques have gained popularity in active fire detection. In particular, convolutional neural networks (CNN) have shown promise in detecting fires of different sizes, where multi-scale kernels were used for feature extraction [32].

Processing satellite data to detect wildfire anomalies and small fires presents a challenge due to the lower spatial resolution of satellite images. In addition, smoke can quickly appear identical to clouds, as shown in Fig. 8 of the Camp Fire, imaged by a MODIS on Terra.

The CubeSat project was started to supply more technology providers with more affordable access to satellite technology. CubeSats feature a standardized design and are launched as part of the payload on other satellite launches [33]. The first example of using CubeSat for wildfire detection was demonstrated by Gangestad et al. [34]. To enable onboard image processing—to eliminate the requirement of transferring images back to earth for processing which significantly slows down the process—Azami et al. proposed k-nearest and CNN Deep Learning (DL) models implemented on a Raspberry Pi [35] using a 6U CubeSat configuration.

The team achieved classification accuracy of better than 95% [35,36]. Another company leveraging the CubeSat architecture is Orora Technologies, which designs 3U CubeSats configured with proprietary infrared (IR) cameras for detecting new wildfires and monitoring existing wildfires.³ The Orora CubeSats have onboard GPUs to assist with wildfire detection without the need to download data to ground stations for processing.

Another emerging trend is constellation approaches, dedicated satellite networks designed explicitly for wildfire detection with high resolution and frequent revisit times. In May 2024, Muon Space,⁴ an end-to-end Space Systems Provider that designs, builds, and operates mission-tailored low-earth orbit (LEO) satellite constellations, in partnership with Earth-Fire Alliance (EFA),⁵ announced the FireSat Constellation. FireSat aims to detect a 5-meter by 5-meter fire anywhere in the world. The first phase of the FireSat Constellation, scheduled to be launched in 2026, will include three Muon Halo™ satellites [37] equipped with 6-band multispectral IR instruments designed and optimized for the wildfire mission. With the first three satellites, the FireSat Constellation will observe every point on Earth at least twice daily, with increased revisit frequencies for high-risk wildfire regions. Once fully operational with over 50 satellites, the revisit times for most of the globe improve to 20 min, with the most wildfire-prone regions benefitting from sampling intervals as short as 9 min.

Airborne-based wildfire detection platforms primarily use various technologies mounted on aircraft, including drones, to identify smoke or heat signatures from fires. This can enable early detection of wildfires from the air, often through thermal cameras that can operate both day and night, providing valuable information for rapid response efforts. This technology can be utilized on both manned aircraft and unmanned aerial vehicles (UAVs) to monitor large areas and pinpoint fire locations. While manned aircraft such as planes and helicopters cover vast areas and provide detailed imagery, drones (UAVs) offer excellent maneuverability and can access hard-to-reach areas for close-up fire monitoring.

As an example of manned aircraft, The Forest Service in Region 5 (California) has two Bell 209 (Cobra) helicopters equipped with FLIR thermal gimbals (Fig. 6) [38]. This technology allows the flight crew to see through smoke, report real-time fire progression, provide tactical IR capabilities to enhance situational awareness for ground resources', and keep incident management informed of critical holding issues. Recent advances in UAVs have led to innovative solutions to reduce response times to wildfires. In November 2024, Sikorsky, a Lockheed Martin company, and Rain, a developer of autonomous aerial wildfire containment technology, demonstrated how an autonomous Black Hawk® helicopter could be commanded to take off, identify the location and size of a small fire, and then accurately drop water to suppress the flames (Fig. 6) [39].

These helicopter-based systems are fairly expensive solutions for wildfire detection and suppression. Much lower-cost drones, comparable in size and cost to regular commercial drones, are also being used to combat wildfires. The main limitation of these smaller, low-cost drone systems is their relatively low-capacity power source which restricts flight time and the ability to search for fires. This limitation arises from structural constraints that prevent the use of heavy, energy-dense batteries. In another study [40], an optimization model was developed to minimize the total amount of energy used by the drone while maintaining the required levels of data quality for transmitting IR feeds to a base station.

UAVs are usually outfitted with various cameras (videos, IR, and imaging) but they may also include additional features depending on the designer [41,42]. For example, some systems are capable of controlling fleets of UAVs. Chi Yuan et al. discuss a system where fleets of different types of semi-autonomous drones are deployed in stages to search, confirm, and observe fires [43]. Although each drone may not be able to handle dynamic and complex tasks like piloted aircraft, the autonomy allows for quick commanding and organization within a fleet regardless of size. In [44], four decentralized algorithms are evaluated to optimize how a swarm of UAVs can cover a large area and detect wildfires. The algorithms were tested with swarms of different sizes to test the spatial coverage of the system in 24 h of simulation time. The simulations showed that the best algorithm can detect 82% of fires using only 20 UAVs and when the swarm consists of 40 or more aircraft; 100% coverage can also be achieved.

In other approaches, drones may be equipped with additional devices, such as a specialized fire sensor or a tank of water [45]. The IGNIS system by Drone Amplified (Fig. 7) is an example of an advanced drone utilizing several features [47]. IGNIS combats wildfires by deploying small, ping-pong ball-sized “ignition spheres” that contain a chemical payload and are dropped from the drone at strategic locations to create controlled fires, which then deprive the wildfire of fuel and prevent its spread by creating firebreaks in advance of the main blaze; this allows firefighters to set controlled burns in hard-to-reach areas while maintaining a safe distance. The drones can be controlled either via a remote controller and/or through preset instructions sent by a mobile app.

Various Machine Learning algorithms have been proposed to process the video feed from airborne systems. Lewicki and Liu designed a Fire Perception Box (FPB) with RGB/IR cameras and an ARM microcontroller can be installed on UAVs in a plug-and-play manner [48]. At a suspected fire scene, the RGB image is first fed to a CNN classifier to calculate an RGB fire score, followed by the IR image being fed to its corresponding pipeline to calculate an IR fire score. Both scores are combined to establish the presence of fire and subsequently find the fire localization hotspots and RGB + IR heatmaps. Researchers have also demonstrated a large-scale YOLOv3- and YOLOv5-based deep-learning network for fire identification from images and video feeds captured by UAVs [49,50]. After training the You Only Look Once (YOLO) network on annotated fire images, the researchers achieved high fire detection accuracies in real-time analysis of UAV video feeds. The main drawback of this method is that the YOLO model needs to be implemented on a ground-based high-performance computer; therefore, the performance of this system is contingent on uninterrupted data transmission between the base station and UAVs. Besides these methods, other deep-learning methods such as recurrent neural networks (RNN), long short-term memory neural networks (LSTM), generative adversarial networks (GAN), deep belief networks (DBN), etc., have also demonstrated good accuracy in identifying wildfires [51,52].

Camera-based systems for wildfire detection utilize networks of strategically placed cameras, often equipped with thermal imaging capabilities, to monitor large areas for signs of smoke or heat. These systems allow for early detection of wildfires by capturing visual data and analyzing it with AI algorithms to identify potential fire outbreaks in real-time; enabling rapid alerts to authorities for a quick response to emerging fires. The most widely deployed example of fixed camera-based solutions, ALERTCalifornia,⁷ has more than 1080 high-definitions, pan-tilt-zoom cameras deployed across California (as of June 2024). This provides a 24-hour backcountry network with near-infrared night vision to monitor disasters such as active wildfires (Fig. 8). ALERTCalifornia cameras can perform 360-degree sweeps approximately every 2 min and can view as far as 60 miles on a clear day and 120 miles on a clear night. Live feeds of these cameras were made publicly available in 2023. ALERTCalifornia collaborated with the California Department of Forestry and Fire Protection (CAL FIRE) and industry partner Digital Path to create a fire detection AI tool to improve firefighting capabilities and response times. When the AI spots a potential fire on ALERTCalifornia's network of cameras, the system alerts firefighters, providing a percentage of certainty and estimated location for the incident. If the incident is vetted and confirmed by trained watch-standers, firefighters respond quickly to extinguish the fire at the incipient phase.

Some other examples of AI-enabled fixed camera solutions designed explicitly for wildfire detection are PanoAI⁸ and IQ-Firewatch.⁹ Several energy and utility companies have announced deployments of PanoAI camera solutions.

Various Machine Learning algorithms are employed to increase the accuracy of detection and minimize false alarms. The details of the algorithms for each vendor are considered proprietary trade secrets. However, the general approach usually involves smoke detection—with the classical approach of feature extraction followed by some form of classifier. Jie Shi et al. reviewed such algorithms and identified their standard extracted features: color, motion, texture, shape, and energy [53]. References [42] and [53] discuss various classifiers typically used in Machine Learning, such as Neural Networks, Logistic Regression, K Nearest Neighbors (KNNs), or Support Vector Machines (SVMs). Usually, the model's output by these algorithms is trained and tested on past fire data offline before deployment. Large data sets of images with and without fire might be needed depending on the type of algorithm used. This allows for continuous improvement of the model even after deployment. For example, in Ref. [54], Darko Stipaničev et al. released improved models for their iForestFire semi-automatic monitoring system even after its commercial deployment and operations in various regions.

The concept behind low-power sensor-based solutions that require no existing infrastructure is to deploy them in areas with a higher risk of ignition (Fig. 9). This increases the likelihood of detecting wildfires at an earlier and smaller stage compared to satellite-based or regional camera-based solutions. This approach is motivated because over 90% of catastrophic fires are man-made. In many cases, firefighting resources are aware of regions of high ignition risk in many scenarios. Thus, if sensors are deployed with enough density to cover those areas, they should be able to detect wildfires during their incipient stage (which would be close to impossible for satellite or regional camera-based solutions).

The general architecture of these sensor networks can be divided into four layers (Fig. 10) [56], and design decisions at each level involving various tradeoffs:

1. Physical Layer: the actual sensors and surrounding electronics, including wireless modem, designed to be very low power.

2. Network Layer: the wireless link can be Cellular(3G/4G/5G), LoRa, BLE, WiFi Halow, or direct to Satellite.

3. Middleware Layer: impacts communication architecture between physical sensors (mesh versus perimeter versus client/server) as well as cloud software.

4. Application Layer: any AI models to process sensor data, dashboards, etc.

In recent years, several vendors have introduced solutions that incorporate gas or smoke sensors capable of “sniffing” a wildfire, potentially detecting a theme at an early stage. Companies such as N5 Sensors,¹⁰ Dryad Networks,¹¹ and LAD Sensor¹² have developed small-form-factor low-power solar solutions (Fig. 11), enabling them to be deployed at any location. Delphire¹³ has designed a heterogeneous multi-sensor solution with the possibility of integrating a camera, thermal, and chemical sensor, providing richer data. However, this comes with a larger form factor (Fig. 11) and higher power consumption than other solutions. EverSense¹⁴[42] claims they can harness energy from the fire, thus preventing the need for any battery or solar panels. By providing a wireless backhaul, these solutions can give rapid notification when a fire is detected. Some solutions have direct satellite capability, but most adopt LoRaWan¹⁵ as their backhaul. While very low power, LoRa supports very low bandwidth data <22 Kbps, and as such, only non-image data can be transmitted. By leveraging AI, models within these sensor networks have been trained to minimize false alarms (e.g., differentiate between a barbecue fire or a car's exhaust and a wildfire). Potential downside of gas-based “sniffing” solutions is their susceptibility to wind patterns, which can influence the direction of wildfire smoke. Especially in the case of small incipient wildfire, there is not yet enough smoke to saturate the region, and the only way these sensors can capture a wildfire is a high deployment density of such sensors, which can be costly and operationally prohibitive.

Various algorithms have been developed to analyze sensor data and distinguish between wildfire and non-wildfire conditions. Varela et al. [57] proposed a simple algorithm to establish base models for temperature and humidity during fire events. Sensor nodes data nodes can be compared against these base functions to determine the occurrence of an actual fire. Khalid et al. used data on flame, smoke, temperature, humidity, and light intensity collected from a base station and applied it to a binary Bayesian classifier achieving a fire detection accuracy of 97.2% [58]. Yan et al. used an artificial neural network (ANN) to identify the combustion phase in real time using CO2, air temperature, and smoke sensors [59]. Their model considered three combustion phases: no fire, smoldering combustion, and flaming combustion, and obtained a consistent accuracy of >82% when multiple sensor input data were considered. Abbassi et al. used a fusion of heterogeneous sensors, smoke, heat, and flame, across a wireless sensor network. They implemented a multi-level alert system using a combination of KNN classifier and fuzzy logic [60]. The first level alert notifies of a fire initiation based on a single sensor measurement supported by collaboration with neighboring sensors using KNN. The second alert estimates the spread of the fire based on aggregated intra-cluster data. Dampage et al. captured humidity, temperature, light intensity, and CO level from sensors under different climate conditions and calculated ratios between current data values and those recorded 30 s ago to indicate a fire. A multiple regression model was used to reduce false alarms [61].

In recent years, end-to-end systems have also been proposed that combine spaceborne, airborne, and terrestrial solutions. In one study [62], a ground-based sensor network monitors temperature, humidity, and wind speed across multiple sites. Octocopter drones are deployed for inspection when the ground-based sensors detect conditions conducive to wildfires, such as hot, dry weather and strong winds. These drones are equipped with specialized packaging tubes for collecting smoke samples and chemical sensors to provide real-time data on CO₂ and smoke particles. This allows for the mapping of chemical fire signatures in complex landscapes before visible detection. In [63], an autonomous and scalable monitoring system architecture is proposed for early detection and spread estimation of wildfires by leveraging low-cost UAVs, satellite data, and ground sensors (Fig. 12). The Internet of Things (IoT) sensors play a vital role in the system as they act as a low-maintenance and cost asset that can cover large areas once carefully placed. Through high levels of autonomy and swarming principles, UAVs will provide on-demand area coverage to identify wildfires as early as possible and receive real-time feedback as the wildfire progresses. Additionally, the constellation of LEO satellites will enable communication exchange beyond the visual line of sight, and through their finer ground sampling distance and hyperspectral sensors, they can assess vegetation health, which results in proactive wildfire management strategies and post-fire rehabilitation efforts.

SenseNet¹⁷ combines gas sensors, thermal imaging, camera streaming, and satellite imagery to enable early wildfire detection with early alerts, minimizing false alarms. SenseNet merges data from diverse sources such as sensors (their platform which can operate in a mesh mode or peripheral mode), cameras (they also offer their own fixed camera), satellites, historical data, weather stations, topography data, railways maps, campsites maps, city infrastructures maps, etc. to provide insights and ultra-early detection, wildfire monitoring, management, and decision making, to protect the communities and assets.

Table 4 provides an overview of the various wildfire detection architectures and tradeoffs associated with each.

Within this chain, the greatest vulnerability, and opportunity for improvement, is the “Notification” component, which is a function of “Detection”. It was outlined here why early detection of an incipient fire is crucial for reducing the “Response Time”. As a result, the majority of our research focused on technologies and approaches that enable wildfire detection, and we evaluated their ability towards detecting wildfires at their incipient stage.

The effectiveness of any wildfire detection solution depends on numerous parameters and providing generalizations can be controversial and inaccurate for any given solution and scenario. Any technology or product should be evaluated based on its own individual merits. However, generally speaking, certain tradeoffs exist between the different architectures outlined in this research.

For comparison purposes, we will be referring to the Rank of a fire as outlined in Fig. 3. While local ground-based sensors are the best solution for detecting wildfires within their incipient stage (Rank 1 or 2), once a wildfire reaches Rank 4, the sensors might have been destroyed. However, that is the stage of wildfire where a regional camera and/or airborne sensors will most likely detect and report a wildfire. If the fire reaches a Rank 6, it is likely so large, and there is so much smoke in the area that these cameras might not have much visibility. Aircraft and satellites might be best for monitoring wildfires and their smoke plumes. Additionally, satellites would be very useful for picking up wildfires in remote environments where no human might be around for tens or hundreds of miles. As such, there would be no reason to install local sensors or even regional cameras.

Each architecture, along with the types of sensors used within those architectures, has its strengths and weaknesses. Figures 13 and 14 provide radar charts we created, attempting to summarize the advantages and limitations of various approaches. As depicted in Fig. 13, terrestrial sensors are the best solution for early detection of incipient fires, their biggest limitation is limited coverage. This is due to the fact that operationally it is just not feasible to cover large remote complicated terrain with such architecture. Furthermore, the ongoing maintenance could be impractical. As such, sensors are the best choice where areas of high ignition risk are generally well known. Since 90% of very damaging wildfires are caused by humans (mostly unintentionally as accidents), an argument can be made that for many scenarios, sensors can provide adequate coverage. However, ultimately, airborne and spaceborne solutions are undoubtedly required when coverage of large and difficult terrain is necessary.

Referring to Fig. 14, the choice of the actual sensor also presents various tradeoffs. Ultimately, our research suggests that in order to maximize the probability of correct and early wildfire detection—while minimizing the false alarms—the optimal choice is to have a heterogeneous set of sensors looking simultaneously at different fire signatures (flame, smoke, heat). In addition, recent trends show a growing emphasis on “systems of systems,” where multiple architectures are integrated and communicate seamlessly to provide the best coverage and enable the fastest possible response.

Another major trend in wildfire detection has been the rapid adoption of AI and Machine Learning (ML), even DL in wildfire detection platforms. The background to this is that over the past decade, ML models, particularly Convolutional Neural Networks (CNN) have shown remarkable capabilities for powerful object detection and classification. Consequently, it was natural that researchers started exploring them to start detecting signatures of fires (smoke or flame). Recent publications suggest that this area of research has exploded, particularly after 2019/2020. As a result, many different ML and DL models and approaches have been proposed for wildfire detection. These models come in a variety of shapes and forms and have demonstrated the potential for providing dramatic improvements in the ability to detect wildfires while simultaneously lowering false alarms. Especially, in recent years, hybrid models that combine traditional CNNs with transfer learning and data augmentation, or with LSTM to deal with time-series data have shown much promise. Other examples are the real-time object detection platforms YOLOv3 and YOLOv5 which have shown excellent detection capabilities for real-world wildfire images. For example, YOLOv3-tiny was tested with wildfire images from a UAV and was able to detect small fires with high accuracy. Furthermore, by merging detection with classification, such as combining YOLO with CNNs, powerful fire monitoring capabilities can be offered. One of the main feedbacks from professional firefighting resources is the need to minimize false alarms, as each false alarm could lead to the deployment of valuable firefighting resources, especially in high-risk meteorological conditions. These resources could then be diverted from potential actual ignitions, ultimately, increasing the response times. As such, minimizing false alarms is crucial. The adoption of effective DL models could be an important way of minimizing recall while maintaining sensitivity.

Advances in DL with improvements in remote sensing technology, and high-performance computation and low-power hardware platforms, have enabled the real-time execution of DL algorithms on the edge of the network, in platforms such as UAVs and cameras and sensors deployed in the field. These trends are crucial toward achieving the objective of lowering response times, as they do not rely on information being sent back to the cloud and processed at a later time in the cloud. This is especially important in scenarios where the system is operating in remote, unpredictable, and fragile environments, and cloud connectivity, especially for high-volume image data, can be unpredictable and unreliable.

A common challenge with many ML and DL approaches is their reliance on large high-quality datasets for model training. This information can be limited or non-existent in real-world scenarios. Furthermore, aggregating data from various platforms such as spaceborne, airborne, and sensor data can be challenging due to lack of standardization, as well as lowering the response times of these typically cloud-based approaches. As such it is our opinion that further research needs to be focused on the following areas:

• Sensing technologies and algorithms that focus on the physical footprints of fire, and attempt to detect it reliably, and as far away as possible.

• Deep Learning algorithms, leveraging transfer learning and data augmentation (or other techniques) that can provide reliable outputs with access to limited, and potentially faulty, training data.

• Hardware/firmware architectures with low enough power requirements that enable inference and potentially even training of DL models in real-time on the edge.

Wildfires are a growing global calamity. Due to a combination of climate change and population growth (which will only increase the Wildland Urban Interface, hence increasing the number of humans at risk of wildfires), all studies suggest an increase in risk and damages associated with wildfires throughout the twenty-first century. It is important and certainly worthwhile to engage wildfire behavioral specialists and firefighting professionals along with scientists and engineers, working together towards leveraging technology in addressing this global issue.

There are no conflicts of interest. This article does not include research in which human participants were involved. Informed consent is not applicable. This article does not include any research in which animal participants were involved.

The datasets generated and supporting the findings of this article are obtainable from the corresponding author upon reasonable request.